Data Heterogeneity and Algorithmic Bias in AI-Based Antimicrobial Resistance Prediction: A Systematic Review and Mitigation Framework

Data Heterogeneity and Algorithmic Bias in AI-Based Antimicrobial Resistance Prediction: A Systematic Review and Mitigation Framework | 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 Data Heterogeneity and Algorithmic Bias in AI-Based Antimicrobial Resistance Prediction: A Systematic Review and Mitigation Framework Jema Kiazolu This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9552638/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 Background Antimicrobial resistance (AMR) represents one of the most critical global public health threats of the contemporary era, contributing to millions of deaths and substantial morbidity across all regions of the world (Murray et al., 2022). Artificial intelligence and machine learning (AI/ML) have emerged as transformative tools for improving the detection, prediction, and management of AMR by capitalizing on large-scale biomedical datasets. Despite their promise, the clinical translation of these methods remains constrained by fundamental methodological challenges, particularly data heterogeneity and algorithmic bias (Anahtar et al., 2021; Bilal et al., 2025). Objective This study aimed to systematically assess the impact of data heterogeneity and algorithmic bias on AI/ML-based AMR prediction and to propose a structured, multi-layered framework to guide their mitigation. Methods A systematic review with descriptive synthesis was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA 2020) guidelines (Page et al.,2021). Searches were performed across multiple databases, including PubMed/MEDLINE, Embase, Scopus, Web of Science, IEEE Xplore, and the ACM Digital Library, covering studies published between January 2020 and April 2026. Eligible studies applied AI/ML approaches to AMR prediction using data modalities such as whole-genome sequencing (WGS), antimicrobial susceptibility testing (AST), electronic health records (EHRs), surveillance datasets, and spectral data. Data were extracted using a standardized charting form and synthesized using descriptive and thematic methods. No meta-analysis was conducted owing to substantial methodological and data heterogeneity across included studies. Results A total of 68 studies published between 2020 and 2026 were included, of which 21 underwent detailed analysis. Included studies demonstrated substantial heterogeneity across data modalities, patient populations, laboratory practices, and geographic settings. WGS-based approaches were the most frequently represented, followed by AST- and EHR-based models. While many models achieved strong internal performance, generalizability across external settings remained limited. Key sources of heterogeneity included variability in data modalities, laboratory protocols, population composition, and temporal and geographic distribution. Major forms of bias identified included sampling bias, label inconsistency, structural confounding, and clinical context bias. Mitigation strategies including data rebalancing, structure-aware modeling, robust evaluation frameworks, and explainability approaches demonstrated partial and context-dependent improvements. Conclusion The performance of AI/ML-based AMR prediction models is constrained primarily by data heterogeneity and algorithmic bias rather than by limitations in algorithmic sophistication. Addressing these challenges requires a deliberate shift toward standardized, representative datasets, rigorous evaluation practices, and governance-aligned deployment. The proposed Heterogeneity Mitigation Framework offers a multi-layered, end-to-end approach to improving model robustness, fairness, and generalizability, providing a foundation for clinically applicable AI systems in antimicrobial resistance. Epidemiology antimicrobial resistance artificial intelligence machine learning data heterogeneity algorithmic bias systematic review predictive modeling Figures Figure 1 1. Introduction Antimicrobial resistance (AMR) has emerged as one of the most pressing global public health challenges of the 21st century. The widespread misuse and overuse of antimicrobial agents in human medicine, agriculture, and animal health have accelerated the emergence and dissemination of resistant pathogens, rendering infections that were once readily treatable increasingly difficult or impossible to manage. Recent global estimates indicate that AMR contributes substantially to both morbidity and mortality, with millions of deaths associated with resistant infections each year (Murray et al., 2022). This burden is disproportionately concentrated in low- and middle-income countries (LMICs), where deficiencies in diagnostic capacity, surveillance infrastructure, and access to effective therapies further amplify adverse outcomes (Lewnard et al., 2024). In response to the growing complexity of AMR, artificial intelligence (AI) and machine learning (ML) have increasingly been applied to enhance the detection, prediction, and clinical management of resistance. These methods leverage large-scale biomedical datasets encompassing whole-genome sequencing (WGS), antimicrobial susceptibility testing (AST), electronic health records (EHRs), surveillance systems, and mass spectrometry to identify complex, non-linear associations with resistance phenotypes (Anahtar et al., 2021; Kim et al., 2022). Commonly employed models, including random forests, support vector machines, gradient boosting algorithms, and deep learning architectures, have demonstrated moderate to high predictive performance across a range of pathogens and clinical contexts (Tang et al., 2022; Bilal et al., 2025). Beyond prediction, AI-driven tools are progressively being incorporated into clinical decision support systems and antimicrobial stewardship programs to guide empiric therapy selection and optimize antibiotic prescribing (Ferrari et al., 2024; Cohen et al., 2025). Despite these advances, the clinical translation of AI/ML-based AMR prediction models remains substantially limited. A central challenge is data heterogeneity, which arises from variability in data sources, formats, laboratory protocols, and patient populations. AMR datasets are inherently multi-modal, integrating genomic, phenotypic, clinical, and surveillance data that each differ in structure, quality, and completeness. Such variability complicates data integration, reduces reproducibility, and undermines model generalizability across settings (Kherabi et al., 2024; Sakagianni et al., 2023). Closely related to heterogeneity is algorithmic bias, whereby models learn systematic distortions embedded in training data rather than true biological or clinical relationships. In healthcare applications, these biases can produce unequal model performance across subpopulations, potentially reinforcing existing health disparities (Chen et al., 2023; Gao et al., 2025). For instance, models trained predominantly on data from high-income settings may perform poorly in LMIC contexts, where resistance patterns, healthcare systems, and data infrastructures differ substantially (Nsubuga et al., 2024; Ali et al., 2023). Furthermore, many AMR prediction models rely on single-center datasets and internal validation strategies, resulting in overestimated performance metrics and limited external validity when applied to new populations or clinical environments (Kim et al., 2022; Tang et al., 2022). These challenges expose a critical gap between methodological development and real-world implementation of AI in AMR prediction. While existing studies have explored the application of ML to AMR prediction, a comprehensive synthesis focusing specifically on the interconnected roles of data heterogeneity and algorithmic bias remains lacking. In particular, the mechanisms through which heterogeneity generates bias, and how these factors jointly influence model robustness, fairness, and generalizability, have not been systematically characterized. This study therefore aims to synthesize current evidence on data heterogeneity and algorithmic bias in AI/ML-based AMR prediction and to propose a structured framework for their mitigation. By shifting focus from predictive performance alone to considerations of robustness, fairness, and real-world applicability, this work seeks to support the development of reliable and clinically deployable AI systems for antimicrobial resistance management. This study uniquely integrates data heterogeneity and algorithmic bias into a unified analytical framework, addressing a critical gap in AI-based AMR prediction research 2. Methods 2.1 Study Design This study employed a systematic review design with descriptive synthesis, conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA 2020) framework (Page et al., 2021). A systematic review approach was selected to provide structured, reproducible, and transparent identification and synthesis of evidence on data heterogeneity and algorithmic bias in AI-based AMR prediction. The review protocol emphasized comprehensive database searching, duplicate-independent study selection, standardized data extraction, and thematic synthesis. No meta-analysis was conducted owing to substantial methodological and outcome heterogeneity across included studies, which precluded statistical pooling. This review was not prospectively registered in PROSPERO. 2.2 Eligibility Criteria Eligibility criteria were defined using the Population–Concept–Context (PCC) framework to ensure structured and transparent study selection (Peters et al., 2020). Inclusion Criteria Studies were included if they: ( 1 ) applied AI, machine learning (ML), or deep learning (DL) methods to predict AMR-related outcomes, including susceptibility classification, minimum inhibitory concentration (MIC) prediction, or resistance risk estimation; ( 2 ) reported model development, validation strategies, or performance metrics such as accuracy, sensitivity, specificity, F1-score, or area under the curve (AUC); ( 3 ) utilized one or more data modalities including AST, WGS, EHRs, surveillance datasets, or spectral data such as MALDI-TOF mass spectrometry; ( 4 ) were conducted in clinical, laboratory, public health, or surveillance contexts; and ( 5 ) were primary research studies, conference proceedings, or review articles providing methodological or conceptual insights. Exclusion Criteria Studies were excluded if they: ( 1 ) used non-AI approaches, including rule-based or traditional statistical models without machine learning components; ( 2 ) were purely experimental laboratory studies without predictive modeling; ( 3 ) did not report predictive outcomes or model performance; or ( 4 ) were editorials, commentaries, or opinion papers without methodological or empirical contribution. These criteria ensured that included studies contributed to both empirical evidence and conceptual understanding of heterogeneity, bias, and generalizability in AI-based AMR prediction. 2.3 Information Sources and Database Search A comprehensive literature search was conducted across multiple electronic databases, including PubMed/MEDLINE, Embase, Scopus, Web of Science, IEEE Xplore, and the ACM Digital Library, to capture interdisciplinary research spanning biomedical, computational, and public health domains. The search was restricted to studies published between January 2020 and April 2026 to reflect recent methodological developments in artificial intelligence and antimicrobial resistance prediction. To enhance comprehensiveness, reference lists of included studies and relevant reviews were manually screened, and database-specific indexing terms, including MeSH and Emtree, were adapted accordingly. The study selection process followed PRISMA 2020 reporting standards (Page et al., 2021), including identification, screening, eligibility assessment, and inclusion of studies. The flow of study selection is illustrated in Fig. 1. 2.4 Search Strategy The search strategy combined controlled vocabulary and free-text terms related to AI and AMR prediction. Key concepts included core terms ("artificial intelligence," "machine learning," "deep learning"), domain terms ("antimicrobial resistance," "AMR," "prediction," "genomics," "whole-genome sequencing"), and clinical terms ("antimicrobial susceptibility testing," "surveillance," "stewardship"). Boolean operators (AND, OR, NOT) were applied to optimize retrieval. Searches were restricted to studies published between January 2020 and April 2026 to capture recent methodological developments in AI and biomedical data integration. 2.5 Study Selection All retrieved records were imported into a reference management system and screened using a systematic review platform. Screening was conducted in two sequential stages: ( 1 ) title and abstract screening, in which two independent reviewers assessed records against eligibility criteria; and ( 2 ) full-text screening, in which potentially eligible studies were evaluated in detail. Discrepancies were resolved through discussion or consultation with a third reviewer. The selection process was documented using a PRISMA 2020 flow diagram (Page et al., 2021). 2.6 Data Extraction A standardized data extraction form was developed and piloted to ensure consistency and completeness. Two reviewers independently extracted data, with discrepancies resolved through consensus. Extracted variables included: pathogen and antibiotic focus; data modality (e.g., WGS, AST, EHR, surveillance); model type (e.g., random forest, support vector machine, neural networks); validation strategy (internal and/or external); performance metrics (e.g., accuracy, AUC, F1-score); sources of heterogeneity; types of bias; and reported mitigation strategies. 2.7 Data Synthesis A combined descriptive and thematic synthesis approach was employed, consistent with established systematic review methodology for evidence that is unsuitable for meta-analysis (Peters et al., 2020). Descriptive synthesis involved summarizing study characteristics using tables and visualizations, including modality distribution, model types, validation strategies, and geographic representation, to enable structured comparison across studies. Thematic synthesis involved a narrative approach to identify key patterns related to dimensions of data heterogeneity (e.g., modality, population, temporal variation), types of algorithmic bias (e.g., sampling bias, label bias, structural confounding), and strategies for their mitigation. Themes were developed inductively and refined through iterative reviewer discussion. Formal meta-analysis was not performed due to the substantial clinical, methodological, and statistical heterogeneity across included studies, which precluded meaningful pooling of effect estimates. 3. Results 3.1 Overview of Included Studies A total of 68 studies published between 2020 and 2026 were included in this systematic review, of which 21 met the criteria for detailed analysis. These 21 studies were selected for in-depth analysis based on their relevance to data heterogeneity and algorithmic bias, as well as the completeness of methodological reporting. This reflects a growing yet methodologically heterogeneous body of literature on AI/ML-based AMR prediction. The characteristics of the 21 studies selected for detailed analysis are presented in Table 1. Table 1. Characteristics of the 21 Studies Selected for Detailed Analysis in AI/ML-Based AMR Prediction (2020–2026). First Author (Year) Study Type Data Modality Model Type Pathogen Scope Setting Region Khaledi (2020) Primary WGS RF, SVM P. aeruginosa Hospital Germany Weis (2020) Primary MALDI-TOF RF, CNN Multiple Clinical lab Switzerland Ren (2022) Primary WGS RF, DNN Multiple Clinical isolates Germany Feretzakis (2021) Primary AST, Surveillance Gradient Boosting Multiple ICU Greece Ardila (2024) Primary WGS + AST RF, Ensemble Priority pathogens Healthcare Colombia Ferrari (2024) Primary EHR LR, RF Bloodstream infections ICU UK Park (2024) Primary MALDI-TOF CNN, RF Multiple Clinical lab Germany Cohen (2025) Primary EHR ML, LLM Sepsis Hospital USA Valavarasu (2025) Primary Surveillance, AST Time-series ML Broad Surveillance India Zou (2025) Primary WGS CNN, DNN Pediatric infections Hospital China Urena (2024) Primary Clinical + Ecology ML Hospital infections Hospital France Hardan (2024) Primary EHR (multimodal) Neural Networks Multiple Hospital Middle East Blechman (2024) Primary EHR Gradient Boosting Broad infections Hospital USA Condorelli (2024) Primary WGS RF, XGBoost K. pneumoniae Clinical isolates Italy Zhao (2024) Primary Surveillance ML Food-animal pathogens Surveillance Global Ardila (2025) Primary AST RF, Gradient Boosting Priority pathogens Healthcare Colombia Nsubuga (2024) Primary WGS ML E. coli Multicountry Africa Prosperi (2022) Primary Genomic ML Multiple Mixed USA Sidorczuk (2022) Primary Peptide/Genomic ML Antimicrobial peptides Computational Europe Kim (2022) Review WGS — Broad Clinical Canada/UK Tang (2022) Review Multi-modal — Broad Mixed China Abbreviations: WGS = Whole-genome sequencing; AST = Antimicrobial susceptibility testing; EHR = Electronic health records; RF = Random Forest; SVM = Support Vector Machine; CNN = Convolutional Neural Network; DNN = Deep Neural Network; LR = Logistic Regression. The included studies utilized diverse data modalities, underscoring the multi-domain nature of AMR prediction. These modalities included WGS, AST phenotypes, EHRs, spectral data (e.g., MALDI-TOF mass spectrometry), and surveillance datasets. WGS-based approaches were the most frequently represented, followed by AST- and EHR-based models, while spectral and surveillance data were comparatively less common (Ardila et al., 2024; Weis et al., 2020; Valavarasu et al., 2025). Geographically, the evidence base was heavily concentrated in high-income regions, particularly North America, Europe, and East Asia, where access to genomic sequencing infrastructure, clinical informatics systems, and large-scale surveillance data is more established (Kim et al., 2022; Zou et al., 2025). In contrast, representation from LMICs remained limited despite their disproportionate burden of AMR (Ali et al., 2023; Nsubuga et al., 2024). Most studies were conducted in hospital-based settings, particularly among inpatient and intensive care unit (ICU) populations, where AMR prevalence is high and data availability is greater (Ferrari et al., 2024; Cohen et al., 2025). Community-level and outpatient datasets were comparatively underrepresented. Although surveillance-based studies partially expanded geographic and population coverage, important gaps persisted in non-hospital and resource-limited settings (Valavarasu et al., 2025). Overall, the included studies demonstrate substantial heterogeneity across data modalities, geographic regions, and clinical settings, with a clear dominance of high-resource, hospital-based datasets an imbalance with important implications for the generalizability and equity of AI/ML-based AMR prediction models. 3.2 Dimensions of Data Heterogeneity Data heterogeneity emerged as a central determinant of model performance, reproducibility, and generalizability. Heterogeneity was observed across multiple interacting dimensions, including data modality, laboratory practices, population characteristics, and temporal and geographic variation (Tang et al., 2022; Sakagianni et al., 2023; Zou et al., 2025). 3.2.1 Data Modality and Measurement Heterogeneity Data modality heterogeneity arises from the integration of fundamentally different data sources, each capturing distinct aspects of AMR. WGS-based approaches depend on sequencing quality, bioinformatics pipelines, resistance gene databases, and genotype–phenotype mapping strategies, resulting in variability in feature representation and predictive outputs (Ardila et al., 2024; Ren et al., 2022). AST-based models are influenced by methodological differences including broth microdilution, disk diffusion, and automated systems, as well as laboratory-specific protocols and class imbalance in resistance phenotypes (Valavarasu et al., 2025; Wenzler et al., 2023). EHR-derived data introduce further heterogeneity through variability in coding systems, documentation practices, and missing data patterns, complicating integration with genomic and phenotypic data (Gianfrancesco & Goldstein, 2021; Boyd et al., 2023). Similarly, MALDI-TOF-based approaches are sensitive to instrument calibration, sample preparation, and spectral preprocessing, generating variability across laboratories (Weis et al., 2020; Park et al., 2024). Collectively, these modality-specific differences create significant interoperability challenges and limit the development of unified, multi-modal prediction models. 3.2.2 Laboratory and Technical Heterogeneity Laboratory protocols introduce systematic variability that directly affects resistance measurement and model performance. Differences in AST methodologies and interpretive breakpoints particularly between CLSI and EUCAST standards lead to inconsistent resistance classification across datasets (Humphries et al., 2021; Wenzler et al., 2023). In WGS-based approaches, sequencing depth, quality control procedures, and bioinformatics pipelines further contribute to heterogeneity, resulting in discordant resistance gene detection and reduced predictive reliability (Doyle et al., 2020). These technical inconsistencies introduce measurement bias and undermine comparability across studies, particularly in multi-center datasets. 3.2.3 Population and Setting Heterogeneity Population-level heterogeneity reflects imbalances in geographic, demographic, and clinical representation within datasets. Most studies were conducted in high-income countries, resulting in limited representation of LMICs where resistance patterns and healthcare systems differ substantially (Ali et al., 2023; Kim et al., 2022). Datasets were predominantly derived from hospital-based populations, while community and outpatient data were underrepresented (Zou et al., 2025). Variability across patient groups, including differences between paediatric and adult populations, further complicates model generalizability. Imbalances in pathogen–antibiotic combinations contribute additionally, with overrepresentation of commonly studied species and underrepresentation of rare or context-specific resistance patterns (Zhao et al., 2024). 3.2.4 Temporal and Geographic Heterogeneity Temporal variation introduces complexity through changes in antimicrobial prescribing practices, resistance prevalence, and laboratory methodologies over time phenomena that can degrade model performance when applied to future datasets (Ali et al., 2023; Kherabi et al., 2024). Geographic heterogeneity further limits model transferability, as resistance patterns vary across regions due to differences in microbial ecology, healthcare systems, and surveillance infrastructure (Nsubuga et al., 2024; Zhao et al., 2024). These dynamics highlight the importance of context-aware model development and validation strategies. 3.2.5 Species and Antibiotic Imbalance AMR datasets exhibit substantial imbalance across pathogen–antibiotic combinations, with frequently studied bacterial species and antimicrobial classes overrepresented while rare pathogens and certain antibiotic groups remain underrepresented (Prosperi et al., 2022). Environmental and livestock surveillance data further reveal geographic biases, with disproportionate representation from specific regions (Zhao et al., 2024). This imbalance biases model training toward dominant patterns and reduces predictive performance for underrepresented species–drug combinations. Across studies, data heterogeneity is multi-dimensional and interdependent, affecting data quality, interoperability, and model performance. These interacting dimensions reduce reproducibility, limit generalizability, and complicate cross-study comparability, thereby constraining the clinical applicability of AI/ML-based AMR prediction models (Tang et al., 2022; Bilal et al., 2025). The key dimensions of data heterogeneity and their implications for AI/ML-based AMR prediction are summarized in Table 2. Table 2. Key Dimensions of Data Heterogeneity and Their Implications for AI/ML-Based AMR Prediction. Dimension Source of Heterogeneity Mechanism Impact on AI/ML Models Data modality WGS, AST, EHR, MALDI-TOF, surveillance data Differences in data structure, format, and feature representation Reduced interoperability; inconsistent feature learning Laboratory practices CLSI vs EUCAST standards; AST methodologies Variability in resistance measurement and classification Label inconsistency; reduced reproducibility Population & setting High-income country dominance; hospital-based datasets Sampling imbalance across populations and clinical settings Poor generalizability; sampling bias Temporal variation Changing resistance patterns over time Dataset shift and evolving epidemiological trends Model performance degradation Geographic variation Regional differences in resistance patterns Context-dependent resistance ecology Limited transferability Species–antibiotic imbalance Uneven representation of pathogen–drug combinations Skewed training distribution Poor performance on underrepresented classes 3.3 Types of Algorithmic Bias Identified While heterogeneity reflects variability in data, algorithmic bias represents the systematic distortions that arise from such variability during model development and deployment. Four major types of bias were consistently identified across studies. 3.3.1 Sampling Bias Sampling bias occurs when training datasets do not adequately represent the populations, geographic regions, or clinical settings in which models are subsequently applied, causing models to learn context-specific correlations rather than generalizable resistance mechanisms (Ali et al., 2023; Nsubuga et al., 2024). Global datasets are often skewed toward high-income countries and hospital-based samples, limiting representation of LMIC contexts and community settings. Temporal shifts in resistance patterns further exacerbate sampling bias, resulting in performance degradation when models are applied to newer datasets (Finlayson et al., 2021; Nsubuga et al., 2024). At the institutional level, models trained on single-center data often capture local prescribing practices and microbial ecology, leading to inflated internal performance but poor external validity (Cohen et al., 2025; Ferrari et al., 2024). 3.3.2 Label Bias Label bias arises from inconsistencies in resistance definitions and classification across datasets. Variability in susceptibility categories and MIC thresholds introduces systematic noise into training labels (Humphries et al., 2021). Differences between guideline systems such as CLSI and EUCAST can result in identical MIC values being classified differently, while aggregation of susceptibility categories may obscure clinically relevant distinctions. These inconsistencies reduce reproducibility and compromise model generalizability. 3.3.3 Structural (Confounding) Bias Structural bias occurs when models exploit non-causal correlations within the data, such as phylogenetic structure or species–drug co-occurrence patterns. Resistance phenotypes are often unevenly distributed across bacterial lineages, leading models to rely on lineage-specific markers as proxies for resistance rather than identifying causal mechanisms (Prosperi et al., 2022; Ren et al., 2022). Similarly, non-random pairing of pathogens and antibiotics in datasets introduces co-occurrence bias, where models learn dataset-specific relationships rather than biologically meaningful associations (Tang et al., 2022; Sidorczuk et al., 2022). 3.3.4 Clinical Context Bias Clinical context bias arises from artifacts embedded in EHR data, including healthcare access patterns, documentation practices, and missing data (Gianfrancesco & Goldstein, 2021; Boyd et al., 2023). Derived clinical variables and comorbidity indices may vary across populations, introducing additional inconsistencies (Gao et al., 2025). Moreover, observational clinical data often reflect prescribing behaviours and healthcare system processes rather than causal relationships between antibiotic use and resistance (Collignon & Beggs, 2025). These factors limit model transportability and raise concerns regarding fairness and reliability. The sources, mechanisms, consequences, and mitigation strategies of algorithmic bias identified across studies are summarized in Table 3. Table 3. Sources, Mechanisms, Consequences, and Mitigation of Algorithmic Bias in AI-Based AMR Prediction. Bias Type Source Mechanism Consequence Mitigation Strategies Sampling bias HIC-dominated datasets; single-center data Non-representative training data Poor external validity; limited generalizability Multi-center datasets; reweighting; stratified sampling Label bias CLSI vs EUCAST breakpoints; MIC variability Inconsistent resistance classification Misclassification; reduced reproducibility Standardized breakpoints; use of raw MIC values Structural bias Phylogeny; species–drug co-occurrence Learning non-causal correlations Spurious predictions; reduced biological validity Structure-aware modeling; phylogenetic splitting Clinical context bias EHR documentation; healthcare access differences Context-dependent artifacts Limited transportability; fairness concerns Feature harmonization; bias auditing Algorithmic bias Class imbalance; subgroup disparities Unequal representation in training Unequal model performance Resampling; cost-sensitive learning 3.4 Mitigation Strategies Used in Studies Studies addressing AMR prediction using AI/ML employ a range of mitigation strategies targeting different stages of the modeling pipeline. These strategies can be broadly categorized into data-level, model-level, and evaluation-level approaches, as well as emerging explainability and causal methods. While these approaches contribute to improved robustness and fairness, their effectiveness remains partial and context-dependent, reflecting the underlying complexity of AMR data (Bilal et al., 2025; Chen et al., 2023). 3.4.1 Data-Level Strategies: Addressing Class Imbalance and Sampling Bias Class imbalance is a pervasive challenge in AMR datasets, where resistant phenotypes or minority classes are frequently underrepresented. Common approaches include oversampling, undersampling, and synthetic data generation techniques such as the Synthetic Minority Over-sampling Technique (SMOTE), alongside cost-sensitive learning and weighted loss functions (Abdelhay et al., 2025). Resampling and reweighting techniques are widely applied in clinical and EHR-based models to improve representation of minority classes; however, their impact on overall performance is inconsistent and may introduce overfitting in small or highly imbalanced datasets (Bilal et al., 2025). Bias-aware sampling strategies have also been used to address imbalances across geographic regions, temporal distributions, and pathogen–antibiotic combinations, including group-based reweighting and targeted data augmentation, though reported improvements are generally modest and highly dependent on dataset characteristics (Prosperi et al., 2022; Bilal et al., 2025). 3.4.2 Structure-Aware Modeling and Data Splitting To mitigate structural and phylogenetic bias, several studies emphasize population structure-aware modeling and evaluation. Conventional random train–test splits allow models to exploit lineage-specific or dataset-specific correlations, leading to inflated performance estimates. In contrast, structure-aware approaches including phylogeny-informed data splitting, clade-based partitioning, and stratified sampling aim to ensure independence between training and testing datasets (Prosperi et al., 2022; Ren et al., 2022). These methods reduce reliance on non-causal correlations and promote the learning of more generalizable resistance mechanisms. However, they often result in lower reported performance metrics, reflecting a more realistic assessment of model capability. This trade-off highlights the tension between optimizing predictive accuracy and ensuring biological and clinical validity. 3.4.3 Robust Evaluation Strategies Robust evaluation is a critical component of bias mitigation, directly influencing the reliability and generalizability of model performance estimates. Key strategies include external validation using independent datasets, temporal validation across different time periods, and multi-center study designs (Chen et al., 2023). Despite their importance, these approaches remain underutilized in the AMR prediction literature. Evidence consistently demonstrates that models evaluated on external datasets exhibit reduced performance compared to internal validation, indicating substantial overfitting to training data (Bilal et al., 2025). Beyond discrimination metrics such as AUC, there is increasing emphasis on reporting calibration, subgroup performance, and fairness-related metrics to provide a more comprehensive assessment of real-world applicability across diverse populations. 3.4.4 Explainability and Causal Approaches Explainability techniques are increasingly incorporated to identify and mitigate bias in AI/ML models. Methods such as feature attribution, model interpretability tools, and fairness metrics including demographic parity and equalized odds enable detection of spurious associations and improve transparency (Xu & Ouyang, 2022; Loh et al., 2022). Causal modeling approaches aim to address confounding by incorporating explicit assumptions about data-generating mechanisms. Techniques such as propensity score adjustment, causal inference frameworks, and Bayesian networks have been applied to disentangle true causal relationships from dataset-specific correlations (Prosperi et al., 2022; Moccia, 2024). While these approaches enhance interpretability and support bias detection, their effectiveness depends on data quality, appropriate model assumptions, and the availability of relevant covariates, and should therefore be considered complementary rather than standalone solutions. Mitigation strategies in AI/ML-based AMR prediction can be conceptualized across four interrelated domains: (1) data rebalancing and sampling strategies; (2) structure-aware modeling and data partitioning; (3) robust and context-aware evaluation frameworks; and (4) explainability and causal inference approaches. No single strategy fully resolves bias or heterogeneity. Effective mitigation requires an integrated, multi-layered approach spanning the entire machine learning pipeline, reinforcing the need for end-to-end design strategies that address data quality, model development, evaluation rigor, and real-world implementation simultaneously (Bilal et al., 2025; Chen et al., 2023). A comparative summary of data domains, modeling approaches, performance, validation strategies, and key limitations across studies is presented in Table 4. Table 4. Data Domains, Algorithms, Performance, and Limitations in AI-Based AMR Prediction. Data Domain Common Algorithms Typical AUC Range Validation Type Key Limitations WGS RF, CNN, DNN 0.80–0.97 Internal + limited external Population bias; sequencing variability AST RF, SVM, Gradient Boosting 0.65–0.90 Mostly internal Laboratory variability; class imbalance EHR LR, RF, Gradient Boosting 0.60–0.85 Mostly internal Missing data; coding heterogeneity Surveillance Time-series ML 0.70–0.92 Limited external Regional bias; reporting inconsistency MALDI-TOF CNN, RF 0.48–0.85 Internal Technical variability; reproducibility issues 4. Proposed Framework: Heterogeneity Mitigation Framework To address the multifaceted sources of data heterogeneity and algorithmic bias identified in this review, we propose a four-layer Heterogeneity Mitigation Framework spanning the full machine learning pipeline, from data generation to real-world deployment. The framework integrates data standardization, bias-aware modeling, robust evaluation, and governance mechanisms, reflecting current best practices in AMR research and computational medicine (Ardila et al., 2025; Bilal et al., 2025). Multiple dimensions of heterogeneity including data modality, laboratory practices, population characteristics, study design, and modeling approaches interact to produce distinct but interrelated forms of bias, such as sampling bias, label bias, structural confounding, and clinical context bias. These interactions collectively influence model performance, fairness, and generalizability. Importantly, mitigation strategies operate across all dimensions, emphasizing the need for an integrated, end-to-end approach rather than isolated interventions. 4.1 Layer 1: Data and Infrastructure Bias mitigation begins at the data level, where heterogeneity is most pronounced. Standardization of AST protocols, including harmonization of laboratory procedures and breakpoint systems, is critical to reduce label inconsistency and improve comparability across datasets (Gajic et al., 2022). Equally, standardized genomic quality control pipelines for WGS including consistent assembly, annotation, and resistance gene detection workflows are necessary to minimize inter-laboratory variability (Bortolaia et al., 2020; Doyle et al., 2020). Addressing global sampling imbalance requires expanding surveillance and data collection in underrepresented regions, particularly LMICs, to improve model generalizability across diverse resistance ecologies (Nsubuga et al., 2024). Development of well-curated benchmarking datasets with balanced representation across species, geography, and time further supports reproducibility and fair model comparison (Bortolaia et al., 2020; Doyle et al., 2020). 4.2 Layer 2: Modeling Approaches At the modeling stage, mitigation strategies focus on improving robustness while reducing reliance on spurious correlations in heterogeneous data. Robust algorithms including ensemble methods such as random forests and gradient boosting have demonstrated resilience to noisy and heterogeneous datasets (Ren et al., 2022). Bias-aware sample weighting can further improve fairness by adjusting for underrepresented groups, including geographic regions, pathogen types, and resistance classes. Structure-aware modeling approaches are particularly important in AMR prediction; incorporating bacterial population structure or phylogenetic information helps prevent models from exploiting lineage as a proxy for resistance, thereby improving biological validity and generalizability (Ardila et al., 2024). Strategies for handling class imbalance, such as resampling and cost-sensitive learning, can enhance detection of rare resistance phenotypes, although their effectiveness remains context-dependent (Bilal et al., 2025). 4.3 Layer 3: Evaluation Standards Robust evaluation is essential to ensure that models generalize beyond their training environment and provide clinically meaningful predictions. External validation using independent datasets and temporal validation across different time periods are critical for assessing real-world performance and identifying overfitting (Anahtar et al., 2021). Beyond traditional performance metrics such as AUROC, evaluation should incorporate calibration, subgroup performance, and fairness-oriented metrics to assess reliability across diverse populations (Chen et al., 2023). Benchmarking against clinical baselines, such as antibiograms or clinician decision-making, is also necessary to confirm that AI/ML models provide meaningful improvements over existing practice (Feretzakis et al., 2021). 4.4 Layer 4: Governance and Implementation Effective deployment of AI/ML-based AMR prediction models requires alignment with clinical workflows, ethical standards, and regulatory frameworks. Integration into antimicrobial stewardship (AMS) programs ensures that predictive models support, rather than replace, clinical decision-making, particularly in empiric therapy selection (Howard et al., 2024; Beaudoin et al., 2016). Ethical considerations, including fairness, transparency, and accountability, are essential to prevent reinforcement of existing healthcare inequities (Dankwa-Mullan, 2024). Ongoing monitoring of model performance is required to detect degradation due to temporal changes in resistance patterns and clinical practices. Adherence to data governance and privacy frameworks is also necessary for responsible and sustainable AI deployment in healthcare (Mennella et al., 2024). The proposed framework conceptualizes bias mitigation in AI/ML-based AMR prediction as a continuous, multi-layered process. Data-level interventions address representativeness and measurement validity; modeling approaches reduce structural and algorithmic bias; evaluation strategies ensure robustness and generalizability; and governance mechanisms enable safe and equitable deployment. Critically, these layers are interdependent—limitations in data quality and representation cannot be fully corrected through downstream modeling or evaluation alone. Effective AMR prediction systems therefore require an end-to-end design approach that integrates epidemiological, microbiological, and computational perspectives. An integrated mapping of data heterogeneity, associated biases, and mitigation strategies across the machine learning pipeline is presented in Table 5. Table 5. Integrated Mapping of Data Heterogeneity, Algorithmic Bias, and Mitigation Strategies Across the Machine Learning Pipeline. Pipeline Stage Source of Heterogeneity Associated Bias Impact Mitigation Strategy Data collection Geographic imbalance; limited LMIC data Sampling bias Poor generalizability Multi-center datasets; inclusive sampling Data labeling Breakpoint variability (CLSI/EUCAST) Label bias Misclassification Standardization; harmonized protocols Feature engineering EHR variability; missing data Measurement bias Noisy features Data preprocessing; validation pipelines Model training Class imbalance; skewed datasets Algorithmic bias Unequal performance Resampling; cost-sensitive learning Model validation Lack of external datasets Evaluation bias Overestimated performance External and temporal validation Deployment Temporal drift; clinical variability Performance degradation Reduced reliability Continuous monitoring; model updating 5. Discussion 5.1 Summary of Key Findings This systematic review using descriptive synthesis demonstrates that AI/ML approaches for AMR prediction consistently achieve strong internal performance, frequently reporting high discrimination metrics such as area under the receiver operating characteristic curve (AUROC). However, this apparent success is substantially constrained by pervasive data heterogeneity and multiple forms of algorithmic bias that collectively limit real-world applicability and clinical translation (Anahtar et al., 2021; Ardila et al., 2025). A central finding is that model performance is highly context-dependent. Many studies rely on internally consistent datasets, enabling models to implicitly learn local epidemiological patterns, institutional practices, or population structure rather than generalizable biological mechanisms. As a result, models often perform well under controlled conditions but fail to maintain accuracy when applied across different geographic regions, healthcare settings, or time periods (Kim et al., 2022; Kherabi et al., 2024). Data heterogeneity emerged as a multi-dimensional challenge encompassing variability in data modalities, laboratory protocols, population composition, and temporal and geographic distribution. These interacting dimensions complicate data integration, reduce reproducibility, and contribute to inconsistent model performance. In particular, the dominance of high-income country datasets and hospital-based populations introduces structural imbalances that further limit generalizability and equity. Algorithmic bias was similarly pervasive, with sampling bias, label inconsistency, structural confounding, and clinical context bias identified as major contributors to model fragility. These biases often cause models to exploit spurious correlations such as phylogenetic structure, institutional prevalence, or documentation patterns rather than learning causal determinants of resistance (Prosperi et al.,2022; Ren et al., 2022). Consequently, high internal accuracy frequently masks poor performance in external or real-world settings. Importantly, external validation remains limited across the literature. Where implemented, it consistently reveals reduced predictive performance compared to internal validation, highlighting the extent to which current evaluation practices overestimate model reliability (Tang et al., 2022; Bilal et al., 2025). This gap underscores a broader methodological issue in AMR machine learning: the prioritization of predictive accuracy over robustness, fairness, and generalizability. Taken together, these findings indicate that the primary constraint in AI/ML-based AMR prediction is not algorithmic capability, but the quality, structure, and representativeness of the underlying data. 5.2 Interpretation and Implications The findings highlight that the limitations of AI-based AMR prediction are fundamentally rooted in data-related challenges rather than insufficient model complexity. While increasingly sophisticated algorithms continue to be developed, their performance remains contingent on the characteristics of the training data. In many cases, models learn context-specific patterns such as institutional prescribing behaviour or dataset-specific correlations rather than biologically meaningful or generalizable relationships. A key insight is the interconnected nature of data heterogeneity and algorithmic bias. Heterogeneity across data modalities, laboratory practices, and populations creates conditions under which bias can emerge and propagate through the modeling pipeline. Geographic and demographic imbalances contribute to sampling bias, while variability in laboratory methods and breakpoint definitions introduces label inconsistency. These factors interact to amplify model instability and reduce external validity. The limited use of external and temporal validation further highlights a critical gap between model development and real-world deployment. Many studies rely on internal validation strategies that fail to capture distributional shifts across time, geography, and clinical settings, resulting in performance metrics that overestimate reliability in practice. This finding aligns with broader evidence in clinical machine learning, where failure to account for dataset shift remains a major barrier to implementation (Finlayson et al., 2021). From a clinical and public health perspective, these limitations carry important implications. Models that perform inconsistently across populations risk reinforcing existing health disparities, particularly in LMICs where AMR burden is highest but data representation is lowest (Nsubuga et al., 2024). In this context, fairness is not only a technical concern but also an ethical imperative (Dankwa-Mullan, 2024). The findings also suggest a necessary shift in research priorities. Rather than focusing solely on improving predictive performance, future work should prioritize data quality, standardization, and representativeness. Development of multi-center, globally representative datasets, standardized laboratory practices, and transparent reporting frameworks is essential to improving reproducibility and comparability. The Heterogeneity Mitigation Framework proposed in this study provides a structured approach to addressing these challenges across the full machine learning pipeline, integrating data-level standardization, bias-aware modeling, robust evaluation, and governance mechanisms. 5.3 Limitations This study has several limitations warranting consideration. First, this systematic review employed descriptive synthesis rather than formal meta-analysis. No pooled estimates of model performance or formal assessment of statistical heterogeneity across studies were produced, as the degree of methodological, clinical, and outcome heterogeneity across included studies precluded meaningful statistical aggregation. Consequently, findings are presented as structured qualitative and descriptive syntheses. Second, the review is dependent on the quality and reporting of included studies. Variability in study design, data sources, outcome definitions, and evaluation metrics limited direct comparability. Inconsistent reporting of validation strategies, preprocessing methods, and bias mitigation techniques may have led to underestimation or overestimation of certain methodological challenges. Third, publication bias may be present, as studies reporting higher predictive performance or novel methodologies are more likely to be published, potentially yielding an overly optimistic representation of AI performance in AMR prediction. Exclusion of grey literature and non-English publications may also have limited comprehensiveness. Fourth, the rapidly evolving nature of AI and AMR research means that emerging methodologies, datasets, and real-world implementation studies may not be fully captured within the review timeframe. Finally, while this study proposes a Heterogeneity Mitigation Framework, it has not yet been empirically validated. Its effectiveness will depend on future application and evaluation across diverse datasets and clinical settings. It should therefore be considered a conceptual guide rather than a definitive solution. 6. Conclusion Artificial intelligence and machine learning hold substantial promise for advancing AMR prediction by enabling rapid, data-driven insights and supporting clinical decision-making. However, this systematic review demonstrates that current AI/ML-based approaches are fundamentally constrained not by algorithmic limitations, but by pervasive data heterogeneity and multiple forms of algorithmic bias that undermine generalizability and real-world applicability (Anahtar et al., 2021; Bilal et al., 2025). The findings highlight that strong internal model performance does not reliably translate into consistent performance across diverse populations, healthcare settings, and time periods. Variability in data modalities, laboratory practices, population representation, and evaluation strategies introduces systematic challenges that reduce reproducibility and external validity. These issues are further compounded by limited use of external validation and the underrepresentation of LMICs in existing datasets, despite their bearing the highest burden of AMR (Nsubuga et al., 2024). Addressing these challenges requires a fundamental shift in research priorities from optimizing predictive accuracy alone to improving data quality, standardization, representativeness, and fairness. The Heterogeneity Mitigation Framework proposed in this study provides a structured, multi-layered approach to addressing these limitations across the machine learning pipeline, encompassing data infrastructure, modeling strategies, evaluation standards, and governance considerations. Future research should prioritize the development of globally representative, multi-center datasets; standardized laboratory and reporting practices; and robust validation frameworks that reflect real-world conditions. Interdisciplinary collaboration among clinicians, microbiologists, data scientists, and policymakers will be essential to ensure that AI systems are both scientifically robust and clinically meaningful. Ultimately, successful translation of AI into AMR prediction depends on addressing foundational challenges related to data heterogeneity, bias, and evaluation rigor. A coordinated, end-to-end approach that integrates data quality, methodological robustness, and ethical considerations is essential for developing reliable, equitable, and clinically deployable AI systems that can meaningfully contribute to antimicrobial stewardship and global health. Declarations Ethics Approval and Consent to Participate: Not applicable. This study is a systematic review based on previously published data and does not involve direct human participants or personal data collection. Consent for Publication: Not applicable. Funding: The author received no external funding for this study. Competing Interests: The author declares no competing interests. Data Availability: All data used in this study are derived from publicly available sources cited in the manuscript. No primary data were generated. References Abdelhay O, Alshoubaki R, Murad S, Abdel-Hafez O, Abdelhay Q, Haddad B et al (2025) Using statistical modelling and machine learning in detecting bone properties: A systematic review protocol. PLoS ONE 20(3):e0319583. 10.1371/journal.pone.0319583 Ali T, Ahmed S, Aslam M (2023) Artificial intelligence for antimicrobial resistance prediction: Challenges and opportunities towards practical implementation. Antibiotics 12(3):523. 10.3390/antibiotics12030523 Anahtar MN, Yang JH, Kanjilal S (2021) Applications of machine learning to the problem of antimicrobial resistance: An emerging model for translational research. J Clin Microbiol 59(7):e01260–e01220. 10.1128/JCM.01260-20 Ardila CM, González-Arroyave D, Tobón S (2025) Machine learning for predicting antimicrobial resistance in critical and high-priority pathogens: A systematic review considering antimicrobial susceptibility tests in real-world healthcare settings. PLoS ONE 20(2):e0319460. 10.1371/journal.pone.0319460 Ardila CM, Yadalam PK, González-Arroyave D (2024) Integrating whole genome sequencing and machine learning for predicting antimicrobial resistance in critical pathogens: A systematic review of antimicrobial susceptibility tests. PeerJ 12:e18213. 10.7717/peerj.18213 Beaudoin M, Kabanza F, Nault V, Valiquette L (2016) Evaluation of a machine learning capability for a clinical decision support system to enhance antimicrobial stewardship programs. Artif Intell Med 68:29–36. 10.1016/j.artmed.2016.01.004 Bilal H, Khan MN, Khan S, Shafiq M, Fang W, Khan RU et al (2025) The role of artificial intelligence and machine learning in predicting and combating antimicrobial resistance. Comput Struct Biotechnol J 27:423–439. 10.1016/j.csbj.2025.01.006 Bortolaia V, Kaas RS, Ruppe E, Roberts MC, Schwarz S, Cattoir V et al (2020) ResFinder 4.0 for predictions of phenotypes from genotypes. J Antimicrob Chemother 75(12):3491–3500. 10.1093/jac/dkaa345 Boyd AD, Gonzalez-Guarda R, Lawrence K, Patil CL, Ezenwa MO, O'Brien EC et al (2023) Equity and bias in electronic health records data. Contemp Clin Trials 130:107238. 10.1016/j.cct.2023.107238 Chen RJ, Wang JJ, Williamson DF, Chen TY, Lipkova J, Lu MY et al (2023) Algorithmic fairness in artificial intelligence for medicine and healthcare. Nat Biomed Eng 7(6):719–742. 10.1038/s41551-023-01056-8 Cohen SA, Chen Z, Bian J, Boucher C, Wu Y, Prosperi M (2025) Comparative evaluation of clinical large language models and machine learning to predict antimicrobial resistance in hospital-onset sepsis. In: International Conference on Artificial Intelligence in Medicine. Cham: Springer Nature Switzerland; pp. 65–76. 10.1007/978-3-031-93528-7_7 Collignon P, Beggs JJ (2025) The persistence of antibiotic resistance in observational studies: Is it really due to differences in sub-populations rather than antibiotic use? Antibiotics 14(1):39. 10.3390/antibiotics14010039 Dankwa-Mullan I (2024) Health equity and ethical considerations in using artificial intelligence in public health and medicine. Prev Chronic Dis 21:E64. 10.5888/pcd21.240245 Doyle RM, O'Sullivan DM, Aller SD, Bruchmann S, Clark T, Coello Pelegrin A et al (2020) Discordant bioinformatic predictions of antimicrobial resistance from whole-genome sequencing data of bacterial isolates: An inter-laboratory study. Microb Genomics 6(2):e000335. 10.1099/mgen.0.000335 Ferrari D, Arina P, Edgeworth J, Curcin V, Guidetti V, Mandreoli F et al (2024) Using interpretable machine learning to predict bloodstream infection and antimicrobial resistance in patients admitted to ICU. PLOS Digit Health 3(10):e0000641. 10.1371/journal.pdig.0000641 Feretzakis G, Loupelis E, Sakagianni A, Kalles D, Martsoukou M, Lykoudis P et al (2021) Using machine learning techniques to aid empirical antibiotic therapy decisions in the intensive care unit of a general hospital in Greece. Antibiotics 10(9):1112. 10.3390/antibiotics10091112 Finlayson SG, Subbaswamy A, Singh K, Bowers J, Kupke A, Zittrain J et al (2021) The clinician and dataset shift in artificial intelligence. N Engl J Med 385(3):283–286. 10.1056/NEJMc2104626 Gajic I, Kabic J, Kekic D, Jovicevic M, Milenkovic M, Mitic Culafic D et al (2022) Antimicrobial susceptibility testing: A comprehensive review of currently used methods. Antibiotics 11(4):427. 10.3390/antibiotics11040427 Gao J, Chou B, McCaw ZR, Thurston H, Varghese P, Hong C et al (2025) What is fair? Defining fairness in machine learning for health. Stat Med 44(20–22):e70234. 10.1002/sim.70234 Gianfrancesco MA, Goldstein ND (2021) A narrative review on the validity of electronic health record-based research in epidemiology. BMC Med Res Methodol 21(1):234. 10.1186/s12874-021-01422-3 Howard A, Aston S, Gerada A, Reza N, Bincalar J, Mwandumba H et al (2024) Antimicrobial learning systems: An implementation blueprint for artificial intelligence to tackle antimicrobial resistance. Lancet Digit Health 6(1):e79–e86. 10.1016/S2589-7500(23)00180-6 Humphries R, Bobenchik AM, Hindler JA, Schuetz AN (2021) Overview of changes to the clinical and laboratory standards institute performance standards for antimicrobial susceptibility testing, M100. J Clin Microbiol 59(12):e00213–e00221. 10.1128/JCM.00213-21 Kherabi Y, Thy M, Bouzid D, Antcliffe DB, Rawson TM, Peiffer-Smadja N (2024) Machine learning to predict antimicrobial resistance: Future applications in clinical practice? Infect Dis Now 54(3):104864. 10.1016/j.idnow.2024.104864 Kim JI, Maguire F, Tsang KK, Gouliouris T, Peacock SJ, McAllister TA et al (2022) Machine learning for antimicrobial resistance prediction: Current practice, limitations, and clinical perspective. Clin Microbiol Rev 35(3):e00179–e00121. 10.1128/cmr.00179-21 Lewnard JA, Charani E, Gleason A, Hsu LY, Khan WA, Karkey A et al (2024) Burden of bacterial antimicrobial resistance in low-income and middle-income countries aversible by existing interventions: An evidence review and modeling analysis. Lancet 403(10442):2439–2454. 10.1016/S0140-6736(24)00474-4 Loh HW, Ooi CP, Seoni S, Barua PD, Molinari F, Acharya UR (2022) Application of explainable artificial intelligence for healthcare: A systematic review of the last decade (2011–2022). Comput Methods Programs Biomed 226:107161. 10.1016/j.cmpb.2022.107161 McGowan J, Straus S, Moher D, Langlois EV, O'Brien KK, Horsley T et al (2020) Reporting scoping reviews—PRISMA ScR extension. J Clin Epidemiol 123:177–179. 10.1016/j.jclinepi.2019.09.016 Mennella C, Maniscalco U, De Pietro G, Esposito M (2024) Ethical and regulatory challenges of AI technologies in healthcare: A narrative review. Heliyon 10(4). 10.1016/j.heliyon.2024.e26297 Moccia S (2024) Trustworthy AI in neuroengineering: From data management to ethics assessment. In: To Understand and Restore the Nervous System. p. 205. 10.1088/978-0-7503-3693-8ch10 Murray CJ, Ikuta KS, Sharara F, Swetschinski L, Aguilar GR, Gray A et al (2022) Global burden of bacterial antimicrobial resistance in 2019: A systematic analysis. Lancet 399(10325):629–655. 10.1016/S0140-6736(21)02724-0 Nsubuga M, Galiwango R, Jjingo D, Mboowa G (2024) Generalizability of machine learning in predicting antimicrobial resistance in E. coli: A multi-country case study in Africa. BMC Genomics 25(1):287. 10.1186/s12864-024-10214-4 Page MJ, McKenzie JE, Bossuyt PM, Boutron I, Hoffmann TC, Mulrow CD et al (2021) The PRISMA 2020 statement: An updated guideline for reporting systematic reviews. BMJ 372. 10.1136/bmj.n71 Park Y, Weig M, Noll C, Bader O, Hauschild AC (2024) Effect of data heterogeneity in clinical MALDI-TOF mass spectra profiles on direct antimicrobial resistance prediction through machine learning. bioRxiv. 10.1101/2024.10.07.617009 Peters MD, Marnie C, Tricco AC, Pollock D, Munn Z, Alexander L et al (2020) Updated methodological guidance for the conduct of scoping reviews. JBI Evid Synth 18(10):2119–2126. 10.11124/JBIES-20-00167 Prosperi M, Boucher C, Bian J, Marini S (2022) Assessing putative bias in prediction of anti-microbial resistance from real-world genotyping data under explicit causal assumptions. Artif Intell Med 130:102326. 10.1016/j.artmed.2022.102326 Ren Y, Chakraborty T, Doijad S, Falgenhauer L, Falgenhauer J, Goesmann A et al (2022) Prediction of antimicrobial resistance based on whole-genome sequencing and machine learning. Bioinformatics 38(2):325–334. 10.1093/bioinformatics/btab681 Sakagianni A, Koufopoulou C, Feretzakis G, Kalles D, Verykios VS, Myrianthefs P (2023) Using machine learning to predict antimicrobial resistance: A literature review. Antibiotics 12(3):452. 10.3390/antibiotics12030452 Sidorczuk K, Gagat P, Pietluch F, Kala J, Rafacz D, Bakala L et al (2022) Benchmarks in antimicrobial peptide prediction are biased due to the selection of negative data. Brief Bioinform 23(5):bbac343. 10.1093/bib/bbac343 Tang R, Luo R, Tang S, Song H, Chen X (2022) Machine learning in predicting antimicrobial resistance: A systematic review and meta-analysis. Int J Antimicrob Agents 60(5–6):106684. 10.1016/j.ijantimicag.2022.106684 Tricco AC, Lillie E, Zarin W, O'Brien KK, Colquhoun H, Levac D et al (2018) PRISMA extension for scoping reviews (PRISMA-ScR): Checklist and explanation. Ann Intern Med 169(7):467–473. 10.7326/M18-0850 Valavarasu S, Sangu Y, Mahapatra T (2025) Prediction of antibiotic resistance from antibiotic susceptibility testing results from surveillance data using machine learning. Sci Rep 15(1):30509. 10.1038/s41598-025-14940-3 Weis C, Cuenod A, Rieck B, Llinares-Lopez F, Dubuis O, Graf S et al (2020) Direct antimicrobial resistance prediction from MALDI-TOF mass spectra profile in clinical isolates through machine learning. bioRxiv 1–35. 10.1101/2020.07.30.228411 Wenzler E, Maximos M, Asempa TE, Biehle L, Schuetz AN, Hirsch EB (2023) Antimicrobial susceptibility testing: An updated primer for clinicians in the era of antimicrobial resistance. Pharmacotherapy 43(4):264–278. 10.1002/phar.2770 Xu W, Ouyang F (2022) The application of AI technologies in STEM education: A systematic review from 2011 to 2021. Int J STEM Educ 9(1):59. 10.1186/s40594-022-00377-5 Zhao C, Wang Y, Mulchandani R, Van Boeckel TP (2024) Global surveillance of antimicrobial resistance in food animals using priority drugs maps. Nat Commun 15(1):763. 10.1038/s41467-024-44713-3 Zou Z, Tang F, Qiao L, Wang S, Zhang H (2025) Integrating sequencing methods with machine learning for antimicrobial susceptibility testing in pediatric infections: Current advances and future insights. Front Microbiol 16:1528696. 10.3389/fmicb.2025.1528696 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-9552638","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Systematic Review","associatedPublications":[],"authors":[{"id":630988077,"identity":"79685c1c-c66c-48f5-94ca-20ee4a8469ac","order_by":0,"name":"Jema Kiazolu","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA20lEQVRIie3OvQrCMBDA8SsHZol0bRH7DCeOBX2VSsEuDoKLg0NBaEfXgk8hvkAhoEvBWVwEV4c8gXhaxK3NKJg/5GPIjwuAzfaDefja54AgUqcEPkEYEeKnsoSaYBuBmvAtMiR+jjetSQnXv1MJy3CSYvfaSHrYoUFBCv3tjEmVMBHUSAIEiiUTujBxMsWk0/yxAIVWLzI+V0weBqSHcrB+T/Ekk9SA+Gu5cApK0Kum8zI6JMOsjXin417rZRi7udpd9Srsb0TWTD7F7z3i1TLj28j0oc1ms/1hTwUUO9/7aKGcAAAAAElFTkSuQmCC","orcid":"https://orcid.org/0009-0004-9603-7111","institution":"Hasanuddin University","correspondingAuthor":true,"prefix":"","firstName":"Jema","middleName":"","lastName":"Kiazolu","suffix":""}],"badges":[],"createdAt":"2026-04-28 10:37:30","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-9552638/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9552638/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":108105552,"identity":"119ce50e-5729-4c2f-a6b8-bc9ea46c6d93","added_by":"auto","created_at":"2026-04-29 11:50:18","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":349551,"visible":true,"origin":"","legend":"\u003cp\u003eSee image above for figure legend\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-9552638/v1/45a3b32ab411eab3671895b8.png"},{"id":108182413,"identity":"50c0c429-c486-498f-9776-f8f7f5a9ec16","added_by":"auto","created_at":"2026-04-30 08:59:21","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":740502,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9552638/v1/ae5a6fe4-c778-4222-b6f3-4a4a5e131e5b.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003e\u003cstrong\u003eData Heterogeneity and Algorithmic Bias in AI-Based Antimicrobial Resistance Prediction: A Systematic Review and Mitigation Framework\u003c/strong\u003e\u003c/p\u003e","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eAntimicrobial resistance (AMR) has emerged as one of the most pressing global public health challenges of the 21st century. The widespread misuse and overuse of antimicrobial agents in human medicine, agriculture, and animal health have accelerated the emergence and dissemination of resistant pathogens, rendering infections that were once readily treatable increasingly difficult or impossible to manage. Recent global estimates indicate that AMR contributes substantially to both morbidity and mortality, with millions of deaths associated with resistant infections each year (Murray et al., 2022). This burden is disproportionately concentrated in low- and middle-income countries (LMICs), where deficiencies in diagnostic capacity, surveillance infrastructure, and access to effective therapies further amplify adverse outcomes (Lewnard et al., 2024).\u003c/p\u003e \u003cp\u003eIn response to the growing complexity of AMR, artificial intelligence (AI) and machine learning (ML) have increasingly been applied to enhance the detection, prediction, and clinical management of resistance. These methods leverage large-scale biomedical datasets encompassing whole-genome sequencing (WGS), antimicrobial susceptibility testing (AST), electronic health records (EHRs), surveillance systems, and mass spectrometry to identify complex, non-linear associations with resistance phenotypes (Anahtar et al., 2021; Kim et al., 2022). Commonly employed models, including random forests, support vector machines, gradient boosting algorithms, and deep learning architectures, have demonstrated moderate to high predictive performance across a range of pathogens and clinical contexts (Tang et al., 2022; Bilal et al., 2025). Beyond prediction, AI-driven tools are progressively being incorporated into clinical decision support systems and antimicrobial stewardship programs to guide empiric therapy selection and optimize antibiotic prescribing (Ferrari et al., 2024; Cohen et al., 2025).\u003c/p\u003e \u003cp\u003eDespite these advances, the clinical translation of AI/ML-based AMR prediction models remains substantially limited. A central challenge is data heterogeneity, which arises from variability in data sources, formats, laboratory protocols, and patient populations. AMR datasets are inherently multi-modal, integrating genomic, phenotypic, clinical, and surveillance data that each differ in structure, quality, and completeness. Such variability complicates data integration, reduces reproducibility, and undermines model generalizability across settings (Kherabi et al., 2024; Sakagianni et al., 2023). Closely related to heterogeneity is algorithmic bias, whereby models learn systematic distortions embedded in training data rather than true biological or clinical relationships. In healthcare applications, these biases can produce unequal model performance across subpopulations, potentially reinforcing existing health disparities (Chen et al., 2023; Gao et al., 2025).\u003c/p\u003e \u003cp\u003eFor instance, models trained predominantly on data from high-income settings may perform poorly in LMIC contexts, where resistance patterns, healthcare systems, and data infrastructures differ substantially (Nsubuga et al., 2024; Ali et al., 2023). Furthermore, many AMR prediction models rely on single-center datasets and internal validation strategies, resulting in overestimated performance metrics and limited external validity when applied to new populations or clinical environments (Kim et al., 2022; Tang et al., 2022). These challenges expose a critical gap between methodological development and real-world implementation of AI in AMR prediction.\u003c/p\u003e \u003cp\u003eWhile existing studies have explored the application of ML to AMR prediction, a comprehensive synthesis focusing specifically on the interconnected roles of data heterogeneity and algorithmic bias remains lacking. In particular, the mechanisms through which heterogeneity generates bias, and how these factors jointly influence model robustness, fairness, and generalizability, have not been systematically characterized.\u003c/p\u003e \u003cp\u003eThis study therefore aims to synthesize current evidence on data heterogeneity and algorithmic bias in AI/ML-based AMR prediction and to propose a structured framework for their mitigation. By shifting focus from predictive performance alone to considerations of robustness, fairness, and real-world applicability, this work seeks to support the development of reliable and clinically deployable AI systems for antimicrobial resistance management. This study uniquely integrates data heterogeneity and algorithmic bias into a unified analytical framework, addressing a critical gap in AI-based AMR prediction research\u003c/p\u003e"},{"header":"2. Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Study Design\u003c/h2\u003e \u003cp\u003eThis study employed a systematic review design with descriptive synthesis, conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA 2020) framework (Page et al., 2021). A systematic review approach was selected to provide structured, reproducible, and transparent identification and synthesis of evidence on data heterogeneity and algorithmic bias in AI-based AMR prediction. The review protocol emphasized comprehensive database searching, duplicate-independent study selection, standardized data extraction, and thematic synthesis. No meta-analysis was conducted owing to substantial methodological and outcome heterogeneity across included studies, which precluded statistical pooling. This review was not prospectively registered in PROSPERO.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Eligibility Criteria\u003c/h2\u003e \u003cp\u003eEligibility criteria were defined using the Population\u0026ndash;Concept\u0026ndash;Context (PCC) framework to ensure structured and transparent study selection (Peters et al., 2020).\u003c/p\u003e \u003cp\u003e \u003cb\u003eInclusion Criteria\u003c/b\u003e \u003c/p\u003e \u003cp\u003eStudies were included if they: (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e) applied AI, machine learning (ML), or deep learning (DL) methods to predict AMR-related outcomes, including susceptibility classification, minimum inhibitory concentration (MIC) prediction, or resistance risk estimation; (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e) reported model development, validation strategies, or performance metrics such as accuracy, sensitivity, specificity, F1-score, or area under the curve (AUC); (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e) utilized one or more data modalities including AST, WGS, EHRs, surveillance datasets, or spectral data such as MALDI-TOF mass spectrometry; (\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e) were conducted in clinical, laboratory, public health, or surveillance contexts; and (\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e) were primary research studies, conference proceedings, or review articles providing methodological or conceptual insights.\u003c/p\u003e \u003cp\u003e \u003cb\u003eExclusion Criteria\u003c/b\u003e \u003c/p\u003e \u003cp\u003eStudies were excluded if they: (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e) used non-AI approaches, including rule-based or traditional statistical models without machine learning components; (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e) were purely experimental laboratory studies without predictive modeling; (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e) did not report predictive outcomes or model performance; or (\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e) were editorials, commentaries, or opinion papers without methodological or empirical contribution. These criteria ensured that included studies contributed to both empirical evidence and conceptual understanding of heterogeneity, bias, and generalizability in AI-based AMR prediction.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3 Information Sources and Database Search\u003c/h2\u003e \u003cp\u003eA comprehensive literature search was conducted across multiple electronic databases, including PubMed/MEDLINE, Embase, Scopus, Web of Science, IEEE Xplore, and the ACM Digital Library, to capture interdisciplinary research spanning biomedical, computational, and public health domains. The search was restricted to studies published between January 2020 and April 2026 to reflect recent methodological developments in artificial intelligence and antimicrobial resistance prediction. To enhance comprehensiveness, reference lists of included studies and relevant reviews were manually screened, and database-specific indexing terms, including MeSH and Emtree, were adapted accordingly. The study selection process followed PRISMA 2020 reporting standards (Page et al., 2021), including identification, screening, eligibility assessment, and inclusion of studies. The flow of study selection is illustrated in Fig.\u0026nbsp;1.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.4 Search Strategy\u003c/h2\u003e \u003cp\u003eThe search strategy combined controlled vocabulary and free-text terms related to AI and AMR prediction. Key concepts included core terms (\"artificial intelligence,\" \"machine learning,\" \"deep learning\"), domain terms (\"antimicrobial resistance,\" \"AMR,\" \"prediction,\" \"genomics,\" \"whole-genome sequencing\"), and clinical terms (\"antimicrobial susceptibility testing,\" \"surveillance,\" \"stewardship\"). Boolean operators (AND, OR, NOT) were applied to optimize retrieval. Searches were restricted to studies published between January 2020 and April 2026 to capture recent methodological developments in AI and biomedical data integration.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e2.5 Study Selection\u003c/h2\u003e \u003cp\u003eAll retrieved records were imported into a reference management system and screened using a systematic review platform. Screening was conducted in two sequential stages: (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e) title and abstract screening, in which two independent reviewers assessed records against eligibility criteria; and (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e) full-text screening, in which potentially eligible studies were evaluated in detail. Discrepancies were resolved through discussion or consultation with a third reviewer. The selection process was documented using a PRISMA 2020 flow diagram (Page et al., 2021).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e2.6 Data Extraction\u003c/h2\u003e \u003cp\u003eA standardized data extraction form was developed and piloted to ensure consistency and completeness. Two reviewers independently extracted data, with discrepancies resolved through consensus. Extracted variables included: pathogen and antibiotic focus; data modality (e.g., WGS, AST, EHR, surveillance); model type (e.g., random forest, support vector machine, neural networks); validation strategy (internal and/or external); performance metrics (e.g., accuracy, AUC, F1-score); sources of heterogeneity; types of bias; and reported mitigation strategies.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e2.7 Data Synthesis\u003c/h2\u003e \u003cp\u003eA combined descriptive and thematic synthesis approach was employed, consistent with established systematic review methodology for evidence that is unsuitable for meta-analysis (Peters et al., 2020). Descriptive synthesis involved summarizing study characteristics using tables and visualizations, including modality distribution, model types, validation strategies, and geographic representation, to enable structured comparison across studies. Thematic synthesis involved a narrative approach to identify key patterns related to dimensions of data heterogeneity (e.g., modality, population, temporal variation), types of algorithmic bias (e.g., sampling bias, label bias, structural confounding), and strategies for their mitigation. Themes were developed inductively and refined through iterative reviewer discussion. Formal meta-analysis was not performed due to the substantial clinical, methodological, and statistical heterogeneity across included studies, which precluded meaningful pooling of effect estimates.\u003c/p\u003e \u003c/div\u003e"},{"header":"3. Results","content":"\u003ch2\u003e3.1 Overview of Included Studies\u003c/h2\u003e\n\u003cp\u003eA total of 68 studies published between 2020 and 2026 were included in this systematic review, of which 21 met the criteria for detailed analysis. These 21 studies were selected for in-depth analysis based on their relevance to data heterogeneity and algorithmic bias, as well as the completeness of methodological reporting. This reflects a growing yet methodologically heterogeneous body of literature on AI/ML-based AMR prediction. The characteristics of the 21 studies selected for detailed analysis are presented in Table 1.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 1. Characteristics of the 21 Studies Selected for Detailed Analysis in AI/ML-Based AMR Prediction (2020\u0026ndash;2026).\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"100%\"\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eFirst Author (Year)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eStudy Type\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eData Modality\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eModel Type\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003ePathogen Scope\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eSetting\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eRegion\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eKhaledi (2020)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePrimary\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eWGS\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eRF, SVM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 18.7783%;\"\u003e\n \u003cp\u003eP. aeruginosa\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\" style=\"width: 12.5565%;\"\u003e\n \u003cp\u003eHospital\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eGermany\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eWeis (2020)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePrimary\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eMALDI-TOF\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eRF, CNN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 18.7783%;\"\u003e\n \u003cp\u003eMultiple\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\" style=\"width: 12.5565%;\"\u003e\n \u003cp\u003eClinical lab\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eSwitzerland\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eRen (2022)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePrimary\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eWGS\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eRF, DNN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 18.7783%;\"\u003e\n \u003cp\u003eMultiple\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\" style=\"width: 12.5565%;\"\u003e\n \u003cp\u003eClinical isolates\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eGermany\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eFeretzakis (2021)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePrimary\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAST, Surveillance\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eGradient Boosting\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 18.7783%;\"\u003e\n \u003cp\u003eMultiple\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\" style=\"width: 12.5565%;\"\u003e\n \u003cp\u003eICU\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eGreece\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eArdila (2024)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePrimary\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eWGS + AST\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eRF, Ensemble\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 18.7783%;\"\u003e\n \u003cp\u003ePriority pathogens\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\" style=\"width: 12.5565%;\"\u003e\n \u003cp\u003eHealthcare\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eColombia\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eFerrari (2024)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePrimary\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eEHR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eLR, RF\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 18.7783%;\"\u003e\n \u003cp\u003eBloodstream infections\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\" style=\"width: 12.5565%;\"\u003e\n \u003cp\u003eICU\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eUK\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePark (2024)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePrimary\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eMALDI-TOF\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eCNN, RF\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 18.7783%;\"\u003e\n \u003cp\u003eMultiple\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\" style=\"width: 12.5565%;\"\u003e\n \u003cp\u003eClinical lab\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eGermany\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCohen (2025)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePrimary\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eEHR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eML, LLM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 18.7783%;\"\u003e\n \u003cp\u003eSepsis\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\" style=\"width: 12.5565%;\"\u003e\n \u003cp\u003eHospital\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eUSA\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eValavarasu (2025)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePrimary\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eSurveillance, AST\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eTime-series ML\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 18.7783%;\"\u003e\n \u003cp\u003eBroad\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\" style=\"width: 12.5565%;\"\u003e\n \u003cp\u003eSurveillance\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eIndia\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eZou (2025)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePrimary\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eWGS\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eCNN, DNN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 18.7783%;\"\u003e\n \u003cp\u003ePediatric infections\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\" style=\"width: 12.5565%;\"\u003e\n \u003cp\u003eHospital\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eChina\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eUrena (2024)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePrimary\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eClinical + Ecology\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eML\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 18.7783%;\"\u003e\n \u003cp\u003eHospital infections\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\" style=\"width: 12.5565%;\"\u003e\n \u003cp\u003eHospital\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eFrance\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eHardan (2024)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePrimary\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eEHR (multimodal)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eNeural Networks\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 18.7783%;\"\u003e\n \u003cp\u003eMultiple\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\" style=\"width: 12.5565%;\"\u003e\n \u003cp\u003eHospital\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eMiddle East\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eBlechman (2024)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePrimary\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eEHR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eGradient Boosting\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 18.7783%;\"\u003e\n \u003cp\u003eBroad infections\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\" style=\"width: 12.5565%;\"\u003e\n \u003cp\u003eHospital\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eUSA\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCondorelli (2024)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePrimary\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eWGS\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eRF, XGBoost\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 18.7783%;\"\u003e\n \u003cp\u003eK. pneumoniae\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\" style=\"width: 12.5565%;\"\u003e\n \u003cp\u003eClinical isolates\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eItaly\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eZhao (2024)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePrimary\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eSurveillance\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eML\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 18.7783%;\"\u003e\n \u003cp\u003eFood-animal pathogens\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\" style=\"width: 12.5565%;\"\u003e\n \u003cp\u003eSurveillance\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eGlobal\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eArdila (2025)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePrimary\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAST\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eRF, Gradient Boosting\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 18.7783%;\"\u003e\n \u003cp\u003ePriority pathogens\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\" style=\"width: 12.5565%;\"\u003e\n \u003cp\u003eHealthcare\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eColombia\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eNsubuga (2024)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePrimary\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eWGS\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eML\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 18.7783%;\"\u003e\n \u003cp\u003eE. coli\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\" style=\"width: 12.5565%;\"\u003e\n \u003cp\u003eMulticountry\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eAfrica\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eProsperi (2022)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePrimary\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eGenomic\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eML\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 18.7783%;\"\u003e\n \u003cp\u003eMultiple\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\" style=\"width: 12.5565%;\"\u003e\n \u003cp\u003eMixed\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eUSA\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eSidorczuk (2022)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePrimary\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePeptide/Genomic\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eML\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 18.7783%;\"\u003e\n \u003cp\u003eAntimicrobial peptides\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\" style=\"width: 12.5565%;\"\u003e\n \u003cp\u003eComputational\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eEurope\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eKim (2022)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eReview\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eWGS\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003e\u0026mdash;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 18.7783%;\"\u003e\n \u003cp\u003eBroad\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\" style=\"width: 12.5565%;\"\u003e\n \u003cp\u003eClinical\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eCanada/UK\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTang (2022)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eReview\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eMulti-modal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003e\u0026mdash;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 18.7783%;\"\u003e\n \u003cp\u003eBroad\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\" style=\"width: 12.5565%;\"\u003e\n \u003cp\u003eMixed\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eChina\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 18.7783%;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 3.3936%;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 9.2132%;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cstrong\u003eAbbreviations:\u003c/strong\u003e WGS = Whole-genome sequencing; AST = Antimicrobial susceptibility testing; EHR = Electronic health records; RF = Random Forest; SVM = Support Vector Machine; CNN = Convolutional Neural Network; DNN = Deep Neural Network; LR = Logistic Regression.\u003c/p\u003e\n\u003cp\u003eThe included studies utilized diverse data modalities, underscoring the multi-domain nature of AMR prediction. These modalities included WGS, AST phenotypes, EHRs, spectral data (e.g., MALDI-TOF mass spectrometry), and surveillance datasets. WGS-based approaches were the most frequently represented, followed by AST- and EHR-based models, while spectral and surveillance data were comparatively less common (Ardila et al., 2024; Weis et al., 2020; Valavarasu et al., 2025).\u003c/p\u003e\n\u003cp\u003eGeographically, the evidence base was heavily concentrated in high-income regions, particularly North America, Europe, and East Asia, where access to genomic sequencing infrastructure, clinical informatics systems, and large-scale surveillance data is more established (Kim et al., 2022; Zou et al., 2025). In contrast, representation from LMICs remained limited despite their disproportionate burden of AMR (Ali et al., 2023; Nsubuga et al., 2024).\u003c/p\u003e\n\u003cp\u003eMost studies were conducted in hospital-based settings, particularly among inpatient and intensive care unit (ICU) populations, where AMR prevalence is high and data availability is greater (Ferrari et al., 2024; Cohen et al., 2025). Community-level and outpatient datasets were comparatively underrepresented. Although surveillance-based studies partially expanded geographic and population coverage, important gaps persisted in non-hospital and resource-limited settings (Valavarasu et al., 2025). Overall, the included studies demonstrate substantial heterogeneity across data modalities, geographic regions, and clinical settings, with a clear dominance of high-resource, hospital-based datasets an imbalance with important implications for the generalizability and equity of AI/ML-based AMR prediction models.\u003c/p\u003e\n\u003ch2\u003e3.2 Dimensions of Data Heterogeneity\u003c/h2\u003e\n\u003cp\u003eData heterogeneity emerged as a central determinant of model performance, reproducibility, and generalizability. Heterogeneity was observed across multiple interacting dimensions, including data modality, laboratory practices, population characteristics, and temporal and geographic variation (Tang et al., 2022; Sakagianni et al., 2023; Zou et al., 2025).\u003c/p\u003e\n\u003ch3\u003e3.2.1 Data Modality and Measurement Heterogeneity\u003c/h3\u003e\n\u003cp\u003eData modality heterogeneity arises from the integration of fundamentally different data sources, each capturing distinct aspects of AMR. WGS-based approaches depend on sequencing quality, bioinformatics pipelines, resistance gene databases, and genotype\u0026ndash;phenotype mapping strategies, resulting in variability in feature representation and predictive outputs (Ardila et al., 2024; Ren et al., 2022). AST-based models are influenced by methodological differences including broth microdilution, disk diffusion, and automated systems, as well as laboratory-specific protocols and class imbalance in resistance phenotypes (Valavarasu et al., 2025; Wenzler et al., 2023). EHR-derived data introduce further heterogeneity through variability in coding systems, documentation practices, and missing data patterns, complicating integration with genomic and phenotypic data (Gianfrancesco \u0026amp; Goldstein, 2021; Boyd et al., 2023). Similarly, MALDI-TOF-based approaches are sensitive to instrument calibration, sample preparation, and spectral preprocessing, generating variability across laboratories (Weis et al., 2020; Park et al., 2024). Collectively, these modality-specific differences create significant interoperability challenges and limit the development of unified, multi-modal prediction models.\u003c/p\u003e\n\u003ch3\u003e3.2.2 Laboratory and Technical Heterogeneity\u003c/h3\u003e\n\u003cp\u003eLaboratory protocols introduce systematic variability that directly affects resistance measurement and model performance. Differences in AST methodologies and interpretive breakpoints particularly between CLSI and EUCAST standards lead to inconsistent resistance classification across datasets (Humphries et al., 2021; Wenzler et al., 2023). In WGS-based approaches, sequencing depth, quality control procedures, and bioinformatics pipelines further contribute to heterogeneity, resulting in discordant resistance gene detection and reduced predictive reliability (Doyle et al., 2020). These technical inconsistencies introduce measurement bias and undermine comparability across studies, particularly in multi-center datasets.\u003c/p\u003e\n\u003ch3\u003e3.2.3 Population and Setting Heterogeneity\u003c/h3\u003e\n\u003cp\u003ePopulation-level heterogeneity reflects imbalances in geographic, demographic, and clinical representation within datasets. Most studies were conducted in high-income countries, resulting in limited representation of LMICs where resistance patterns and healthcare systems differ substantially (Ali et al., 2023; Kim et al., 2022). Datasets were predominantly derived from hospital-based populations, while community and outpatient data were underrepresented (Zou et al., 2025). Variability across patient groups, including differences between paediatric and adult populations, further complicates model generalizability. Imbalances in pathogen\u0026ndash;antibiotic combinations contribute additionally, with overrepresentation of commonly studied species and underrepresentation of rare or context-specific resistance patterns (Zhao et al., 2024).\u003c/p\u003e\n\u003ch3\u003e3.2.4 Temporal and Geographic Heterogeneity\u003c/h3\u003e\n\u003cp\u003eTemporal variation introduces complexity through changes in antimicrobial prescribing practices, resistance prevalence, and laboratory methodologies over time phenomena that can degrade model performance when applied to future datasets (Ali et al., 2023; Kherabi et al., 2024). Geographic heterogeneity further limits model transferability, as resistance patterns vary across regions due to differences in microbial ecology, healthcare systems, and surveillance infrastructure (Nsubuga et al., 2024; Zhao et al., 2024). These dynamics highlight the importance of context-aware model development and validation strategies.\u003c/p\u003e\n\u003ch3\u003e3.2.5 Species and Antibiotic Imbalance\u003c/h3\u003e\n\u003cp\u003eAMR datasets exhibit substantial imbalance across pathogen\u0026ndash;antibiotic combinations, with frequently studied bacterial species and antimicrobial classes overrepresented while rare pathogens and certain antibiotic groups remain underrepresented (Prosperi et al., 2022). Environmental and livestock surveillance data further reveal geographic biases, with disproportionate representation from specific regions (Zhao et al., 2024). This imbalance biases model training toward dominant patterns and reduces predictive performance for underrepresented species\u0026ndash;drug combinations.\u003c/p\u003e\n\u003cp\u003eAcross studies, data heterogeneity is multi-dimensional and interdependent, affecting data quality, interoperability, and model performance. These interacting dimensions reduce reproducibility, limit generalizability, and complicate cross-study comparability, thereby constraining the clinical applicability of AI/ML-based AMR prediction models (Tang et al., 2022; Bilal et al., 2025). The key dimensions of data heterogeneity and their implications for AI/ML-based AMR prediction are summarized in Table 2.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 2. Key Dimensions of Data Heterogeneity and Their Implications for AI/ML-Based AMR Prediction.\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"100%\"\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eDimension\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eSource of Heterogeneity\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eMechanism\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eImpact on AI/ML Models\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eData modality\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eWGS, AST, EHR, MALDI-TOF, surveillance data\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eDifferences in data structure, format, and feature representation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eReduced interoperability; inconsistent feature learning\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eLaboratory practices\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCLSI vs EUCAST standards; AST methodologies\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eVariability in resistance measurement and classification\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eLabel inconsistency; reduced reproducibility\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePopulation \u0026amp; setting\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eHigh-income country dominance; hospital-based datasets\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eSampling imbalance across populations and clinical settings\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePoor generalizability; sampling bias\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTemporal variation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eChanging resistance patterns over time\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eDataset shift and evolving epidemiological trends\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eModel performance degradation\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eGeographic variation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eRegional differences in resistance patterns\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eContext-dependent resistance ecology\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eLimited transferability\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eSpecies\u0026ndash;antibiotic imbalance\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eUneven representation of pathogen\u0026ndash;drug combinations\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eSkewed training distribution\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePoor performance on underrepresented classes\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003ch2\u003e3.3 Types of Algorithmic Bias Identified\u003c/h2\u003e\n\u003cp\u003eWhile heterogeneity reflects variability in data, algorithmic bias represents the systematic distortions that arise from such variability during model development and deployment. Four major types of bias were consistently identified across studies.\u003c/p\u003e\n\u003ch3\u003e3.3.1 Sampling Bias\u003c/h3\u003e\n\u003cp\u003eSampling bias occurs when training datasets do not adequately represent the populations, geographic regions, or clinical settings in which models are subsequently applied, causing models to learn context-specific correlations rather than generalizable resistance mechanisms (Ali et al., 2023; Nsubuga et al., 2024). Global datasets are often skewed toward high-income countries and hospital-based samples, limiting representation of LMIC contexts and community settings. Temporal shifts in resistance patterns further exacerbate sampling bias, resulting in performance degradation when models are applied to newer datasets (Finlayson et al., 2021; Nsubuga et al., 2024). At the institutional level, models trained on single-center data often capture local prescribing practices and microbial ecology, leading to inflated internal performance but poor external validity (Cohen et al., 2025; Ferrari et al., 2024).\u003c/p\u003e\n\u003ch3\u003e3.3.2 Label Bias\u003c/h3\u003e\n\u003cp\u003eLabel bias arises from inconsistencies in resistance definitions and classification across datasets. Variability in susceptibility categories and MIC thresholds introduces systematic noise into training labels (Humphries et al., 2021). Differences between guideline systems such as CLSI and EUCAST can result in identical MIC values being classified differently, while aggregation of susceptibility categories may obscure clinically relevant distinctions. These inconsistencies reduce reproducibility and compromise model generalizability.\u003c/p\u003e\n\u003ch3\u003e3.3.3 Structural (Confounding) Bias\u003c/h3\u003e\n\u003cp\u003eStructural bias occurs when models exploit non-causal correlations within the data, such as phylogenetic structure or species\u0026ndash;drug co-occurrence patterns. Resistance phenotypes are often unevenly distributed across bacterial lineages, leading models to rely on lineage-specific markers as proxies for resistance rather than identifying causal mechanisms (Prosperi et al., 2022; Ren et al., 2022). Similarly, non-random pairing of pathogens and antibiotics in datasets introduces co-occurrence bias, where models learn dataset-specific relationships rather than biologically meaningful associations (Tang et al., 2022; Sidorczuk et al., 2022).\u003c/p\u003e\n\u003ch3\u003e3.3.4 Clinical Context Bias\u003c/h3\u003e\n\u003cp\u003eClinical context bias arises from artifacts embedded in EHR data, including healthcare access patterns, documentation practices, and missing data (Gianfrancesco \u0026amp; Goldstein, 2021; Boyd et al., 2023). Derived clinical variables and comorbidity indices may vary across populations, introducing additional inconsistencies (Gao et al., 2025). Moreover, observational clinical data often reflect prescribing behaviours and healthcare system processes rather than causal relationships between antibiotic use and resistance (Collignon \u0026amp; Beggs, 2025). These factors limit model transportability and raise concerns regarding fairness and reliability. The sources, mechanisms, consequences, and mitigation strategies of algorithmic bias identified across studies are summarized in Table 3.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 3. Sources, Mechanisms, Consequences, and Mitigation of Algorithmic Bias in AI-Based AMR Prediction.\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"100%\"\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eBias Type\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eSource\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eMechanism\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eConsequence\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eMitigation Strategies\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eSampling bias\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eHIC-dominated datasets; single-center data\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eNon-representative training data\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePoor external validity; limited generalizability\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eMulti-center datasets; reweighting; stratified sampling\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eLabel bias\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCLSI vs EUCAST breakpoints; MIC variability\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eInconsistent resistance classification\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eMisclassification; reduced reproducibility\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eStandardized breakpoints; use of raw MIC values\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eStructural bias\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePhylogeny; species\u0026ndash;drug co-occurrence\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eLearning non-causal correlations\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eSpurious predictions; reduced biological validity\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eStructure-aware modeling; phylogenetic splitting\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eClinical context bias\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eEHR documentation; healthcare access differences\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eContext-dependent artifacts\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eLimited transportability; fairness concerns\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eFeature harmonization; bias auditing\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAlgorithmic bias\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eClass imbalance; subgroup disparities\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eUnequal representation in training\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eUnequal model performance\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eResampling; cost-sensitive learning\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003ch2\u003e3.4 Mitigation Strategies Used in Studies\u003c/h2\u003e\n\u003cp\u003eStudies addressing AMR prediction using AI/ML employ a range of mitigation strategies targeting different stages of the modeling pipeline. These strategies can be broadly categorized into data-level, model-level, and evaluation-level approaches, as well as emerging explainability and causal methods. While these approaches contribute to improved robustness and fairness, their effectiveness remains partial and context-dependent, reflecting the underlying complexity of AMR data (Bilal et al., 2025; Chen et al., 2023).\u003c/p\u003e\n\u003ch3\u003e3.4.1 Data-Level Strategies: Addressing Class Imbalance and Sampling Bias\u003c/h3\u003e\n\u003cp\u003eClass imbalance is a pervasive challenge in AMR datasets, where resistant phenotypes or minority classes are frequently underrepresented. Common approaches include oversampling, undersampling, and synthetic data generation techniques such as the Synthetic Minority Over-sampling Technique (SMOTE), alongside cost-sensitive learning and weighted loss functions (Abdelhay et al., 2025). Resampling and reweighting techniques are widely applied in clinical and EHR-based models to improve representation of minority classes; however, their impact on overall performance is inconsistent and may introduce overfitting in small or highly imbalanced datasets (Bilal et al., 2025). Bias-aware sampling strategies have also been used to address imbalances across geographic regions, temporal distributions, and pathogen\u0026ndash;antibiotic combinations, including group-based reweighting and targeted data augmentation, though reported improvements are generally modest and highly dependent on dataset characteristics (Prosperi et al., 2022; Bilal et al., 2025).\u003c/p\u003e\n\u003ch3\u003e3.4.2 Structure-Aware Modeling and Data Splitting\u003c/h3\u003e\n\u003cp\u003eTo mitigate structural and phylogenetic bias, several studies emphasize population structure-aware modeling and evaluation. Conventional random train\u0026ndash;test splits allow models to exploit lineage-specific or dataset-specific correlations, leading to inflated performance estimates. In contrast, structure-aware approaches including phylogeny-informed data splitting, clade-based partitioning, and stratified sampling aim to ensure independence between training and testing datasets (Prosperi et al., 2022; Ren et al., 2022). These methods reduce reliance on non-causal correlations and promote the learning of more generalizable resistance mechanisms. However, they often result in lower reported performance metrics, reflecting a more realistic assessment of model capability. This trade-off highlights the tension between optimizing predictive accuracy and ensuring biological and clinical validity.\u003c/p\u003e\n\u003ch3\u003e3.4.3 Robust Evaluation Strategies\u003c/h3\u003e\n\u003cp\u003eRobust evaluation is a critical component of bias mitigation, directly influencing the reliability and generalizability of model performance estimates. Key strategies include external validation using independent datasets, temporal validation across different time periods, and multi-center study designs (Chen et al., 2023). Despite their importance, these approaches remain underutilized in the AMR prediction literature. Evidence consistently demonstrates that models evaluated on external datasets exhibit reduced performance compared to internal validation, indicating substantial overfitting to training data (Bilal et al., 2025). Beyond discrimination metrics such as AUC, there is increasing emphasis on reporting calibration, subgroup performance, and fairness-related metrics to provide a more comprehensive assessment of real-world applicability across diverse populations.\u003c/p\u003e\n\u003ch3\u003e3.4.4 Explainability and Causal Approaches\u003c/h3\u003e\n\u003cp\u003eExplainability techniques are increasingly incorporated to identify and mitigate bias in AI/ML models. Methods such as feature attribution, model interpretability tools, and fairness metrics including demographic parity and equalized odds enable detection of spurious associations and improve transparency (Xu \u0026amp; Ouyang, 2022; Loh et al., 2022). Causal modeling approaches aim to address confounding by incorporating explicit assumptions about data-generating mechanisms. Techniques such as propensity score adjustment, causal inference frameworks, and Bayesian networks have been applied to disentangle true causal relationships from dataset-specific correlations (Prosperi et al., 2022; Moccia, 2024). While these approaches enhance interpretability and support bias detection, their effectiveness depends on data quality, appropriate model assumptions, and the availability of relevant covariates, and should therefore be considered complementary rather than standalone solutions.\u003c/p\u003e\n\u003cp\u003eMitigation strategies in AI/ML-based AMR prediction can be conceptualized across four interrelated domains: (1) data rebalancing and sampling strategies; (2) structure-aware modeling and data partitioning; (3) robust and context-aware evaluation frameworks; and (4) explainability and causal inference approaches. No single strategy fully resolves bias or heterogeneity. Effective mitigation requires an integrated, multi-layered approach spanning the entire machine learning pipeline, reinforcing the need for end-to-end design strategies that address data quality, model development, evaluation rigor, and real-world implementation simultaneously (Bilal et al., 2025; Chen et al., 2023). A comparative summary of data domains, modeling approaches, performance, validation strategies, and key limitations across studies is presented in Table 4.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 4. Data Domains, Algorithms, Performance, and Limitations in AI-Based AMR Prediction.\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"100%\"\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eData Domain\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eCommon Algorithms\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eTypical AUC Range\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eValidation Type\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eKey Limitations\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eWGS\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eRF, CNN, DNN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.80\u0026ndash;0.97\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eInternal + limited external\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePopulation bias; sequencing variability\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAST\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eRF, SVM, Gradient Boosting\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.65\u0026ndash;0.90\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eMostly internal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eLaboratory variability; class imbalance\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eEHR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eLR, RF, Gradient Boosting\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.60\u0026ndash;0.85\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eMostly internal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eMissing data; coding heterogeneity\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eSurveillance\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTime-series ML\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.70\u0026ndash;0.92\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eLimited external\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eRegional bias; reporting inconsistency\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eMALDI-TOF\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCNN, RF\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.48\u0026ndash;0.85\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eInternal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTechnical variability; reproducibility issues\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e"},{"header":"4. Proposed Framework: Heterogeneity Mitigation Framework","content":"\u003cp\u003eTo address the multifaceted sources of data heterogeneity and algorithmic bias identified in this review, we propose a four-layer Heterogeneity Mitigation Framework spanning the full machine learning pipeline, from data generation to real-world deployment. The framework integrates data standardization, bias-aware modeling, robust evaluation, and governance mechanisms, reflecting current best practices in AMR research and computational medicine (Ardila et al., 2025; Bilal et al., 2025).\u003c/p\u003e\n\u003cp\u003eMultiple dimensions of heterogeneity including data modality, laboratory practices, population characteristics, study design, and modeling approaches interact to produce distinct but interrelated forms of bias, such as sampling bias, label bias, structural confounding, and clinical context bias. These interactions collectively influence model performance, fairness, and generalizability. Importantly, mitigation strategies operate across all dimensions, emphasizing the need for an integrated, end-to-end approach rather than isolated interventions.\u003c/p\u003e\n\u003ch2\u003e\u003cem\u003e4.1 Layer 1: Data and Infrastructure\u003c/em\u003e\u003c/h2\u003e\n\u003cp\u003eBias mitigation begins at the data level, where heterogeneity is most pronounced. Standardization of AST protocols, including harmonization of laboratory procedures and breakpoint systems, is critical to reduce label inconsistency and improve comparability across datasets (Gajic et al., 2022). Equally, standardized genomic quality control pipelines for WGS including consistent assembly, annotation, and resistance gene detection workflows are necessary to minimize inter-laboratory variability (Bortolaia et al., 2020; Doyle et al., 2020). Addressing global sampling imbalance requires expanding surveillance and data collection in underrepresented regions, particularly LMICs, to improve model generalizability across diverse resistance ecologies (Nsubuga et al., 2024). Development of well-curated benchmarking datasets with balanced representation across species, geography, and time further supports reproducibility and fair model comparison (Bortolaia et al., 2020; Doyle et al., 2020).\u003c/p\u003e\n\u003ch2\u003e\u003cem\u003e4.2 Layer 2: Modeling Approaches\u003c/em\u003e\u003c/h2\u003e\n\u003cp\u003eAt the modeling stage, mitigation strategies focus on improving robustness while reducing reliance on spurious correlations in heterogeneous data. Robust algorithms including ensemble methods such as random forests and gradient boosting have demonstrated resilience to noisy and heterogeneous datasets (Ren et al., 2022). Bias-aware sample weighting can further improve fairness by adjusting for underrepresented groups, including geographic regions, pathogen types, and resistance classes. Structure-aware modeling approaches are particularly important in AMR prediction; incorporating bacterial population structure or phylogenetic information helps prevent models from exploiting lineage as a proxy for resistance, thereby improving biological validity and generalizability (Ardila et al., 2024). Strategies for handling class imbalance, such as resampling and cost-sensitive learning, can enhance detection of rare resistance phenotypes, although their effectiveness remains context-dependent (Bilal et al., 2025).\u003c/p\u003e\n\u003ch2\u003e\u003cem\u003e4.3 Layer 3: Evaluation Standards\u003c/em\u003e\u003c/h2\u003e\n\u003cp\u003eRobust evaluation is essential to ensure that models generalize beyond their training environment and provide clinically meaningful predictions. External validation using independent datasets and temporal validation across different time periods are critical for assessing real-world performance and identifying overfitting (Anahtar et al., 2021). Beyond traditional performance metrics such as AUROC, evaluation should incorporate calibration, subgroup performance, and fairness-oriented metrics to assess reliability across diverse populations (Chen et al., 2023). Benchmarking against clinical baselines, such as antibiograms or clinician decision-making, is also necessary to confirm that AI/ML models provide meaningful improvements over existing practice (Feretzakis et al., 2021).\u003c/p\u003e\n\u003ch2\u003e\u003cem\u003e4.4 Layer 4: Governance and Implementation\u003c/em\u003e\u003c/h2\u003e\n\u003cp\u003eEffective deployment of AI/ML-based AMR prediction models requires alignment with clinical workflows, ethical standards, and regulatory frameworks. Integration into antimicrobial stewardship (AMS) programs ensures that predictive models support, rather than replace, clinical decision-making, particularly in empiric therapy selection (Howard et al., 2024; Beaudoin et al., 2016). Ethical considerations, including fairness, transparency, and accountability, are essential to prevent reinforcement of existing healthcare inequities (Dankwa-Mullan, 2024). Ongoing monitoring of model performance is required to detect degradation due to temporal changes in resistance patterns and clinical practices. Adherence to data governance and privacy frameworks is also necessary for responsible and sustainable AI deployment in healthcare (Mennella et al., 2024).\u003c/p\u003e\n\u003cp\u003eThe proposed framework conceptualizes bias mitigation in AI/ML-based AMR prediction as a continuous, multi-layered process. Data-level interventions address representativeness and measurement validity; modeling approaches reduce structural and algorithmic bias; evaluation strategies ensure robustness and generalizability; and governance mechanisms enable safe and equitable deployment. Critically, these layers are interdependent\u0026mdash;limitations in data quality and representation cannot be fully corrected through downstream modeling or evaluation alone. Effective AMR prediction systems therefore require an end-to-end design approach that integrates epidemiological, microbiological, and computational perspectives. An integrated mapping of data heterogeneity, associated biases, and mitigation strategies across the machine learning pipeline is presented in Table 5.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 5. Integrated Mapping of Data Heterogeneity, Algorithmic Bias, and Mitigation Strategies Across the Machine Learning Pipeline.\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"100%\"\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e\u003cstrong\u003ePipeline Stage\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 107px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSource of Heterogeneity\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 93px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAssociated Bias\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 107px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eImpact\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 120px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eMitigation Strategy\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003eData collection\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 107px;\"\u003e\n \u003cp\u003eGeographic imbalance; limited LMIC data\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 93px;\"\u003e\n \u003cp\u003eSampling bias\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 107px;\"\u003e\n \u003cp\u003ePoor generalizability\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 120px;\"\u003e\n \u003cp\u003eMulti-center datasets; inclusive sampling\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003eData labeling\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 107px;\"\u003e\n \u003cp\u003eBreakpoint variability (CLSI/EUCAST)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 93px;\"\u003e\n \u003cp\u003eLabel bias\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 107px;\"\u003e\n \u003cp\u003eMisclassification\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 120px;\"\u003e\n \u003cp\u003eStandardization; harmonized protocols\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003eFeature engineering\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 107px;\"\u003e\n \u003cp\u003eEHR variability; missing data\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 93px;\"\u003e\n \u003cp\u003eMeasurement bias\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 107px;\"\u003e\n \u003cp\u003eNoisy features\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 120px;\"\u003e\n \u003cp\u003eData preprocessing; validation pipelines\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003eModel training\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 107px;\"\u003e\n \u003cp\u003eClass imbalance; skewed datasets\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 93px;\"\u003e\n \u003cp\u003eAlgorithmic bias\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 107px;\"\u003e\n \u003cp\u003eUnequal performance\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 120px;\"\u003e\n \u003cp\u003eResampling; cost-sensitive learning\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003eModel validation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 107px;\"\u003e\n \u003cp\u003eLack of external datasets\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 93px;\"\u003e\n \u003cp\u003eEvaluation bias\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 107px;\"\u003e\n \u003cp\u003eOverestimated performance\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 120px;\"\u003e\n \u003cp\u003eExternal and temporal validation\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003eDeployment\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 107px;\"\u003e\n \u003cp\u003eTemporal drift; clinical variability\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 93px;\"\u003e\n \u003cp\u003ePerformance degradation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 107px;\"\u003e\n \u003cp\u003eReduced reliability\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 120px;\"\u003e\n \u003cp\u003eContinuous monitoring; model updating\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e"},{"header":"5. Discussion","content":"\u003ch2\u003e\u003cem\u003e5.1 Summary of Key Findings\u003c/em\u003e\u003c/h2\u003e\n\u003cp\u003eThis systematic review using descriptive synthesis demonstrates that AI/ML approaches for AMR prediction consistently achieve strong internal performance, frequently reporting high discrimination metrics such as area under the receiver operating characteristic curve (AUROC). However, this apparent success is substantially constrained by pervasive data heterogeneity and multiple forms of algorithmic bias that collectively limit real-world applicability and clinical translation (Anahtar et al., 2021; Ardila et al., 2025).\u003c/p\u003e\n\u003cp\u003eA central finding is that model performance is highly context-dependent. Many studies rely on internally consistent datasets, enabling models to implicitly learn local epidemiological patterns, institutional practices, or population structure rather than generalizable biological mechanisms. As a result, models often perform well under controlled conditions but fail to maintain accuracy when applied across different geographic regions, healthcare settings, or time periods (Kim et al., 2022; Kherabi et al., 2024).\u003c/p\u003e\n\u003cp\u003eData heterogeneity emerged as a multi-dimensional challenge encompassing variability in data modalities, laboratory protocols, population composition, and temporal and geographic distribution. These interacting dimensions complicate data integration, reduce reproducibility, and contribute to inconsistent model performance. In particular, the dominance of high-income country datasets and hospital-based populations introduces structural imbalances that further limit generalizability and equity.\u003c/p\u003e\n\u003cp\u003eAlgorithmic bias was similarly pervasive, with sampling bias, label inconsistency, structural confounding, and clinical context bias identified as major contributors to model fragility. These biases often cause models to exploit spurious correlations such as phylogenetic structure, institutional prevalence, or documentation patterns rather than learning causal determinants of resistance (Prosperi et al.,2022; Ren et al., 2022). Consequently, high internal accuracy frequently masks poor performance in external or real-world settings.\u003c/p\u003e\n\u003cp\u003eImportantly, external validation remains limited across the literature. Where implemented, it consistently reveals reduced predictive performance compared to internal validation, highlighting the extent to which current evaluation practices overestimate model reliability (Tang et al., 2022; Bilal et al., 2025). This gap underscores a broader methodological issue in AMR machine learning: the prioritization of predictive accuracy over robustness, fairness, and generalizability. Taken together, these findings indicate that the primary constraint in AI/ML-based AMR prediction is not algorithmic capability, but the quality, structure, and representativeness of the underlying data.\u003c/p\u003e\n\u003ch2\u003e5.2 Interpretation and Implications\u003c/h2\u003e\n\u003cp\u003eThe findings highlight that the limitations of AI-based AMR prediction are fundamentally rooted in data-related challenges rather than insufficient model complexity. While increasingly sophisticated algorithms continue to be developed, their performance remains contingent on the characteristics of the training data. In many cases, models learn context-specific patterns such as institutional prescribing behaviour or dataset-specific correlations rather than biologically meaningful or generalizable relationships.\u003c/p\u003e\n\u003cp\u003eA key insight is the interconnected nature of data heterogeneity and algorithmic bias. Heterogeneity across data modalities, laboratory practices, and populations creates conditions under which bias can emerge and propagate through the modeling pipeline. Geographic and demographic imbalances contribute to sampling bias, while variability in laboratory methods and breakpoint definitions introduces label inconsistency. These factors interact to amplify model instability and reduce external validity.\u003c/p\u003e\n\u003cp\u003eThe limited use of external and temporal validation further highlights a critical gap between model development and real-world deployment. Many studies rely on internal validation strategies that fail to capture distributional shifts across time, geography, and clinical settings, resulting in performance metrics that overestimate reliability in practice. This finding aligns with broader evidence in clinical machine learning, where failure to account for dataset shift remains a major barrier to implementation (Finlayson et al., 2021).\u003c/p\u003e\n\u003cp\u003eFrom a clinical and public health perspective, these limitations carry important implications. Models that perform inconsistently across populations risk reinforcing existing health disparities, particularly in LMICs where AMR burden is highest but data representation is lowest (Nsubuga et al., 2024). In this context, fairness is not only a technical concern but also an ethical imperative (Dankwa-Mullan, 2024).\u003c/p\u003e\n\u003cp\u003eThe findings also suggest a necessary shift in research priorities. Rather than focusing solely on improving predictive performance, future work should prioritize data quality, standardization, and representativeness. Development of multi-center, globally representative datasets, standardized laboratory practices, and transparent reporting frameworks is essential to improving reproducibility and comparability. The Heterogeneity Mitigation Framework proposed in this study provides a structured approach to addressing these challenges across the full machine learning pipeline, integrating data-level standardization, bias-aware modeling, robust evaluation, and governance mechanisms.\u003c/p\u003e\n\u003ch2\u003e5.3 Limitations\u003c/h2\u003e\n\u003cp\u003eThis study has several limitations warranting consideration. First, this systematic review employed descriptive synthesis rather than formal meta-analysis. No pooled estimates of model performance or formal assessment of statistical heterogeneity across studies were produced, as the degree of methodological, clinical, and outcome heterogeneity across included studies precluded meaningful statistical aggregation. Consequently, findings are presented as structured qualitative and descriptive syntheses.\u003c/p\u003e\n\u003cp\u003eSecond, the review is dependent on the quality and reporting of included studies. Variability in study design, data sources, outcome definitions, and evaluation metrics limited direct comparability. Inconsistent reporting of validation strategies, preprocessing methods, and bias mitigation techniques may have led to underestimation or overestimation of certain methodological challenges.\u003c/p\u003e\n\u003cp\u003eThird, publication bias may be present, as studies reporting higher predictive performance or novel methodologies are more likely to be published, potentially yielding an overly optimistic representation of AI performance in AMR prediction. Exclusion of grey literature and non-English publications may also have limited comprehensiveness.\u003c/p\u003e\n\u003cp\u003eFourth, the rapidly evolving nature of AI and AMR research means that emerging methodologies, datasets, and real-world implementation studies may not be fully captured within the review timeframe.\u003c/p\u003e\n\u003cp\u003eFinally, while this study proposes a Heterogeneity Mitigation Framework, it has not yet been empirically validated. Its effectiveness will depend on future application and evaluation across diverse datasets and clinical settings. It should therefore be considered a conceptual guide rather than a definitive solution.\u003c/p\u003e"},{"header":"6. Conclusion","content":"\u003cp\u003eArtificial intelligence and machine learning hold substantial promise for advancing AMR prediction by enabling rapid, data-driven insights and supporting clinical decision-making. However, this systematic review demonstrates that current AI/ML-based approaches are fundamentally constrained not by algorithmic limitations, but by pervasive data heterogeneity and multiple forms of algorithmic bias that undermine generalizability and real-world applicability (Anahtar et al., 2021; Bilal et al., 2025).\u003c/p\u003e\n\u003cp\u003eThe findings highlight that strong internal model performance does not reliably translate into consistent performance across diverse populations, healthcare settings, and time periods. Variability in data modalities, laboratory practices, population representation, and evaluation strategies introduces systematic challenges that reduce reproducibility and external validity. These issues are further compounded by limited use of external validation and the underrepresentation of LMICs in existing datasets, despite their bearing the highest burden of AMR (Nsubuga et al., 2024).\u003c/p\u003e\n\u003cp\u003eAddressing these challenges requires a fundamental shift in research priorities from optimizing predictive accuracy alone to improving data quality, standardization, representativeness, and fairness. The Heterogeneity Mitigation Framework proposed in this study provides a structured, multi-layered approach to addressing these limitations across the machine learning pipeline, encompassing data infrastructure, modeling strategies, evaluation standards, and governance considerations.\u003c/p\u003e\n\u003cp\u003eFuture research should prioritize the development of globally representative, multi-center datasets; standardized laboratory and reporting practices; and robust validation frameworks that reflect real-world conditions. Interdisciplinary collaboration among clinicians, microbiologists, data scientists, and policymakers will be essential to ensure that AI systems are both scientifically robust and clinically meaningful.\u003c/p\u003e\n\u003cp\u003eUltimately, successful translation of AI into AMR prediction depends on addressing foundational challenges related to data heterogeneity, bias, and evaluation rigor. A coordinated, end-to-end approach that integrates data quality, methodological robustness, and ethical considerations is essential for developing reliable, equitable, and clinically deployable AI systems that can meaningfully contribute to antimicrobial stewardship and global health.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eEthics Approval and Consent to Participate:\u0026nbsp;\u003c/strong\u003eNot applicable. This study is a systematic review based on previously published data and does not involve direct human participants or personal data collection.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for Publication:\u0026nbsp;\u003c/strong\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding:\u0026nbsp;\u003c/strong\u003eThe author received no external funding for this study.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting Interests:\u0026nbsp;\u003c/strong\u003eThe author declares no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData Availability:\u0026nbsp;\u003c/strong\u003eAll data used in this study are derived from publicly available sources cited in the manuscript. No primary data were generated.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eAbdelhay O, Alshoubaki R, Murad S, Abdel-Hafez O, Abdelhay Q, Haddad B et al (2025) Using statistical modelling and machine learning in detecting bone properties: A systematic review protocol. PLoS ONE 20(3):e0319583. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1371/journal.pone.0319583\u003c/span\u003e\u003cspan address=\"10.1371/journal.pone.0319583\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAli T, Ahmed S, Aslam M (2023) Artificial intelligence for antimicrobial resistance prediction: Challenges and opportunities towards practical implementation. Antibiotics 12(3):523. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3390/antibiotics12030523\u003c/span\u003e\u003cspan address=\"10.3390/antibiotics12030523\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAnahtar MN, Yang JH, Kanjilal S (2021) Applications of machine learning to the problem of antimicrobial resistance: An emerging model for translational research. J Clin Microbiol 59(7):e01260\u0026ndash;e01220. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1128/JCM.01260-20\u003c/span\u003e\u003cspan address=\"10.1128/JCM.01260-20\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eArdila CM, Gonz\u0026aacute;lez-Arroyave D, Tob\u0026oacute;n S (2025) Machine learning for predicting antimicrobial resistance in critical and high-priority pathogens: A systematic review considering antimicrobial susceptibility tests in real-world healthcare settings. PLoS ONE 20(2):e0319460. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1371/journal.pone.0319460\u003c/span\u003e\u003cspan address=\"10.1371/journal.pone.0319460\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eArdila CM, Yadalam PK, Gonz\u0026aacute;lez-Arroyave D (2024) Integrating whole genome sequencing and machine learning for predicting antimicrobial resistance in critical pathogens: A systematic review of antimicrobial susceptibility tests. PeerJ 12:e18213. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.7717/peerj.18213\u003c/span\u003e\u003cspan address=\"10.7717/peerj.18213\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBeaudoin M, Kabanza F, Nault V, Valiquette L (2016) Evaluation of a machine learning capability for a clinical decision support system to enhance antimicrobial stewardship programs. Artif Intell Med 68:29\u0026ndash;36. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.artmed.2016.01.004\u003c/span\u003e\u003cspan address=\"10.1016/j.artmed.2016.01.004\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBilal H, Khan MN, Khan S, Shafiq M, Fang W, Khan RU et al (2025) The role of artificial intelligence and machine learning in predicting and combating antimicrobial resistance. Comput Struct Biotechnol J 27:423\u0026ndash;439. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.csbj.2025.01.006\u003c/span\u003e\u003cspan address=\"10.1016/j.csbj.2025.01.006\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBortolaia V, Kaas RS, Ruppe E, Roberts MC, Schwarz S, Cattoir V et al (2020) ResFinder 4.0 for predictions of phenotypes from genotypes. J Antimicrob Chemother 75(12):3491\u0026ndash;3500. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1093/jac/dkaa345\u003c/span\u003e\u003cspan address=\"10.1093/jac/dkaa345\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBoyd AD, Gonzalez-Guarda R, Lawrence K, Patil CL, Ezenwa MO, O'Brien EC et al (2023) Equity and bias in electronic health records data. Contemp Clin Trials 130:107238. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.cct.2023.107238\u003c/span\u003e\u003cspan address=\"10.1016/j.cct.2023.107238\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen RJ, Wang JJ, Williamson DF, Chen TY, Lipkova J, Lu MY et al (2023) Algorithmic fairness in artificial intelligence for medicine and healthcare. Nat Biomed Eng 7(6):719\u0026ndash;742. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41551-023-01056-8\u003c/span\u003e\u003cspan address=\"10.1038/s41551-023-01056-8\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCohen SA, Chen Z, Bian J, Boucher C, Wu Y, Prosperi M (2025) Comparative evaluation of clinical large language models and machine learning to predict antimicrobial resistance in hospital-onset sepsis. In: International Conference on Artificial Intelligence in Medicine. Cham: Springer Nature Switzerland; pp. 65\u0026ndash;76. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/978-3-031-93528-7_7\u003c/span\u003e\u003cspan address=\"10.1007/978-3-031-93528-7_7\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCollignon P, Beggs JJ (2025) The persistence of antibiotic resistance in observational studies: Is it really due to differences in sub-populations rather than antibiotic use? Antibiotics 14(1):39. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3390/antibiotics14010039\u003c/span\u003e\u003cspan address=\"10.3390/antibiotics14010039\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDankwa-Mullan I (2024) Health equity and ethical considerations in using artificial intelligence in public health and medicine. Prev Chronic Dis 21:E64. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.5888/pcd21.240245\u003c/span\u003e\u003cspan address=\"10.5888/pcd21.240245\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDoyle RM, O'Sullivan DM, Aller SD, Bruchmann S, Clark T, Coello Pelegrin A et al (2020) Discordant bioinformatic predictions of antimicrobial resistance from whole-genome sequencing data of bacterial isolates: An inter-laboratory study. Microb Genomics 6(2):e000335. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1099/mgen.0.000335\u003c/span\u003e\u003cspan address=\"10.1099/mgen.0.000335\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFerrari D, Arina P, Edgeworth J, Curcin V, Guidetti V, Mandreoli F et al (2024) Using interpretable machine learning to predict bloodstream infection and antimicrobial resistance in patients admitted to ICU. PLOS Digit Health 3(10):e0000641. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1371/journal.pdig.0000641\u003c/span\u003e\u003cspan address=\"10.1371/journal.pdig.0000641\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFeretzakis G, Loupelis E, Sakagianni A, Kalles D, Martsoukou M, Lykoudis P et al (2021) Using machine learning techniques to aid empirical antibiotic therapy decisions in the intensive care unit of a general hospital in Greece. Antibiotics 10(9):1112. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3390/antibiotics10091112\u003c/span\u003e\u003cspan address=\"10.3390/antibiotics10091112\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFinlayson SG, Subbaswamy A, Singh K, Bowers J, Kupke A, Zittrain J et al (2021) The clinician and dataset shift in artificial intelligence. N Engl J Med 385(3):283\u0026ndash;286. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1056/NEJMc2104626\u003c/span\u003e\u003cspan address=\"10.1056/NEJMc2104626\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGajic I, Kabic J, Kekic D, Jovicevic M, Milenkovic M, Mitic Culafic D et al (2022) Antimicrobial susceptibility testing: A comprehensive review of currently used methods. Antibiotics 11(4):427. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3390/antibiotics11040427\u003c/span\u003e\u003cspan address=\"10.3390/antibiotics11040427\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGao J, Chou B, McCaw ZR, Thurston H, Varghese P, Hong C et al (2025) What is fair? Defining fairness in machine learning for health. Stat Med 44(20\u0026ndash;22):e70234. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1002/sim.70234\u003c/span\u003e\u003cspan address=\"10.1002/sim.70234\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGianfrancesco MA, Goldstein ND (2021) A narrative review on the validity of electronic health record-based research in epidemiology. BMC Med Res Methodol 21(1):234. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/s12874-021-01422-3\u003c/span\u003e\u003cspan address=\"10.1186/s12874-021-01422-3\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHoward A, Aston S, Gerada A, Reza N, Bincalar J, Mwandumba H et al (2024) Antimicrobial learning systems: An implementation blueprint for artificial intelligence to tackle antimicrobial resistance. Lancet Digit Health 6(1):e79\u0026ndash;e86. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/S2589-7500(23)00180-6\u003c/span\u003e\u003cspan address=\"10.1016/S2589-7500(23)00180-6\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHumphries R, Bobenchik AM, Hindler JA, Schuetz AN (2021) Overview of changes to the clinical and laboratory standards institute performance standards for antimicrobial susceptibility testing, M100. J Clin Microbiol 59(12):e00213\u0026ndash;e00221. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1128/JCM.00213-21\u003c/span\u003e\u003cspan address=\"10.1128/JCM.00213-21\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKherabi Y, Thy M, Bouzid D, Antcliffe DB, Rawson TM, Peiffer-Smadja N (2024) Machine learning to predict antimicrobial resistance: Future applications in clinical practice? Infect Dis Now 54(3):104864. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.idnow.2024.104864\u003c/span\u003e\u003cspan address=\"10.1016/j.idnow.2024.104864\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKim JI, Maguire F, Tsang KK, Gouliouris T, Peacock SJ, McAllister TA et al (2022) Machine learning for antimicrobial resistance prediction: Current practice, limitations, and clinical perspective. Clin Microbiol Rev 35(3):e00179\u0026ndash;e00121. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1128/cmr.00179-21\u003c/span\u003e\u003cspan address=\"10.1128/cmr.00179-21\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLewnard JA, Charani E, Gleason A, Hsu LY, Khan WA, Karkey A et al (2024) Burden of bacterial antimicrobial resistance in low-income and middle-income countries aversible by existing interventions: An evidence review and modeling analysis. Lancet 403(10442):2439\u0026ndash;2454. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/S0140-6736(24)00474-4\u003c/span\u003e\u003cspan address=\"10.1016/S0140-6736(24)00474-4\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLoh HW, Ooi CP, Seoni S, Barua PD, Molinari F, Acharya UR (2022) Application of explainable artificial intelligence for healthcare: A systematic review of the last decade (2011\u0026ndash;2022). Comput Methods Programs Biomed 226:107161. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.cmpb.2022.107161\u003c/span\u003e\u003cspan address=\"10.1016/j.cmpb.2022.107161\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMcGowan J, Straus S, Moher D, Langlois EV, O'Brien KK, Horsley T et al (2020) Reporting scoping reviews\u0026mdash;PRISMA ScR extension. J Clin Epidemiol 123:177\u0026ndash;179. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.jclinepi.2019.09.016\u003c/span\u003e\u003cspan address=\"10.1016/j.jclinepi.2019.09.016\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMennella C, Maniscalco U, De Pietro G, Esposito M (2024) Ethical and regulatory challenges of AI technologies in healthcare: A narrative review. Heliyon 10(4). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.heliyon.2024.e26297\u003c/span\u003e\u003cspan address=\"10.1016/j.heliyon.2024.e26297\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMoccia S (2024) Trustworthy AI in neuroengineering: From data management to ethics assessment. In: To Understand and Restore the Nervous System. p. 205. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1088/978-0-7503-3693-8ch10\u003c/span\u003e\u003cspan address=\"10.1088/978-0-7503-3693-8ch10\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMurray CJ, Ikuta KS, Sharara F, Swetschinski L, Aguilar GR, Gray A et al (2022) Global burden of bacterial antimicrobial resistance in 2019: A systematic analysis. Lancet 399(10325):629\u0026ndash;655. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/S0140-6736(21)02724-0\u003c/span\u003e\u003cspan address=\"10.1016/S0140-6736(21)02724-0\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNsubuga M, Galiwango R, Jjingo D, Mboowa G (2024) Generalizability of machine learning in predicting antimicrobial resistance in E. coli: A multi-country case study in Africa. BMC Genomics 25(1):287. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/s12864-024-10214-4\u003c/span\u003e\u003cspan address=\"10.1186/s12864-024-10214-4\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePage MJ, McKenzie JE, Bossuyt PM, Boutron I, Hoffmann TC, Mulrow CD et al (2021) The PRISMA 2020 statement: An updated guideline for reporting systematic reviews. BMJ 372. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1136/bmj.n71\u003c/span\u003e\u003cspan address=\"10.1136/bmj.n71\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePark Y, Weig M, Noll C, Bader O, Hauschild AC (2024) Effect of data heterogeneity in clinical MALDI-TOF mass spectra profiles on direct antimicrobial resistance prediction through machine learning. bioRxiv. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1101/2024.10.07.617009\u003c/span\u003e\u003cspan address=\"10.1101/2024.10.07.617009\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePeters MD, Marnie C, Tricco AC, Pollock D, Munn Z, Alexander L et al (2020) Updated methodological guidance for the conduct of scoping reviews. JBI Evid Synth 18(10):2119\u0026ndash;2126. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.11124/JBIES-20-00167\u003c/span\u003e\u003cspan address=\"10.11124/JBIES-20-00167\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eProsperi M, Boucher C, Bian J, Marini S (2022) Assessing putative bias in prediction of anti-microbial resistance from real-world genotyping data under explicit causal assumptions. Artif Intell Med 130:102326. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.artmed.2022.102326\u003c/span\u003e\u003cspan address=\"10.1016/j.artmed.2022.102326\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRen Y, Chakraborty T, Doijad S, Falgenhauer L, Falgenhauer J, Goesmann A et al (2022) Prediction of antimicrobial resistance based on whole-genome sequencing and machine learning. Bioinformatics 38(2):325\u0026ndash;334. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1093/bioinformatics/btab681\u003c/span\u003e\u003cspan address=\"10.1093/bioinformatics/btab681\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSakagianni A, Koufopoulou C, Feretzakis G, Kalles D, Verykios VS, Myrianthefs P (2023) Using machine learning to predict antimicrobial resistance: A literature review. Antibiotics 12(3):452. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3390/antibiotics12030452\u003c/span\u003e\u003cspan address=\"10.3390/antibiotics12030452\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSidorczuk K, Gagat P, Pietluch F, Kala J, Rafacz D, Bakala L et al (2022) Benchmarks in antimicrobial peptide prediction are biased due to the selection of negative data. Brief Bioinform 23(5):bbac343. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1093/bib/bbac343\u003c/span\u003e\u003cspan address=\"10.1093/bib/bbac343\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTang R, Luo R, Tang S, Song H, Chen X (2022) Machine learning in predicting antimicrobial resistance: A systematic review and meta-analysis. Int J Antimicrob Agents 60(5\u0026ndash;6):106684. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.ijantimicag.2022.106684\u003c/span\u003e\u003cspan address=\"10.1016/j.ijantimicag.2022.106684\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTricco AC, Lillie E, Zarin W, O'Brien KK, Colquhoun H, Levac D et al (2018) PRISMA extension for scoping reviews (PRISMA-ScR): Checklist and explanation. Ann Intern Med 169(7):467\u0026ndash;473. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.7326/M18-0850\u003c/span\u003e\u003cspan address=\"10.7326/M18-0850\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eValavarasu S, Sangu Y, Mahapatra T (2025) Prediction of antibiotic resistance from antibiotic susceptibility testing results from surveillance data using machine learning. Sci Rep 15(1):30509. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41598-025-14940-3\u003c/span\u003e\u003cspan address=\"10.1038/s41598-025-14940-3\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWeis C, Cuenod A, Rieck B, Llinares-Lopez F, Dubuis O, Graf S et al (2020) Direct antimicrobial resistance prediction from MALDI-TOF mass spectra profile in clinical isolates through machine learning. bioRxiv 1\u0026ndash;35. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1101/2020.07.30.228411\u003c/span\u003e\u003cspan address=\"10.1101/2020.07.30.228411\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWenzler E, Maximos M, Asempa TE, Biehle L, Schuetz AN, Hirsch EB (2023) Antimicrobial susceptibility testing: An updated primer for clinicians in the era of antimicrobial resistance. Pharmacotherapy 43(4):264\u0026ndash;278. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1002/phar.2770\u003c/span\u003e\u003cspan address=\"10.1002/phar.2770\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eXu W, Ouyang F (2022) The application of AI technologies in STEM education: A systematic review from 2011 to 2021. Int J STEM Educ 9(1):59. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/s40594-022-00377-5\u003c/span\u003e\u003cspan address=\"10.1186/s40594-022-00377-5\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhao C, Wang Y, Mulchandani R, Van Boeckel TP (2024) Global surveillance of antimicrobial resistance in food animals using priority drugs maps. Nat Commun 15(1):763. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41467-024-44713-3\u003c/span\u003e\u003cspan address=\"10.1038/s41467-024-44713-3\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZou Z, Tang F, Qiao L, Wang S, Zhang H (2025) Integrating sequencing methods with machine learning for antimicrobial susceptibility testing in pediatric infections: Current advances and future insights. Front Microbiol 16:1528696. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/fmicb.2025.1528696\u003c/span\u003e\u003cspan address=\"10.3389/fmicb.2025.1528696\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","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":"antimicrobial resistance, artificial intelligence, machine learning, data heterogeneity, algorithmic bias, systematic review, predictive modeling","lastPublishedDoi":"10.21203/rs.3.rs-9552638/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9552638/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e \u003cp\u003eAntimicrobial resistance (AMR) represents one of the most critical global public health threats of the contemporary era, contributing to millions of deaths and substantial morbidity across all regions of the world (Murray et al., 2022). Artificial intelligence and machine learning (AI/ML) have emerged as transformative tools for improving the detection, prediction, and management of AMR by capitalizing on large-scale biomedical datasets. Despite their promise, the clinical translation of these methods remains constrained by fundamental methodological challenges, particularly data heterogeneity and algorithmic bias (Anahtar et al., 2021; Bilal et al., 2025).\u003c/p\u003e\u003ch2\u003eObjective\u003c/h2\u003e \u003cp\u003eThis study aimed to systematically assess the impact of data heterogeneity and algorithmic bias on AI/ML-based AMR prediction and to propose a structured, multi-layered framework to guide their mitigation.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e \u003cp\u003eA systematic review with descriptive synthesis was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA 2020) guidelines (Page et al.,2021). Searches were performed across multiple databases, including PubMed/MEDLINE, Embase, Scopus, Web of Science, IEEE Xplore, and the ACM Digital Library, covering studies published between January 2020 and April 2026. Eligible studies applied AI/ML approaches to AMR prediction using data modalities such as whole-genome sequencing (WGS), antimicrobial susceptibility testing (AST), electronic health records (EHRs), surveillance datasets, and spectral data. Data were extracted using a standardized charting form and synthesized using descriptive and thematic methods. No meta-analysis was conducted owing to substantial methodological and data heterogeneity across included studies.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003eA total of 68 studies published between 2020 and 2026 were included, of which 21 underwent detailed analysis. Included studies demonstrated substantial heterogeneity across data modalities, patient populations, laboratory practices, and geographic settings. WGS-based approaches were the most frequently represented, followed by AST- and EHR-based models. While many models achieved strong internal performance, generalizability across external settings remained limited. Key sources of heterogeneity included variability in data modalities, laboratory protocols, population composition, and temporal and geographic distribution. Major forms of bias identified included sampling bias, label inconsistency, structural confounding, and clinical context bias. Mitigation strategies including data rebalancing, structure-aware modeling, robust evaluation frameworks, and explainability approaches demonstrated partial and context-dependent improvements.\u003c/p\u003e\u003ch2\u003eConclusion\u003c/h2\u003e \u003cp\u003eThe performance of AI/ML-based AMR prediction models is constrained primarily by data heterogeneity and algorithmic bias rather than by limitations in algorithmic sophistication. Addressing these challenges requires a deliberate shift toward standardized, representative datasets, rigorous evaluation practices, and governance-aligned deployment. The proposed Heterogeneity Mitigation Framework offers a multi-layered, end-to-end approach to improving model robustness, fairness, and generalizability, providing a foundation for clinically applicable AI systems in antimicrobial resistance.\u003c/p\u003e","manuscriptTitle":"Data Heterogeneity and Algorithmic Bias in AI-Based Antimicrobial Resistance Prediction: A Systematic Review and Mitigation Framework","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-29 11:50:15","doi":"10.21203/rs.3.rs-9552638/v1","editorialEvents":[{"type":"communityComments","content":4}],"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":"b7aac6eb-2b4c-4559-8eb8-b63c467cd965","owner":[],"postedDate":"April 29th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":67161156,"name":"Epidemiology"}],"tags":[],"updatedAt":"2026-05-09T16:55:17+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-29 11:50:15","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9552638","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9552638","identity":"rs-9552638","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}