Machine Learning on Microcontrollers for Biological Sensing: A Systematic Review

Machine Learning on Microcontrollers for Biological Sensing: A Systematic Review | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Systematic Review Machine Learning on Microcontrollers for Biological Sensing: A Systematic Review Hulisani Mukwevho, Unarine Mulaudzi, Bokang Motala, Sibusiso Moyo This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6844035/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Microcontroller-class devices, when integrated with machine learning (ML) models, offer transformative potential for biological sensing in resource-constrained environments. However, the deployment of such systems demands a careful balance between computational limitations, sensor integration, and ecological relevance. This systematic review evaluates trends, architectures, constraints, and applications of ML deployed on microcontroller-class hardware for biological sensing between 2015 and 2025. A systematic search across Google Scholar (n = 142), Web of Science (n = 22), and Scopus (n = 4,266) yielded 4,430 records. After screening and eligibility assessment using PRISMA guidelines, 60 studies were included. The review focused on temporal trends, research types, ML toolchains, hardware platforms, task types, model architectures, dataset sources, system constraints, performance metrics, and domain-specific applications. Publication activity surged after 2019, peaking again in 2024. Most studies employed empirical and applied research methods (Fig. 8), with a majority using embedded platforms like Arduino and TinyML (32.61%) and lightweight frameworks such as TensorFlow Lite. ARM-based processors (34%) and AI-focused SoCs (22%) were the most common hardware platforms. Classification tasks dominated (56.36%), followed by monitoring (25.45%) and regression (18.18%). Deep learning architectures (CNNs, LSTMs, VAEs) accounted for 55.56% of models used. Most studies utilized custom, real-world datasets (67.27%) (Fig. 13) and emphasized performance constraints such as low latency (< 500 ms, 52%) and memory optimization (36%). Hardware limitations were primarily memory-based (44%) or unspecified (32%) (Fig. 15). Real-time inference (38.18%) and edge-device suitability (16.36%) were the most reported performance goals. Application areas were led by healthcare monitoring (25.45%) and water quality analysis (23.64%). Dominant toolchains included Arduino (29.09%), TensorFlow Lite (18.18%), and Edge Impulse (12.73%). Machine learning on microcontroller-class hardware is gaining traction in biological sensing, particularly in health and environmental monitoring. Despite progress, challenges persist in standardized benchmarking, performance reporting, and balancing system constraints. This review offers a detailed synthesis of implementation trends and practical bottlenecks, guiding future development of robust, low-power, and domain-specific ML sensing platforms. Microcontroller-based sensing TinyML Biological monitoring Embedded machine learning Resource-constrained systems Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 Figure 10 Figure 11 Figure 12 Figure 13 Figure 14 Figure 15 Figure 16 Figure 17 Figure 18 1. Introduction Water quality analytics linked with sensor networks represents a innovative shift that improves the assessment of water ecosystems such as in fresh and coastal water systems. These technologies change how researchers analyze and provide crucial monitoring tools for water data science and they create environmental data which enables better safeguarding choices to achieveenhanced environmental protection (Philips, 2023 ). The modern technique of water resource management depends on sensors that identify ecosystem conditions while foreseeing biological responses. Modern technology has produced significant changes in different water settings according to studies (Uniwill-Lily Desun, 2023 ). Things like toxic algal blooms get detected early through permanent monitoring equipment while systems forecast fish deaths and follow aquatic wildlife in sensitive areas. These techniques have proven to deliver considerable improvements in the discovery of pollution sources while also contributing strategies for restoration plans as well as habitat safety. Machine learning in water quality assessment uncovers its importance for real-time monitoring of water environments through its ability to interpret water quality data. The research on the biological effects of water quality parameters is still incomplete in developing areas despite the scientific progress (Zhang & Zhang, 2015). The protection of aquatic ecosystems has become vital to sustainable development because these systems sustain 40% of fish species along with 12% of world protein consumption from animals (Tedesco et al., 2017 ). Water quality requires effective monitoring and data interpretation systems to address distinct problems facing freshwater ecosystems caused by humans. Biological impact monitoring systems consist of sensor networks and data transmission structures as their fundamental operational units. Most articles on IoT system-based biological impact assessment seem fictional resulting in most research listing laboratory-based methods (Miller et al., 2023 ). The rise in pollution and climate change places pressure on aquatic environments compared to pre-industrial conditions therefore requiring the implementation of monitoring techniques to analyze biological ecosystems (Dibyanshu et al., 2024). The practical usage of water quality index models across various ecosystems faces obstacles during full-scale monitoring system implementation. The implementation of monitoring systems becomes problematic due to their high cost and complex technology that prevents widespread (Lupi et al., 2023 ). The field of water quality monitoring research now focuses on two areas regarding low-cost sensor networks and their technical advancements and ecological considerations tied with framework development to monitor biological impacts in developing nations in real time (Moudou et al., 2024). Sensor networks running during pollution occurrences generate insights about significant variables which influence how ecosystems respond. (Fascista, 2022 ) This review systematizes eleven years of studies investigating biological water quality effects whilst documenting technological monitoring trends and obstacles. This review documents the biological effects and technological water quality monitoring solutions for comparison in Table 1 against existing review methods. The review discusses sensor deployment with analytical programs which enables improved assessments of aquatic systems to detect biological responses. This review reveals data patterns from technology adoption through specific case studies and the resulting analysis which establishes necessary foundation for scientific investigation and practical applications improvement. Research will enhance knowledge about continuous monitoring's in aquatic ecosystem management techniques that are evolving. Table 1 Comparative Analysis of the Existing Review Works and Proposed Systematic Review on the Advantages and Application of Machine Learning on Microcontroller-Class Hardware for Biological Sensing Ref. Contribution Pros Cons (Krokidis et al., 2022 ) Reviews sensor-based approaches for diagnoses, with ML integration and data monitoring. Outlines wearable sensor’s potential for early detection, the role of ML in diagnostic accuracy. Requires real-world validation, and sensor data standardization. (Cascales et al., 2021 ) Studies how ML models compensate for external effects Low cost alternates while ML enhances the toughness The system need validation and the accuracy as well as the scalability depends on the training of Machine learning. (Shabani et al.,2021.) Studies Energy-autonomous water quality sensor using MFCs. Requires no external power and COD detection has high accuracy Validation required, and the operation is limited to controlled conditions (Gerós et al., 2022 ) Studeis Integration of cameras and Arduino as well as rodent/wildlife detection Not biased to human and works in the dark. It is a solution that is open source to external aspects Was not peer reviewed. (Diab & Rodriguez-Villegas, 2022 ) Reviewed the TinyML wearables, outlining MCU and hardware/software compromises. Open access licenced and the structure of the TinyML is practical Less protical impemations with more theory based focus,and the recommendations are outdated. (Schizas et al., 2022 ) Brief Machine Learning review that integrates 5G and cloud. Analysis of network integration, and descriptive ML model creation Consists primarily of theory and lacks depth in areas. (Haque Zim et al., 2021 ) Proposed Presents the development of a network structure, briefs the usage of microcontroller with three operating modes. Demostrates application of neural networks in robotics for certain tasks. Limiteddetails and advantages and disadvantgaes of the approach found in the abstract (Sinha et al., 2021 ) Explores the enhancements of pH sensors in its temperature,etc characteristics Drives innovation, supports product development strategies. The abstract is concise and doesn't provide extensive details on the specific ML algorithms used or the extent of performance improvement. (Venkatachalam et al., 2024 ) Proposes an efficient four-layered convolutional neural network (CNN) model for real-time person identification through gait analysis using sensors on edge devices. The model, trained on a public and custom dataset, achieves high accuracy with a small size storagr Presents a compact and efficient CNN model suitable for real-time gait-based person identification on resource-constrained devices. The abstract does not detail the specific architecture or layers of the proposed CNN. (Merenda et al., 2020 ) Reviews the main techniques for executing machine learning models on resource-constrained hardware in the Internet of Things Has a comprehensive review of machine learning techniques for IoT devices. The abstract is broad and doesn't delve into the specifics of the reviewed techniques or their proportional performance. We found several research needs in biological water quality monitoring via sensor technologies in our systematic review. While many studies consider sensor networks for water quality, not many focus on low-cost, scalable systems tailored to limited resource settings. In general, current studies tend to focus on improvements of sensors and data transmission and disregard human, ecological, and operational aspects that are also key for a sustainable monitoring system. The current systems lack proper connectivity between biological impact assessment measurements and sensor-based data systems. The ecological significance of research findings weakens because physical-chemical parameter assessments like DO and pH continue to be studied separately from biological response measurements (Emerson, 2024 ). Long-term changes within real ecosystems cannot be properly studied because researchers rely on laboratory data and cross-sectional data (Chen et al., 2022 ). Research should focus on developing long-term research approaches united with ecological indicators from sensors which will enhance ecosystem defense and management strategies. 1.1. Research questions This review investigates the implementation of machine learning models on microcontroller-class hardware for biological sensing across varied environmental and health contexts. While numerous studies highlight advancements in embedded sensing systems, gaps remain in performance optimization, hardware selection, dataset standardization, and application-specific integration. This review addresses the following questions: How are ML models optimized for low-latency and memory-efficient execution on microcontroller platforms in biological sensing applications? What are the dominant toolchains, hardware architectures, and datasets used in deploying embedded ML for biological tasks, and how do they influence system performance? In what ways are ML-enabled microcontrollers applied across domains such as health monitoring, water quality, and biometrics, and what domain-specific constraints emerge? What technical and methodological limitations (e.g., underreporting of constraints, performance metrics) hinder effective deployment in resource-constrained environments? How do real-time inference capabilities and edge suitability influence the design and evaluation of embedded ML systems for biological sensing? 1.2. Hypotheses Development The research questions explored in this systematic review give rise to several hypotheses centered on system constraints, methodological practices, and application performance in embedded ML for biological sensing. These hypotheses reflect observed trends across healthcare, environmental, and biometric monitoring systems using microcontroller-class hardware: H1: Inconsistent reporting of performance metrics and hardware limitations (e.g., RAM, power, latency) leads to reduced reproducibility and deployment inefficiency in embedded ML systems. H2: The dominance of custom datasets in biological sensing (67.27%) limits cross-study comparability and hinders the creation of generalized ML deployment benchmarks. H3: Real-time inference and edge deployment suitability are critical drivers in the selection of ML models and hardware stacks, particularly in healthcare and environmental sensing applications. H4: The integration of deep learning models on constrained platforms (e.g., CNNs on Arduino/ESP32) is feasible but demands tailored optimization strategies such as quantization and pruning. H5: The limited use of accuracy as a primary evaluation metric (7.27%) and the absence of standardized constraints (32%) undermine confidence in reported system performance. 1.3. Rationale The rationale for this systematic review is to critically assess the technological landscape of machine learning applications on microcontroller-class devices for biological sensing. Unlike traditional water quality studies, this review expands the scope to include diverse domains—such as healthcare, biometrics, and environmental monitoring—where embedded systems are increasingly deployed. With the rise of TinyML and resource-constrained AI, it becomes essential to evaluate how these systems manage performance, cost, and contextual applicability. By synthesizing findings from 2015 to 2025, the review uncovers implementation patterns, identifies deployment barriers, and informs future design of low-power, domain-specific ML solutions for real-time biological monitoring. 1.4. Objectives The primary objective of this systematic review is to evaluate the state of machine learning deployment on microcontroller-class hardware for biological sensing applications. Specifically, the review aims to: Analyze the types of ML models, hardware platforms, and toolchains used in resource-constrained biological monitoring. Assess the distribution and characteristics of datasets employed in ML model training and evaluation. Document the performance constraints and optimization strategies implemented in real-world systems. Examine the application areas (e.g., healthcare, water monitoring, biometrics) and the unique challenges each domain presents. Identify methodological gaps in system reporting, performance benchmarking, and dataset standardization. Provide data-driven recommendations to enhance system design, reproducibility, and scalability of embedded ML for biological tasks. 1.5. Research Contributions This review makes several key contributions to the field of embedded biological sensing: It provides a detailed breakdown of 60 studies based on ML model types, toolchains, dataset usage, hardware selection, and application domains, offering a holistic understanding of system architectures and deployment patterns. The review highlights critical issues such as the underreporting of constraints (32%), low emphasis on performance benchmarking (6% with no metrics), and overreliance on non-standardized custom datasets (67.27%). It introduces an analytical framework for evaluating embedded ML systems in biological sensing, facilitating cross-domain assessment and guiding optimization efforts for real-world use cases. Examining multiple application domains, the review demonstrates how system design varies across healthcare, water quality, environmental monitoring, and biometrics, establishing a multi-sector perspective for future integration efforts. 1.6. Research Novelty To the best of the authors’ knowledge, this is the first systematic review to comprehensively examine the integration of machine learning models with microcontroller-class hardware specifically for biological sensing applications across multiple domains. The novelty of this work lies in: Unlike previous reviews limited to water quality or wearable health monitoring, this review spans a broader spectrum including environmental sensing, biometric authentication, and infrastructure monitoring. It provides a three-dimensional analysis connecting ML model types, hardware constraints, and dataset usage, offering practical insight into how performance trade-offs are managed in real-world deployments. The review uniquely emphasizes real-time inference and edge readiness as central metrics, revealing how system performance goals drive design across use cases. Based on findings, the review proposes an integration roadmap that aligns model selection, toolchains, and performance strategies with resource-constrained deployment scenarios. 2. Materials and Methods In this section, we describe the approach we used to carry out our systematic review on the use of machine learning on microcontroller-class hardware for biological sensing applications. Our review focuses on research published between 2015 and 2025, a period that captures major developments in TinyML, edge AI, and wearable sensing technologies. To the best of our knowledge, there has not yet been a dedicated review that specifically brings together both the embedded hardware and biological sensing perspectives within this timeframe, which makes our contribution both relevant and timely. We collected our literature by searching trusted online databases, including Scopus, Google Scholar, and Web of Science. To make sure we captured the most relevant studies, we carefully chose a set of keywords related to machine learning, microcontrollers, biological sensing, embedded systems, and low-power design. After gathering the initial results, we took the time to review each paper and selected only those that directly focused on practical implementations of ML models on microcontroller-class devices for biological sensing tasks. This process helped us build a strong and focused collection of studies that could genuinely inform and enrich our findings. 2.1. Eligibility criteria For this review, we systematically considered all peer-reviewed and published research works that were directly relevant to the study of machine learning applications on microcontroller-class hardware for biological sensing tasks. To keep the review focused and consistent, we only included articles that were published in English between 2015 and 2025. We applied clear inclusion criteria to make sure that only the most relevant and high-quality studies were selected. Specifically, we only considered research papers that focused on deploying ML models on resource-constrained, low-power hardware and involved biological sensing tasks, such as health monitoring, environmental sensing, or wearable devices. Papers that did not include practical implementations, experimental results, or did not focus on embedded ML systems were excluded from our final analysis. This careful selection process helped us ensure that our review stayed aligned with the core aims of our study (Khanyi et al., 2024; Thobejane et al., 2024; Skosana et al, 2024; Mkhize et al., 2025). The inclusion and exclusion criteria for this study are tabulated as in Table 2. Table 2. Proposed Inclusion and Exclusion Criteria. Criteria Inclusion Exclusion Topic Article papers focusing on the application of machine learning on microcontroller-class hardware for biological sensing tasks Article papers not focusing on machine learning applications on microcontroller-class hardware for biological sensing Research Framework Articles must include a research framework, methodology, or experimental setup related to embedded ML models for biological sensing Articles without a clear research framework, methodology, or experimental validation for embedded ML in biological sensing Language Must be written in English Articles published in languages other than English Period Articles between 2015 to 2025 Articles outside 2015 and 2025 2.2. Information sources A systematic search of online databases was conducted to identify relevant studies for this review. The databases Scopus, Google Scholar, and Web of Science were utilized due to their comprehensive coverage of peer-reviewed literature (Khanyi et al., 2024; Thobejane et al., 2024; Skosana et al, 2024; Mkhize et al., 2025). Each database was thoroughly searched using a combination of keywords related to the study topic, ensuring that the most pertinent research articles were captured. Scopus provided access to a broad range of scientific journals and conference papers, while Google Scholar enabled the inclusion of gray literature and dissertations that might not be indexed elsewhere. Web of Science was used to cross-reference and ensure the robustness of the selected studies by providing citation data and impact factors of the journals. The search results from these databases formed the core of the literature review, ensuring a well-rounded and exhaustive collection of research works. 2.3. Search strategy For this review, we collected research articles from well-known online databases, focusing on how machine learning is being used on microcontroller-class hardware for biological sensing. We mainly searched through three platforms: Google Scholar, Scopus and Web of Science (Khanyi et al., 2024; Thobejane et al., 2024; Skosana et al, 2024; Mkhize et al., 2025). To make sure we found papers closely related to our topic, we used a specific set of keywords: ("Machine Learning" AND "Microcontroller" AND ("Biological Sensing" OR "Health Monitoring" OR "Agricultural Technology") AND ("Embedded Systems" OR "TinyML" OR "IoT") AND ("Low Power" OR "Resource-Constrained Optimization") AND "Algorithm Development" AND "Benchmarking"). These keywords were chosen based on the areas we identified during our initial research, which are shown in the heatmap in Figure 1. The heatmap helped us focus on the most common and important topics. We limited the search to papers published between 2015 and 2025 to make sure we covered the latest work in this fast-moving field. Our search returned about 142 papers from Google Scholar, 4266 from Scopus, and 22 from Web of Science. After getting all the results, we carefully filtered the papers by checking if they involved real-world biological sensing applications and if they tested machine learning models on actual microcontroller-class devices. We left out papers that were only theoretical or not relevant to our topic. The Bibliometric Analysis of Study Search Keywords is illustrated in Figure 1. Table 3. Results Achieved from Literature Search. No. Online Repository Number of results 1 Google Scholar 142 2 Web of Science 22 3 Scopus 4266 Total 4430 2.4. Selection process We used a careful step-by-step process to pick the best studies. First, four researchers read the titles and summaries of the first 60 articles separately. When we disagreed about an article, we talked about it until we all agreed. Then we worked together to review the rest of the articles the same way. Next, we read the full articles that made it past the first round to make sure they really talked about sustainable practices in project management (Khanyi et al., 2024; Thobejane et al., 2024; Skosana et al, 2024; Mkhize et al., 2025). If we still couldn't agree on an article, a fifth researcher helped decide. This way, we made sure we only used the most relevant and useful studies. This systematic approach, illustrated in Figure 2, ensured methodological consistency while accommodating multiple perspectives, ultimately yielding a robust final selection of studies directly relevant to our research objectives 2.5. Data collection process To ensure the accuracy and reliability of the extracted data, we implemented a rigorous, multi-stage validation process. Three independent reviewers systematically collected data from each study under the supervision of a fourth reviewer, who acted as both an arbitrator and subject matter expert. To maintain consistency, we employed a standardized data extraction form. All data were manually extracted (without automation tools) and cross-verified by reviewers to minimize transcription errors. Discrepancies in data interpretation such as conflicting outcomes or methodological ambiguities were resolved through iterative team discussions until consensus was reached (Khanyi et al., 2024; Thobejane et al., 2024; Skosana et al, 2024; Mkhize et al., 2025). For missing information, we conducted in-depth reviews of external materials (appendices, datasets, or linked publications) and consulted external experts when necessary. To address duplicate or overlapping studies (e.g., multiple reports from the same project), we prioritized the most recent and comprehensive publications (2015–2025) and reconciled inconsistencies by comparing methodologies and results. Non-English studies were excluded to prevent language-related biases, ensuring uniformity in analysis. This process, illustrated in Figure 3, guaranteed that only high-quality, verifiable data informed our findings. 2.6. Data items In this section We're looking at all the key details that make machine learning on tiny chips for biological sensing. First, we will check the basics of how accurate these systems are, how fast they run and how much power they consume. We will note exactly which microcontrollers are being used and how they handle sensors. But we are also digging into what really matters such as, can these devices reliably detect viruses or glucose levels as used in biological sensing systems? Do they work outside a lab? Crucially, what is the price tag for small businesses wanting to use this technology? Beyond the specs, we are tracking how these studies were done. Did researchers share their code? Were the tests realistic? This way we can separate good lab results from technology that works in the real world. Mapping out all these pieces, we'll show where this field is really delivering and where it’s still just under review (Myataza et al, 2024; Gumede et al, 2024; Mudau et al., 2024; Mtjilibe et al, 2024). 2.6.1 Data Collection Method We wanted to understand whether these microcontroller-based biosensors deliver in real-world conditions, not just in pristine labs. Our review focused on four key questions. How reliably do they detect biological signals when faced with real-world complexities like variable temperatures or imperfect samples? Can they operate long enough on limited power to be genuinely useful in field applications? Do their results lead to better decisions, like earlier disease detection or more accurate environmental monitoring? Importantly, are they designed for real users, with intuitive interfaces that don't require specialist training? By applying these practical tests, we separated truly impactful innovations from those only working under ideal conditions, highlighting systems that deliver both technical excellence and real-world usability (Myataza et al, 2024; Gumede et al, 2024; Mudau et al., 2024; Mtjilibe et al, 2024). We did not just count technical specs, we wanted to know if these microcontroller biosensors improve outcomes. For performance, we examined how consistently they detect real biological signals when faced with imperfect field conditions like whether a glucose monitor works as well after eight hours in a pocket as it does in a climate-controlled lab. We tracked operational durability too: Can it maintain accuracy through temperature swings? Does it stay reliable when running on limited power for weeks? Most importantly, we looked at real-world utility, do the results help users make better decisions? A system that detects water contamination 12 hours faster is only valuable if communities can act on that data. We also prioritized designs that real people can use, where intuitive interfaces matter as much as algorithmic brilliance. This approach helped us spotlight innovations that deliver both scientific merit and practical value. We assessed how these technologies reduce operational expenditures while improving diagnostic outcomes, with particular attention to cost-benefit ratios in real-world healthcare applications. Studies demonstrating measurable improvements in resource utilization and clinical workflow efficiency were prioritized, revealing how embedded machine learning creates value beyond technical performance metrics alone. 2.6.2 Definition of Collected Data Variables In addition to these primary outcomes, we collected and categorized key variables to enable comprehensive analysis. The extracted data encompassed three primary dimensions: technical specifications (including sensor integration methods, computational constraints, and model optimization techniques), performance metrics such as detection accuracy, latency and power efficiency under varying conditions, and implementation factors such as notably calibration requirements, environmental robustness and usability considerations. Emphasis was placed on capturing both quantitative measurements and qualitative implementation challenges to assess real-world viability beyond controlled laboratory results. These variables were systematically documented through rigorous examination of peer-reviewed studies from Google Scholar, SCOPUS and Web of Science, with inclusion criteria prioritizing studies that reported empirical results from functional prototypes. Maintaining this structured yet nuanced approach to data collection, we ensure our review provides both technical depth and practical insights for researchers and practitioners developing embedded biosensing solutions (Myataza et al, 2024; Gumede et al, 2024; Mudau et al., 2024; Mtjilibe et al, 2024). Table 4. Data Variables Collected. Field Description Study characteristics Geographic location of testing application domain (healthcare, agriculture), and scale of deployment (lab prototype, field trial, commercial product). Includes publication year and study duration. Device specifications Microcontroller model (e.g., ESP32), sensor type (electrochemical, optical), and connectivity (Bluetooth, Wifi). Power source (battery, solar) and energy consumption metrics. ML implementation Machine learning techniques used (e.g., CNN, RNN), model optimization methods (quantization, pruning) and framework (TensorFlow Lite, Edge Impulse). Includes training dataset size and diversity. Economic factors Development costs, production scalability, and maintenance requirements (calibration frequency and expected lifespan). Development costs, production scalability, and maintenance requirements (calibration frequency and expected lifespan). External influences Technical skill requirements for deployment(e.g., FDA approval for medical use), and market competition with existing solutions. 2.7. Study risk of bias assessment To ensure the integrity of our findings, we established rigorous quality standards for study inclusion in this systematic review. Each publication underwent careful evaluation based on several critical factors such as the completeness of technical specifications provided, the validity of biological testing protocols, the thoroughness of power efficiency reporting, the robustness of environmental stress testing, and the availability of materials needed for replication. Scrutiny was applied to studies with industry sponsorship to safeguard against potential performance inflation (Myataza et al, 2024; Gumede et al, 2024; Mudau et al., 2024; Mtjilibe et al, 2024). Through this meticulous screening process which resulted in the exclusion of approximately twenty percent of initially identified studies—we prioritized research demonstrating both technical excellence and practical applicability. The selected studies collectively provide reliable insights into the real-world performance of machine learning implementations on microcontroller platforms for biological sensing applications. 2.8. Synthesis methods This flow chart below in Figure 5 illustrates the systematic approach used in our review of machine learning on microcontroller-class hardware for biological sensing Starting with Study Selection Process, we identify, and screen studies based on set eligibility criteria. Next, Data Standardization involves converting, cleaning, aligning sensor data and handling any missing values. In the Data Analysis phase, we present the data in tables or graphs and perform initial analyses. The flow then moves to Heterogeneity Assessment, where we evaluate variability through subgroup or sensitivity analyses. Finally, Bias Assessment ensures we identify potential biases and maintain transparency in our methods. This structured approach ensures a thorough and reliable review process (Myataza et al, 2024; Gumede et al, 2024; Mudau et al., 2024; Mtjilibe et al, 2024). In this systematic review on the machine learning on microcontroller-class hardware for biological sensing we employed rigorous synthesis methods to ensure that our results were robust, transparent, and reproducible. To determine the eligibility of studies for synthesis, we meticulously tabulated the characteristics of each study and compared them against our predefined synthesis groups on hardware class, machine learning technique employed and biological sensing type. This approach allowed us to include only the most relevant studies, ensuring that our findings were both valid and aligned with the review's objectives. In preparing the data for synthesis, we addressed missing summary statistics through imputation techniques and conducted necessary data conversions to maintain consistency across studies in terms of performance metrics, hardware specifications and sensor modalities. The results were then presented using a combination of structured tables and forest plots, which provided a clear visual representation of the energy consumption, accuracy rate, inference and memory footprints allowing effective identification of trends and outliers. The synthesis of results was conducted using a random-effects meta-analysis model, with subgroup analyses explicitly focusing on geographic and economic contexts to understand their influence on system performance. This approach provided nuanced insights into how different deployment context and hardware limitations interact with success of machine learning application for biological sensing, which was further explored through subgroup analyses and meta-regressions. These analyses helped us identify potential sources of heterogeneity, such as type of microcontroller used, the complexity of biological data, or machine learning model architecture employed. Additionally, sensitivity analyses were performed to assess the robustness of the synthesized results, ensuring that our conclusions were well-supported by stable and reliable evidence. Through this comprehensive approach, we were able to provide a meaningful aggregation of the evidence, offering valuable insights for researchers, practitioners and developers interested in leveraging machine learning for efficient deployment on microcontroller-class hardware in biological sensing applications. To determine study eligibility for inclusion in our systematic review on machine learning application on microcontroller-class hardware for biological sensing, each study was carefully evaluated for its relevance and alignment with the review’s objectives. We manually assessed and compared each study’s characteristics—such as hardware specification, biological sensing target, machine learning model used and performance outcomes—against our predefined synthesis groups. A matrix was created to visually compare the scope and methodologies of the studies with our inclusion criteria, ensuring a comprehensive and objective evaluation. This process ensured that only studies directly pertinent to the review topic were included, thus enhancing the review’s overall rigor and reliability of the review. 2.8.2. Data Preparation for Synthesis In this review, the methods used involved converting or standardizing data collected from various studies to ensure consistency before synthesis. For example, when performance metrics were reported differently across studies, algebraic manipulations were employed to convert these into a uniform scale, such as converting odds ratios to risk ratios where appropriate. Additionally, handling missing data was a critical aspect of the analysis. Missing summary statistics, such as standard deviations, energy consumption or memory usage, were imputed using established statistical methods like multiple imputation. This approach ensured that the dataset was comprehensive and robust, allowing for a more accurate and reliable analysis on machine learning performance on microcontroller-class hardware in biological sensing application (Myataza et al, 2024; Gumede et al, 2024; Mudau et al., 2024; Mtjilibe et al, 2024). 2.8.3. Tabulation and Visual Display of Results Results from individual studies and synthesis efforts were organized using both tabular and graphical methods to enhance clarity and facilitate comparison. Tabular structures were employed to present the data in a structured format, where outcomes were organized by performance—such as accuracy, memory used, and efficiency, and within each category, studies were ordered from lowest to highest risk of bias. This organization allowed for easy comparison across studies and highlighted the most reliable evidence. Additionally, graphical methods, specifically forest plots, were used as the principal tool for visually displaying meta-analysis results (Myataza et al, 2024; Gumede et al, 2024; Mudau et al., 2024; Mtjilibe et al, 2024). These plots showcased effect estimates and confidence intervals for each study alongside a summary estimate. The studies in the forest plots were ordered based on effect size, for example classification accuracy or year of publication, helping to reveal trends over time and across different research focuses. 2.8.4. Synthesis of Results During our manual search on online repositories such as Google Scholar, Scopus, and Web of Science, we carefully reviewed and synthesized the results of relevant studies. The approach to data synthesis was guided by the nature of the data and the degree of variability observed across studies. Based on the findings from our search, we manually assessed the applicability of both fixed-effects and random-effects models, depending on the level of heterogeneity among study results. The selection of the model was determined by the characteristics of the data and our assumptions about the consistency of machine learning performance metrics across different microcontroller platforms and biological sensing tasks. After exporting the data to Excel, we created charts to visually inspect the performance metrics such as accuracy, memory usage and power consumption, allowing us to identify patterns of variability and potential heterogeneity across the studies. This initial visual inspection provided an overview of how study results differed from one another, facilitating a more nuanced analysis. 2.8.5. Exploring Causes of Heterogeneity Subgroup analyses and meta-regression were performed to explore potential sources of heterogeneity, such as differences in study settings, hardware configuration, machine learning model types, or biological sensing targets. Specific analyses focused on factors like the type of microcontroller used such as ESP32, Arduino so on, the biological sensing application such as pathogen detection, and the complexity of the machine learning model, all of which were examined to assess their impact on the performance outcomes such as accuracy, energy efficiency and inference time. These methods helped to identify underlying patterns and relationships that contributed to the overall variability observed across the studies. 2.8.6. Sensitivity Analyses Sensitivity analyses were employed to evaluate the robustness of the synthesis results in relation to various assumptions and methodological decisions made during the review process. These analyses included testing the impact of excluding studies at high risk of bias and using alternative statistical models to ensure that the conclusions were not unduly influenced by specific studies or analytical techniques. This approach helped to confirm the reliability and validity of the findings by addressing potential sources of bias and ensuring that the results were consistent across different analytical scenarios, strengthening confidence in trends identified for machine learning deployment on microcontroller-class hardware on biological sensing application (Myataza et al, 2024; Gumede et al, 2024; Mudau et al., 2024; Mtjilibe et al, 2024). 2.9. Reporting bias assessment In conducting our systematic review on the deployment and performance of machine learning on microcontroller-class hardware for biological sensing application, it was crucial to assess the risk of bias due to potentially missing results, particularly those arising from reporting biases such as selective publication or selective reporting of outcomes. We recognized that these biases could significantly impact the validity and reliability of our synthesis, and thus, we employed a thorough and methodical approach to address this concern. Our assessment of reporting bias was conducted using a combination of well-established statistical and graphical methods. We opted for the use of contour-enhanced funnel plots, a powerful visual tool that allowed us to detect asymmetries in the data. These plots were carefully inspected to identify any potential publication bias by highlighting areas where studies might be missing due to bias versus those missing due to chance. The inclusion of statistical significance contours provided us with a clear and intuitive way to differentiate between these two scenarios, offering a robust visual representation of potential biases. For this assessment, we chose not to develop new tools but rather relied on standard, proven techniques documented extensively in the literature. The methodological rigor of the tools we used was integral to our process. Contour-enhanced funnel plots provided a straightforward yet effective method to visually assess the distribution of studies, allowing us to identify and account for potential biases in our synthesis. The assessment process was designed to minimize subjective bias, multiple independent reviewers were involved in evaluating the studies, and any discrepancies between their assessments were resolved through consensus discussions or, when necessary, by consulting a methodological expert. This collaborative approach ensured that the interpretation of results was balanced and unbiased. We intentionally did not use automation tools for assessing reporting bias in this review. Instead, we opted for a manual approach, utilizing tools such as Excel for creating charts and plots. This hands-on method allowed us to carefully analyse and visualize the data, ensuring a detailed and thorough examination. By manually inspecting the data, we ensured that no subtle patterns or potential biases were overlooked. To further strengthen the review, we conducted comprehensive manual searches across multiple online repositories, including Google Scholar, Scopus, and Web of Science. This approach enabled us to cross-reference data from different studies and sources, addressing any discrepancies and reinforcing the robustness of our conclusions. These manual searches were critical in ensuring that our synthesis was based on the most complete and accurate data available. Given the unique context of machine learning application on resource-constrained microcontroller, we adapted the standard methods for assessing reporting bias to fit this specific field. Studies in embedded biological sensing often exhibit different reporting and validation patterns compared to fields such computer vision or clinical trials, necessitating this adaptation to ensure relevance and accuracy. By tailoring our methods to align with the characteristics of the studies we reviewed, we ensured that our analysis was both contextually appropriate and methodologically sound. To promote transparency and replicability, all methods and approaches used in our assessment have been thoroughly documented and made publicly accessible in the supplementary materials of this review (Myataza et al, 2024; Gumede et al, 2024; Mudau et al., 2024; Mtjilibe et al, 2024). This commitment to openness allows other researchers to replicate our analysis or build upon it in future studies, thereby contributing to the overall rigor and reliability of research in the field of machine learning for biological sensing on microcontroller-class hardware. 2.10. Certainty assessment The reviewed literature was evaluated based on five quality assessment (QA) criteria to ensure rigor and relevance (Myataza et al, 2024; Gumede et al, 2024; Mudau et al., 2024; Mtjilibe et al, 2024): QA1: The clarity and explicitness of the research aim related to machine learning implementation on microcontroller. QA2: The specification and transparency of hardware configuration, sensor types and data collection methods. QA3: The clear definition and explanation of the machine learning model, training processes, and deployment techniques. QA4: The application of a well-defined and appropriate research methodology for evaluating biological sensing performance. QA5: The contribution of the research findings to the enhancement of existing literature in deploying machine learning on resource constrained hardware for biological applications. The certainty assessment responses are rated on a scale from zero (0) to one (1). A 'No' response is assigned '0' points, a score of '0.5' is given if the criterion is 'Partially' met, and '1' point is assigned for a 'Yes' response. All five criteria are scored using this scale. Each piece of literature under review can receive a total score between 0 and 5 points. The results of the certainty assessment for the collected literature on the deployment on machine on microcontroller-class hardware for biological sensing are presented in Table 5. Table 5. Certainty Assessment Results for Collected Literature on Machine Learning on Microcontroller-Class Hardware for Biological Sensing. Ref. QA1 QA2 QA3 QA4 QA5 Total % grading (Fernandez et al., 2017; Saravanan et al., 2018; Cruz et al., 2019; Gokulanathan et al., 2019; Covaciu & Iordan, 2022; Dai & Tan, 2024) 1 0 0.5 0 1 2.5 50 (Chen et al., 2016; Bakar et al., 2016; Osnaya et al., 2016; Hosseini et al., 2017; Osman et al., 2018; Moparthi et al., 2018; Ab Aziz et al., 2019; Bates et al., 2019; Blackstock et al., 2019; Eskin et al., 2019; Gokulanathan et al., 2019; Indrasari et al., 2019; Melo et al., 2019; Mendes et al., 2019; Raji & Thasleena, 2019; Salam et al., 2019; Ahmed & Hassan, 2022; Albanese et al., 2022; Ghimire et al., 2022; Manor & Greenberg, 2022; Wang et al., 2022; Shabani et al., 2021; Chamola, 2021; Zim et al., 2021; Fragner et al., 2024; Kargar et al., 2024; Alves et al., 2024; Saha et al., 2024) 0.5 0.5 0.5 0.5 1 3 60 (Page et al., 2015; Caro et al., 2015; Jadhav & Nerkar, 2015; Duarte et al., 2016; Jafari et al., 2017; Wieclaw et al., 2017; Cascales et al., 2021; Hu et al., 2023; Sarkar & Ghosh, 2024; Venkatachalam et al., 2024; Pushpalatha & Vimalathithan, n.d.) 1 0.5 0.5 1 0.5 3.5 70 (Despins, 2017; Krokidis et al., 2022; Giordano et al., 2024; Tsakanika et al., 2025) 1 0.5 1 1 0.5 4 80 (Merenda et al., 2020; Diab & Rodriguez-Villegas, 2022) 1 1 1 1 0.5 4.5 90 (Saha et al., 2022; Abadade et al., 2023) 1 1 1 1 1 5 100 To support the conclusions of this systematic review on the applications of machine learning on microcontroller-class hardware for biological sensing, we undertook a rigorous assessment of the certainty of the evidence. The strength and reliability of our findings depend on a systematic evaluation process, which we carried out using the GRADE (Grading of Recommendations, Assessment, Development, and Evaluations) framework. GRADE is a globally recognized system that offers a comprehensive and transparent approach to assessing the quality of evidence, ensuring that the conclusions drawn are both credible and well founded. The certainty of the evidence across key outcomes was meticulously evaluated using several critical factors. First, we closely examined the precision of performance metrics by considering sample sizes and the width of confidence intervals in the studies. Narrow confidence intervals coupled with large sample sizes were indicative of a high level of certainty in the evidence, as they suggest more reliable and precise effect estimates. We also assessed the consistency of findings by comparing results across the included studies. When studies demonstrated similar performance results and behavior across different hardware platform and biological sensing application, it contributed to a greater certainty of evidence. Any observed heterogeneity such as difference in sensor type or microcontroller specification was thoroughly analysed to understand its sources and potential impact on the overall findings. Furthermore, the potential for bias was evaluated using an adapted version of the Cochrane Risk of Bias tool. Studies with a low risk of bias were considered to contribute more significantly to the overall certainty of the evidence. We also judged directness based on the alignment of study populations, interventions specifically machine learning models, hardware platform, and biological sensing with the research questions of this review. High directness strengthened the support for our conclusions, leading to greater confidence in the evidence. Based on these factors, the certainty of evidence was categorized as follows: High certainty was assigned when studies were consistent, precise, directly applicable, and exhibited a low risk of bias. Moderate certainty was applied when there were minor concerns about one factor, such as some inconsistency in performance results or a moderate risk of bias. Low certainty was given when significant concerns existed in multiple areas, including imprecision in measurement results, variability in microcontroller hardware performance, or methodological limitation. Very low certainty was assigned when critical issues were present across all factors, significantly undermining confidence in the results. To ensure the relevance of the GRADE approach to this review, we adapted it specifically for outcomes related to the deployment of machine learning algorithms on microcontroller-class hardware for biological sensing application. Multiple independent reviewers assessed the certainty of evidence for each outcome. Disagreements were resolved through consensus discussions, ensuring a balanced and thorough evaluation. Additionally, where possible, we sought additional data or clarification from study authors to strengthen our certainty assessments. 3. Results 3.1. Study selection For this systematic review, we implemented a rigorous study selection process to identify the most relevant and high-quality research. We began by systematically searching three major academic databases: Google Scholar (yielding 142 records), Web of Science (22 records), and Scopus (4,266 records), resulting in an initial pool of 4,430 publications. After removing duplicate entries, we were left with 1,078 unique studies for preliminary screening. Through careful evaluation of titles and abstracts, we narrowed this down to 60 potentially relevant full-text articles. Following a thorough examination of these complete texts, we ultimately included 60 studies that met all our eligibility criteria. The final selection comprised 38 journal articles, 2 book chapters, 18 conference papers, and 2 theses or dissertations. The entire selection methodology, including exclusion reasons at each stage, is visually represented in Figure 6 using a PRISMA flowchart, providing transparent documentation of our systematic approach. 3.2. Study results Figure 7 illustrates the temporal trend of publications from 2015 to 2024 that focus on machine learning applications on microcontroller-class hardware for biological sensing. The data shows a marked increase in research output starting in 2019, with a noticeable peak in both 2019 and 2024. This upward trend suggests growing interest and technological advancement in the integration of embedded systems and biological sensing frameworks during this period. These peaks may correspond with key developments in TinyML and edge AI, reinforcing the relevance and urgency of such research in low-power, real-time monitoring applications. Figure 8 presents the classification of reviewed studies by research type and corresponding research category. A significant proportion of the literature (40 papers) falls under experimental research with an empirical orientation, demonstrating a strong emphasis on real-world implementation and performance testing. Applied research also accounts for a notable share, aligning with the practical nature of microcontroller-based biological sensing. Meanwhile, review studies and conceptual frameworks show a mix of theoretical and design/technical orientations, highlighting the evolving academic groundwork in this emerging field. Figure 9 displays the spectrum of tools and frameworks employed across the included studies, classified by their functional categories. Embedded platforms, such as Arduino and TinyML, accounted for the largest share (32.61%), reflecting the central role of resource-constrained hardware in biological sensing applications. Deep learning models and classical ML libraries also featured prominently, with PyTorch, TensorFlow Lite, and Sklearn among the most utilized. Notably, platforms like Edge Impulse and X-CUBE-AI bridged multiple categories, acting as both lifecycle and deployment tools. This diverse ecosystem of tools underscores the multidisciplinary integration required in deploying ML models on microcontroller-class hardware. Figure 10 categorizes the hardware platforms used across the selected studies, mapping each to its functional classification and percentage contribution. High-end ARM processors, such as the ARM Cortex A15 and STM32 series, represented the largest portion of platforms (34%), followed by AI/parallel system-on-chip (SoC) solutions (22%) like the GAP8 and ESP32. Microcontrollers—including Arduino-based variants—remained prevalent, though their share was distributed across subcategories such as hybrid systems and unspecified configurations. Vision AI development boards and reconfigurable logic devices like FPGAs also featured in specialized applications. This distribution highlights the growing complexity and diversity of platforms supporting machine learning in embedded biological sensing tasks. Figure 11 illustrates the distribution of machine learning task types applied in biological sensing studies and their alignment with broader analytical categories. Classification is the most commonly applied task (56.36%), reflecting the widespread use of ML models for detecting predefined biological conditions. Monitoring tasks (25.45%) and regression-based analyses (18.18%) follow, often applied in dynamic tracking of environmental variables and quantitative modeling. These tasks align with categories such as descriptive/operational, multi-task, and predictive modeling, indicating a blend of analytical depth and real-time adaptability in the reviewed systems. A smaller subset of studies focused on detection and quantification or hybrid approaches combining classification and monitoring. Figure 12 compares the types of machine learning models employed across reviewed studies, categorized both by specific architectures and general methodological groupings. As shown in Figure 12a, deep learning architectures such as convolutional neural networks (CNNs), long short-term memory (LSTM), and variational autoencoders (VAE) dominate the field, comprising 55.56% of implementations. These models outperform rule-based systems (25.93%) and classical ML techniques like SVM and decision trees (14.81%) in adaptive and high-dimensional biological tasks. Figure 12b further consolidates these findings by grouping models into broader categories, reaffirming that deep learning approaches are the most frequently adopted, followed by rule-based/signal-driven methods, classical machine learning, and pattern mining or heuristic techniques. This trend underscores the field’s shift toward data-driven architectures capable of handling complex, nonlinear biological data streams. Figure 13 presents the types of datasets used across the reviewed literature, categorized by origin and data accessibility. A significant majority (67.27%) of studies utilized custom datasets, often derived from specific biological, biomedical, or environmental sensing scenarios. This reflects the need for application-specific data in microcontroller-based biological sensing research. Open-source benchmarking datasets accounted for 25.45%, supporting model comparability across studies. Empirical/lab-based data (5.45%), hybrid dataset sources (3.64%), and synthetic/simulated data (also 3.64%) were used less frequently, suggesting a limited reliance on generalized or artificial data in favor of domain-specific, real-world inputs. Figure 14 categorizes the performance constraints observed in the selected studies, highlighting trade-offs between latency, model size, and memory efficiency. Over half of the studies (52%) emphasized low-latency operation (<500 ms) using compact models, which is essential for real-time biological sensing. A significant portion (36%) employed model compression techniques, such as reducing memory usage to below 100 KB, enabling deployment on constrained microcontroller-class hardware. Ultra-low latency configurations (<50 ms) were reported in 10% of cases, while moderate-latency and variable performance models made up a smaller share. Notably, 6% of studies did not report performance metrics, reflecting a need for standardized benchmarking practices. Figure 15 presents the distribution of hardware and architectural constraints reported in the reviewed literature. The most frequently cited limitation was unspecified (32%), reflecting a gap in detailed system reporting. Among specific constraints, RAM and Flash limitations (combined or standalone) accounted for 44%, highlighting memory as a primary concern for deploying ML models on microcontroller-class hardware. Power constraints (14%) and processor performance-related mentions (12%) were also notable, emphasizing the need for optimized inference under low-resource conditions. Communication and integration complexities appeared in fewer studies, yet they remain significant barriers in real-time biological sensing deployments. Figure 16 summarizes the key performance evaluation themes found in the reviewed literature, indicating a strong emphasis on real-time monitoring and inference capabilities, which appeared in 38.18% of the studies. This reflects the critical need for responsive, low-latency systems in biological sensing. Suitability for edge/on-device tasks was the second most frequent focus (16.36%), underscoring the field’s pivot toward decentralized processing. Other important performance considerations included latency/power efficiency (12.73%), domain-specific functionality (10.91%), and comparative benchmarking (10.91%). Notably, only 7.27% of studies were explicitly accuracy-focused, while 3.64% lacked clearly stated performance criteria, indicating opportunities for improved reporting consistency. Figure 17 outlines the application domains in which ML-integrated microcontroller systems have been applied. The most common use case was healthcare and medical monitoring (25.45%), followed closely by water quality monitoring and management (23.64%), reflecting the critical societal need for health and environmental tracking. Other notable domains included biometric authentication (10.91%), industrial and infrastructure monitoring (9.09%), and general environmental sensing (7.27%). Emerging areas such as gesture recognition, educational tools, and smart city infrastructure also featured, though with lower representation. The spread of applications illustrates the versatility of embedded ML across diverse real-world scenarios. Figure 18 presents the toolchains and software stacks employed in the development of embedded ML systems. Arduino and Arduino-compatible platforms dominated the space, accounting for 29.09% of all implementations—highlighting their accessibility and broad hardware support. TensorFlow Lite and Lite Micro followed at 18.18%, confirming the growing adoption of lightweight ML frameworks for on-device learning. Edge Impulse (12.73%) and various custom embedded development tools (9.09%) were also frequently used, supporting real-time deployment and optimization. Additional stacks included simulation tools like Unity (7.27%), Python-based programming environments, and MySQL data systems. The diversity of stacks illustrates the field’s cross-disciplinary nature and the balance between hardware simplicity and software flexibility. 3.3. Study reearch questions results 1. How are ML models optimized for low-latency and memory-efficient execution on microcontroller platforms in biological sensing applications? The review highlights strong efforts in optimizing machine learning (ML) models for low-latency and memory-efficient deployment in biological sensing systems. Over half of the reviewed studies (52%) prioritized models with inference times below 500 milliseconds, a critical threshold for real-time monitoring (Figure 14). Additionally, 36% of studies implemented memory compression techniques, such as quantization and pruning, allowing models to operate within <100 KB memory constraints—ideal for microcontroller-class devices. Only 10% of systems reported achieving ultra-low latency (<50 ms), which may be required for time-sensitive healthcare or gesture recognition tasks. However, a notable 6% of studies did not report any performance metrics at all, signaling an ongoing lack of standardization in benchmarking latency, memory, and computational performance. This variability underscores the importance of defining clear evaluation frameworks for embedded ML deployment, especially in constrained environments where trade-offs between latency, power, and accuracy are inevitable. 2. What are the dominant toolchains, hardware architectures, and datasets used in deploying embedded ML for biological tasks, and how do they influence system performance? A clear pattern emerges in toolchain and hardware usage across the studies. Arduino and its compatible toolsets lead implementation strategies, appearing in 29.09% of systems, largely due to their accessibility, affordability, and extensive hardware integration support (Figure 18). Lightweight ML frameworks, particularly TensorFlow Lite and Lite Micro (18.18%), were preferred for deploying models on edge devices, while Edge Impulse (12.73%) played a key role in automating the optimization and deployment pipeline. On the hardware side, high-performance ARM processors (34%), including STM32 and Cortex variants, were most commonly used, offering a balance of power efficiency and computational capability (Figure 10). AI/parallel SoCs like ESP32 and GAP8 (22%) were also widely deployed, indicating growing reliance on specialized AI microcontrollers for inference-heavy workloads. In terms of datasets, 67.27% of studies used custom-built, domain-specific datasets (Figure 13). While this enhances contextual accuracy, it limits model generalizability and reproducibility. The remaining studies relied on open/public datasets (25.45%), with simulated or hybrid sources comprising under 8%. Together, these findings reveal a field focused on tailor-made solutions, but one that must advance toward standardized, sharable pipelines to ensure scalability and reproducibility. 3. In what ways are ML-enabled microcontrollers applied across domains such as health monitoring, water quality, and biometrics, and what domain-specific constraints emerge? The systematic review found that ML-enabled microcontrollers have been adopted across a wide array of biological sensing domains, with notable focus areas emerging. Healthcare and medical monitoring was the most common application domain, cited in 25.45% of the studies, often involving wearable biosensors and physiological data monitoring systems (Figure 17). Water quality monitoring and environmental sensing followed closely at 23.64%, leveraging microcontroller-based systems for continuous ecosystem health tracking, particularly in low-resource contexts. Biometric authentication (10.91%), such as gait or gesture recognition, and industrial/infrastructure monitoring (9.09%) also featured significantly. Each domain introduced distinct system requirements and constraints. For instance, healthcare systems required low latency and high reliability, while water sensing systems emphasized robustness and long-term power efficiency. These domain-specific constraints directly influenced hardware choices, dataset characteristics, and model complexity, leading to differing deployment strategies. For example, CNNs and LSTMs were commonly used in gesture and health recognition due to their superior pattern detection capabilities, while simpler, rule-based systems sufficed for steady-state environmental measurements. 4. What technical and methodological limitations (e.g., underreporting of constraints, performance metrics) hinder effective deployment in resource-constrained environments? Despite advancements in embedded ML systems, your review identified significant technical and methodological limitations that hinder widespread, reliable deployment. Most notably, 32% of reviewed studies did not report key system-level constraints, such as available RAM, flash memory, or CPU load (Figure 15). This lack of transparency severely affects reproducibility and makes it difficult to compare system efficiency across studies. Additionally, 6% of papers failed to report any performance benchmarks, such as latency, accuracy, or energy usage (Figure 14), further weakening the practical applicability of these findings. Even among those that did report metrics, only 7.27% made accuracy their primary evaluation focus, suggesting a need for more balanced assessments that weigh accuracy, speed, power, and reliability (Figure 16). Such methodological inconsistencies create gaps between prototype-level performance and real-world deployment. To bridge this, future research must emphasize standardized evaluation protocols that include detailed constraint reporting, real-world testing, and comprehensive performance characterization under variable environmental conditions. 5. How do real-time inference capabilities and edge suitability influence the design and evaluation of embedded ML systems for biological sensing? Real-time capability and edge deployment readiness are now central to the design of ML-integrated microcontroller systems. In 38.18% of the studies, real-time inference was explicitly listed as a performance goal (Figure 16), often influencing model architecture (e.g., CNNs over RNNs) and latency optimization methods. Another 16.36% of studies prioritized edge deployment suitability, which encompasses local data processing, intermittent connectivity tolerance, and energy-aware operation. These priorities drove the adoption of compressed model formats, lightweight frameworks like TensorFlow Lite, and power-efficient hardware such as the ESP32 (Figure 10 and Figure 18). Domains such as water quality monitoring and health diagnostics particularly benefited from these features, as they require fast decision-making in low-bandwidth, remote, or mobile settings. Additionally, studies employing custom datasets (67.27%) often tailored their model training to edge constraints, ensuring compatibility with limited onboard memory and real-time event detection needs (Figure 13). These trends confirm that edge-readiness is no longer optional—it is a defining requirement shaping both system architecture and evaluation strategies. 4. Conclusion This systematic review presents a comprehensive analysis of machine learning applications on microcontroller-class hardware for biological sensing, synthesizing findings from studies published between 2015 and 2025. The review identifies a growing research interest in this domain, marked by sharp increases in publication activity beginning in 2019 and peaking again in 2024 (Fig. 7 ). The dominance of empirical and applied studies (Fig. 8 ) underscores a clear shift from theoretical exploration to real-world deployment. Findings indicate that embedded ML systems are increasingly optimized for low-latency performance and memory efficiency, with 52% of studies prioritizing sub-500 ms inference and 36% adopting compression methods to meet memory constraints (Fig. 14 ). Hardware selection favored high-performance ARM microcontrollers (34%) and AI-focused SoCs (22%), while Arduino and TensorFlow Lite emerged as the most widely adopted development stacks (Figs. 10 and 18 ). Despite the breadth of hardware and tools, reporting inconsistencies remain: 32% of studies omitted constraint specifications, and 6% lacked performance benchmarks entirely (Figs. 15 and 14 ). Model architectures leaned heavily toward deep learning, particularly CNNs, LSTMs, and hybrid architectures, which accounted for 55.56% of usage (Fig. 12 ). However, reliance on custom, domain-specific datasets (67.27%) presents challenges for reproducibility and cross-study comparability (Fig. 13 ). In terms of application, biological sensing was most prevalent in healthcare (25.45%) and water quality (23.64%) systems, with increasing exploration in biometrics, infrastructure monitoring, and environmental sensing (Fig. 17 ). Edge readiness and real-time capability have become defining factors in the system design process, influencing the selection of models, datasets, and performance goals. With 38.18% of studies emphasizing real-time inference and 16.36% focusing on edge deployment (Fig. 16 ), the field is clearly shifting toward autonomous, decentralized sensing platforms. References Chen, J., Zhang, Y., Kuzyakov, Y., Wang, D., & Olesen, J. E. (2022). Challenges in upscaling laboratory studies to ecosystems in soil microbiology research. Global Change Biology, 29(3), 569–588. https://doi.org/10.1111/gcb.16537 Diab, M. S., & Rodriguez-Villegas, E. (2022). Embedded machine learning using microcontrollers in wearable and ambulatory systems for health and care applications: A review. IEEE Access, 10, 98450–98474. https://doi.org/10.1109/ACCESS.2022.3206782 Dibyanshu, Kern, M., & Scheytt, T. (2024). Trace compounds in the urban water cycle in the Freiberg region, Germany. Frontiers in Water, 6. https://doi.org/10.3389/frwa.2024.1335766 Emerson, B. (2024). Beyond boundaries: Integrating biosensors and bioelectronics for environmental sustainability. Biosensors & Bioelectronics, 15(1), 1–2. https://doi.org/10.37421/2155-6210.2024.15.431 Fascista, A. (2022). Toward integrated large-scale environmental monitoring using WSN/UAV/crowdsensing: A review of applications, signal processing, and future perspectives. Sensors, 22(5), 1824. https://doi.org/10.3390/s22051824 Gerós, A., Cruz, R., de Chaumont, F., Cardoso, J. S., & Aguiar, P. (2022). Deep learning-based system for real-time behavior recognition and closed-loop control of behavioral mazes using depth sensing. bioRxiv. https://doi.org/10.1101/2022.02.22.481410 Haque Zim, M. Z., Sazzad Sarkar, M., Nisi, Z. F., & Das, N. C. (2021). LFNNR: Light follower neural network robot conducted by machine learning technique. In Proceedings of the 2021 IEEE Conference of Russian Young Researchers in Electrical and Electronic Engineering (ElConRus) (pp. 1149–1154). IEEE. https://doi.org/10.1109/ElConRus51938.2021.9396503 Krokidis, M. G., Dimitrakopoulos, G. N., Vrahatis, A. G., Tzouvelekis, C., Drakoulis, D., Papavassileiou, F., Exarchos, T. P., & Vlamos, P. (2022). A sensor-based perspective in early-stage Parkinson’s disease: Current state and the need for machine learning processes. Sensors, 22(2), 409. https://doi.org/10.3390/s22020409 Krokidis, M. G., Dimitrakopoulos, G. N., Vrahatis, A. G., Tzouvelekis, C., Drakoulis, D., Papavassileiou, F., Exarchos, T. P., & Vlamos, P. (2022). A sensor-based perspective in early-stage Parkinson’s disease: Current state and the need for machine learning processes. Sensors, 22(2), 409. https://doi.org/10.3390/s22020409 Lupi, F., Herriges, J. A., Kim, H., & Stevenson, R. J. (2023). Getting off the ladder: Disentangling water quality indices to enhance the valuation of divergent ecosystem services. Proceedings of the National Academy of Sciences of the United States of America, 120(18), e2120261120. https://doi.org/10.1073/pnas.2120261120 Mao, F., Khamis, K., Krause, S., Clark, J., & Hannah, D. M. (2019). Low-cost environmental sensor networks: Recent advances and future directions. Frontiers in Earth Science, 7, 221. https://doi.org/10.3389/feart.2019.00221 Merenda, M., Porcaro, C., & Iero, D. (2020). Edge machine learning for AI-enabled IoT devices: A review. Sensors, 20(9), 2533. https://doi.org/10.3390/s20092533 Miller, M., Kisiel, A., Cembrowska-Lech, D., Durlik, I., & Miller, T. (2023). IoT in water quality monitoring: Are we really here? Sensors, 23(2), 960. https://doi.org/10.3390/s23020960 Philips, J. (2023, November 23). The future of water pollution monitoring: Advanced technologies for better data. Sustainable Living. https://www.environmentalconsortium.org/the-future-of-water-pollution-monitoring-advanced-technologies-for-better-data/ Schizas, N., Karras, A., Karras, C., & Sioutas, S. (2022). TinyML for ultra-low power AI and large scale IoT deployments: A systematic review. Future Internet, 14(12), 363. https://doi.org/10.3390/fi14120363 Shabani, F., Philamore, H., & Matsuno, F. (n.d.). An energy-autonomous chemical oxygen demand sensor using a microbial fuel cell and embedded machine learning. IEEE Access. https://doi.org/10.1109/ACCESS.2021.3101496 Sinha, S., Sahu, N., Bhardwaj, R., Mehta, A., Ahuja, H., Srivastava, S., Elhence, A., & Chamola, V. (2021). Machine learning on FPGA for robust Si3N4-gate ISFET pH sensor in industrial IoT applications. IEEE Transactions on Industry Applications, 57(6), 6700–6712. https://doi.org/10.1109/TIA.2021.3117233 Tedesco, P. A., Beauchard, O., Bigorne, R., Blanchet, S., Buisson, L., Conti, L., Cornu, J. F., Dias, M. S., Grenouillet, G., Hugueny, B., Jézéquel, C., Leprieur, F., Brosse, S., & Oberdorff, T. (2017). A global database on freshwater fish species occurrence in drainage basins. Scientific Data, 4(1), 170141. https://doi.org/10.1038/sdata.2017.141 Uniwill-Lily Desun. (2023, September 13). Applications of water quality sensors in environmental monitoring. Water Quality Sensors. https://disen-sensor.com/applications-of-water-quality-sensors-in-environmental-monitoring/ Venkatachalam, S., Nair, H., Vellaisamy, P., Zhou, Y., Youssfi, Z., & Shen, J. P. (2024). Realtime person identification via gait analysis. arXiv. https://arxiv.org/pdf/2404.15312 Zhang, C., & Zhang, J. (2014). Current techniques for detecting and monitoring algal toxins and causative harmful algal blooms. Journal of Environmental & Analytical Toxicology, 4(2), 123. https://doi.org/10.4172/2380-2391.1000123 Krokidis, M. G., Dimitrakopoulos, G. N., Vrahatis, A. G., Tzouvelekis, C., Drakoulis, D., Papavassileiou, F., Exarchos, T. P., & Vlamos, P. (2022). A Sensor-Based Perspective in Early-Stage Parkinson’s Disease: Current State and the Need for Machine Learning Processes. Sensors, 22(2), 409. https://doi.org/10.3390/s22020409 Myataza, A., Mafunga, M., Mkhulisi, N. S., & Thango, B. A. A Systematic Review of ERP, CRM, and HRM Systems for SMEs: Managerial and Employee Support. Preprints 2024, 2024100384. Gumede, T. T., Chiworeka, J. M., Magoda, A. S., & Thango, B. Building Effective Social Media Strategies for Business: A Systematic Review. Preprints 2024, 2024100379. Cascales, J. P., Greenfield, D. A., Roussakis, E., Witthauer, L., Li, X., Goss, A., & Evans, C. L. (2021). Wireless wearable sensor paired with machine learning for the quantification of tissue oxygenation. IEEE Internet of Things Journal, 8(24), 17557-17567. Shabani, F., Philamore, H., & Matsuno, F. (2021). An energy-autonomous chemical oxygen demand sensor using a microbial fuel cell and embedded machine learning. IEEE Access, 9, 108689-108701. Gerós, A., Cruz, R., de Chaumont, F., Cardoso, J. S., & Aguiar, P. (2022). Deep learning-based system for real-time behavior recognition and closed-loop control of behavioral mazes using depth sensing. Biorxiv, 2022-02. Diab, M. S., & Rodriguez-Villegas, E. (2022). Embedded machine learning using microcontrollers in wearable and ambulatory systems for health and care applications: A review. IEEE Access, 10, 98450-98474. Schizas, N., Karras, A., Karras, C., & Sioutas, S. (2022). TinyML for ultra-low power AI and large scale IoT deployments: A systematic review. Future Internet, 14(12), 363. Zim, M. Z. H., Sarkar, M. S., Nisi, Z. F., & Das, N. C. (2021, January). LFNNR: Light Follower Neural Network Robot Conducted by Machine Learning Technique. In 2021 IEEE Conference of Russian Young Researchers in Electrical and Electronic Engineering (ElConRus) (pp. 1149-1154). IEEE. Chamola, V. (2021). Machine Learning on FPGA for Robust Si3N4-Gate ISFET pH Sensor in Industrial IoT Applications. Venkatachalam, S., Nair, H., Vellaisamy, P., Zhou, Y., Youssfi, Z., & Shen, J. P. (2024, July). Realtime Person Identification via Gait Analysis Using IMU Sensors on Edge Devices. In 2024 International Conference on Neuromorphic Systems (ICONS) (pp. 371-375). IEEE. Merenda, M., Porcaro, C., & Iero, D. (2020). Edge machine learning for ai-enabled iot devices: A review. Sensors, 20(9), 2533. Ahmed, K., & Hassan, M. (2022, June). TinyCare: a TinyML-based Low-Cost continuous blood pressure estimation on the extreme edge. In 2022 IEEE 10th International Conference on Healthcare Informatics (ICHI) (pp. 264-275). IEEE. Covaciu, F., & Iordan, A. E. (2022). Control of a drone in virtual reality using MEMS sensor technology and machine learning. Micromachines, 13(4), 521. Mudau, M. C., Moshapo, L. W., Monyela, T. M., & Thango, B. A. The Role of Manufacturing Operations in SMEs Performance: A Systematic Review. Preprints 2024, 2024100539. Mtjilibe, T., Rameetse, E., Mgwenya, N., & Thango, B. (2024). Exploring the Challenges and Opportunities of Social Media for Organizational Engagement in SMEs: A Comprehensive Systematic Review. Available at SSRN 4998542. Wang, T., Fang, K., Wei, W., Tian, J., Pan, Y., & Li, J. (2022). Microcontroller unit chip temperature fingerprint informed machine learning for IIoT intrusion detection. IEEE Transactions on Industrial Informatics, 19(2), 2219-2227. Albanese, A., Nardello, M., Fiacco, G., & Brunelli, D. (2022). Tiny machine learning for high accuracy product quality inspection. IEEE Sensors Journal, 23(2), 1575-1583. Manor, E., & Greenberg, S. (2022). Custom hardware inference accelerator for tensorflow lite for microcontrollers. IEEE Access, 10, 73484-73493. Abadade, Y., Temouden, A., Bamoumen, H., Benamar, N., Chtouki, Y., & Hafid, A. S. (2023). A comprehensive survey on tinyml. IEEE Access, 11, 96892-96922. Saha, S. S., Sandha, S. S., & Srivastava, M. (2022). Machine learning for microcontroller-class hardware: A review. IEEE Sensors Journal, 22(22), 21362-21390. Fragner, C., Krutzler, C., Weiss, A. P., & Leitgeb, E. (2024). LEDPOS: Indoor visible light positioning based on LED as sensor and machine learning. IEEE Access. Giordano, M., Dheman, K., & Magno, M. (2024). SepAl: Sepsis Alerts On Low Power Wearables With Digital Biomarkers and On-Device Tiny Machine Learning. IEEE Sensors Journal. Tsakanika, E., Tsoukas, V., Kakarountas, A., & Kokkinos, V. (2025). High Accuracy of Epileptic Seizure Detection Using Tiny Machine Learning Technology for Implantable Closed-Loop Neurostimulation Systems. BioMedInformatics, 5(1), 14. Kargar, A., Zorbas, D., Tedesco, S., Gaffney, M., & O’Flynn, B. (2024). Detecting Halyomorpha halys using a low-power edge-based monitoring system. Computers and Electronics in Agriculture, 221, 108935. Dai, Y., & Tan, R. (2024, May). FedCFC: On-Device Personalized Federated Learning with Closed-Form Continuous-Time Neural Networks. In 2024 23rd ACM/IEEE International Conference on Information Processing in Sensor Networks (IPSN) (pp. 14-26). IEEE. Sarkar, S., & Ghosh, A. (2024). Schrödinger spectrum and slim CNN architecture-based signal quality estimation for Photoplethysmogram signals. Biomedical Signal Processing and Control, 94, 106240. Hu, J. R., Park, D. Y., Agarwal, N., Herzig, M., Ormseth, G., Kaushik, M., Giao, D. M., Turkson-Ocran, R. N., & Juraschek, S. P. (2023). The promise and illusion of continuous, cuffless blood pressure monitoring. Current Cardiology Reports, 25(10), 1139–1149. https://doi.org/10.1007/s11886-023-01932-4 Ghimire, D., Kil, D., & Kim, S. H. (2022). A survey on efficient convolutional neural networks and hardware acceleration. Electronics, 11(6), 945. Khanyi, M. B., Xaba, S. N., Mlotshwa, N. A., Thango, B., & Matshaka, L. (2024). A Roadmap to Systematic Review: Evaluating the Role of Data Networks and Application Programming Interfaces in Enhancing Operational Efficiency in Small and Medium Enterprises. Sustainability, 16(23), 10192. https://doi.org/10.3390/su162310192 Thobejane, L. T., & Thango, B. A. (2024). Partial Discharge Source Classification in Power Transformers: A Systematic Literature Review. Applied Sciences, 14(14), 6097. Fu, Z., He, X., Wang, E., Huo, J., Huang, J., & Wu, D. (2021). Personalized human activity recognition based on integrated wearable sensor and transfer learning. Sensors, 21(3), 885. Alves, D. N., Barboza, E. D. A., & Vieira, T. F. (2024, November). TinyPPE: Ensuring Workplace Safety Helmet Compliance with Tiny Machine Learning. In 2024 XIV Brazilian Symposium on Computing Systems Engineering (SBESC) (pp. 1-6). IEEE. Saha, S. S., Sandha, S. S., Aggarwal, M., Wang, B., Han, L., Briseno, J. D. G., & Srivastava, M. (2024). TinyNS: Platform-aware neurosymbolic auto tiny machine learning. ACM Transactions on Embedded Computing Systems, 23(3), 1-48. Mohaidat, T., & Khalil, K. (2024). A survey on neural network hardware accelerators. IEEE Transactions on Artificial Intelligence. Saha, S. S., Davis, C., Sandha, S. S., Park, J., Geronimo, J., Garcia, L. A., & Srivastava, M. B. (2024). Locomote: Ai-driven sensor tags for fine-grained undersea localization and sensing. IEEE sensors journal, 24(10), 16999-17018. Cruz, I. A., de Melo, L., Leite, A. N., Sátiro, J. V. M., Andrade, L. R. S., Torres, N. H., ... & Ferreira, L. F. R. (2019). A new approach using an open-source low cost system for monitoring and controlling biogas production from dairy wastewater. Journal of Cleaner Production, 241, 118284. Mendes, J. P., Coelho, L., Kovacs, B., Almeida, J. D., Pereira, C. M., Jorge, P. A., & Borges, M. T. (2019). Dissolved carbon dioxide sensing platform for freshwater and saline water applications: Characterization and validation in aquaculture environments. Sensors, 19(24), 5513. Salam, A. E. U., Sadjad, R. S. O., SAMMAN, F. A., & FAKHRI, M. (2019). Online Hydraulic Parameter Monitoring System in Water Distribution Network. ICIC express letters. Part B, Applications: an international journal of research and surveys, 10(12), 1037-1045. Raji, C. G., & Thasleena, V. A. (2019, December). IOT based water quality monitoring with android application. In 2019 Third International conference on I-SMAC (IoT in Social, Mobile, Analytics and Cloud)(I-SMAC) (pp. 446-451). IEEE. Ab Aziz, M. A., Abas, M. F., Bashri, M. K. A. A., Saad, N. M., & Ariff, M. H. (2019). Evaluating IoT based passive water catchment monitoring system data acquisition and analysis. Bulletin of Electrical Engineering and Informatics, 8(4), 1373-1382. Blackstock, J. M., Covington, M. D., Perne, M., & Myre, J. M. (2019). Monitoring atmospheric, soil, and dissolved CO2 using a low-cost, Arduino monitoring platform (CO2-LAMP): Theory, fabrication, and operation. Frontiers in Earth Science, 7, 313. Bates, H., Zavafer, A., Szabó, M., & Ralph, P. J. (2019). A guide to Open-JIP, a low-cost open-source chlorophyll fluorometer. Photosynthesis research, 142, 361-368. Indrasari, W., Budi, E., Alayya, S. R., & Ramli, R. (2019, October). Measurement of water polluted quality based on turbidity, pH, magnetic property, and dissolved solid. In Journal of Physics: Conference Series (Vol. 1317, No. 1, p. 012060). IOP Publishing. Eskin, M. G., Torabfam, M., Psillakis, E., Cincinelli, A., Kurt, H., & Yüce, M. (2019). Real-time water quality monitoring of an artificial lake using a portable, affordable, simple, arduino-based open source sensor. Environmental Engineering-Inženjerstvo Okoliša, 6(1), 7-14. Saravanan, K., Anusuya, E., Kumar, R., & Son, L. H. (2018). Real-time water quality monitoring using Internet of Things in SCADA. Environmental monitoring and assessment, 190(9), 556. Osman, S. O., Mohamed, M. Z., Suliman, A. M., & Mohammed, A. A. (2018, August). Design and implementation of a low-cost real-time in-situ drinking water quality monitoring system using arduino. In 2018 International Conference on Computer, Control, Electrical, and Electronics Engineering (ICCCEEE) (pp. 1-7). IEEE. Melo, M., Mota, F., Albuquerque, V., & Alexandria, A. (2019). Development of a robotic airboat for online water quality monitoring in lakes. Robotics, 8(1), 19. Skosana, S., Mlambo, S., Madiope, T., & Thango, B. Evaluating Wireless Network Technologies (3G, 4G, 5G) and Their Infrastructure: A Systematic Review. Preprints 2024, 2024101331. Mkhize, A., Mokhothu, K. D., Tshikhotho, M., & Thango, B. A. (2025). Evaluating the Impact of Cloud Computing on SME Performance: A Systematic Review. Businesses, 5(2), 23. https://doi.org/10.3390/businesses5020023. Parra, L., Rocher, J., Escrivá, J., & Lloret, J. (2018). Design and development of low cost smart turbidity sensor for water quality monitoring in fish farms. Aquacultural Engineering, 81, 10-18. Moparthi, N. R., Mukesh, C., & Sagar, P. V. (2018, February). Water quality monitoring system using IoT. In 2018 Fourth international conference on advances in electrical, electronics, information, communication and bio-informatics (AEEICB) (pp. 1-5). IEEE. Gokulanathan, S., Manivasagam, P., Prabu, N., & Venkatesh, T. (2019). A GSM based water quality monitoring system using Arduino. Shanlax International Journal of Arts, Science and Humanities, 6(4), 22-26. Despins, L. A. (2017). Automated detection of sepsis using electronic medical record data: a systematic review. The Journal for Healthcare Quality (JHQ), 39(6), 322-333. Page, A., Kulkarni, A., & Mohsenin, T. (2015, October). Utilizing deep neural nets for an embedded ECG-based biometric authentication system. In 2015 IEEE biomedical circuits and systems conference (BioCAS) (pp. 1-4). IEEE. Wieclaw, L., Khoma, Y., Fałat, P., Sabodashko, D., & Herasymenko, V. (2017, September). Biometrie identification from raw ECG signal using deep learning techniques. In 2017 9th IEEE International Conference on Intelligent Data Acquisition and Advanced Computing Systems: Technology and Applications (IDAACS) (Vol. 1, pp. 129-133). IEEE. Caro, L. A., Silva, C., Peralta, B., Herrera, O. A., & Barrientos, S. (2015). Real-time recognition of arm motion using artificial neural network multi-perceptron with arduino one microcontroller and EKG/EMG shield sensor. In Ambient Intelligence for Health: First International Conference, AmIHEALTH 2015, Puerto Varas, Chile, December 1-4, 2015, Proceedings 1 (pp. 3-14). Springer International Publishing. Chen, A. T. Y., Biglari-Abhari, M., Wang, K. I. K., Bouzerdoum, A., & Tivive, F. H. C. (2016). Hardware/software co-design for a gender recognition embedded system. In Trends in Applied Knowledge-Based Systems and Data Science: 29th International Conference on Industrial Engineering and Other Applications of Applied Intelligent Systems, IEA/AIE 2016, Morioka, Japan, August 2-4, 2016, Proceedings 29 (pp. 541-552). Springer International Publishing. Osnaya, M. R., Rocha, M. F., Flores, Z. C., Navarrete, R., Andrade, E., Canto, C., & Rocha, M. E. (2016). MatlabSoftware in PC to Replace an Embedded System Used for Face Recognition. Journal of Industrial and Intelligent Information Vol, 4(2). Jadhav, M., & Nerkar, P. (2015). A Biometric System for Person Authentication Based on Finger Vein Patterns on FPGA. International Journal of Advanced Research in Electrical, Electronics and Instrumentation Engineering, 4(6). Pushpalatha, B., Scholar, P. G., & Vimalathithan, S. Implementing the Concept of Biometrics Recognition System in Voting Machine. Fernandez, C., Rusk, S., Glattard, N., & Shokoueinejad, M. (2017). 0067 COMPUTATIONAL PHENOTYPING IN POLYSOMNOGRAPHY: USING INTERPRETABLE PHYSIOLOGY-BASED MACHINE LEARNING MODELS TO PREDICT HEALTH OUTCOMES. Journal of Sleep and Sleep Disorders Research, 40(suppl_1), A26-A26. Duarte, T., Pimentão, J. P., Sousa, P., & Onofre, S. (2016, September). Biometric access control systems: A review on technologies to improve their efficiency. In 2016 IEEE International Power Electronics and Motion Control Conference Banbury, C. R., et al. (2021). "Benchmarking TinyML Systems: Challenges and Directions." IEEE Design & Test. (PEMC) (pp. 795-800). IEEE. Jafari, A., Gandhi, S., Konuru, S. H., Hairston, W. D., Oates, T., & Mohsenin, T. (2017, May). An EEG artifact identification embedded system using ICA and multi-instance learning. In 2017 IEEE international symposium on circuits and systems (ISCAS) (pp. 1-4). IEEE. Hosseini, M. P., Tran, T. X., Pompili, D., Elisevich, K., & Soltanian-Zadeh, H. (2017, July). Deep learning with edge computing for localization of epileptogenicity using multimodal rs-fMRI and EEG big data. In 2017 IEEE international conference on autonomic computing (ICAC) (pp. 83-92). IEEE. Bakar, W. A. B. W. A., Abdullah, Z. B., Md Saman, M. Y. B., Jalil, M. M. B. A., Man, M. B., Herawan, T., & Hamdan, A. R. (2016). Incremental-eclat model: an implementation via benchmark case study. In Advances in Machine Learning and Signal Processing: Proceedings of MALSIP 2015 (pp. 35-46). Springer International Publishing. David, R., et al. (2020). "TensorFlow Lite Micro: Embedded Machine Learning for TinyML Systems." arXiv. Han, S., et al. (2016). "Deep Compression: Compressing Deep Neural Networks with Pruning, Trained Quantization and Huffman Coding." ICLR. Lin, J., et al. (2020). "MCUNet: Tiny Deep Learning on IoT Devices." NeurIPS. Sze, V., et al. (2017). "Efficient Processing of Deep Neural Networks: A Tutorial and Survey." Proceedings of the IEEE. Warden, P., & Situnayake, D. (2019). "TinyML: Machine Learning with TensorFlow Lite on Arduino and Ultra-Low-Power Microcontrollers." O'Reilly Media. Zhang, Y., et al. (2019). "An Intelligent Water Quality Monitoring System Using Low-Power Microcontrollers and Machine Learning." Journal of Environmental Informatics. Additional Declarations The authors declare no competing interests. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6844035","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Systematic Review","associatedPublications":[],"authors":[{"id":467977277,"identity":"349b48ae-f753-43bd-b333-66c9e209b031","order_by":0,"name":"Hulisani Mukwevho","email":"data:image/png;base64,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","orcid":"","institution":"University of Johannesburg","correspondingAuthor":true,"prefix":"","firstName":"Hulisani","middleName":"","lastName":"Mukwevho","suffix":""},{"id":467977278,"identity":"c96677a9-31d2-4cdd-a3fb-3126c718f7b1","order_by":1,"name":"Unarine Mulaudzi","email":"","orcid":"","institution":"University of Johannesburg","correspondingAuthor":false,"prefix":"","firstName":"Unarine","middleName":"","lastName":"Mulaudzi","suffix":""},{"id":467977279,"identity":"ce4400e3-76d4-4670-9015-fc7ae0db0162","order_by":2,"name":"Bokang Motala","email":"","orcid":"","institution":"University of Johannesburg","correspondingAuthor":false,"prefix":"","firstName":"Bokang","middleName":"","lastName":"Motala","suffix":""},{"id":467977280,"identity":"8254422a-a594-4a14-b4b6-1ac84b1b1718","order_by":3,"name":"Sibusiso Moyo","email":"","orcid":"","institution":"University of Johannesburg","correspondingAuthor":false,"prefix":"","firstName":"Sibusiso","middleName":"","lastName":"Moyo","suffix":""}],"badges":[],"createdAt":"2025-06-07 16:56:00","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":true,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":true},"doi":"10.21203/rs.3.rs-6844035/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6844035/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":84389404,"identity":"85325aee-d566-4982-9d7d-2781dcde59f5","added_by":"auto","created_at":"2025-06-11 11:02:58","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":432866,"visible":true,"origin":"","legend":"\u003cp\u003eBibliometric Analysis of Study Search Keywords: (a) Overlay Visualization. (b) Network Visualization. (c) Density Visualization.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-6844035/v1/9b199cb158b5ce15235cf9c3.png"},{"id":84389223,"identity":"0635f82d-38fe-407f-a6f5-76fac9f48d86","added_by":"auto","created_at":"2025-06-11 10:54:58","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":113478,"visible":true,"origin":"","legend":"\u003cp\u003eProcedures and Stages of the Review.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-6844035/v1/e390433dae8838b6beab886c.png"},{"id":84388659,"identity":"481442f9-5d29-46ec-a467-47e3accf3360","added_by":"auto","created_at":"2025-06-11 10:46:57","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":33943,"visible":true,"origin":"","legend":"\u003cp\u003eFlow of Data Selection and Extraction.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-6844035/v1/a188592b676e965bf395a33e.png"},{"id":84388661,"identity":"9ac037cf-de46-4ed4-8dce-3b480fab2a9c","added_by":"auto","created_at":"2025-06-11 10:46:58","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":49497,"visible":true,"origin":"","legend":"\u003cp\u003eRisk of Bias Assessment Process for Non-Randomized Studies.\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-6844035/v1/543041e939d81feeb2629e01.png"},{"id":84389226,"identity":"5769034b-bb68-4491-8102-5ffa0cb97324","added_by":"auto","created_at":"2025-06-11 10:54:58","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":56722,"visible":true,"origin":"","legend":"\u003cp\u003eSystematic Review Process for machine learning on microcontroller-class hardware for biological sensing.\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-6844035/v1/862890ef79d7c1392da47559.png"},{"id":84388663,"identity":"5e3dc979-d76c-4a52-b474-852c6c217ae3","added_by":"auto","created_at":"2025-06-11 10:46:58","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":107243,"visible":true,"origin":"","legend":"\u003cp\u003eProposed PRISMA Flowchart.\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-6844035/v1/31e22c4457be10c0d04b839d.png"},{"id":84389224,"identity":"64ac27f2-1573-4276-833b-b9509b2ec2bb","added_by":"auto","created_at":"2025-06-11 10:54:58","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":28230,"visible":true,"origin":"","legend":"\u003cp\u003eAnnual distribution of included studies (2015–2024).\u003c/p\u003e","description":"","filename":"7.png","url":"https://assets-eu.researchsquare.com/files/rs-6844035/v1/66146c1d925ab79faddafd28.png"},{"id":84388667,"identity":"3021ede8-0831-4400-b081-e506c6324ab0","added_by":"auto","created_at":"2025-06-11 10:46:58","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":62081,"visible":true,"origin":"","legend":"\u003cp\u003eDistribution of study types and corresponding research categories.\u003c/p\u003e","description":"","filename":"8.png","url":"https://assets-eu.researchsquare.com/files/rs-6844035/v1/14eec479df89608a51cf122a.png"},{"id":84389227,"identity":"a403084e-3471-476c-9573-3327708b89f5","added_by":"auto","created_at":"2025-06-11 10:54:58","extension":"png","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":205802,"visible":true,"origin":"","legend":"\u003cp\u003eDistribution of tools and frameworks used in reviewed studies.\u003c/p\u003e","description":"","filename":"9.png","url":"https://assets-eu.researchsquare.com/files/rs-6844035/v1/aae707ee7b4f27e3d697039f.png"},{"id":84389229,"identity":"85174e2f-2acb-44f1-ab2d-eb0413b7d631","added_by":"auto","created_at":"2025-06-11 10:54:58","extension":"png","order_by":10,"title":"Figure 10","display":"","copyAsset":false,"role":"figure","size":291828,"visible":true,"origin":"","legend":"\u003cp\u003eClassification of \u003cstrong\u003ehardware \u003c/strong\u003eplatforms by category and usage share.\u003c/p\u003e","description":"","filename":"10.png","url":"https://assets-eu.researchsquare.com/files/rs-6844035/v1/3770f141d057a6b0c1fb1efb.png"},{"id":84389405,"identity":"7eb166b1-34a3-4c22-88f6-56c1e8c34078","added_by":"auto","created_at":"2025-06-11 11:02:58","extension":"png","order_by":11,"title":"Figure 11","display":"","copyAsset":false,"role":"figure","size":106545,"visible":true,"origin":"","legend":"\u003cp\u003eDistribution of ML task types and their corresponding categories.\u003c/p\u003e","description":"","filename":"11.png","url":"https://assets-eu.researchsquare.com/files/rs-6844035/v1/062ad90c6789b011e88f4347.png"},{"id":84388668,"identity":"ad089619-c6d6-4733-9d3e-69519edd7e77","added_by":"auto","created_at":"2025-06-11 10:46:58","extension":"png","order_by":12,"title":"Figure 12","display":"","copyAsset":false,"role":"figure","size":90654,"visible":true,"origin":"","legend":"\u003cp\u003eDistribution of ML model categories used in \u003cstrong\u003ebiological \u003c/strong\u003esensing systems. \u003cstrong\u003e(a)\u003c/strong\u003eSpecific architecture types, with deep learning architectures.\u003cstrong\u003e (b)\u003c/strong\u003e General classification of models by methodological group.\u003c/p\u003e","description":"","filename":"12.png","url":"https://assets-eu.researchsquare.com/files/rs-6844035/v1/8e5de5c5c599c000f5f9bf29.png"},{"id":84388672,"identity":"59605a0e-7702-4e9e-8d94-c68322826b0b","added_by":"auto","created_at":"2025-06-11 10:46:58","extension":"png","order_by":13,"title":"Figure 13","display":"","copyAsset":false,"role":"figure","size":67439,"visible":true,"origin":"","legend":"\u003cp\u003eDistribution of dataset types used in reviewed studies.\u003c/p\u003e","description":"","filename":"13.png","url":"https://assets-eu.researchsquare.com/files/rs-6844035/v1/3b9fc8eb60220c2cb4367452.png"},{"id":84388678,"identity":"3e76699c-76cc-4200-a895-269da3ae63ff","added_by":"auto","created_at":"2025-06-11 10:46:58","extension":"png","order_by":14,"title":"Figure 14","display":"","copyAsset":false,"role":"figure","size":105550,"visible":true,"origin":"","legend":"\u003cp\u003eDistribution \u003cstrong\u003eof\u003c/strong\u003eperformance constraints across reviewed studies.\u003c/p\u003e","description":"","filename":"14.png","url":"https://assets-eu.researchsquare.com/files/rs-6844035/v1/663877b14780109a3d13182c.png"},{"id":84388676,"identity":"ee7241db-7ed9-4baa-a9f4-3e230c572f5c","added_by":"auto","created_at":"2025-06-11 10:46:58","extension":"png","order_by":15,"title":"Figure 15","display":"","copyAsset":false,"role":"figure","size":179721,"visible":true,"origin":"","legend":"\u003cp\u003eReported hardware/system constraints across reviewed studies.\u003c/p\u003e","description":"","filename":"15.png","url":"https://assets-eu.researchsquare.com/files/rs-6844035/v1/56e4d6ba22d4b1c1a9ca63b1.png"},{"id":84389232,"identity":"ff1fa31f-fdfc-439d-9034-41508dd4e0d2","added_by":"auto","created_at":"2025-06-11 10:54:58","extension":"png","order_by":16,"title":"Figure 16","display":"","copyAsset":false,"role":"figure","size":183148,"visible":true,"origin":"","legend":"\u003cp\u003ePerformance evaluation themes across studies.\u003c/p\u003e","description":"","filename":"16.png","url":"https://assets-eu.researchsquare.com/files/rs-6844035/v1/bff2b3985738c4e291660408.png"},{"id":84388680,"identity":"8dfc15f9-31b1-46e4-80ed-5452b7f27493","added_by":"auto","created_at":"2025-06-11 10:46:58","extension":"png","order_by":17,"title":"Figure 17","display":"","copyAsset":false,"role":"figure","size":154996,"visible":true,"origin":"","legend":"\u003cp\u003eApplication domains for ML-enabled microcontroller systems.\u003c/p\u003e","description":"","filename":"17.png","url":"https://assets-eu.researchsquare.com/files/rs-6844035/v1/d12f991a8295f61eb0a0e9fa.png"},{"id":84389238,"identity":"cd0bca65-bef7-4eec-a407-fadd2c4e731b","added_by":"auto","created_at":"2025-06-11 10:54:59","extension":"png","order_by":18,"title":"Figure 18","display":"","copyAsset":false,"role":"figure","size":224398,"visible":true,"origin":"","legend":"\u003cp\u003eToolchains and software stacks used in reviewed studies.\u003c/p\u003e","description":"","filename":"18.png","url":"https://assets-eu.researchsquare.com/files/rs-6844035/v1/5b48f5fff0c55457bd6efca2.png"},{"id":84390162,"identity":"f1d70eee-6c58-489b-8b12-80f4aea399db","added_by":"auto","created_at":"2025-06-11 11:11:00","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":3417989,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6844035/v1/f8cba6cf-9b73-4c5f-8929-92301f4b40fc.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003eMachine Learning on Microcontrollers for Biological Sensing: A Systematic Review\u003c/p\u003e","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eWater quality analytics linked with sensor networks represents a innovative shift that improves the assessment of water ecosystems such as in fresh and coastal water systems. These technologies change how researchers analyze and provide crucial monitoring tools for water data science and they create environmental data which enables better safeguarding choices to achieveenhanced environmental protection (Philips, \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). The modern technique of water resource management depends on sensors that identify ecosystem conditions while foreseeing biological responses. Modern technology has produced significant changes in different water settings according to studies (Uniwill-Lily Desun, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). Things like toxic algal blooms get detected early through permanent monitoring equipment while systems forecast fish deaths and follow aquatic wildlife in sensitive areas. These techniques have proven to deliver considerable improvements in the discovery of pollution sources while also contributing strategies for restoration plans as well as habitat safety. Machine learning in water quality assessment uncovers its importance for real-time monitoring of water environments through its ability to interpret water quality data. The research on the biological effects of water quality parameters is still incomplete in developing areas despite the scientific progress (Zhang \u0026amp; Zhang, 2015). The protection of aquatic ecosystems has become vital to sustainable development because these systems sustain 40% of fish species along with 12% of world protein consumption from animals (Tedesco et al., \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2017\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eWater quality requires effective monitoring and data interpretation systems to address distinct problems facing freshwater ecosystems caused by humans. Biological impact monitoring systems consist of sensor networks and data transmission structures as their fundamental operational units. Most articles on IoT system-based biological impact assessment seem fictional resulting in most research listing laboratory-based methods (Miller et al., \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2023\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe rise in pollution and climate change places pressure on aquatic environments compared to pre-industrial conditions therefore requiring the implementation of monitoring techniques to analyze biological ecosystems (Dibyanshu et al., 2024). The practical usage of water quality index models across various ecosystems faces obstacles during full-scale monitoring system implementation. The implementation of monitoring systems becomes problematic due to their high cost and complex technology that prevents widespread (Lupi et al., \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). The field of water quality monitoring research now focuses on two areas regarding low-cost sensor networks and their technical advancements and ecological considerations tied with framework development to monitor biological impacts in developing nations in real time (Moudou et al., 2024). Sensor networks running during pollution occurrences generate insights about significant variables which influence how ecosystems respond. (Fascista, \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) This review systematizes eleven years of studies investigating biological water quality effects whilst documenting technological monitoring trends and obstacles.\u003c/p\u003e \u003cp\u003eThis review documents the biological effects and technological water quality monitoring solutions for comparison in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e against existing review methods. The review discusses sensor deployment with analytical programs which enables improved assessments of aquatic systems to detect biological responses. This review reveals data patterns from technology adoption through specific case studies and the resulting analysis which establishes necessary foundation for scientific investigation and practical applications improvement. Research will enhance knowledge about continuous monitoring's in aquatic ecosystem management techniques that are evolving.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eComparative Analysis of the Existing Review Works and Proposed Systematic Review on the Advantages and Application of Machine Learning on Microcontroller-Class Hardware for Biological Sensing\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRef.\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eContribution\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePros\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eCons\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e(Krokidis et al., \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2022\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eReviews sensor-based approaches for diagnoses, with ML integration and data monitoring.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eOutlines wearable sensor\u0026rsquo;s potential for early detection, the role of ML in diagnostic accuracy.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRequires real-world validation, and sensor data standardization.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e(Cascales et al., \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2021\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eStudies how ML models compensate for external effects\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eLow cost alternates while ML enhances the toughness\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eThe system need validation and the accuracy as well as the scalability depends on the training of Machine learning.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e(Shabani et al.,2021.)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eStudies Energy-autonomous water quality sensor using MFCs.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eRequires no external power and COD detection has high accuracy\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eValidation required, and the operation is limited to controlled conditions\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e(Ger\u0026oacute;s et al., \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2022\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eStudeis Integration of cameras and Arduino as well as rodent/wildlife detection\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNot biased to human and works in the dark. It is a solution that is open source to external aspects\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eWas not peer reviewed.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e(Diab \u0026amp; Rodriguez-Villegas, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2022\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eReviewed the TinyML wearables, outlining MCU and hardware/software compromises.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eOpen access licenced and the structure of the TinyML is practical\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eLess protical impemations with more theory based focus,and the recommendations are outdated.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e(Schizas et al., \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2022\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eBrief Machine Learning review that integrates 5G and cloud.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAnalysis of network integration, and descriptive ML model creation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eConsists primarily of theory and lacks depth in areas.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e(Haque Zim et al., \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2021\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eProposed Presents the development of a network structure, briefs the usage of microcontroller with three operating modes.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eDemostrates application of neural networks in robotics for certain tasks.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eLimiteddetails and advantages and disadvantgaes of the approach found in the abstract\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e(Sinha et al., \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2021\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eExplores the enhancements of pH sensors in its temperature,etc characteristics\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eDrives innovation, supports product development strategies.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eThe abstract is concise and doesn't provide extensive details on the specific ML algorithms used or the extent of performance improvement.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e(Venkatachalam et al., \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2024\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eProposes an efficient four-layered convolutional neural network (CNN) model for real-time person identification through gait analysis using sensors on edge devices. The model, trained on a public and custom dataset, achieves high accuracy with a small size storagr\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePresents a compact and efficient CNN model suitable for real-time gait-based person identification on resource-constrained devices.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eThe abstract does not detail the specific architecture or layers of the proposed CNN.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e(Merenda et al., \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2020\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eReviews the main techniques for executing machine learning models on resource-constrained hardware in the Internet of Things\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eHas a comprehensive review of machine learning techniques for IoT devices.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eThe abstract is broad and doesn't delve into the specifics of the reviewed techniques or their proportional performance.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eWe found several research needs in biological\u0026ensp;water quality monitoring via sensor technologies in our systematic review. While many studies consider\u0026ensp;sensor networks for water quality, not many focus on low-cost, scalable systems tailored to limited resource settings. In general, current studies\u0026ensp;tend to focus on improvements of sensors and data transmission and disregard human, ecological, and operational aspects that are also key for a sustainable monitoring system.\u003c/p\u003e \u003cp\u003eThe current systems lack proper connectivity between biological impact assessment measurements and sensor-based data systems. The ecological significance of research findings weakens because physical-chemical parameter assessments like DO and pH continue to be studied separately from biological response measurements (Emerson, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Long-term changes within real ecosystems cannot be properly studied because researchers rely on laboratory data and cross-sectional data (Chen et al., \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). Research should focus on developing long-term research approaches united with ecological indicators from sensors which will enhance ecosystem defense and management strategies.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cdiv id=\"Sec2\" class=\"Section2\"\u003e \u003ch2\u003e1.1. Research questions\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThis review investigates the implementation of machine learning models on microcontroller-class hardware for biological sensing across varied environmental and health contexts. While numerous studies highlight advancements in embedded sensing systems, gaps remain in performance optimization, hardware selection, dataset standardization, and application-specific integration. This review addresses the following questions:\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eHow are ML models optimized for low-latency and memory-efficient execution on microcontroller platforms in biological sensing applications?\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eWhat are the dominant toolchains, hardware architectures, and datasets used in deploying embedded ML for biological tasks, and how do they influence system performance?\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eIn what ways are ML-enabled microcontrollers applied across domains such as health monitoring, water quality, and biometrics, and what domain-specific constraints emerge?\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eWhat technical and methodological limitations (e.g., underreporting of constraints, performance metrics) hinder effective deployment in resource-constrained environments?\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eHow do real-time inference capabilities and edge suitability influence the design and evaluation of embedded ML systems for biological sensing?\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e1.2. Hypotheses Development\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThe research questions explored in this systematic review give rise to several hypotheses centered on system constraints, methodological practices, and application performance in embedded ML for biological sensing. These hypotheses reflect observed trends across healthcare, environmental, and biometric monitoring systems using microcontroller-class hardware:\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eH1: Inconsistent reporting of performance metrics and hardware limitations (e.g., RAM, power, latency) leads to reduced reproducibility and deployment inefficiency in embedded ML systems.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eH2: The dominance of custom datasets in biological sensing (67.27%) limits cross-study comparability and hinders the creation of generalized ML deployment benchmarks.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eH3: Real-time inference and edge deployment suitability are critical drivers in the selection of ML models and hardware stacks, particularly in healthcare and environmental sensing applications.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eH4: The integration of deep learning models on constrained platforms (e.g., CNNs on Arduino/ESP32) is feasible but demands tailored optimization strategies such as quantization and pruning.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eH5: The limited use of accuracy as a primary evaluation metric (7.27%) and the absence of standardized constraints (32%) undermine confidence in reported system performance.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e1.3. Rationale\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThe rationale for this systematic review is to critically assess the technological landscape of machine learning applications on microcontroller-class devices for biological sensing. Unlike traditional water quality studies, this review expands the scope to include diverse domains\u0026mdash;such as healthcare, biometrics, and environmental monitoring\u0026mdash;where embedded systems are increasingly deployed. With the rise of TinyML and resource-constrained AI, it becomes essential to evaluate how these systems manage performance, cost, and contextual applicability. By synthesizing findings from 2015 to 2025, the review uncovers implementation patterns, identifies deployment barriers, and informs future design of low-power, domain-specific ML solutions for real-time biological monitoring.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e1.4. Objectives\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThe primary objective of this systematic review is to evaluate the state of machine learning deployment on microcontroller-class hardware for biological sensing applications. Specifically, the review aims to:\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eAnalyze the types of ML models, hardware platforms, and toolchains used in resource-constrained biological monitoring.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eAssess the distribution and characteristics of datasets employed in ML model training and evaluation.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eDocument the performance constraints and optimization strategies implemented in real-world systems.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eExamine the application areas (e.g., healthcare, water monitoring, biometrics) and the unique challenges each domain presents.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eIdentify methodological gaps in system reporting, performance benchmarking, and dataset standardization.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eProvide data-driven recommendations to enhance system design, reproducibility, and scalability of embedded ML for biological tasks.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e1.5. Research Contributions\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThis review makes several key contributions to the field of embedded biological sensing:\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eIt provides a detailed breakdown of 60 studies based on ML model types, toolchains, dataset usage, hardware selection, and application domains, offering a holistic understanding of system architectures and deployment patterns.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eThe review highlights critical issues such as the underreporting of constraints (32%), low emphasis on performance benchmarking (6% with no metrics), and overreliance on non-standardized custom datasets (67.27%).\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eIt introduces an analytical framework for evaluating embedded ML systems in biological sensing, facilitating cross-domain assessment and guiding optimization efforts for real-world use cases.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eExamining multiple application domains, the review demonstrates how system design varies across healthcare, water quality, environmental monitoring, and biometrics, establishing a multi-sector perspective for future integration efforts.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e1.6. Research Novelty\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eTo the best of the authors\u0026rsquo; knowledge, this is the first systematic review to comprehensively examine the integration of machine learning models with microcontroller-class hardware specifically for biological sensing applications across multiple domains. The novelty of this work lies in:\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eUnlike previous reviews limited to water quality or wearable health monitoring, this review spans a broader spectrum including environmental sensing, biometric authentication, and infrastructure monitoring.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eIt provides a three-dimensional analysis connecting ML model types, hardware constraints, and dataset usage, offering practical insight into how performance trade-offs are managed in real-world deployments.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eThe review uniquely emphasizes real-time inference and edge readiness as central metrics, revealing how system performance goals drive design across use cases.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eBased on findings, the review proposes an integration roadmap that aligns model selection, toolchains, and performance strategies with resource-constrained deployment scenarios.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"2. Materials and Methods","content":"\u003cp\u003eIn this section, we describe the approach we used to carry out our systematic review on the use of machine learning on microcontroller-class hardware for biological sensing applications. Our review focuses on research published between 2015 and 2025, a period that captures major developments in TinyML, edge AI, and wearable sensing technologies. To the best of our knowledge, there has not yet been a dedicated review that specifically brings together both the embedded hardware and biological sensing perspectives within this timeframe, which makes our contribution both relevant and timely.\u003c/p\u003e\n\u003cp\u003eWe collected our literature by searching trusted online databases, including Scopus, Google Scholar, and Web of Science. To make sure we captured the most relevant studies, we carefully chose a set of keywords related to machine learning, microcontrollers, biological sensing, embedded systems, and low-power design. After gathering the initial results, we took the time to review each paper and selected only those that directly focused on practical implementations of ML models on microcontroller-class devices for biological sensing tasks. This process helped us build a strong and focused collection of studies that could genuinely inform and enrich our findings.\u003c/p\u003e\n\u003cp\u003e2.1. Eligibility criteria\u003c/p\u003e\n\u003cp\u003eFor this review, we systematically considered all peer-reviewed and published research works that were directly relevant to the study of machine learning applications on microcontroller-class hardware for biological sensing tasks. To keep the review focused and consistent, we only included articles that were published in English between 2015 and 2025. We applied clear inclusion criteria to make sure that only the most relevant and high-quality studies were selected. Specifically, we only considered research papers that focused on deploying ML models on resource-constrained, low-power hardware and involved biological sensing tasks, such as health monitoring, environmental sensing, or wearable devices. Papers that did not include practical implementations, experimental results, or did not focus on embedded ML systems were excluded from our final analysis. This careful selection process helped us ensure that our review stayed aligned with the core aims of our study (Khanyi et al., 2024; Thobejane et al., 2024; Skosana et al, 2024; Mkhize et al., 2025). The inclusion and exclusion criteria for this study are tabulated as in Table 2.\u003c/p\u003e\n\u003cp\u003eTable 2. Proposed Inclusion and Exclusion Criteria.\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"696\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 18.9655%;\"\u003e\n \u003cp\u003eCriteria\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 41.092%;\"\u003e\n \u003cp\u003eInclusion\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 39.9425%;\"\u003e\n \u003cp\u003eExclusion\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 18.9655%;\"\u003e\n \u003cp\u003eTopic\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 41.092%;\"\u003e\n \u003cp\u003eArticle papers focusing on the application of machine learning on microcontroller-class hardware for biological sensing tasks\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 39.9425%;\"\u003e\n \u003cp\u003eArticle papers not focusing on machine learning applications on microcontroller-class hardware for biological sensing\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 18.9655%;\"\u003e\n \u003cp\u003eResearch Framework\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 41.092%;\"\u003e\n \u003cp\u003eArticles must include a research framework, methodology, or experimental setup related to embedded ML models for biological sensing\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 39.9425%;\"\u003e\n \u003cp\u003eArticles without a clear research framework, methodology, or experimental validation for embedded ML in biological sensing\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 18.9655%;\"\u003e\n \u003cp\u003eLanguage\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 41.092%;\"\u003e\n \u003cp\u003eMust be written in English\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 39.9425%;\"\u003e\n \u003cp\u003eArticles published in languages other than English\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 18.9655%;\"\u003e\n \u003cp\u003ePeriod\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 41.092%;\"\u003e\n \u003cp\u003eArticles between 2015 to 2025\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 39.9425%;\"\u003e\n \u003cp\u003eArticles outside 2015 and 2025\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e2.2. Information sources\u003c/p\u003e\n\u003cp\u003eA systematic search of online databases was conducted to identify relevant studies for this review. The databases Scopus, Google Scholar, and Web of Science were utilized due to their comprehensive coverage of peer-reviewed literature (Khanyi et al., 2024; Thobejane et al., 2024; Skosana et al, 2024; Mkhize et al., 2025). Each database was thoroughly searched using a combination of keywords related to the study topic, ensuring that the most pertinent research articles were captured. Scopus provided access to a broad range of scientific journals and conference papers, while Google Scholar enabled the inclusion of gray literature and dissertations that might not be indexed elsewhere. Web of Science was used to cross-reference and ensure the robustness of the selected studies by providing citation data and impact factors of the journals. The search results from these databases formed the core of the literature review, ensuring a well-rounded and exhaustive collection of research works.\u003c/p\u003e\n\u003cp\u003e2.3. Search strategy\u003c/p\u003e\n\u003cp\u003eFor this review, we collected research articles from well-known online databases, focusing on how machine learning is being used on microcontroller-class hardware for biological sensing. We mainly searched through three platforms: Google Scholar, Scopus and Web of Science (Khanyi et al., 2024; Thobejane et al., 2024; Skosana et al, 2024; Mkhize et al., 2025).\u003c/p\u003e\n\u003cp\u003eTo make sure we found papers closely related to our topic, we used a specific set of keywords: (\u0026quot;Machine Learning\u0026quot; AND \u0026quot;Microcontroller\u0026quot; AND (\u0026quot;Biological Sensing\u0026quot; OR \u0026quot;Health Monitoring\u0026quot; OR \u0026quot;Agricultural Technology\u0026quot;) AND (\u0026quot;Embedded Systems\u0026quot; OR \u0026quot;TinyML\u0026quot; OR \u0026quot;IoT\u0026quot;) AND (\u0026quot;Low Power\u0026quot; OR \u0026quot;Resource-Constrained Optimization\u0026quot;) AND \u0026quot;Algorithm Development\u0026quot; AND \u0026quot;Benchmarking\u0026quot;). These keywords were chosen based on the areas we identified during our initial research, which are shown in the heatmap in Figure 1. The heatmap helped us focus on the most common and important topics.\u003c/p\u003e\n\u003cp\u003eWe limited the search to papers published between 2015 and 2025 to make sure we covered the latest work in this fast-moving field. Our search returned about 142 papers from Google Scholar, 4266 from Scopus, and 22 from Web of Science. After getting all the results, we carefully filtered the papers by checking if they involved real-world biological sensing applications and if they tested machine learning models on actual microcontroller-class devices. We left out papers that were only theoretical or not relevant to our topic. The Bibliometric Analysis of Study Search Keywords is illustrated in Figure 1.\u003c/p\u003e\n\u003cp\u003eTable 3.\u0026nbsp;Results Achieved from Literature Search.\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"524\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 33.3333%;\"\u003e\n \u003cp\u003eNo.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 33.3333%;\"\u003e\n \u003cp\u003eOnline Repository \u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 33.3333%;\"\u003e\n \u003cp\u003eNumber of results\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 33.3333%;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 33.3333%;\"\u003e\n \u003cp\u003eGoogle Scholar\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 33.3333%;\"\u003e\n \u003cp\u003e142\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 33.3333%;\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 33.3333%;\"\u003e\n \u003cp\u003eWeb of Science\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 33.3333%;\"\u003e\n \u003cp\u003e22\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 33.3333%;\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 33.3333%;\"\u003e\n \u003cp\u003eScopus\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 33.3333%;\"\u003e\n \u003cp\u003e4266\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 33.3333%;\"\u003e\n \u003cp\u003eTotal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 33.3333%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 33.3333%;\"\u003e\n \u003cp\u003e4430\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e2.4. Selection process\u003c/p\u003e\n\u003cp\u003eWe used a careful step-by-step process to pick the best studies. First, four researchers read the titles and summaries of the first 60 articles separately. When we disagreed about an article, we talked about it until we all agreed. Then we worked together to review the rest of the articles the same way. Next, we read the full articles that made it past the first round to make sure they really talked about sustainable practices in project management (Khanyi et al., 2024; Thobejane et al., 2024; Skosana et al, 2024; Mkhize et al., 2025). If we still couldn\u0026apos;t agree on an article, a fifth researcher helped decide. This way, we made sure we only used the most relevant and useful studies. This systematic approach, illustrated in Figure 2, ensured methodological consistency while accommodating multiple perspectives, ultimately yielding a robust final selection of studies directly relevant to our research objectives\u003c/p\u003e\n\u003cp\u003e2.5. Data collection process\u003c/p\u003e\n\u003cp\u003eTo ensure the accuracy and reliability of the extracted data, we implemented a rigorous, multi-stage validation process. Three independent reviewers systematically collected data from each study under the supervision of a fourth reviewer, who acted as both an arbitrator and subject matter expert. To maintain consistency, we employed a standardized data extraction form. All data were manually extracted (without automation tools) and cross-verified by reviewers to minimize transcription errors.\u003c/p\u003e\n\u003cp\u003eDiscrepancies in data interpretation such as conflicting outcomes or methodological ambiguities were resolved through iterative team discussions until consensus was reached (Khanyi et al., 2024; Thobejane et al., 2024; Skosana et al, 2024; Mkhize et al., 2025). For missing information, we conducted in-depth reviews of external materials (appendices, datasets, or linked publications) and consulted external experts when necessary. To address duplicate or overlapping studies (e.g., multiple reports from the same project), we prioritized the most recent and comprehensive publications (2015\u0026ndash;2025) and reconciled inconsistencies by comparing methodologies and results. Non-English studies were excluded to prevent language-related biases, ensuring uniformity in analysis. This process, illustrated in Figure 3, guaranteed that only high-quality, verifiable data informed our findings.\u003c/p\u003e\n\u003cp\u003e2.6. Data items\u003c/p\u003e\n\u003cp\u003eIn this section We\u0026apos;re looking at all the key details that make machine learning on tiny chips for biological sensing. First, we will check the basics of how accurate these systems are, how fast they run and how much power they consume. We will note exactly which microcontrollers are being used and how they handle sensors. But we are also digging into what really matters such as, can these devices reliably detect viruses or glucose levels as used in biological sensing systems? Do they work outside a lab? Crucially, what is the price tag for small businesses wanting to use this technology? Beyond the specs, we are tracking how these studies were done. Did researchers share their code? Were the tests realistic? This way we can separate good lab results from technology that works in the real world. Mapping out all these pieces, we\u0026apos;ll show where this field is really delivering and where it\u0026rsquo;s still just under review (Myataza et al, 2024; Gumede et al, 2024; Mudau et al., 2024; Mtjilibe et al, 2024).\u003c/p\u003e\n\u003cp\u003e2.6.1 Data Collection Method\u003c/p\u003e\n\u003cp\u003eWe wanted to understand whether these microcontroller-based biosensors deliver in real-world conditions, not just in pristine labs. Our review focused on four key questions. How reliably do they detect biological signals when faced with real-world complexities like variable temperatures or imperfect samples? Can they operate long enough on limited power to be genuinely useful in field applications? Do their results lead to better decisions, like earlier disease detection or more accurate environmental monitoring? Importantly, are they designed for real users, with intuitive interfaces that don\u0026apos;t require specialist training? By applying these practical tests, we separated truly impactful innovations from those only working under ideal conditions, highlighting systems that deliver both technical excellence and real-world usability (Myataza et al, 2024; Gumede et al, 2024; Mudau et al., 2024; Mtjilibe et al, 2024).\u003c/p\u003e\n\u003cp\u003eWe did not just count technical specs, we wanted to know if these microcontroller biosensors improve outcomes. For performance, we examined how consistently they detect real biological signals when faced with imperfect field conditions like whether a glucose monitor works as well after eight hours in a pocket as it does in a climate-controlled lab. We tracked operational durability too: Can it maintain accuracy through temperature swings? Does it stay reliable when running on limited power for weeks? Most importantly, we looked at real-world utility, do the results help users make better decisions? A system that detects water contamination 12 hours faster is only valuable if communities can act on that data. We also prioritized designs that real people can use, where intuitive interfaces matter as much as algorithmic brilliance. This approach helped us spotlight innovations that deliver both scientific merit and practical value.\u003c/p\u003e\n\u003cp\u003eWe assessed how these technologies reduce operational expenditures while improving diagnostic outcomes, with particular attention to cost-benefit ratios in real-world healthcare applications. Studies demonstrating measurable improvements in resource utilization and clinical workflow efficiency were prioritized, revealing how embedded machine learning creates value beyond technical performance metrics alone.\u003c/p\u003e\n\u003cp\u003e2.6.2 Definition of Collected Data Variables\u003c/p\u003e\n\u003cp\u003eIn addition to these primary outcomes, we collected and categorized key variables to enable comprehensive analysis. The extracted data encompassed three primary dimensions: technical specifications (including sensor integration methods, computational constraints, and model optimization techniques), performance metrics such as detection accuracy, latency and power efficiency under varying conditions, and implementation factors such as notably calibration requirements, environmental robustness and usability considerations. Emphasis was placed on capturing both quantitative measurements and qualitative implementation challenges to assess real-world viability beyond controlled laboratory results. These variables were systematically documented through rigorous examination of peer-reviewed studies from Google Scholar, SCOPUS and Web of Science, with inclusion criteria prioritizing studies that reported empirical results from functional prototypes. Maintaining this structured yet nuanced approach to data collection, we ensure our review provides both technical depth and practical insights for researchers and practitioners developing embedded biosensing solutions (Myataza et al, 2024; Gumede et al, 2024; Mudau et al., 2024; Mtjilibe et al, 2024).\u003c/p\u003e\n\u003cp\u003eTable 4. Data Variables Collected.\u003c/p\u003e\n\u003cdiv align=\"\"\u003e\n \u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"699\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 25.0358%;\"\u003e\n \u003cp\u003eField\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 74.9642%;\"\u003e\n \u003cp\u003eDescription\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 25.0358%;\"\u003e\n \u003cp\u003eStudy characteristics\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 74.9642%;\"\u003e\n \u003cp\u003eGeographic location of testing application domain (healthcare, agriculture), and scale of deployment (lab prototype, field trial, commercial product). Includes publication year and study duration.\u003c/p\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 25.0358%;\"\u003e\n \u003cp\u003eDevice specifications\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 74.9642%;\"\u003e\n \u003cp\u003eMicrocontroller model (e.g., ESP32), sensor type (electrochemical, optical), and connectivity (Bluetooth, Wifi). Power source (battery, solar) and energy consumption metrics.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 25.0358%;\"\u003e\n \u003cp\u003eML implementation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 74.9642%;\"\u003e\n \u003cp\u003eMachine learning techniques used (e.g., CNN, RNN), model optimization methods (quantization, pruning) and framework (TensorFlow Lite, Edge Impulse). Includes training dataset size and diversity.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 25.0358%;\"\u003e\n \u003cp\u003eEconomic factors\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 74.9642%;\"\u003e\n \u003cp\u003eDevelopment costs, production scalability, and maintenance requirements (calibration frequency and expected lifespan).\u003c/p\u003e\n \u003cp\u003eDevelopment costs, production scalability, and maintenance requirements (calibration frequency and expected lifespan).\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 25.0358%;\"\u003e\n \u003cp\u003eExternal influences\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 74.9642%;\"\u003e\n \u003cp\u003eTechnical skill requirements for deployment(e.g., FDA approval for medical use), and market competition with existing solutions.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003e2.7. Study risk of bias assessment\u003c/p\u003e\n\u003cp\u003eTo ensure the integrity of our findings, we established rigorous quality standards for study inclusion in this systematic review. Each publication underwent careful evaluation based on several critical factors such as the completeness of technical specifications provided, the validity of biological testing protocols, the thoroughness of power efficiency reporting, the robustness of environmental stress testing, and the availability of materials needed for replication. Scrutiny was applied to studies with industry sponsorship to safeguard against potential performance inflation (Myataza et al, 2024; Gumede et al, 2024; Mudau et al., 2024; Mtjilibe et al, 2024). Through this meticulous screening process which resulted in the exclusion of approximately twenty percent of initially identified studies\u0026mdash;we prioritized research demonstrating both technical excellence and practical applicability. The selected studies collectively provide reliable insights into the real-world performance of machine learning implementations on microcontroller platforms for biological sensing applications.\u003c/p\u003e\n\u003cp\u003e2.8. Synthesis methods\u003c/p\u003e\n\u003cp\u003eThis flow chart below in Figure 5 illustrates the systematic approach used in our review of machine learning on microcontroller-class hardware for biological sensing Starting with Study Selection Process, we identify, and screen studies based on set eligibility criteria. Next, Data Standardization involves converting, cleaning, aligning sensor data and handling any missing values. In the Data Analysis phase, we present the data in tables or graphs and perform initial analyses. The flow then moves to Heterogeneity Assessment, where we evaluate variability through subgroup or sensitivity analyses. Finally, Bias Assessment ensures we identify potential biases and maintain transparency in our methods. This structured approach ensures a thorough and reliable review process (Myataza et al, 2024; Gumede et al, 2024; Mudau et al., 2024; Mtjilibe et al, 2024).\u003c/p\u003e\n\u003cp\u003eIn this systematic review on the machine learning on microcontroller-class hardware for biological sensing we employed rigorous synthesis methods to ensure that our results were robust, transparent, and reproducible. To determine the eligibility of studies for synthesis, we meticulously tabulated the characteristics of each study and compared them against our predefined synthesis groups on hardware class, machine learning technique employed and biological sensing type. This approach allowed us to include only the most relevant studies, ensuring that our findings were both valid and aligned with the review\u0026apos;s objectives. In preparing the data for synthesis, we addressed missing summary statistics through imputation techniques and conducted necessary data conversions to maintain consistency across studies in terms of performance metrics, hardware specifications and sensor modalities. The results were then presented using a combination of structured tables and forest plots, which provided a clear visual representation of the energy consumption, accuracy rate, inference and memory footprints allowing effective identification of trends and outliers.\u003c/p\u003e\n\u003cp\u003eThe synthesis of results was conducted using a random-effects meta-analysis model, with subgroup analyses explicitly focusing on geographic and economic contexts to understand their influence on system performance. This approach provided nuanced insights into how different deployment context and hardware limitations interact with success of machine learning application for biological sensing, which was further explored through subgroup analyses and meta-regressions. These analyses helped us identify potential sources of heterogeneity, such as type of microcontroller used, the complexity of biological data, or machine learning model architecture employed. \u0026nbsp;Additionally, sensitivity analyses were performed to assess the robustness of the synthesized results, ensuring that our conclusions were well-supported by stable and reliable evidence. Through this comprehensive approach, we were able to provide a meaningful aggregation of the evidence, offering valuable insights for researchers, practitioners and developers interested in leveraging machine learning for efficient deployment on microcontroller-class hardware in biological sensing applications.\u003c/p\u003e\n\u003cp\u003eTo determine study eligibility for inclusion in our systematic review on machine learning application on microcontroller-class hardware for biological sensing, each study was carefully evaluated for its relevance and alignment with the review\u0026rsquo;s objectives. We manually assessed and compared each study\u0026rsquo;s characteristics\u0026mdash;such as hardware specification, biological sensing target, machine learning model used and performance outcomes\u0026mdash;against our predefined synthesis groups. A matrix was created to visually compare the scope and methodologies of the studies with our inclusion criteria, ensuring a comprehensive and objective evaluation. This process ensured that only studies directly pertinent to the review topic were included, thus enhancing the review\u0026rsquo;s overall rigor and reliability of the review.\u003c/p\u003e\n\u003cp\u003e2.8.2. Data Preparation for Synthesis\u003c/p\u003e\n\u003cp\u003eIn this review, the methods used involved converting or standardizing data collected from various studies to ensure consistency before synthesis. For example, when performance metrics were reported differently across studies, algebraic manipulations were employed to convert these into a uniform scale, such as converting odds ratios to risk ratios where appropriate. Additionally, handling missing data was a critical aspect of the analysis. Missing summary statistics, such as standard deviations, energy consumption or memory usage, were imputed using established statistical methods like multiple imputation. This approach ensured that the dataset was comprehensive and robust, allowing for a more accurate and reliable analysis on machine learning performance on microcontroller-class hardware in biological sensing application (Myataza et al, 2024; Gumede et al, 2024; Mudau et al., 2024; Mtjilibe et al, 2024).\u003c/p\u003e\n\u003cp\u003e2.8.3. Tabulation and Visual Display of Results\u003c/p\u003e\n\u003cp\u003eResults from individual studies and synthesis efforts were organized using both tabular and graphical methods to enhance clarity and facilitate comparison. Tabular structures were employed to present the data in a structured format, where outcomes were organized by performance\u0026mdash;such as accuracy, memory used, and efficiency, and within each category, studies were ordered from lowest to highest risk of bias. This organization allowed for easy comparison across studies and highlighted the most reliable evidence. Additionally, graphical methods, specifically forest plots, were used as the principal tool for visually displaying meta-analysis results (Myataza et al, 2024; Gumede et al, 2024; Mudau et al., 2024; Mtjilibe et al, 2024). These plots showcased effect estimates and confidence intervals for each study alongside a summary estimate. The studies in the forest plots were ordered based on effect size, for example classification accuracy or year of publication, helping to reveal trends over time and across different research focuses.\u003c/p\u003e\n\u003cp\u003e2.8.4. Synthesis of Results\u003c/p\u003e\n\u003cp\u003eDuring our manual search on online repositories such as Google Scholar, Scopus, and Web of Science, we carefully reviewed and synthesized the results of relevant studies. The approach to data synthesis was guided by the nature of the data and the degree of variability observed across studies. Based on the findings from our search, we manually assessed the applicability of both fixed-effects and random-effects models, depending on the level of heterogeneity among study results. The selection of the model was determined by the characteristics of the data and our assumptions about the consistency of machine learning performance metrics across different microcontroller platforms and biological sensing tasks. After exporting the data to Excel, we created charts to visually inspect the performance metrics such as accuracy, memory usage and power consumption, allowing us to identify patterns of variability and potential heterogeneity across the studies. This initial visual inspection provided an overview of how study results differed from one another, facilitating a more nuanced analysis.\u003c/p\u003e\n\u003cp\u003e2.8.5. Exploring Causes of Heterogeneity\u003c/p\u003e\n\u003cp\u003eSubgroup analyses and meta-regression were performed to explore potential sources of heterogeneity, such as differences in study settings, hardware configuration, machine learning model types, or biological sensing targets. Specific analyses focused on factors like the type of microcontroller used such as ESP32, Arduino so on, the biological sensing application such as pathogen detection, and the complexity of the machine learning model, all of which were examined to assess their impact on the performance outcomes such as accuracy, energy efficiency and inference time. These methods helped to identify underlying patterns and relationships that contributed to the overall variability observed across the studies.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e2.8.6. Sensitivity Analyses\u003c/p\u003e\n\u003cp\u003eSensitivity analyses were employed to evaluate the robustness of the synthesis results in relation to various assumptions and methodological decisions made during the review process. These analyses included testing the impact of excluding studies at high risk of bias and using alternative statistical models to ensure that the conclusions were not unduly influenced by specific studies or analytical techniques. This approach helped to confirm the reliability and validity of the findings by addressing potential sources of bias and ensuring that the results were consistent across different analytical scenarios, strengthening confidence in trends identified for machine learning deployment on microcontroller-class hardware on biological sensing application (Myataza et al, 2024; Gumede et al, 2024; Mudau et al., 2024; Mtjilibe et al, 2024).\u003c/p\u003e\n\u003cp\u003e2.9. Reporting bias assessment\u003c/p\u003e\n\u003cp\u003eIn conducting our systematic review on the deployment and performance of machine learning on microcontroller-class hardware for biological sensing application, it was crucial to assess the risk of bias due to potentially missing results, particularly those arising from reporting biases such as selective publication or selective reporting of outcomes. We recognized that these biases could significantly impact the validity and reliability of our synthesis, and thus, we employed a thorough and methodical approach to address this concern. Our assessment of reporting bias was conducted using a combination of well-established statistical and graphical methods. We opted for the use of contour-enhanced funnel plots, a powerful visual tool that allowed us to detect asymmetries in the data. These plots were carefully inspected to identify any potential publication bias by highlighting areas where studies might be missing due to bias versus those missing due to chance. The inclusion of statistical significance contours provided us with a clear and intuitive way to differentiate between these two scenarios, offering a robust visual representation of potential biases.\u003c/p\u003e\n\u003cp\u003eFor this assessment, we chose not to develop new tools but rather relied on standard, proven techniques documented extensively in the literature. The methodological rigor of the tools we used was integral to our process. Contour-enhanced funnel plots provided a straightforward yet effective method to visually assess the distribution of studies, allowing us to identify and account for potential biases in our synthesis. The assessment process was designed to minimize subjective bias, multiple independent reviewers were involved in evaluating the studies, and any discrepancies between their assessments were resolved through consensus discussions or, when necessary, by consulting a methodological expert. This collaborative approach ensured that the interpretation of results was balanced and unbiased. We intentionally did not use automation tools for assessing reporting bias in this review. Instead, we opted for a manual approach, utilizing tools such as Excel for creating charts and plots. This hands-on method allowed us to carefully analyse and visualize the data, ensuring a detailed and thorough examination. By manually inspecting the data, we ensured that no subtle patterns or potential biases were overlooked.\u003c/p\u003e\n\u003cp\u003eTo further strengthen the review, we conducted comprehensive manual searches across multiple online repositories, including Google Scholar, Scopus, and Web of Science. This approach enabled us to cross-reference data from different studies and sources, addressing any discrepancies and reinforcing the robustness of our conclusions. These manual searches were critical in ensuring that our synthesis was based on the most complete and accurate data available. Given the unique context of machine learning application on resource-constrained microcontroller, we adapted the standard methods for assessing reporting bias to fit this specific field. Studies in embedded biological sensing often exhibit different reporting and validation patterns compared to fields such computer vision or clinical trials, necessitating this adaptation to ensure relevance and accuracy. By tailoring our methods to align with the characteristics of the studies we reviewed, we ensured that our analysis was both contextually appropriate and methodologically sound. To promote transparency and replicability, all methods and approaches used in our assessment have been thoroughly documented and made publicly accessible in the supplementary materials of this review (Myataza et al, 2024; Gumede et al, 2024; Mudau et al., 2024; Mtjilibe et al, 2024). This commitment to openness allows other researchers to replicate our analysis or build upon it in future studies, thereby contributing to the overall rigor and reliability of research in the field of machine learning for biological sensing on microcontroller-class hardware.\u003c/p\u003e\n\u003cp\u003e2.10. Certainty assessment\u003c/p\u003e\n\u003cp\u003eThe reviewed literature was evaluated based on five quality assessment (QA) criteria to ensure rigor and relevance (Myataza et al, 2024; Gumede et al, 2024; Mudau et al., 2024; Mtjilibe et al, 2024):\u0026nbsp;\u003c/p\u003e\n\u003cul\u003e\n \u003cli\u003eQA1: The clarity and explicitness of the research aim related to machine learning implementation on microcontroller.\u003c/li\u003e\n \u003cli\u003eQA2: The specification and transparency of hardware configuration, sensor types and data collection methods.\u003c/li\u003e\n \u003cli\u003eQA3: The clear definition and explanation of the machine learning model, training processes, and deployment techniques.\u003c/li\u003e\n \u003cli\u003eQA4: The application of a well-defined and appropriate research methodology for evaluating biological sensing performance.\u003c/li\u003e\n \u003cli\u003eQA5: The contribution of the research findings to the enhancement of existing literature in deploying machine learning on resource constrained hardware for biological applications.\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eThe certainty assessment responses are rated on a scale from zero (0) to one (1). A \u0026apos;No\u0026apos; response is assigned \u0026apos;0\u0026apos; points, a score of \u0026apos;0.5\u0026apos; is given if the criterion is \u0026apos;Partially\u0026apos; met, and \u0026apos;1\u0026apos; point is assigned for a \u0026apos;Yes\u0026apos; response. All five criteria are scored using this scale. Each piece of literature under review can receive a total score between 0 and 5 points. The results of the certainty assessment for the collected literature on the deployment on machine on microcontroller-class hardware for biological sensing are presented in Table 5.\u003c/p\u003e\n\u003cp\u003eTable 5. Certainty Assessment Results for Collected Literature on Machine Learning on Microcontroller-Class Hardware for Biological Sensing.\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"100%\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 54.7368%;\"\u003e\n \u003cp\u003eRef.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 5.26316%;\"\u003e\n \u003cp\u003eQA1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6.31579%;\"\u003e\n \u003cp\u003eQA2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 5.26316%;\"\u003e\n \u003cp\u003eQA3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6.31579%;\"\u003e\n \u003cp\u003eQA4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 5.26316%;\"\u003e\n \u003cp\u003eQA5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6.31579%;\"\u003e\n \u003cp\u003eTotal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10.5263%;\"\u003e\n \u003cp\u003e% grading\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 54.7368%;\"\u003e\n \u003cp\u003e(Fernandez et al., 2017; Saravanan et al., 2018; Cruz et al., 2019; Gokulanathan et al., 2019; Covaciu \u0026amp; Iordan, 2022; Dai \u0026amp; Tan, 2024)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 5.26316%;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6.31579%;\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 5.26316%;\"\u003e\n \u003cp\u003e0.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6.31579%;\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 5.26316%;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6.31579%;\"\u003e\n \u003cp\u003e2.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10.5263%;\"\u003e\n \u003cp\u003e50\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 54.7368%;\"\u003e\n \u003cp\u003e(Chen et al., 2016; Bakar et al., 2016; Osnaya et al., 2016; Hosseini et al., 2017; Osman et al., 2018; Moparthi et al., 2018; Ab Aziz et al., 2019; Bates et al., 2019; Blackstock et al., 2019; Eskin et al., 2019; Gokulanathan et al., 2019; Indrasari et al., 2019; Melo et al., 2019; Mendes et al., 2019; Raji \u0026amp; Thasleena, 2019; Salam et al., 2019; Ahmed \u0026amp; Hassan, 2022; Albanese et al., 2022; Ghimire et al., 2022; Manor \u0026amp; Greenberg, 2022; Wang et al., 2022; Shabani et al., 2021; Chamola, 2021; Zim et al., 2021; Fragner et al., 2024; Kargar et al., 2024; Alves et al., 2024; Saha et al., 2024)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 5.26316%;\"\u003e\n \u003cp\u003e0.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6.31579%;\"\u003e\n \u003cp\u003e0.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 5.26316%;\"\u003e\n \u003cp\u003e0.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6.31579%;\"\u003e\n \u003cp\u003e0.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 5.26316%;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6.31579%;\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10.5263%;\"\u003e\n \u003cp\u003e60\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 54.7368%;\"\u003e\n \u003cp\u003e(Page et al., 2015; Caro et al., 2015; Jadhav \u0026amp; Nerkar, 2015; Duarte et al., 2016; Jafari et al., 2017; Wieclaw et al., 2017; Cascales et al., 2021; Hu et al., 2023; Sarkar \u0026amp; Ghosh, 2024; Venkatachalam et al., 2024; Pushpalatha \u0026amp; Vimalathithan, n.d.)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 5.26316%;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6.31579%;\"\u003e\n \u003cp\u003e0.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 5.26316%;\"\u003e\n \u003cp\u003e0.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6.31579%;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 5.26316%;\"\u003e\n \u003cp\u003e0.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6.31579%;\"\u003e\n \u003cp\u003e3.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10.5263%;\"\u003e\n \u003cp\u003e70\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 54.7368%;\"\u003e\n \u003cp\u003e(Despins, 2017; Krokidis et al., 2022; Giordano et al., 2024; Tsakanika et al., 2025)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 5.26316%;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6.31579%;\"\u003e\n \u003cp\u003e0.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 5.26316%;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6.31579%;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 5.26316%;\"\u003e\n \u003cp\u003e0.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6.31579%;\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10.5263%;\"\u003e\n \u003cp\u003e80\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 54.7368%;\"\u003e\n \u003cp\u003e(Merenda et al., 2020; Diab \u0026amp; Rodriguez-Villegas, 2022)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 5.26316%;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6.31579%;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 5.26316%;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6.31579%;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 5.26316%;\"\u003e\n \u003cp\u003e0.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6.31579%;\"\u003e\n \u003cp\u003e4.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10.5263%;\"\u003e\n \u003cp\u003e90\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 54.7368%;\"\u003e\n \u003cp\u003e(Saha et al., 2022; Abadade et al., 2023)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 5.26316%;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6.31579%;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 5.26316%;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6.31579%;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 5.26316%;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6.31579%;\"\u003e\n \u003cp\u003e5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10.5263%;\"\u003e\n \u003cp\u003e100\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eTo support the conclusions of this systematic review on the applications of machine learning on microcontroller-class hardware for biological sensing, we undertook a rigorous assessment of the certainty of the evidence. The strength and reliability of our findings depend on a systematic evaluation process, which we carried out using the GRADE (Grading of Recommendations, Assessment, Development, and Evaluations) framework. GRADE is a globally recognized system that offers a comprehensive and transparent approach to assessing the quality of evidence, ensuring that the conclusions drawn are both credible and well founded. The certainty of the evidence across key outcomes was meticulously evaluated using several critical factors. First, we closely examined the precision of performance metrics by considering sample sizes and the width of confidence intervals in the studies. Narrow confidence intervals coupled with large sample sizes were indicative of a high level of certainty in the evidence, as they suggest more reliable and precise effect estimates. We also assessed the consistency of findings by comparing results across the included studies. When studies demonstrated similar performance results and behavior across different hardware platform and biological sensing application, it contributed to a greater certainty of evidence. Any observed heterogeneity such as difference in sensor type or microcontroller specification was thoroughly analysed to understand its sources and potential impact on the overall findings.\u003c/p\u003e\n\u003cp\u003eFurthermore, the potential for bias was evaluated using an adapted version of the Cochrane Risk of Bias tool. Studies with a low risk of bias were considered to contribute more significantly to the overall certainty of the evidence. We also judged directness based on the alignment of study populations, interventions specifically machine learning models, hardware platform, and biological sensing with the research questions of this review. High directness strengthened the support for our conclusions, leading to greater confidence in the evidence. Based on these factors, the certainty of evidence was categorized as follows: High certainty was assigned when studies were consistent, precise, directly applicable, and exhibited a low risk of bias. Moderate certainty was applied when there were minor concerns about one factor, such as some inconsistency in performance results or a moderate risk of bias. Low certainty was given when significant concerns existed in multiple areas, including imprecision in measurement results, variability in microcontroller hardware performance, or methodological limitation. Very low certainty was assigned when critical issues were present across all factors, significantly undermining confidence in the results. To ensure the relevance of the GRADE approach to this review, we adapted it specifically for outcomes related to the deployment of machine learning algorithms on microcontroller-class hardware for biological sensing application. Multiple independent reviewers assessed the certainty of evidence for each outcome. Disagreements were resolved through consensus discussions, ensuring a balanced and thorough evaluation. Additionally, where possible, we sought additional data or clarification from study authors to strengthen our certainty assessments.\u003c/p\u003e"},{"header":"3. Results","content":"\u003cp\u003e3.1. Study selection\u003c/p\u003e\n\u003cp\u003eFor this systematic review, we implemented a rigorous study selection process to identify the most relevant and high-quality research. We began by systematically searching three major academic databases: Google Scholar (yielding 142 records), Web of Science (22 records), and Scopus (4,266 records), resulting in an initial pool of 4,430 publications. After removing duplicate entries, we were left with 1,078 unique studies for preliminary screening. Through careful evaluation of titles and abstracts, we narrowed this down to 60 potentially relevant full-text articles. Following a thorough examination of these complete texts, we ultimately included 60 studies that met all our eligibility criteria. The final selection comprised 38 journal articles, 2 book chapters, 18 conference papers, and 2 theses or dissertations. The entire selection methodology, including exclusion reasons at each stage, is visually represented in Figure 6 using a PRISMA flowchart, providing transparent documentation of our systematic approach.\u003c/p\u003e\n\u003cp\u003e3.2. Study results\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eFigure 7 illustrates the temporal trend of publications from 2015 to 2024 that focus on machine learning applications on microcontroller-class hardware for biological sensing. The data shows a marked increase in research output starting in 2019, with a noticeable peak in both 2019 and 2024. This upward trend suggests growing interest and technological advancement in the integration of embedded systems and biological sensing frameworks during this period. These peaks may correspond with key developments in TinyML and edge AI, reinforcing the relevance and urgency of such research in low-power, real-time monitoring applications.\u003c/p\u003e\n\u003cp\u003eFigure 8 presents the classification of reviewed studies by research type and corresponding research category. A significant proportion of the literature (40 papers) falls under experimental research with an empirical orientation, demonstrating a strong emphasis on real-world implementation and performance testing. Applied research also accounts for a notable share, aligning with the practical nature of microcontroller-based biological sensing. Meanwhile, review studies and conceptual frameworks show a mix of theoretical and design/technical orientations, highlighting the evolving academic groundwork in this emerging field.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eFigure 9 displays the spectrum of tools and frameworks employed across the included studies, classified by their functional categories. Embedded platforms, such as Arduino and TinyML, accounted for the largest share (32.61%), reflecting the central role of resource-constrained hardware in biological sensing applications. Deep learning models and classical ML libraries also featured prominently, with PyTorch, TensorFlow Lite, and Sklearn among the most utilized. Notably, platforms like Edge Impulse and X-CUBE-AI bridged multiple categories, acting as both lifecycle and deployment tools. This diverse ecosystem of tools underscores the multidisciplinary integration required in deploying ML models on microcontroller-class hardware.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eFigure 10 categorizes the hardware platforms used across the selected studies, mapping each to its functional classification and percentage contribution. High-end ARM processors, such as the ARM Cortex A15 and STM32 series, represented the largest portion of platforms (34%), followed by AI/parallel system-on-chip (SoC) solutions (22%) like the GAP8 and ESP32. Microcontrollers\u0026mdash;including Arduino-based variants\u0026mdash;remained prevalent, though their share was distributed across subcategories such as hybrid systems and unspecified configurations. Vision AI development boards and reconfigurable logic devices like FPGAs also featured in specialized applications. This distribution highlights the growing complexity and diversity of platforms supporting machine learning in embedded biological sensing tasks.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eFigure 11 illustrates the distribution of machine learning task types applied in biological sensing studies and their alignment with broader analytical categories. Classification is the most commonly applied task (56.36%), reflecting the widespread use of ML models for detecting predefined biological conditions. Monitoring tasks (25.45%) and regression-based analyses (18.18%) follow, often applied in dynamic tracking of environmental variables and quantitative modeling. These tasks align with categories such as descriptive/operational, multi-task, and predictive modeling, indicating a blend of analytical depth and real-time adaptability in the reviewed systems. A smaller subset of studies focused on detection and quantification or hybrid approaches combining classification and monitoring. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003eFigure 12 compares the types of machine learning models employed across reviewed studies, categorized both by specific architectures and general methodological groupings. As shown in Figure 12a, deep learning architectures such as convolutional neural networks (CNNs), long short-term memory (LSTM), and variational autoencoders (VAE) dominate the field, comprising 55.56% of implementations. These models outperform rule-based systems (25.93%) and classical ML techniques like SVM and decision trees (14.81%) in adaptive and high-dimensional biological tasks.\u003c/p\u003e\n\u003cp\u003eFigure 12b further consolidates these findings by grouping models into broader categories, reaffirming that deep learning approaches are the most frequently adopted, followed by rule-based/signal-driven methods, classical machine learning, and pattern mining or heuristic techniques. This trend underscores the field\u0026rsquo;s shift toward data-driven architectures capable of handling complex, nonlinear biological data streams.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eFigure 13 presents the types of datasets used across the reviewed literature, categorized by origin and data accessibility. A significant majority (67.27%) of studies utilized custom datasets, often derived from specific biological, biomedical, or environmental sensing scenarios. This reflects the need for application-specific data in microcontroller-based biological sensing research. Open-source benchmarking datasets accounted for 25.45%, supporting model comparability across studies. Empirical/lab-based data (5.45%), hybrid dataset sources (3.64%), and synthetic/simulated data (also 3.64%) were used less frequently, suggesting a limited reliance on generalized or artificial data in favor of domain-specific, real-world inputs.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eFigure 14 categorizes the performance constraints observed in the selected studies, highlighting trade-offs between latency, model size, and memory efficiency. Over half of the studies (52%) emphasized low-latency operation (\u0026lt;500 ms) using compact models, which is essential for real-time biological sensing. A significant portion (36%) employed model compression techniques, such as reducing memory usage to below 100 KB, enabling deployment on constrained microcontroller-class hardware. Ultra-low latency configurations (\u0026lt;50 ms) were reported in 10% of cases, while moderate-latency and variable performance models made up a smaller share. Notably, 6% of studies did not report performance metrics, reflecting a need for standardized benchmarking practices.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eFigure 15 presents the distribution of hardware and architectural constraints reported in the reviewed literature. The most frequently cited limitation was unspecified (32%), reflecting a gap in detailed system reporting. Among specific constraints, RAM and Flash limitations (combined or standalone) accounted for 44%, highlighting memory as a primary concern for deploying ML models on microcontroller-class hardware. Power constraints (14%) and processor performance-related mentions (12%) were also notable, emphasizing the need for optimized inference under low-resource conditions. Communication and integration complexities appeared in fewer studies, yet they remain significant barriers in real-time biological sensing deployments.\u003c/p\u003e\n\u003cp\u003eFigure 16 summarizes the key performance evaluation themes found in the reviewed literature, indicating a strong emphasis on real-time monitoring and inference capabilities, which appeared in 38.18% of the studies. This reflects the critical need for responsive, low-latency systems in biological sensing. Suitability for edge/on-device tasks was the second most frequent focus (16.36%), underscoring the field\u0026rsquo;s pivot toward decentralized processing. Other important performance considerations included latency/power efficiency (12.73%), domain-specific functionality (10.91%), and comparative benchmarking (10.91%). Notably, only 7.27% of studies were explicitly accuracy-focused, while 3.64% lacked clearly stated performance criteria, indicating opportunities for improved reporting consistency.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eFigure 17 outlines the application domains in which ML-integrated microcontroller systems have been applied. The most common use case was healthcare and medical monitoring (25.45%), followed closely by water quality monitoring and management (23.64%), reflecting the critical societal need for health and environmental tracking. Other notable domains included biometric authentication (10.91%), industrial and infrastructure monitoring (9.09%), and general environmental sensing (7.27%). Emerging areas such as gesture recognition, educational tools, and smart city infrastructure also featured, though with lower representation. The spread of applications illustrates the versatility of embedded ML across diverse real-world scenarios.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eFigure 18 presents the toolchains and software stacks employed in the development of embedded ML systems. Arduino and Arduino-compatible platforms dominated the space, accounting for 29.09% of all implementations\u0026mdash;highlighting their accessibility and broad hardware support. TensorFlow Lite and Lite Micro followed at 18.18%, confirming the growing adoption of lightweight ML frameworks for on-device learning. Edge Impulse (12.73%) and various custom embedded development tools (9.09%) were also frequently used, supporting real-time deployment and optimization. Additional stacks included simulation tools like Unity (7.27%), Python-based programming environments, and MySQL data systems. The diversity of stacks illustrates the field\u0026rsquo;s cross-disciplinary nature and the balance between hardware simplicity and software flexibility.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e3.3. Study reearch questions results\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e1. How are ML models optimized for low-latency and memory-efficient execution on microcontroller platforms in biological sensing applications?\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe review highlights strong efforts in optimizing machine learning (ML) models for low-latency and memory-efficient deployment in biological sensing systems. Over half of the reviewed studies (52%) prioritized models with inference times below 500 milliseconds, a critical threshold for real-time monitoring (Figure 14). Additionally, 36% of studies implemented memory compression techniques, such as quantization and pruning, allowing models to operate within \u0026lt;100 KB memory constraints\u0026mdash;ideal for microcontroller-class devices. Only 10% of systems reported achieving ultra-low latency (\u0026lt;50 ms), which may be required for time-sensitive healthcare or gesture recognition tasks. However, a notable 6% of studies did not report any performance metrics at all, signaling an ongoing lack of standardization in benchmarking latency, memory, and computational performance. This variability underscores the importance of defining clear evaluation frameworks for embedded ML deployment, especially in constrained environments where trade-offs between latency, power, and accuracy are inevitable.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2. What are the dominant toolchains, hardware architectures, and datasets used in deploying embedded ML for biological tasks, and how do they influence system performance?\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eA clear pattern emerges in toolchain and hardware usage across the studies. Arduino and its compatible toolsets lead implementation strategies, appearing in 29.09% of systems, largely due to their accessibility, affordability, and extensive hardware integration support (Figure 18). Lightweight ML frameworks, particularly TensorFlow Lite and Lite Micro (18.18%), were preferred for deploying models on edge devices, while Edge Impulse (12.73%) played a key role in automating the optimization and deployment pipeline. On the hardware side, high-performance ARM processors (34%), including STM32 and Cortex variants, were most commonly used, offering a balance of power efficiency and computational capability (Figure 10). AI/parallel SoCs like ESP32 and GAP8 (22%) were also widely deployed, indicating growing reliance on specialized AI microcontrollers for inference-heavy workloads. In terms of datasets, 67.27% of studies used custom-built, domain-specific datasets (Figure 13). While this enhances contextual accuracy, it limits model generalizability and reproducibility. The remaining studies relied on open/public datasets (25.45%), with simulated or hybrid sources comprising under 8%. Together, these findings reveal a field focused on tailor-made solutions, but one that must advance toward standardized, sharable pipelines to ensure scalability and reproducibility.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3. In what ways are ML-enabled microcontrollers applied across domains such as health monitoring, water quality, and biometrics, and what domain-specific constraints emerge?\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe systematic review found that ML-enabled microcontrollers have been adopted across a wide array of biological sensing domains, with notable focus areas emerging. Healthcare and medical monitoring was the most common application domain, cited in 25.45% of the studies, often involving wearable biosensors and physiological data monitoring systems (Figure 17). Water quality monitoring and environmental sensing followed closely at 23.64%, leveraging microcontroller-based systems for continuous ecosystem health tracking, particularly in low-resource contexts. Biometric authentication (10.91%), such as gait or gesture recognition, and industrial/infrastructure monitoring (9.09%) also featured significantly. Each domain introduced distinct system requirements and constraints. For instance, healthcare systems required low latency and high reliability, while water sensing systems emphasized robustness and long-term power efficiency. These domain-specific constraints directly influenced hardware choices, dataset characteristics, and model complexity, leading to differing deployment strategies. For example, CNNs and LSTMs were commonly used in gesture and health recognition due to their superior pattern detection capabilities, while simpler, rule-based systems sufficed for steady-state environmental measurements.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4. What technical and methodological limitations (e.g., underreporting of constraints, performance metrics) hinder effective deployment in resource-constrained environments?\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eDespite advancements in embedded ML systems, your review identified significant technical and methodological limitations that hinder widespread, reliable deployment. Most notably, 32% of reviewed studies did not report key system-level constraints, such as available RAM, flash memory, or CPU load (Figure 15). This lack of transparency severely affects reproducibility and makes it difficult to compare system efficiency across studies. Additionally, 6% of papers failed to report any performance benchmarks, such as latency, accuracy, or energy usage (Figure 14), further weakening the practical applicability of these findings. Even among those that did report metrics, only 7.27% made accuracy their primary evaluation focus, suggesting a need for more balanced assessments that weigh accuracy, speed, power, and reliability (Figure 16). Such methodological inconsistencies create gaps between prototype-level performance and real-world deployment. To bridge this, future research must emphasize standardized evaluation protocols that include detailed constraint reporting, real-world testing, and comprehensive performance characterization under variable environmental conditions.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e5. How do real-time inference capabilities and edge suitability influence the design and evaluation of embedded ML systems for biological sensing?\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eReal-time capability and edge deployment readiness are now central to the design of ML-integrated microcontroller systems. In 38.18% of the studies, real-time inference was explicitly listed as a performance goal (Figure 16), often influencing model architecture (e.g., CNNs over RNNs) and latency optimization methods. Another 16.36% of studies prioritized edge deployment suitability, which encompasses local data processing, intermittent connectivity tolerance, and energy-aware operation. These priorities drove the adoption of compressed model formats, lightweight frameworks like TensorFlow Lite, and power-efficient hardware such as the ESP32 (Figure 10 and Figure 18). Domains such as water quality monitoring and health diagnostics particularly benefited from these features, as they require fast decision-making in low-bandwidth, remote, or mobile settings. Additionally, studies employing custom datasets (67.27%) often tailored their model training to edge constraints, ensuring compatibility with limited onboard memory and real-time event detection needs (Figure 13). These trends confirm that edge-readiness is no longer optional\u0026mdash;it is a defining requirement shaping both system architecture and evaluation strategies.\u0026nbsp;\u003c/p\u003e"},{"header":"4. Conclusion","content":"\u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThis systematic review presents a comprehensive analysis of machine learning applications on microcontroller-class hardware for biological sensing, synthesizing findings from studies published between 2015 and 2025. The review identifies a growing research interest in this domain, marked by sharp increases in publication activity beginning in 2019 and peaking again in 2024 (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e7\u003c/span\u003e). The dominance of empirical and applied studies (Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e8\u003c/span\u003e) underscores a clear shift from theoretical exploration to real-world deployment.\u003c/p\u003e \u003cp\u003eFindings indicate that embedded ML systems are increasingly optimized for low-latency performance and memory efficiency, with 52% of studies prioritizing sub-500 ms inference and 36% adopting compression methods to meet memory constraints (Fig.\u0026nbsp;\u003cspan refid=\"Fig13\" class=\"InternalRef\"\u003e14\u003c/span\u003e). Hardware selection favored high-performance ARM microcontrollers (34%) and AI-focused SoCs (22%), while Arduino and TensorFlow Lite emerged as the most widely adopted development stacks (Figs.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e10\u003c/span\u003e and \u003cspan refid=\"Fig17\" class=\"InternalRef\"\u003e18\u003c/span\u003e). Despite the breadth of hardware and tools, reporting inconsistencies remain: 32% of studies omitted constraint specifications, and 6% lacked performance benchmarks entirely (Figs.\u0026nbsp;\u003cspan refid=\"Fig14\" class=\"InternalRef\"\u003e15\u003c/span\u003e and \u003cspan refid=\"Fig13\" class=\"InternalRef\"\u003e14\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eModel architectures leaned heavily toward deep learning, particularly CNNs, LSTMs, and hybrid architectures, which accounted for 55.56% of usage (Fig.\u0026nbsp;\u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e12\u003c/span\u003e). However, reliance on custom, domain-specific datasets (67.27%) presents challenges for reproducibility and cross-study comparability (Fig.\u0026nbsp;\u003cspan refid=\"Fig12\" class=\"InternalRef\"\u003e13\u003c/span\u003e). In terms of application, biological sensing was most prevalent in healthcare (25.45%) and water quality (23.64%) systems, with increasing exploration in biometrics, infrastructure monitoring, and environmental sensing (Fig.\u0026nbsp;\u003cspan refid=\"Fig16\" class=\"InternalRef\"\u003e17\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eEdge readiness and real-time capability have become defining factors in the system design process, influencing the selection of models, datasets, and performance goals. With 38.18% of studies emphasizing real-time inference and 16.36% focusing on edge deployment (Fig.\u0026nbsp;\u003cspan refid=\"Fig15\" class=\"InternalRef\"\u003e16\u003c/span\u003e), the field is clearly shifting toward autonomous, decentralized sensing platforms.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eChen, J., Zhang, Y., Kuzyakov, Y., Wang, D., \u0026amp; Olesen, J. E. (2022). Challenges in upscaling laboratory studies to ecosystems in soil microbiology research. Global Change Biology, 29(3), 569\u0026ndash;588. https://doi.org/10.1111/gcb.16537\u003c/li\u003e\n\u003cli\u003eDiab, M. S., \u0026amp; Rodriguez-Villegas, E. (2022). Embedded machine learning using microcontrollers in wearable and ambulatory systems for health and care applications: A review. IEEE Access, 10, 98450\u0026ndash;98474. https://doi.org/10.1109/ACCESS.2022.3206782\u003c/li\u003e\n\u003cli\u003eDibyanshu, Kern, M., \u0026amp; Scheytt, T. (2024). Trace compounds in the urban water cycle in the Freiberg region, Germany. Frontiers in Water, 6. https://doi.org/10.3389/frwa.2024.1335766\u003c/li\u003e\n\u003cli\u003eEmerson, B. (2024). Beyond boundaries: Integrating biosensors and bioelectronics for environmental sustainability. Biosensors \u0026amp; Bioelectronics, 15(1), 1\u0026ndash;2. https://doi.org/10.37421/2155-6210.2024.15.431\u003c/li\u003e\n\u003cli\u003eFascista, A. (2022). Toward integrated large-scale environmental monitoring using WSN/UAV/crowdsensing: A review of applications, signal processing, and future perspectives. Sensors, 22(5), 1824. https://doi.org/10.3390/s22051824\u003c/li\u003e\n\u003cli\u003eGer\u0026oacute;s, A., Cruz, R., de Chaumont, F., Cardoso, J. S., \u0026amp; Aguiar, P. (2022). Deep learning-based system for real-time behavior recognition and closed-loop control of behavioral mazes using depth sensing. bioRxiv. https://doi.org/10.1101/2022.02.22.481410\u003c/li\u003e\n\u003cli\u003eHaque Zim, M. Z., Sazzad Sarkar, M., Nisi, Z. F., \u0026amp; Das, N. C. (2021). LFNNR: Light follower neural network robot conducted by machine learning technique. In Proceedings of the 2021 IEEE Conference of Russian Young Researchers in Electrical and Electronic Engineering (ElConRus) (pp. 1149\u0026ndash;1154). IEEE. https://doi.org/10.1109/ElConRus51938.2021.9396503\u003c/li\u003e\n\u003cli\u003eKrokidis, M. G., Dimitrakopoulos, G. N., Vrahatis, A. G., Tzouvelekis, C., Drakoulis, D., Papavassileiou, F., Exarchos, T. P., \u0026amp; Vlamos, P. (2022). A sensor-based perspective in early-stage Parkinson\u0026rsquo;s disease: Current state and the need for machine learning processes. Sensors, 22(2), 409. https://doi.org/10.3390/s22020409\u003c/li\u003e\n\u003cli\u003eKrokidis, M. G., Dimitrakopoulos, G. N., Vrahatis, A. G., Tzouvelekis, C., Drakoulis, D., Papavassileiou, F., Exarchos, T. P., \u0026amp; Vlamos, P. (2022). A sensor-based perspective in early-stage Parkinson\u0026rsquo;s disease: Current state and the need for machine learning processes. Sensors, 22(2), 409. https://doi.org/10.3390/s22020409\u003c/li\u003e\n\u003cli\u003eLupi, F., Herriges, J. A., Kim, H., \u0026amp; Stevenson, R. J. (2023). Getting off the ladder: Disentangling water quality indices to enhance the valuation of divergent ecosystem services. Proceedings of the National Academy of Sciences of the United States of America, 120(18), e2120261120. https://doi.org/10.1073/pnas.2120261120\u003c/li\u003e\n\u003cli\u003eMao, F., Khamis, K., Krause, S., Clark, J., \u0026amp; Hannah, D. M. (2019). Low-cost environmental sensor networks: Recent advances and future directions. Frontiers in Earth Science, 7, 221. https://doi.org/10.3389/feart.2019.00221\u003c/li\u003e\n\u003cli\u003eMerenda, M., Porcaro, C., \u0026amp; Iero, D. (2020). Edge machine learning for AI-enabled IoT devices: A review. Sensors, 20(9), 2533. https://doi.org/10.3390/s20092533\u003c/li\u003e\n\u003cli\u003eMiller, M., Kisiel, A., Cembrowska-Lech, D., Durlik, I., \u0026amp; Miller, T. (2023). IoT in water quality monitoring: Are we really here? Sensors, 23(2), 960. https://doi.org/10.3390/s23020960\u003c/li\u003e\n\u003cli\u003ePhilips, J. (2023, November 23). The future of water pollution monitoring: Advanced technologies for better data. Sustainable Living. https://www.environmentalconsortium.org/the-future-of-water-pollution-monitoring-advanced-technologies-for-better-data/\u003c/li\u003e\n\u003cli\u003eSchizas, N., Karras, A., Karras, C., \u0026amp; Sioutas, S. (2022). TinyML for ultra-low power AI and large scale IoT deployments: A systematic review. Future Internet, 14(12), 363. https://doi.org/10.3390/fi14120363\u003c/li\u003e\n\u003cli\u003eShabani, F., Philamore, H., \u0026amp; Matsuno, F. (n.d.). An energy-autonomous chemical oxygen demand sensor using a microbial fuel cell and embedded machine learning. IEEE Access. https://doi.org/10.1109/ACCESS.2021.3101496\u003c/li\u003e\n\u003cli\u003eSinha, S., Sahu, N., Bhardwaj, R., Mehta, A., Ahuja, H., Srivastava, S., Elhence, A., \u0026amp; Chamola, V. (2021). Machine learning on FPGA for robust Si3N4-gate ISFET pH sensor in industrial IoT applications. IEEE Transactions on Industry Applications, 57(6), 6700\u0026ndash;6712. https://doi.org/10.1109/TIA.2021.3117233\u003c/li\u003e\n\u003cli\u003eTedesco, P. A., Beauchard, O., Bigorne, R., Blanchet, S., Buisson, L., Conti, L., Cornu, J. F., Dias, M. S., Grenouillet, G., Hugueny, B., J\u0026eacute;z\u0026eacute;quel, C., Leprieur, F., Brosse, S., \u0026amp; Oberdorff, T. (2017). A global database on freshwater fish species occurrence in drainage basins. Scientific Data, 4(1), 170141. https://doi.org/10.1038/sdata.2017.141\u003c/li\u003e\n\u003cli\u003eUniwill-Lily Desun. (2023, September 13). Applications of water quality sensors in environmental monitoring. Water Quality Sensors. https://disen-sensor.com/applications-of-water-quality-sensors-in-environmental-monitoring/\u003c/li\u003e\n\u003cli\u003eVenkatachalam, S., Nair, H., Vellaisamy, P., Zhou, Y., Youssfi, Z., \u0026amp; Shen, J. P. (2024). Realtime person identification via gait analysis. arXiv. https://arxiv.org/pdf/2404.15312\u003c/li\u003e\n\u003cli\u003eZhang, C., \u0026amp; Zhang, J. (2014). Current techniques for detecting and monitoring algal toxins and causative harmful algal blooms. Journal of Environmental \u0026amp; Analytical Toxicology, 4(2), 123. https://doi.org/10.4172/2380-2391.1000123\u003c/li\u003e\n\u003cli\u003eKrokidis, M. G., Dimitrakopoulos, G. N., Vrahatis, A. G., Tzouvelekis, C., Drakoulis, D., Papavassileiou, F., Exarchos, T. P., \u0026amp; Vlamos, P. (2022). A Sensor-Based Perspective in Early-Stage Parkinson\u0026rsquo;s Disease: Current State and the Need for Machine Learning Processes. Sensors, 22(2), 409. https://doi.org/10.3390/s22020409 \u003c/li\u003e\n\u003cli\u003eMyataza, A., Mafunga, M., Mkhulisi, N. S., \u0026amp; Thango, B. A. A Systematic Review of ERP, CRM, and HRM Systems for SMEs: Managerial and Employee Support. Preprints 2024, 2024100384.\u003c/li\u003e\n\u003cli\u003eGumede, T. T., Chiworeka, J. M., Magoda, A. S., \u0026amp; Thango, B. Building Effective Social Media Strategies for Business: A Systematic Review. Preprints 2024, 2024100379.\u003c/li\u003e\n\u003cli\u003eCascales, J. P., Greenfield, D. A., Roussakis, E., Witthauer, L., Li, X., Goss, A., \u0026amp; Evans, C. L. (2021). Wireless wearable sensor paired with machine learning for the quantification of tissue oxygenation. IEEE Internet of Things Journal, 8(24), 17557-17567.\u003c/li\u003e\n\u003cli\u003eShabani, F., Philamore, H., \u0026amp; Matsuno, F. (2021). An energy-autonomous chemical oxygen demand sensor using a microbial fuel cell and embedded machine learning. IEEE Access, 9, 108689-108701.\u003c/li\u003e\n\u003cli\u003eGer\u0026oacute;s, A., Cruz, R., de Chaumont, F., Cardoso, J. S., \u0026amp; Aguiar, P. (2022). Deep learning-based system for real-time behavior recognition and closed-loop control of behavioral mazes using depth sensing. Biorxiv, 2022-02.\u003c/li\u003e\n\u003cli\u003eDiab, M. S., \u0026amp; Rodriguez-Villegas, E. (2022). Embedded machine learning using microcontrollers in wearable and ambulatory systems for health and care applications: A review. IEEE Access, 10, 98450-98474.\u003c/li\u003e\n\u003cli\u003eSchizas, N., Karras, A., Karras, C., \u0026amp; Sioutas, S. (2022). TinyML for ultra-low power AI and large scale IoT deployments: A systematic review. Future Internet, 14(12), 363.\u003c/li\u003e\n\u003cli\u003eZim, M. Z. H., Sarkar, M. S., Nisi, Z. F., \u0026amp; Das, N. C. (2021, January). LFNNR: Light Follower Neural Network Robot Conducted by Machine Learning Technique. In 2021 IEEE Conference of Russian Young Researchers in Electrical and Electronic Engineering (ElConRus) (pp. 1149-1154). IEEE.\u003c/li\u003e\n\u003cli\u003eChamola, V. (2021). Machine Learning on FPGA for Robust Si3N4-Gate ISFET pH Sensor in Industrial IoT Applications.\u003c/li\u003e\n\u003cli\u003eVenkatachalam, S., Nair, H., Vellaisamy, P., Zhou, Y., Youssfi, Z., \u0026amp; Shen, J. P. (2024, July). Realtime Person Identification via Gait Analysis Using IMU Sensors on Edge Devices. In 2024 International Conference on Neuromorphic Systems (ICONS) (pp. 371-375). IEEE.\u003c/li\u003e\n\u003cli\u003eMerenda, M., Porcaro, C., \u0026amp; Iero, D. (2020). Edge machine learning for ai-enabled iot devices: A review. Sensors, 20(9), 2533.\u003c/li\u003e\n\u003cli\u003eAhmed, K., \u0026amp; Hassan, M. (2022, June). TinyCare: a TinyML-based Low-Cost continuous blood pressure estimation on the extreme edge. In 2022 IEEE 10th International Conference on Healthcare Informatics (ICHI) (pp. 264-275). IEEE.\u003c/li\u003e\n\u003cli\u003eCovaciu, F., \u0026amp; Iordan, A. E. (2022). Control of a drone in virtual reality using MEMS sensor technology and machine learning. Micromachines, 13(4), 521.\u003c/li\u003e\n\u003cli\u003eMudau, M. C., Moshapo, L. W., Monyela, T. M., \u0026amp; Thango, B. A. The Role of Manufacturing Operations in SMEs Performance: A Systematic Review. Preprints 2024, 2024100539.\u003c/li\u003e\n\u003cli\u003eMtjilibe, T., Rameetse, E., Mgwenya, N., \u0026amp; Thango, B. (2024). Exploring the Challenges and Opportunities of Social Media for Organizational Engagement in SMEs: A Comprehensive Systematic Review. Available at SSRN 4998542.\u003c/li\u003e\n\u003cli\u003eWang, T., Fang, K., Wei, W., Tian, J., Pan, Y., \u0026amp; Li, J. (2022). Microcontroller unit chip temperature fingerprint informed machine learning for IIoT intrusion detection. IEEE Transactions on Industrial Informatics, 19(2), 2219-2227.\u003c/li\u003e\n\u003cli\u003eAlbanese, A., Nardello, M., Fiacco, G., \u0026amp; Brunelli, D. (2022). Tiny machine learning for high accuracy product quality inspection. IEEE Sensors Journal, 23(2), 1575-1583.\u003c/li\u003e\n\u003cli\u003eManor, E., \u0026amp; Greenberg, S. (2022). Custom hardware inference accelerator for tensorflow lite for microcontrollers. IEEE Access, 10, 73484-73493.\u003c/li\u003e\n\u003cli\u003eAbadade, Y., Temouden, A., Bamoumen, H., Benamar, N., Chtouki, Y., \u0026amp; Hafid, A. S. (2023). A comprehensive survey on tinyml. IEEE Access, 11, 96892-96922.\u003c/li\u003e\n\u003cli\u003eSaha, S. S., Sandha, S. S., \u0026amp; Srivastava, M. (2022). Machine learning for microcontroller-class hardware: A review. IEEE Sensors Journal, 22(22), 21362-21390.\u003c/li\u003e\n\u003cli\u003eFragner, C., Krutzler, C., Weiss, A. P., \u0026amp; Leitgeb, E. (2024). LEDPOS: Indoor visible light positioning based on LED as sensor and machine learning. IEEE Access.\u003c/li\u003e\n\u003cli\u003eGiordano, M., Dheman, K., \u0026amp; Magno, M. (2024). SepAl: Sepsis Alerts On Low Power Wearables With Digital Biomarkers and On-Device Tiny Machine Learning. IEEE Sensors Journal.\u003c/li\u003e\n\u003cli\u003eTsakanika, E., Tsoukas, V., Kakarountas, A., \u0026amp; Kokkinos, V. (2025). High Accuracy of Epileptic Seizure Detection Using Tiny Machine Learning Technology for Implantable Closed-Loop Neurostimulation Systems. BioMedInformatics, 5(1), 14.\u003c/li\u003e\n\u003cli\u003eKargar, A., Zorbas, D., Tedesco, S., Gaffney, M., \u0026amp; O\u0026rsquo;Flynn, B. (2024). Detecting Halyomorpha halys using a low-power edge-based monitoring system. Computers and Electronics in Agriculture, 221, 108935.\u003c/li\u003e\n\u003cli\u003eDai, Y., \u0026amp; Tan, R. (2024, May). FedCFC: On-Device Personalized Federated Learning with Closed-Form Continuous-Time Neural Networks. In 2024 23rd ACM/IEEE International Conference on Information Processing in Sensor Networks (IPSN) (pp. 14-26). IEEE.\u003c/li\u003e\n\u003cli\u003eSarkar, S., \u0026amp; Ghosh, A. (2024). Schr\u0026ouml;dinger spectrum and slim CNN architecture-based signal quality estimation for Photoplethysmogram signals. Biomedical Signal Processing and Control, 94, 106240.\u003c/li\u003e\n\u003cli\u003eHu, J. R., Park, D. Y., Agarwal, N., Herzig, M., Ormseth, G., Kaushik, M., Giao, D. M., Turkson-Ocran, R. N., \u0026amp; Juraschek, S. P. (2023). The promise and illusion of continuous, cuffless blood pressure monitoring. Current Cardiology Reports, 25(10), 1139\u0026ndash;1149. https://doi.org/10.1007/s11886-023-01932-4 \u003c/li\u003e\n\u003cli\u003eGhimire, D., Kil, D., \u0026amp; Kim, S. H. (2022). A survey on efficient convolutional neural networks and hardware acceleration. Electronics, 11(6), 945.\u003c/li\u003e\n\u003cli\u003eKhanyi, M. B., Xaba, S. N., Mlotshwa, N. A., Thango, B., \u0026amp; Matshaka, L. (2024). A Roadmap to Systematic Review: Evaluating the Role of Data Networks and Application Programming Interfaces in Enhancing Operational Efficiency in Small and Medium Enterprises. Sustainability, 16(23), 10192. https://doi.org/10.3390/su162310192\u003c/li\u003e\n\u003cli\u003eThobejane, L. T., \u0026amp; Thango, B. A. (2024). Partial Discharge Source Classification in Power Transformers: A Systematic Literature Review. Applied Sciences, 14(14), 6097.\u003c/li\u003e\n\u003cli\u003eFu, Z., He, X., Wang, E., Huo, J., Huang, J., \u0026amp; Wu, D. (2021). Personalized human activity recognition based on integrated wearable sensor and transfer learning. Sensors, 21(3), 885.\u003c/li\u003e\n\u003cli\u003eAlves, D. N., Barboza, E. D. A., \u0026amp; Vieira, T. F. (2024, November). TinyPPE: Ensuring Workplace Safety Helmet Compliance with Tiny Machine Learning. In 2024 XIV Brazilian Symposium on Computing Systems Engineering (SBESC) (pp. 1-6). IEEE.\u003c/li\u003e\n\u003cli\u003eSaha, S. S., Sandha, S. S., Aggarwal, M., Wang, B., Han, L., Briseno, J. D. G., \u0026amp; Srivastava, M. (2024). TinyNS: Platform-aware neurosymbolic auto tiny machine learning. ACM Transactions on Embedded Computing Systems, 23(3), 1-48.\u003c/li\u003e\n\u003cli\u003eMohaidat, T., \u0026amp; Khalil, K. (2024). A survey on neural network hardware accelerators. IEEE Transactions on Artificial Intelligence.\u003c/li\u003e\n\u003cli\u003eSaha, S. S., Davis, C., Sandha, S. S., Park, J., Geronimo, J., Garcia, L. A., \u0026amp; Srivastava, M. B. (2024). Locomote: Ai-driven sensor tags for fine-grained undersea localization and sensing. IEEE sensors journal, 24(10), 16999-17018.\u003c/li\u003e\n\u003cli\u003eCruz, I. A., de Melo, L., Leite, A. N., S\u0026aacute;tiro, J. V. M., Andrade, L. R. S., Torres, N. H., ... \u0026amp; Ferreira, L. F. R. (2019). A new approach using an open-source low cost system for monitoring and controlling biogas production from dairy wastewater. Journal of Cleaner Production, 241, 118284.\u003c/li\u003e\n\u003cli\u003eMendes, J. P., Coelho, L., Kovacs, B., Almeida, J. D., Pereira, C. M., Jorge, P. A., \u0026amp; Borges, M. T. (2019). Dissolved carbon dioxide sensing platform for freshwater and saline water applications: Characterization and validation in aquaculture environments. Sensors, 19(24), 5513.\u003c/li\u003e\n\u003cli\u003eSalam, A. E. U., Sadjad, R. S. O., SAMMAN, F. A., \u0026amp; FAKHRI, M. (2019). Online Hydraulic Parameter Monitoring System in Water Distribution Network. ICIC express letters. Part B, Applications: an international journal of research and surveys, 10(12), 1037-1045.\u003c/li\u003e\n\u003cli\u003eRaji, C. G., \u0026amp; Thasleena, V. A. (2019, December). IOT based water quality monitoring with android application. In 2019 Third International conference on I-SMAC (IoT in Social, Mobile, Analytics and Cloud)(I-SMAC) (pp. 446-451). IEEE.\u003c/li\u003e\n\u003cli\u003eAb Aziz, M. A., Abas, M. F., Bashri, M. K. A. A., Saad, N. M., \u0026amp; Ariff, M. H. (2019). Evaluating IoT based passive water catchment monitoring system data acquisition and analysis. Bulletin of Electrical Engineering and Informatics, 8(4), 1373-1382.\u003c/li\u003e\n\u003cli\u003eBlackstock, J. M., Covington, M. D., Perne, M., \u0026amp; Myre, J. M. (2019). Monitoring atmospheric, soil, and dissolved CO2 using a low-cost, Arduino monitoring platform (CO2-LAMP): Theory, fabrication, and operation. Frontiers in Earth Science, 7, 313.\u003c/li\u003e\n\u003cli\u003eBates, H., Zavafer, A., Szab\u0026oacute;, M., \u0026amp; Ralph, P. J. (2019). A guide to Open-JIP, a low-cost open-source chlorophyll fluorometer. Photosynthesis research, 142, 361-368.\u003c/li\u003e\n\u003cli\u003eIndrasari, W., Budi, E., Alayya, S. R., \u0026amp; Ramli, R. (2019, October). Measurement of water polluted quality based on turbidity, pH, magnetic property, and dissolved solid. In Journal of Physics: Conference Series (Vol. 1317, No. 1, p. 012060). IOP Publishing.\u003c/li\u003e\n\u003cli\u003eEskin, M. G., Torabfam, M., Psillakis, E., Cincinelli, A., Kurt, H., \u0026amp; Y\u0026uuml;ce, M. (2019). Real-time water quality monitoring of an artificial lake using a portable, affordable, simple, arduino-based open source sensor. Environmental Engineering-Inženjerstvo Okoli\u0026scaron;a, 6(1), 7-14.\u003c/li\u003e\n\u003cli\u003eSaravanan, K., Anusuya, E., Kumar, R., \u0026amp; Son, L. H. (2018). Real-time water quality monitoring using Internet of Things in SCADA. Environmental monitoring and assessment, 190(9), 556.\u003c/li\u003e\n\u003cli\u003eOsman, S. O., Mohamed, M. Z., Suliman, A. M., \u0026amp; Mohammed, A. A. (2018, August). Design and implementation of a low-cost real-time in-situ drinking water quality monitoring system using arduino. In 2018 International Conference on Computer, Control, Electrical, and Electronics Engineering (ICCCEEE) (pp. 1-7). IEEE.\u003c/li\u003e\n\u003cli\u003eMelo, M., Mota, F., Albuquerque, V., \u0026amp; Alexandria, A. (2019). Development of a robotic airboat for online water quality monitoring in lakes. Robotics, 8(1), 19.\u003c/li\u003e\n\u003cli\u003eSkosana, S., Mlambo, S., Madiope, T., \u0026amp; Thango, B. Evaluating Wireless Network Technologies (3G, 4G, 5G) and Their Infrastructure: A Systematic Review. Preprints 2024, 2024101331. \u003c/li\u003e\n\u003cli\u003eMkhize, A., Mokhothu, K. D., Tshikhotho, M., \u0026amp; Thango, B. A. (2025). Evaluating the Impact of Cloud Computing on SME Performance: A Systematic Review. Businesses, 5(2), 23. https://doi.org/10.3390/businesses5020023. \u003c/li\u003e\n\u003cli\u003eParra, L., Rocher, J., Escriv\u0026aacute;, J., \u0026amp; Lloret, J. (2018). Design and development of low cost smart turbidity sensor for water quality monitoring in fish farms. Aquacultural Engineering, 81, 10-18.\u003c/li\u003e\n\u003cli\u003eMoparthi, N. R., Mukesh, C., \u0026amp; Sagar, P. V. (2018, February). Water quality monitoring system using IoT. In 2018 Fourth international conference on advances in electrical, electronics, information, communication and bio-informatics (AEEICB) (pp. 1-5). IEEE.\u003c/li\u003e\n\u003cli\u003eGokulanathan, S., Manivasagam, P., Prabu, N., \u0026amp; Venkatesh, T. (2019). A GSM based water quality monitoring system using Arduino. Shanlax International Journal of Arts, Science and Humanities, 6(4), 22-26.\u003c/li\u003e\n\u003cli\u003eDespins, L. A. (2017). Automated detection of sepsis using electronic medical record data: a systematic review. The Journal for Healthcare Quality (JHQ), 39(6), 322-333.\u003c/li\u003e\n\u003cli\u003ePage, A., Kulkarni, A., \u0026amp; Mohsenin, T. (2015, October). Utilizing deep neural nets for an embedded ECG-based biometric authentication system. In 2015 IEEE biomedical circuits and systems conference (BioCAS) (pp. 1-4). IEEE.\u003c/li\u003e\n\u003cli\u003eWieclaw, L., Khoma, Y., Fałat, P., Sabodashko, D., \u0026amp; Herasymenko, V. (2017, September). Biometrie identification from raw ECG signal using deep learning techniques. In 2017 9th IEEE International Conference on Intelligent Data Acquisition and Advanced Computing Systems: Technology and Applications (IDAACS) (Vol. 1, pp. 129-133). IEEE.\u003c/li\u003e\n\u003cli\u003eCaro, L. A., Silva, C., Peralta, B., Herrera, O. A., \u0026amp; Barrientos, S. (2015). Real-time recognition of arm motion using artificial neural network multi-perceptron with arduino one microcontroller and EKG/EMG shield sensor. In Ambient Intelligence for Health: First International Conference, AmIHEALTH 2015, Puerto Varas, Chile, December 1-4, 2015, Proceedings 1 (pp. 3-14). Springer International Publishing.\u003c/li\u003e\n\u003cli\u003eChen, A. T. Y., Biglari-Abhari, M., Wang, K. I. K., Bouzerdoum, A., \u0026amp; Tivive, F. H. C. (2016). Hardware/software co-design for a gender recognition embedded system. In Trends in Applied Knowledge-Based Systems and Data Science: 29th International Conference on Industrial Engineering and Other Applications of Applied Intelligent Systems, IEA/AIE 2016, Morioka, Japan, August 2-4, 2016, Proceedings 29 (pp. 541-552). Springer International Publishing.\u003c/li\u003e\n\u003cli\u003eOsnaya, M. R., Rocha, M. F., Flores, Z. C., Navarrete, R., Andrade, E., Canto, C., \u0026amp; Rocha, M. E. (2016). MatlabSoftware in PC to Replace an Embedded System Used for Face Recognition. Journal of Industrial and Intelligent Information Vol, 4(2).\u003c/li\u003e\n\u003cli\u003eJadhav, M., \u0026amp; Nerkar, P. (2015). A Biometric System for Person Authentication Based on Finger Vein Patterns on FPGA. International Journal of Advanced Research in Electrical, Electronics and Instrumentation Engineering, 4(6).\u003c/li\u003e\n\u003cli\u003ePushpalatha, B., Scholar, P. G., \u0026amp; Vimalathithan, S. Implementing the Concept of Biometrics Recognition System in Voting Machine.\u003c/li\u003e\n\u003cli\u003eFernandez, C., Rusk, S., Glattard, N., \u0026amp; Shokoueinejad, M. (2017). 0067 COMPUTATIONAL PHENOTYPING IN POLYSOMNOGRAPHY: USING INTERPRETABLE PHYSIOLOGY-BASED MACHINE LEARNING MODELS TO PREDICT HEALTH OUTCOMES. Journal of Sleep and Sleep Disorders Research, 40(suppl_1), A26-A26.\u003c/li\u003e\n\u003cli\u003eDuarte, T., Piment\u0026atilde;o, J. P., Sousa, P., \u0026amp; Onofre, S. (2016, September). Biometric access control systems: A review on technologies to improve their efficiency. In 2016 IEEE International Power Electronics and Motion Control Conference Banbury, C. R., et al. (2021). \u0026quot;Benchmarking TinyML Systems: Challenges and Directions.\u0026quot; IEEE Design \u0026amp; Test.\u003c/li\u003e\n\u003cli\u003e(PEMC) (pp. 795-800). IEEE.\u003c/li\u003e\n\u003cli\u003eJafari, A., Gandhi, S., Konuru, S. H., Hairston, W. D., Oates, T., \u0026amp; Mohsenin, T. (2017, May). An EEG artifact identification embedded system using ICA and multi-instance learning. In 2017 IEEE international symposium on circuits and systems (ISCAS) (pp. 1-4). IEEE.\u003c/li\u003e\n\u003cli\u003eHosseini, M. P., Tran, T. X., Pompili, D., Elisevich, K., \u0026amp; Soltanian-Zadeh, H. (2017, July). Deep learning with edge computing for localization of epileptogenicity using multimodal rs-fMRI and EEG big data. In 2017 IEEE international conference on autonomic computing (ICAC) (pp. 83-92). IEEE.\u003c/li\u003e\n\u003cli\u003eBakar, W. A. B. W. A., Abdullah, Z. B., Md Saman, M. Y. B., Jalil, M. M. B. A., Man, M. B., Herawan, T., \u0026amp; Hamdan, A. R. (2016). Incremental-eclat model: an implementation via benchmark case study. In Advances in Machine Learning and Signal Processing: Proceedings of MALSIP 2015 (pp. 35-46). Springer International Publishing.\u003c/li\u003e\n\u003cli\u003eDavid, R., et al. (2020). \u0026quot;TensorFlow Lite Micro: Embedded Machine Learning for TinyML Systems.\u0026quot; arXiv.\u003c/li\u003e\n\u003cli\u003eHan, S., et al. (2016). \u0026quot;Deep Compression: Compressing Deep Neural Networks with Pruning, Trained Quantization and Huffman Coding.\u0026quot; ICLR.\u003c/li\u003e\n\u003cli\u003eLin, J., et al. (2020). \u0026quot;MCUNet: Tiny Deep Learning on IoT Devices.\u0026quot; NeurIPS.\u003c/li\u003e\n\u003cli\u003eSze, V., et al. (2017). \u0026quot;Efficient Processing of Deep Neural Networks: A Tutorial and Survey.\u0026quot; Proceedings of the IEEE.\u003c/li\u003e\n\u003cli\u003eWarden, P., \u0026amp; Situnayake, D. (2019). \u0026quot;TinyML: Machine Learning with TensorFlow Lite on Arduino and Ultra-Low-Power Microcontrollers.\u0026quot; O\u0026apos;Reilly Media.\u003c/li\u003e\n\u003cli\u003eZhang, Y., et al. (2019). \u0026quot;An Intelligent Water Quality Monitoring System Using Low-Power Microcontrollers and Machine Learning.\u0026quot; Journal of Environmental Informatics.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"University of Johannesburg","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Microcontroller-based sensing, TinyML, Biological monitoring, Embedded machine learning, Resource-constrained systems","lastPublishedDoi":"10.21203/rs.3.rs-6844035/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6844035/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eMicrocontroller-class devices, when integrated with machine learning (ML) models, offer transformative potential for biological sensing in resource-constrained environments. However, the deployment of such systems demands a careful balance between computational limitations, sensor integration, and ecological relevance. This systematic review evaluates trends, architectures, constraints, and applications of ML deployed on microcontroller-class hardware for biological sensing between 2015 and 2025. A systematic search across Google Scholar (n\u0026thinsp;=\u0026thinsp;142), Web of Science (n\u0026thinsp;=\u0026thinsp;22), and Scopus (n\u0026thinsp;=\u0026thinsp;4,266) yielded 4,430 records. After screening and eligibility assessment using PRISMA guidelines, 60 studies were included. The review focused on temporal trends, research types, ML toolchains, hardware platforms, task types, model architectures, dataset sources, system constraints, performance metrics, and domain-specific applications. Publication activity surged after 2019, peaking again in 2024. Most studies employed empirical and applied research methods (Fig.\u0026nbsp;8), with a majority using embedded platforms like Arduino and TinyML (32.61%) and lightweight frameworks such as TensorFlow Lite. ARM-based processors (34%) and AI-focused SoCs (22%) were the most common hardware platforms. Classification tasks dominated (56.36%), followed by monitoring (25.45%) and regression (18.18%). Deep learning architectures (CNNs, LSTMs, VAEs) accounted for 55.56% of models used. Most studies utilized custom, real-world datasets (67.27%) (Fig.\u0026nbsp;13) and emphasized performance constraints such as low latency (\u0026lt;\u0026thinsp;500 ms, 52%) and memory optimization (36%). Hardware limitations were primarily memory-based (44%) or unspecified (32%) (Fig.\u0026nbsp;15). Real-time inference (38.18%) and edge-device suitability (16.36%) were the most reported performance goals. Application areas were led by healthcare monitoring (25.45%) and water quality analysis (23.64%). Dominant toolchains included Arduino (29.09%), TensorFlow Lite (18.18%), and Edge Impulse (12.73%). Machine learning on microcontroller-class hardware is gaining traction in biological sensing, particularly in health and environmental monitoring. Despite progress, challenges persist in standardized benchmarking, performance reporting, and balancing system constraints. This review offers a detailed synthesis of implementation trends and practical bottlenecks, guiding future development of robust, low-power, and domain-specific ML sensing platforms.\u003c/p\u003e","manuscriptTitle":"Machine Learning on Microcontrollers for Biological Sensing: A Systematic Review","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-06-11 10:46:53","doi":"10.21203/rs.3.rs-6844035/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"64b2bc7b-79dc-4c67-8d0e-ae4054e286a0","owner":[],"postedDate":"June 11th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-06-11T10:46:53+00:00","versionOfRecord":[],"versionCreatedAt":"2025-06-11 10:46:53","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6844035","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6844035","identity":"rs-6844035","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}