Robustness vs. Efficiency in ECG Classification: A Comparative Study of Deep Learning and Classical Machine Learning Architectures

Robustness vs. Efficiency in ECG Classification: A Comparative Study of Deep Learning and Classical Machine Learning Architectures | 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 Research Article Robustness vs. Efficiency in ECG Classification: A Comparative Study of Deep Learning and Classical Machine Learning Architectures Amit Kumar Auddy, Jagannibas Paul Choudhury This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8987933/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 Purpose: Cardiovascular diseases remain the predominant cause of global mortality, necessitating reliable automated tools for early diagnosis. While Deep Learning (DL) has shown superiority in medical imaging, recent studies suggest that Classical Machine Learning models may offer comparable accuracy with lower computational costs. This study investigates this accuracy versus robustness trade-off. Methods: We present a rigorous comparison between modern Efficient Deep Learning architectures (MobileNetV4, GhostNetV2) and established Classical Baselines (CatBoost, Random Forest) for Myocardial Infarction detection using the PTB-XL dataset. Models were first evaluated on 2D scalograms generated via Continuous Wavelet Transform (CWT). To test structural robustness and artifact reliance, we subsequently introduced a 1D Raw Signal Benchmark evaluating the models on unaligned, raw time-series data. Results: On aligned 2D image data, Classical models achieved competitive F1-scores (~0.80). However, the 1D Raw Signal Benchmark revealed a severe performance degradation for Classical methods (F1 drop to ~0.04) on unaligned data, highlighting a heavy reliance on spatial artifacts. Conversely, Convolutional Neural Networks (CNNs) maintained high performance (F1 ~0.84) on raw signals. Conclusion: While Classical models are computationally efficient for highly structured data, they are structurally brittle. Efficient Deep Learning architectures, specifically MobileNetV4, offer the superior translation-invariant feature learning required for clinical signal analysis, delivering a 2.0x speedup over heavy CNN baselines with negligible accuracy loss. Deep Learning Electrocardiogram MobileNetV4 GhostNetV2 Robustness Edge Computing Figures Figure 1 Figure 2 Introduction Cardiovascular diseases (CVDs) remain the leading cause of death globally, necessitating continuous and accurate monitoring of cardiac health [1]. The 12-lead Electrocardiogram (ECG) is the standard diagnostic tool, but its manual interpretation is time-consuming, prone to inter-observer variability, and requires specialized expertise [2]. Consequently, the development of automated diagnostic systems has become a major focus in medical informatics, leading to numerous systematic reviews evaluating the efficacy of artificial intelligence in cardiology [3,4,5,6]. However, the medical AI community is currently divided between two distinct paradigms, each prioritizing different computational methodologies. The first paradigm relies on Classical Machine Learning (ML) algorithms, such as Support Vector Machines (SVMs), Random Forests, and Gradient Boosting frameworks including XGBoost and CatBoost [7,8]. These models are highly favored in clinical tabular data analysis due to their low computational overhead and inherent interpretability [9,10,11]. Recent literature demonstrates their capability to achieve high diagnostic accuracy when applied to ECG classification tasks [12]. However, their application to physiological signals typically requires extensive manual feature extraction, dimensionality reduction, or rigorous spatial alignment [13,14,15]. This makes classical algorithms highly sensitive to phase shifts, baseline wander, and the physiological noise inherent in real-world clinical recordings. To bypass the limitations of handcrafted features, the field has increasingly pivoted toward Deep Learning (DL), particularly Convolutional Neural Networks (CNNs) and recurrent architectures [16,17,18]. Standard heavy CNNs have achieved cardiologist-level performance by automatically learning hierarchical representations from either raw 1D waveforms or 2D time-frequency transformations like scalograms [19,20]. Despite their high accuracy, these over-parameterized models, such as DenseNet and standard ResNet configurations, act as "black boxes" that introduce significant inference latency and memory constraints [21,22]. This computational burden severely limits their deployability in resource-constrained environments, such as wearable Holter monitors and embedded edge computing devices [23,24]. This has driven recent interest toward Efficient Deep Learning architectures such as MobileNet configurations and GhostNet variants which seek to optimize operation types for mobile hardware [25,26,27]. Studies implementing these lightweight networks for ECG and related physiological signal classification have demonstrated massive reductions in computational overhead without sacrificing predictive power [28,29,30]. These architectures heavily utilize techniques like depthwise separable convolutions and feature redundancy operations to drastically reduce parameter counts [31]. Despite these advancements, a critical gap remains in the literature regarding the trade-off between algorithmic efficiency and structural robustness [32,33]. Most comparative studies benchmark accuracy strictly on static, pre-processed, and perfectly aligned datasets, potentially masking the fragility of the underlying models [34]. This study addresses this gap by rigorously comparing modern Efficient Deep Learning architectures against highly optimized Classical Baselines. Using the comprehensive PTB-XL dataset [35], we specifically investigate an "Accuracy vs. Robustness Trade-off," evaluating whether the statistical superiority of Classical ML on aligned 2D data translates to robust performance on raw, unaligned physiological signals. Methodology Dataset and Experimental Validation Split We utilized the PTB-XL dataset [35], containing 21,837 clinical 12-lead ECGs from 18,885 patients. To address data heterogeneity, we utilized the standard 100 Hz versions of the records, decimated using a polyphase anti-aliasing filter. The task was formulated as a binary classification problem: Normal vs. Myocardial Infarction (MI)/Other. To ensure robust evaluation and completely prevent data leakage, we strictly adhered to the pre-defined, patient-stratified 10-fold split provided in the PTB-XL metadata. Specifically, Folds 1–8 (~80%, approx. 15,100 patients) were utilized for training, Fold 9 (~10%) was used for validation and hyperparameter early stopping, and Fold 10 (~10%) was reserved exclusively as an unseen test set for final benchmarking. This patient-stratified approach guarantees that records from the same patient do not simultaneously appear in both the training and testing phases. Preprocessing: Signal-to-Image Conversion To leverage Transfer Learning from ImageNet-pretrained backbones, we implemented a Signal-to-Image conversion pipeline [17]. 1D time-series signals were converted into 2D scalograms using the Continuous Wavelet Transform (CWT) with a Complex Morlet wavelet. Scalograms were resized to 224x224 pixels for Deep Learning models. Biologically-plausible augmentations, including random time shifts (+/- 50ms) and additive Gaussian noise, were applied to ensure physiological realism. Model Architectures We benchmarked four distinct architectures: DenseNet121 [21]: High-capacity baseline representing traditional heavy CNNs. MobileNetV4 [25]: Representative of Neural Architecture Search (NAS) optimization for mobile efficiency. GhostNetV2 [26]: Focuses on cheap feature redundancy operations. MobileOne [31]: Utilizes inference-time reparameterization. To adapt these natively 2D vision architectures (e.g., MobileNetV4, GhostNetV2) for processing raw 1D physiological signals, significant architectural modifications were implemented. We systematically replaced all standard 2D Convolutional layers (Conv2d) with 1D Convolutional layers (Conv1d) and adjusted the input channel depth to 12 to accurately accommodate the 12-lead ECG format. Furthermore, the initial 2D spatial downsampling strides—which assume square image inputs—were modified to preserve the temporal resolution. This adaptation allowed the networks to directly ingest the 12 × 1000 waveform tensor and successfully learn temporal filters via a 1D sliding window. Classical Machine Learning Pipeline To provide a rigorous baseline, we implemented a parallel pipeline using CatBoost [ 8 ], XGBoost [ 9 ], and Random Forest. Inputs were either flattened 2D scalograms (128x128) or raw 1D signals. Dimensionality reduction algorithms were evaluated alongside raw pixel inputs. The 1D Stress Test A key contribution of this study is the 1D Raw Signal Benchmark. To quantify "artifact reliance," we evaluated all models on the raw 1D ECG signal without the spatial alignment provided by 2D image conversion. Deep Learning models (adapted to 1D-CNNs as described above) were trained to predict pathology directly from the waveform. For the Classical ML baselines, the raw input matrix for a single sample comprising 12 leads and 1000 timepoints was simply flattened into a single 12,000-dimensional feature vector per patient. These flattened arrays were fed directly into the Random Forest and XGBoost classifiers without any intermediate feature engineering (such as RR-interval extraction or wavelet transforms) to act as a true stress test of the algorithms' native pattern recognition capabilities on raw temporal data. Results Efficiency vs. Accuracy Table 1 presents the performance metrics. The heavy baseline (DenseNet121) achieved an F1-score of 0.778 with an inference latency of 18.1ms. The proposed efficient model, MobileNetV4, achieved a comparable F1-score of 0.764 while reducing latency to 9.0ms (a 2.0x speedup). Notably, the Classical Baseline (CatBoost) achieved the highest F1-score of 0.803 on the aligned 2D image task. Table 1. Benchmarking of proposed architectures vs. baselines. Inference time measured on RTX3080 GPU (DL) and Ryzen 9 CPU (Classical) . Model Type Architecture F1-Score (2D) Inference (ms) Robustness Feature Heavy Baseline DenseNet121 0.778 18.1 Deep Feature Reuse Proposed Efficient MobileNetV4 0.764 9.0 Spatial Invariance GhostNetV2 0.784 16.3 Feature Redundancy Classical Baseline CatBoost 0.803 < 1.0 Pixel Correlation (Artifacts) Structural Robustness (The 1D Stress Test) While Classical models outperformed on aligned 2D images, the 1D Raw Signal analysis revealed a substantial performance degradation. Table 2 demonstrates that when stripped of spatial alignment cues, Classical models failed to extract meaningful patterns. Table 2. Performance collapse of Classical models on raw 1D signals vs. robustness of Deep Learning. Model Input Type F1-Score Status MobileNetV4 1D Raw Signal 0.844 ROBUST GhostNetV2 1D Raw Signal 0.821 ROBUST XGBoost 1D Raw Signal 0.310 FAILED Random Forest 1D Raw Signal 0.042 COLLAPSED Notably, the Random Forest model collapsed to an F1-score of 0.042 on 1D data, exhibiting extreme majority class bias (Sensitivity ~2%). In contrast, the MobileNetV4-1D architecture maintained high performance (F1 0.844), validating its ability to learn translation-invariant temporal filters. Explainability analysis via Gradient-weighted Class Activation Mapping (Grad-CAM) further confirmed these findings [33]. Discussion The Accuracy Paradox and The Curse of Dimensionality Our results highlight an "Accuracy Paradox." Classical Machine Learning models technically achieved higher scores on the static 2D dataset, but this performance was artifact-driven. The collapse of Classical models on the 1D benchmark confirms their reliance on spatial alignment artifacts (e.g., the fixed position of the QRS complex in the pre-processed image) rather than morphological features. When the structured data alignment was removed, their predictive performance vanished. Feeding 12,000 highly correlated, raw temporal data points directly into Random Forest or XGBoost models triggered the "curse of dimensionality." Without hand-crafted features or the structural invariance provided by a CNN's sliding window, the tree-based algorithms failed to establish meaningful decision boundaries across the 12,000 raw voltage points, causing them to collapse and default to majority class predictions. Deep Learning Suitability Convolutional Neural Networks, specifically efficient variations like MobileNetV4, demonstrated inherent Spatial Invariance. Their sliding-window filters successfully detected the shape of the QRS complex regardless of its temporal position in the input tensor. This structural robustness confirms that Deep Learning is the mathematically safer architectural choice for clinical diagnostics, where physiological signals invariably contain phase shifts and temporal misalignment. Conclusion This study compared the efficacy of Efficient Deep Learning against Classical Machine Learning for ECG analysis. We conclude that while Classical models offer superior speed and performance on rigidly structured data, they are structurally brittle when applied to unaligned physiological signals. Efficient Deep Learning architectures, specifically MobileNetV4, resolve this trade-off by delivering robust, translation-invariant feature extraction at a fraction of the computational cost of traditional heavy models. Future work will focus on prospective clinical validation and quantization for embedded edge devices. Declarations Funding: The authors declare that no funds, grants, or other support were received during the preparation of this manuscript. Competing Interests: The authors have no relevant financial or non-financial interests to disclose. Data Availability: The PTB-XL dataset analysed during the current study is openly available in the PhysioNet repository [36]. Ethics Approval: Not applicable. This study utilized a publicly available, fully de-identified retrospective dataset; therefore, institutional ethics approval was not required. Consent to Participate/Publish: Not applicable. References World Health Organization (WHO), "Cardiovascular diseases (CVDs)," 2025. [Online]. https://www.who.int/news-room/fact-sheets/detail/cardiovascular-diseases-(cvds) Pranav Rajpurkar, Masoumeh Haghpanahi, Geoffrey H. Tison, Codie Bourn, Mintu P. Turakhia & Andrew Y. Ng Awni Y. Hannun, "Cardiologist-level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural network," nature medicine, vol. 25, no. January 2019, pp. 65–69, January 2009. [Online]. https://doi.org/10.1038/s41591-018-0268-3 Abdelkader Nasreddine Belkacem, Ibrahim M. Elfadel, Herbert F. Jelinek, Mohamed Atef Allam Jaya Prakash, "Advances in machine and deep learning for ECG beat classification: a systematic review," vol. 7, 2025. [Online]. https://doi.org/10.3389/fdgth.2025.1649923 A. Shahbahrami M. Sarkaleh, "CLASSIFICATION OF ECG ARRHYTHMIAS USING DISCRETE WAVELET TRANSFORM AND NEURAL NETWORKS," International Journal of Computer Science, Engineering and Applications (IJCSEA), vol. 2, 2012. [Online]. https://doi.org/10.5121/ijcsea.2012.2101 Gleb Gusev, Aleksandr Vorobev, Anna Veronika Dorogush, Andrey Gulin Liudmila Prokhorenkova, "CatBoost: unbiased boosting with categorical features," ARXIV CS, 2019. [Online]. https://doi.org/10.48550/arXiv.1706.09516 Salih R. Marwa and Mohammed A. Shakir, "Improving ECG signals classification by using deep learning techniques: A review," 2nd International Conference on Applied Computing & Smart Cities (ICACS24), vol. 64, p. 01023, 2024. [Online]. https://doi.org/10.1051/itmconf/20246401023 M. Abubakar et al., "ECGNet: High-Precision ECG Classification Using Deep Learning and Advanced Activation Functions," IEEE Access, vol. 13, pp. 133191-133207, 2025. [Online]. https://doi.org/10.1109/ACCESS.2025.3592240 Utkarsh et al. Gupta, "A comprehensive review on efficient artificial intelligence models for classification of abnormal cardiac rhythms using electrocardiograms," vol. 10, no. 5, p. e26787. [Online]. https://doi.org/10.1016/j.heliyon.2024.e26787 Carlos Guestrin Tianqi Chen, "XGBoost: A Scalable Tree Boosting System," ARXIV CS, 2016. [Online]. https://doi.org/10.48550/arXiv.1603.02754 Xiaohong Ye et al., "Automatic Multichannel Electrocardiogram Record Classification Using XGBoost Fusion Model," Frontiers in Physiology, vol. 13, 2022. [Online]. https://doi.org/10.3389/fphys.2022.840011 Diah Asmawati et al., "Arrhythmia Classification with ECG Signal using Extreme Gradient Boosting (XGBoost) Algorithm," Journal of Technology and Informatics JOTI, vol. 6, 2024. [Online]. https://doi.org/10.37802/joti.v6i1.792 Mansi Andotra et al., "Early Detection of Arrhythmia Using XGBoost Model," IEEE Xplore, 2024. [Online]. https://doi.org/10.1109/ICCCNT61001.2024.10724550 M. Sam Navin et al., "ECG Feature-Based Classification of Heart Disease Using Hybrid Ensemble Models," ATLANTIS PRESS ICISD , 2025. [Online]. https://doi.org/10.2991/978-94-6463-866-0_78 Huy Pham et al., "Machine learning-based detection of cardiovascular disease using ECG signals: performance vs. complexity," Frontiers in Cardiovascular Medicine, vol. 10, 2023. [Online]. https://doi.org/10.3389/fcvm.2023.1229743 Brindha V Naveen Kumar Kancharla, "Single and Multi-Lead ECG Heartbeat Classification Using XGBoost on Pre-trained CNN Feature Layer," 2024. [Online]. https://doi.org/10.21203/rs.3.rs-4028614/v1 U. Rajendra Acharya et al., "Application of deep convolutional neural network for automated detection of myocardial infarction using ECG signals," vol. 415–416, pp. 190-198, 2017. [Online]. https://doi.org/10.1016/j.ins.2017.06.027 Naz M et al., "From ECG signals to images: a transformation based approach for deep learning," PeerJ Computer Science, 2021. [Online]. https://doi.org/10.7717/peerj-cs.386 H. Anu et al., "Enhanced ECG Signal Classification with CNN-LSTM Networks using Aquila Optimization," Engineering, Technology & Applied Science Research, vol. 15, no. 3, pp. 23461-23466, June 2025. [Online]. https://doi.org/10.48084/etasr.10492 Ebrahim Alboghbaish Alaa Eleyan, "Electrocardiogram Signals Classification Using Deep-Learning-Based Incorporated Convolutional Neural Network and Long Short-Term Memory Framework," MDPI, June 2023. [Online]. https://doi.org/10.3390/computers13020055 Ahmed M Al Rajeh Manjur Kolhar, "Deep learning hybrid model ECG classification using AlexNet and parallel dual branch fusion network model," Scientific Reports, 2024. [Online]. https://doi.org/10.1038/s41598-024-78028-8 Gao Huang et al., "Densely Connected Convolutional Networks," 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), November 2017. [Online]. https://doi.org/10.1109/CVPR.2017.243 Shenda Hong et al., "Opportunities and challenges of deep learning methods for electrocardiogram data: A systematic review," Computers in Biology and Medicine, vol. 122, July 2020. [Online]. https://doi.org/10.1016/j.compbiomed.2020.103801 Yu Li et al., "ECG Classification with Dual Models: XGBoost Voting and Deep Learning with Attention," 2023 16th International Conference on Advanced Computer Theory and Engineering (ICACTE), September 2023. [Online]. https://doi.org/10.1109/ICACTE59887.2023.10335476 Tao Hua et al., "GEMM-Accelerated ECG Classification via Depthwise Separable 1D-CNN," 2025 IEEE Biomedical Circuits and Systems Conference (BioCAS), October 2025. [Online]. https://doi.org/10.1109/BioCAS67066.2025.00024 Danfeng Qin et al., "MobileNetV4 -- Universal Models for the Mobile Ecosystem," arXiv Computer Science - Computer Vision and Pattern Recognition, September 2024. [Online]. https://doi.org/10.48550/arXiv.2404.10518 Kai Han, Jianyuan Guo, Chang Xu, Chao Xu, Yunhe Wang Yehui Tang, "GhostNetV2: Enhance Cheap Operation with Long-Range Attention," arXiv Computer Vision and Pattern Recognition, 2022. [Online]. https://doi.org/10.48550/arXiv.2211.12905 Aditi Jambotkar, "Automatic Detection of Cardiovascular Abnormalities in ECG Images: CNN and MobileNet," 3rd International Conference on Futuristic Technology (INCOFT 2025), vol. 2, pp. 754-761, 2025. [Online]. https://doi.org/10.5220/0013601600004664 Kusuma Sreenivasa et al., "ECG based cardiac disorder classification using MobileNetV3 and LSTM," Asia-Pacific Journal of Science and Technology, vol. 28, p. 7, 2023. [Online]. https://doi.org/10.14456/apst.2023.82 Amita Dessai et al., "Emotion Classification Based on CWT of ECG and GSR Signals Using Various CNN Models," MDPI, 2023. [Online]. https://doi.org/10.3390/electronics12132795 Thi Diem Tran et al., "Efficient ECG Classification with Light Weight Shuffle GhostNet Architecture," 2023 International Conference on Advanced Technologies for Communications (ATC), 2023. [Online]. https://doi.org/10.1109/ATC58710.2023.10318918 James Gabriel, Jeff Zhu, Oncel Tuzel, Anurag Ranjan Pavan Kumar Anasosalu Vasu, "MobileOne: An Improved One millisecond Mobile Backbone," arXiv Computer Vision and Pattern Recognition , 2023. [Online]. https://doi.org/10.48550/arXiv.2206.04040 Sameer Singh, Carlos Guestrin Marco Tulio Ribeiro, ""Why Should I Trust You?": Explaining the Predictions of Any Classifier," arXiv Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Machine Learning (stat.ML), 2016. [Online]. https://doi.org/10.48550/arXiv.1602.04938 Ramprasaath R. Selvaraju et al., "Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization," International Journal of Computer Vision, 2019. [Online]. https://doi.org/10.48550/arXiv.1610.02391 Pang-Shuo Huang et al., "An Artificial Intelligence-Enabled ECG Algorithm for the Prediction and Localization of Angiography-Proven Coronary Artery Disease," MDPI, 2022. [Online]. https://doi.org/10.3390/biomedicines10020394 Patrick Wagner et al., "PTB-XL, a large publicly available electrocardiography dataset," Scientific Data, 2020. [Online]. https://doi.org/10.1038/s41597-020-0495-6 Goldberger et al., "PTB-XL, a large publicly available electrocardiography dataset," PhysioNet, 2022. [Online]. https://doi.org/10.13026/kfzx-aw45 Additional Declarations No competing interests reported. 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-8987933","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":598560030,"identity":"60d459da-17af-4845-b15e-bd2374253fc8","order_by":0,"name":"Amit Kumar Auddy","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA+UlEQVRIiWNgGAWjYFACxgbGBihTgofBBiTSeICAlsZGJC1pYEMIaAHqQdJyGMzAq4V/2uH2hzPbtsmbt58xvPG27bzd2vbDQFtqbKJxaZG4ndjYuLHttuGcMznGlnPbbidvO5MI1HIsLbcBlx6QlodttxlnMOSYSfMCtZgdAGphbDiMU4s8VIv9DP43IC3nks3OP8SvxQDqsMQZEmBbDtiZ3SBgiyFQy8wZ524nz5B4Vmw551xygtkNoC0JePwidzv9wceestu2M/iTN954U2Znb3Y+/eGDDzU2uL2PAhjZGBLBKhOIUg4GfxjsiVc8CkbBKBgFIwUAAP03bYL0r9PpAAAAAElFTkSuQmCC","orcid":"","institution":"Narula Institute of Technology","correspondingAuthor":true,"prefix":"","firstName":"Amit","middleName":"Kumar","lastName":"Auddy","suffix":""},{"id":598560031,"identity":"bfd97d92-4cd3-4807-8dc8-685590b366bb","order_by":1,"name":"Jagannibas Paul Choudhury","email":"","orcid":"","institution":"Narula Institute of Technology","correspondingAuthor":false,"prefix":"","firstName":"Jagannibas","middleName":"Paul","lastName":"Choudhury","suffix":""}],"badges":[],"createdAt":"2026-02-27 12:23:10","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8987933/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8987933/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":104175775,"identity":"f455def9-a3ea-4c11-a484-4b7a5ca18884","added_by":"auto","created_at":"2026-03-08 16:32:43","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":204719,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eExplainability comparison on 1D signals. (Top) XGBoost SHAP values show scattered focus on noise, explaining its failure on raw data. (Bottom) MobileNetV4 Grad-CAM correctly localizes the QRS complex, demonstrating morphological learning.\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-8987933/v1/02e7d5c990051a1346c1c527.png"},{"id":104175776,"identity":"e437e9c4-f882-4fb4-8247-14be3baa2939","added_by":"auto","created_at":"2026-03-08 16:32:43","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":177314,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eGradient-weighted Class Activation Mapping (Grad-CAM) for MobileNetV4. The model demonstrates high-fidelity localization of the T-wave and QRS complex despite having significantly fewer parameters than the baseline.\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-8987933/v1/19d677db5a249dda82e174d8.png"},{"id":108947580,"identity":"eb527a22-f8fb-48e8-b211-8ea8fca4f23d","added_by":"auto","created_at":"2026-05-11 06:29:57","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":730185,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8987933/v1/649a43a6-d43c-425d-8a04-34d121946f9a.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Robustness vs. Efficiency in ECG Classification: A Comparative Study of Deep Learning and Classical Machine Learning Architectures","fulltext":[{"header":"Introduction","content":"\u003cp\u003eCardiovascular diseases (CVDs) remain the leading cause of death globally, necessitating continuous and accurate monitoring of cardiac health [1]. The 12-lead Electrocardiogram (ECG) is the standard diagnostic tool, but its manual interpretation is time-consuming, prone to inter-observer variability, and requires specialized expertise [2]. Consequently, the development of automated diagnostic systems has become a major focus in medical informatics, leading to numerous systematic reviews evaluating the efficacy of artificial intelligence in cardiology [3,4,5,6]. However, the medical AI community is currently divided between two distinct paradigms, each prioritizing different computational methodologies.\u003c/p\u003e\n\u003cp\u003eThe first paradigm relies on Classical Machine Learning (ML) algorithms, such as Support Vector Machines (SVMs), Random Forests, and Gradient Boosting frameworks including XGBoost and CatBoost [7,8]. These models are highly favored in clinical tabular data analysis due to their low computational overhead and inherent interpretability [9,10,11]. Recent literature demonstrates their capability to achieve high diagnostic accuracy when applied to ECG classification tasks [12]. However, their application to physiological signals typically requires extensive manual feature extraction, dimensionality reduction, or rigorous spatial alignment [13,14,15]. This makes classical algorithms highly sensitive to phase shifts, baseline wander, and the physiological noise inherent in real-world clinical recordings.\u003c/p\u003e\n\u003cp\u003eTo bypass the limitations of handcrafted features, the field has increasingly pivoted toward Deep Learning (DL), particularly Convolutional Neural Networks (CNNs) and recurrent architectures [16,17,18]. Standard heavy CNNs have achieved cardiologist-level performance by automatically learning hierarchical representations from either raw 1D waveforms or 2D time-frequency transformations like scalograms [19,20]. Despite their high accuracy, these over-parameterized models, such as DenseNet and standard ResNet configurations, act as \"black boxes\" that introduce significant inference latency and memory constraints [21,22]. This computational burden severely limits their deployability in resource-constrained environments, such as wearable Holter monitors and embedded edge computing devices [23,24].\u003c/p\u003e\n\u003cp\u003eThis has driven recent interest toward Efficient Deep Learning architectures such as MobileNet configurations and GhostNet variants which seek to optimize operation types for mobile hardware [25,26,27]. Studies implementing these lightweight networks for ECG and related physiological signal classification have demonstrated massive reductions in computational overhead without sacrificing predictive power [28,29,30]. These architectures heavily utilize techniques like depthwise separable convolutions and feature redundancy operations to drastically reduce parameter counts [31].\u003c/p\u003e\n\u003cp\u003eDespite these advancements, a critical gap remains in the literature regarding the trade-off between algorithmic efficiency and structural robustness [32,33]. Most comparative studies benchmark accuracy strictly on static, pre-processed, and perfectly aligned datasets, potentially masking the fragility of the underlying models [34].\u003c/p\u003e\n\u003cp\u003eThis study addresses this gap by rigorously comparing modern Efficient Deep Learning architectures against highly optimized Classical Baselines. Using the comprehensive PTB-XL dataset [35], we specifically investigate an \"Accuracy vs. Robustness Trade-off,\" evaluating whether the statistical superiority of Classical ML on aligned 2D data translates to robust performance on raw, unaligned physiological signals.\u003c/p\u003e"},{"header":"Methodology","content":"\u003ch2\u003eDataset and Experimental Validation Split\u003c/h2\u003e\n\u003cp\u003eWe utilized the PTB-XL dataset [35], containing 21,837 clinical 12-lead ECGs from 18,885 patients. To address data heterogeneity, we utilized the standard 100 Hz versions of the records, decimated using a polyphase anti-aliasing filter. The task was formulated as a binary classification problem: Normal vs. Myocardial Infarction (MI)/Other. To ensure robust evaluation and completely prevent data leakage, we strictly adhered to the pre-defined, patient-stratified 10-fold split provided in the PTB-XL metadata. Specifically, Folds 1–8 (~80%, approx. 15,100 patients) were utilized for training, Fold 9 (~10%) was used for validation and hyperparameter early stopping, and Fold 10 (~10%) was reserved exclusively as an unseen test set for final benchmarking. This patient-stratified approach guarantees that records from the same patient do not simultaneously appear in both the training and testing phases.\u003c/p\u003e\n\u003ch2\u003ePreprocessing: Signal-to-Image Conversion\u003c/h2\u003e\n\u003cp\u003eTo leverage Transfer Learning from ImageNet-pretrained backbones, we implemented a Signal-to-Image conversion pipeline [17]. 1D time-series signals were converted into 2D scalograms using the Continuous Wavelet Transform (CWT) with a Complex Morlet wavelet. Scalograms were resized to 224x224 pixels for Deep Learning models. Biologically-plausible augmentations, including random time shifts (+/- 50ms) and additive Gaussian noise, were applied to ensure physiological realism.\u003c/p\u003e\n\u003ch2\u003eModel Architectures\u003c/h2\u003e\n\u003cp\u003eWe benchmarked four distinct architectures:\u003c/p\u003e\n\u003cul type=\"disc\"\u003e\n \u003cli\u003e\u003cstrong\u003eDenseNet121 [21]:\u003c/strong\u003e High-capacity baseline representing traditional heavy CNNs.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eMobileNetV4 [25]:\u003c/strong\u003e Representative of Neural Architecture Search (NAS) optimization for mobile efficiency.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eGhostNetV2 [26]:\u003c/strong\u003e Focuses on cheap feature redundancy operations.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eMobileOne [31]:\u003c/strong\u003e Utilizes inference-time reparameterization.\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eTo adapt these natively 2D vision architectures (e.g., MobileNetV4, GhostNetV2) for processing raw 1D physiological signals, significant architectural modifications were implemented. We systematically replaced all standard 2D Convolutional layers (Conv2d) with 1D Convolutional layers (Conv1d) and adjusted the input channel depth to 12 to accurately accommodate the 12-lead ECG format. Furthermore, the initial 2D spatial downsampling strides—which assume square image inputs—were modified to preserve the temporal resolution. This adaptation allowed the networks to directly ingest the 12 × 1000 waveform tensor and successfully learn temporal filters via a 1D sliding window.\u003c/p\u003e\n\u003ch2\u003eClassical Machine Learning Pipeline\u003c/h2\u003e\n\u003cp\u003eTo provide a rigorous baseline, we implemented a parallel pipeline using CatBoost [\u003ca href=\"#Gup\"\u003e8\u003c/a\u003e], XGBoost [\u003ca href=\"#Tia16\"\u003e9\u003c/a\u003e], and Random Forest. Inputs were either flattened 2D scalograms (128x128) or raw 1D signals. Dimensionality reduction algorithms were evaluated alongside raw pixel inputs.\u003c/p\u003e\n\u003ch2\u003eThe 1D Stress Test\u003c/h2\u003e\n\u003cp\u003eA key contribution of this study is the 1D Raw Signal Benchmark. To quantify \"artifact reliance,\" we evaluated all models on the raw 1D ECG signal without the spatial alignment provided by 2D image conversion. Deep Learning models (adapted to 1D-CNNs as described above) were trained to predict pathology directly from the waveform. For the Classical ML baselines, the raw input matrix for a single sample comprising 12 leads and 1000 timepoints was simply flattened into a single 12,000-dimensional feature vector per patient. These flattened arrays were fed directly into the Random Forest and XGBoost classifiers without any intermediate feature engineering (such as RR-interval extraction or wavelet transforms) to act as a true stress test of the algorithms' native pattern recognition capabilities on raw temporal data.\u003c/p\u003e"},{"header":"Results","content":"\u003ch2\u003eEfficiency vs. Accuracy\u003c/h2\u003e\n\u003cp\u003eTable 1 presents the performance metrics. The heavy baseline (DenseNet121) achieved an F1-score of 0.778 with an inference latency of 18.1ms. The proposed efficient model, MobileNetV4, achieved a comparable F1-score of 0.764 while reducing latency to 9.0ms (a 2.0x speedup). Notably, the Classical Baseline (CatBoost) achieved the highest F1-score of 0.803 on the aligned 2D image task.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 1.\u003c/strong\u003e \u003cstrong\u003eBenchmarking of proposed architectures vs. baselines. Inference time measured on RTX3080 GPU (DL) and Ryzen 9 CPU (Classical)\u003c/strong\u003e.\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"614\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 99px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eModel Type\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 130px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eArchitecture\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 116px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eF1-Score (2D)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 118px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eInference (ms)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 151px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eRobustness Feature\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 99px;\"\u003e\n \u003cp\u003eHeavy Baseline\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 130px;\"\u003e\n \u003cp\u003eDenseNet121\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 116px;\"\u003e\n \u003cp\u003e0.778\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 118px;\"\u003e\n \u003cp\u003e18.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 151px;\"\u003e\n \u003cp\u003eDeep Feature Reuse\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 99px;\"\u003e\n \u003cp\u003eProposed Efficient\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 130px;\"\u003e\n \u003cp\u003eMobileNetV4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 116px;\"\u003e\n \u003cp\u003e0.764\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 118px;\"\u003e\n \u003cp\u003e9.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 151px;\"\u003e\n \u003cp\u003eSpatial Invariance\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 99px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 130px;\"\u003e\n \u003cp\u003eGhostNetV2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 116px;\"\u003e\n \u003cp\u003e0.784\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 118px;\"\u003e\n \u003cp\u003e16.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 151px;\"\u003e\n \u003cp\u003eFeature Redundancy\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 99px;\"\u003e\n \u003cp\u003eClassical Baseline \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 130px;\"\u003e\n \u003cp\u003eCatBoost\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 116px;\"\u003e\n \u003cp\u003e0.803 \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 118px;\"\u003e\n \u003cp\u003e\u0026lt; 1.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 151px;\"\u003e\n \u003cp\u003ePixel Correlation (Artifacts)\u003c/p\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003ch2\u003eStructural Robustness (The 1D Stress Test)\u003c/h2\u003e\n\u003cp\u003eWhile Classical models outperformed on aligned 2D images, the 1D Raw Signal analysis revealed a substantial performance degradation. Table 2 demonstrates that when stripped of spatial alignment cues, Classical models failed to extract meaningful patterns.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 2. Performance collapse of Classical models on raw 1D signals vs. robustness of Deep Learning.\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"604\" class=\"fr-table-selection-hover\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 216px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eModel\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 129px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eInput Type\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 108px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eF1-Score\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 151px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eStatus\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 216px;\"\u003e\n \u003cp\u003eMobileNetV4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 129px;\"\u003e\n \u003cp\u003e1D Raw Signal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 108px;\"\u003e\n \u003cp\u003e0.844\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 151px;\"\u003e\n \u003cp\u003eROBUST\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 216px;\"\u003e\n \u003cp\u003eGhostNetV2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 129px;\"\u003e\n \u003cp\u003e1D Raw Signal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 108px;\"\u003e\n \u003cp\u003e0.821\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 151px;\"\u003e\n \u003cp\u003eROBUST\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 216px;\"\u003e\n \u003cp\u003eXGBoost\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 129px;\"\u003e\n \u003cp\u003e1D Raw Signal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 108px;\"\u003e\n \u003cp\u003e0.310\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 151px;\"\u003e\n \u003cp\u003eFAILED\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 216px;\"\u003e\n \u003cp\u003eRandom Forest\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 129px;\"\u003e\n \u003cp\u003e1D Raw Signal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 108px;\"\u003e\n \u003cp\u003e0.042\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 151px;\"\u003e\n \u003cp\u003eCOLLAPSED\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eNotably, the Random Forest model collapsed to an F1-score of 0.042 on 1D data, exhibiting extreme majority class bias (Sensitivity ~2%). In contrast, the MobileNetV4-1D architecture maintained high performance (F1 0.844), validating its ability to learn translation-invariant temporal filters. Explainability analysis via Gradient-weighted Class Activation Mapping (Grad-CAM) further confirmed these findings [33].\u003c/p\u003e"},{"header":"Discussion","content":"\u003ch2\u003eThe Accuracy Paradox and The Curse of Dimensionality\u003c/h2\u003e\n\u003cp\u003eOur results highlight an \"Accuracy Paradox.\" Classical Machine Learning models technically achieved higher scores on the static 2D dataset, but this performance was artifact-driven. The collapse of Classical models on the 1D benchmark confirms their reliance on spatial alignment artifacts (e.g., the fixed position of the QRS complex in the pre-processed image) rather than morphological features. When the structured data alignment was removed, their predictive performance vanished. Feeding 12,000 highly correlated, raw temporal data points directly into Random Forest or XGBoost models triggered the \"curse of dimensionality.\" Without hand-crafted features or the structural invariance provided by a CNN's sliding window, the tree-based algorithms failed to establish meaningful decision boundaries across the 12,000 raw voltage points, causing them to collapse and default to majority class predictions.\u003c/p\u003e\n\u003ch2\u003eDeep Learning Suitability\u003c/h2\u003e\n\u003cp\u003eConvolutional Neural Networks, specifically efficient variations like MobileNetV4, demonstrated inherent Spatial Invariance. Their sliding-window filters successfully detected the shape of the QRS complex regardless of its temporal position in the input tensor. This structural robustness confirms that Deep Learning is the mathematically safer architectural choice for clinical diagnostics, where physiological signals invariably contain phase shifts and temporal misalignment.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eThis study compared the efficacy of Efficient Deep Learning against Classical Machine Learning for ECG analysis. We conclude that while Classical models offer superior speed and performance on rigidly structured data, they are structurally brittle when applied to unaligned physiological signals. Efficient Deep Learning architectures, specifically MobileNetV4, resolve this trade-off by delivering robust, translation-invariant feature extraction at a fraction of the computational cost of traditional heavy models. Future work will focus on prospective clinical validation and quantization for embedded edge devices.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cul\u003e\n \u003cli\u003e\u003cstrong\u003eFunding:\u003c/strong\u003e The authors declare that no funds, grants, or other support were received during the preparation of this manuscript.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eCompeting Interests:\u003c/strong\u003e The authors have no relevant financial or non-financial interests to disclose.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eData Availability:\u003c/strong\u003e The PTB-XL dataset analysed during the current study is openly available in the PhysioNet repository [36].\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eEthics Approval:\u003c/strong\u003e Not applicable. This study utilized a publicly available, fully de-identified retrospective dataset; therefore, institutional ethics approval was not required.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eConsent to Participate/Publish:\u003c/strong\u003e Not applicable.\u0026nbsp;\u003c/li\u003e\n\u003c/ul\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eWorld Health Organization (WHO), \u0026quot;Cardiovascular diseases (CVDs),\u0026quot; 2025. [Online]. https://www.who.int/news-room/fact-sheets/detail/cardiovascular-diseases-(cvds)\u003c/li\u003e\n\u003cli\u003ePranav Rajpurkar, Masoumeh Haghpanahi, Geoffrey H. Tison, Codie Bourn, Mintu P. Turakhia \u0026amp; Andrew Y. Ng Awni Y. Hannun, \u0026quot;Cardiologist-level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural network,\u0026quot; nature medicine, vol. 25, no. January 2019, pp. 65\u0026ndash;69, January 2009. [Online]. https://doi.org/10.1038/s41591-018-0268-3\u003c/li\u003e\n\u003cli\u003eAbdelkader Nasreddine Belkacem, Ibrahim M. Elfadel, Herbert F. Jelinek, Mohamed Atef Allam Jaya Prakash, \u0026quot;Advances in machine and deep learning for ECG beat classification: a systematic review,\u0026quot; vol. 7, 2025. [Online]. https://doi.org/10.3389/fdgth.2025.1649923\u003c/li\u003e\n\u003cli\u003eA. Shahbahrami M. Sarkaleh, \u0026quot;CLASSIFICATION OF ECG ARRHYTHMIAS USING DISCRETE WAVELET TRANSFORM AND NEURAL NETWORKS,\u0026quot; International Journal of Computer Science, Engineering and Applications (IJCSEA), vol. 2, 2012. [Online]. https://doi.org/10.5121/ijcsea.2012.2101\u003c/li\u003e\n\u003cli\u003eGleb Gusev, Aleksandr Vorobev, Anna Veronika Dorogush, Andrey Gulin Liudmila Prokhorenkova, \u0026quot;CatBoost: unbiased boosting with categorical features,\u0026quot; ARXIV CS, 2019. [Online]. https://doi.org/10.48550/arXiv.1706.09516\u003c/li\u003e\n\u003cli\u003eSalih R. Marwa and Mohammed A. Shakir, \u0026quot;Improving ECG signals classification by using deep learning techniques: A review,\u0026quot; 2nd International Conference on Applied Computing \u0026amp; Smart Cities (ICACS24), vol. 64, p. 01023, 2024. [Online]. https://doi.org/10.1051/itmconf/20246401023\u003c/li\u003e\n\u003cli\u003eM. Abubakar et al., \u0026quot;ECGNet: High-Precision ECG Classification Using Deep Learning and Advanced Activation Functions,\u0026quot; IEEE Access, vol. 13, pp. 133191-133207, 2025. [Online]. https://doi.org/10.1109/ACCESS.2025.3592240\u003c/li\u003e\n\u003cli\u003eUtkarsh et al. Gupta, \u0026quot;A comprehensive review on efficient artificial intelligence models for classification of abnormal cardiac rhythms using electrocardiograms,\u0026quot; vol. 10, no. 5, p. e26787. [Online]. https://doi.org/10.1016/j.heliyon.2024.e26787\u003c/li\u003e\n\u003cli\u003eCarlos Guestrin Tianqi Chen, \u0026quot;XGBoost: A Scalable Tree Boosting System,\u0026quot; ARXIV CS, 2016. [Online]. https://doi.org/10.48550/arXiv.1603.02754\u003c/li\u003e\n\u003cli\u003eXiaohong Ye et al., \u0026quot;Automatic Multichannel Electrocardiogram Record Classification Using XGBoost Fusion Model,\u0026quot; Frontiers in Physiology, vol. 13, 2022. [Online]. https://doi.org/10.3389/fphys.2022.840011\u003c/li\u003e\n\u003cli\u003eDiah Asmawati et al., \u0026quot;Arrhythmia Classification with ECG Signal using Extreme Gradient Boosting (XGBoost) Algorithm,\u0026quot; Journal of Technology and Informatics JOTI, vol. 6, 2024. [Online]. https://doi.org/10.37802/joti.v6i1.792\u003c/li\u003e\n\u003cli\u003eMansi Andotra et al., \u0026quot;Early Detection of Arrhythmia Using XGBoost Model,\u0026quot; IEEE Xplore, 2024. [Online]. https://doi.org/10.1109/ICCCNT61001.2024.10724550\u003c/li\u003e\n\u003cli\u003eM. Sam Navin et al., \u0026quot;ECG Feature-Based Classification of Heart Disease Using Hybrid Ensemble Models,\u0026quot; ATLANTIS PRESS ICISD , 2025. [Online]. https://doi.org/10.2991/978-94-6463-866-0_78\u003c/li\u003e\n\u003cli\u003eHuy Pham et al., \u0026quot;Machine learning-based detection of cardiovascular disease using ECG signals: performance vs. complexity,\u0026quot; Frontiers in Cardiovascular Medicine, vol. 10, 2023. [Online]. https://doi.org/10.3389/fcvm.2023.1229743\u003c/li\u003e\n\u003cli\u003eBrindha V Naveen Kumar Kancharla, \u0026quot;Single and Multi-Lead ECG Heartbeat Classification Using XGBoost on Pre-trained CNN Feature Layer,\u0026quot; 2024. [Online]. https://doi.org/10.21203/rs.3.rs-4028614/v1\u003c/li\u003e\n\u003cli\u003eU. Rajendra Acharya et al., \u0026quot;Application of deep convolutional neural network for automated detection of myocardial infarction using ECG signals,\u0026quot; vol. 415\u0026ndash;416, pp. 190-198, 2017. [Online]. https://doi.org/10.1016/j.ins.2017.06.027\u003c/li\u003e\n\u003cli\u003eNaz M et al., \u0026quot;From ECG signals to images: a transformation based approach for deep learning,\u0026quot; PeerJ Computer Science, 2021. [Online]. https://doi.org/10.7717/peerj-cs.386\u003c/li\u003e\n\u003cli\u003eH. Anu et al., \u0026quot;Enhanced ECG Signal Classification with CNN-LSTM Networks using Aquila Optimization,\u0026quot; Engineering, Technology \u0026amp; Applied Science Research, vol. 15, no. 3, pp. 23461-23466, June 2025. [Online]. https://doi.org/10.48084/etasr.10492\u003c/li\u003e\n\u003cli\u003eEbrahim Alboghbaish Alaa Eleyan, \u0026quot;Electrocardiogram Signals Classification Using Deep-Learning-Based Incorporated Convolutional Neural Network and Long Short-Term Memory Framework,\u0026quot; MDPI, June 2023. [Online]. https://doi.org/10.3390/computers13020055\u003c/li\u003e\n\u003cli\u003eAhmed M Al Rajeh Manjur Kolhar, \u0026quot;Deep learning hybrid model ECG classification using AlexNet and parallel dual branch fusion network model,\u0026quot; Scientific Reports, 2024. [Online]. https://doi.org/10.1038/s41598-024-78028-8\u003c/li\u003e\n\u003cli\u003eGao Huang et al., \u0026quot;Densely Connected Convolutional Networks,\u0026quot; 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), November 2017. [Online]. https://doi.org/10.1109/CVPR.2017.243\u003c/li\u003e\n\u003cli\u003eShenda Hong et al., \u0026quot;Opportunities and challenges of deep learning methods for electrocardiogram data: A systematic review,\u0026quot; Computers in Biology and Medicine, vol. 122, July 2020. [Online]. https://doi.org/10.1016/j.compbiomed.2020.103801\u003c/li\u003e\n\u003cli\u003eYu Li et al., \u0026quot;ECG Classification with Dual Models: XGBoost Voting and Deep Learning with Attention,\u0026quot; 2023 16th International Conference on Advanced Computer Theory and Engineering (ICACTE), September 2023. [Online]. https://doi.org/10.1109/ICACTE59887.2023.10335476\u003c/li\u003e\n\u003cli\u003eTao Hua et al., \u0026quot;GEMM-Accelerated ECG Classification via Depthwise Separable 1D-CNN,\u0026quot; 2025 IEEE Biomedical Circuits and Systems Conference (BioCAS), October 2025. [Online]. https://doi.org/10.1109/BioCAS67066.2025.00024\u003c/li\u003e\n\u003cli\u003eDanfeng Qin et al., \u0026quot;MobileNetV4 -- Universal Models for the Mobile Ecosystem,\u0026quot; arXiv Computer Science - Computer Vision and Pattern Recognition, September 2024. [Online]. https://doi.org/10.48550/arXiv.2404.10518\u003c/li\u003e\n\u003cli\u003eKai Han, Jianyuan Guo, Chang Xu, Chao Xu, Yunhe Wang Yehui Tang, \u0026quot;GhostNetV2: Enhance Cheap Operation with Long-Range Attention,\u0026quot; arXiv Computer Vision and Pattern Recognition, 2022. [Online]. https://doi.org/10.48550/arXiv.2211.12905\u003c/li\u003e\n\u003cli\u003eAditi Jambotkar, \u0026quot;Automatic Detection of Cardiovascular Abnormalities in ECG Images: CNN and MobileNet,\u0026quot; 3rd International Conference on Futuristic Technology (INCOFT 2025), vol. 2, pp. 754-761, 2025. [Online]. https://doi.org/10.5220/0013601600004664\u003c/li\u003e\n\u003cli\u003eKusuma Sreenivasa et al., \u0026quot;ECG based cardiac disorder classification using MobileNetV3 and LSTM,\u0026quot; Asia-Pacific Journal of Science and Technology, vol. 28, p. 7, 2023. [Online]. https://doi.org/10.14456/apst.2023.82\u003c/li\u003e\n\u003cli\u003eAmita Dessai et al., \u0026quot;Emotion Classification Based on CWT of ECG and GSR Signals Using Various CNN Models,\u0026quot; MDPI, 2023. [Online]. https://doi.org/10.3390/electronics12132795\u003c/li\u003e\n\u003cli\u003eThi Diem Tran et al., \u0026quot;Efficient ECG Classification with Light Weight Shuffle GhostNet Architecture,\u0026quot; 2023 International Conference on Advanced Technologies for Communications (ATC), 2023. [Online]. https://doi.org/10.1109/ATC58710.2023.10318918\u003c/li\u003e\n\u003cli\u003eJames Gabriel, Jeff Zhu, Oncel Tuzel, Anurag Ranjan Pavan Kumar Anasosalu Vasu, \u0026quot;MobileOne: An Improved One millisecond Mobile Backbone,\u0026quot; arXiv Computer Vision and Pattern Recognition , 2023. [Online]. https://doi.org/10.48550/arXiv.2206.04040\u003c/li\u003e\n\u003cli\u003eSameer Singh, Carlos Guestrin Marco Tulio Ribeiro, \u0026quot;\u0026quot;Why Should I Trust You?\u0026quot;: Explaining the Predictions of Any Classifier,\u0026quot; arXiv Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Machine Learning (stat.ML), 2016. [Online]. https://doi.org/10.48550/arXiv.1602.04938\u003c/li\u003e\n\u003cli\u003eRamprasaath R. Selvaraju et al., \u0026quot;Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization,\u0026quot; International Journal of Computer Vision, 2019. [Online]. https://doi.org/10.48550/arXiv.1610.02391\u003c/li\u003e\n\u003cli\u003ePang-Shuo Huang et al., \u0026quot;An Artificial Intelligence-Enabled ECG Algorithm for the Prediction and Localization of Angiography-Proven Coronary Artery Disease,\u0026quot; MDPI, 2022. [Online]. https://doi.org/10.3390/biomedicines10020394\u003c/li\u003e\n\u003cli\u003ePatrick Wagner et al., \u0026quot;PTB-XL, a large publicly available electrocardiography dataset,\u0026quot; Scientific Data, 2020. [Online]. https://doi.org/10.1038/s41597-020-0495-6\u003c/li\u003e\n\u003cli\u003eGoldberger et al., \u0026quot;PTB-XL, a large publicly available electrocardiography dataset,\u0026quot; PhysioNet, 2022. [Online]. https://doi.org/10.13026/kfzx-aw45\u003c/li\u003e\n\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":"Deep Learning, Electrocardiogram, MobileNetV4, GhostNetV2, Robustness, Edge Computing","lastPublishedDoi":"10.21203/rs.3.rs-8987933/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8987933/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003ePurpose: \u003c/strong\u003eCardiovascular diseases remain the predominant cause of global mortality, necessitating reliable automated tools for early diagnosis. While Deep Learning (DL) has shown superiority in medical imaging, recent studies suggest that Classical Machine Learning models may offer comparable accuracy with lower computational costs. This study investigates this accuracy versus robustness trade-off.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods:\u003c/strong\u003e We present a rigorous comparison between modern Efficient Deep Learning architectures (MobileNetV4, GhostNetV2) and established Classical Baselines (CatBoost, Random Forest) for Myocardial Infarction detection using the PTB-XL dataset. Models were first evaluated on 2D scalograms generated via Continuous Wavelet Transform (CWT). To test structural robustness and artifact reliance, we subsequently introduced a 1D Raw Signal Benchmark evaluating the models on unaligned, raw time-series data.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults:\u003c/strong\u003e On aligned 2D image data, Classical models achieved competitive F1-scores (~0.80). However, the 1D Raw Signal Benchmark revealed a severe performance degradation for Classical methods (F1 drop to ~0.04) on unaligned data, highlighting a heavy reliance on spatial artifacts. Conversely, Convolutional Neural Networks (CNNs) maintained high performance (F1 ~0.84) on raw signals.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusion:\u003c/strong\u003e While Classical models are computationally efficient for highly structured data, they are structurally brittle. Efficient Deep Learning architectures, specifically MobileNetV4, offer the superior translation-invariant feature learning required for clinical signal analysis, delivering a 2.0x speedup over heavy CNN baselines with negligible accuracy loss.\u003c/p\u003e","manuscriptTitle":"Robustness vs. Efficiency in ECG Classification: A Comparative Study of Deep Learning and Classical Machine Learning Architectures","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-03-08 16:32:38","doi":"10.21203/rs.3.rs-8987933/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"24df90e4-1c40-476f-8b42-ef59c623f5c7","owner":[],"postedDate":"March 8th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-05-11T06:27:15+00:00","versionOfRecord":[],"versionCreatedAt":"2026-03-08 16:32:38","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8987933","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8987933","identity":"rs-8987933","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}