ArrhythmiaVision: resource-conscious deep learning models with visual explanations for ECG arrhythmia classification

Cardiac arrhythmias are a leading cause of life-threatening cardiac events, highlighting the urgent need for accurate and timely detection. Electrocardiography (ECG) remains the clinical gold standard for arrhythmia diagnosis; however, manual interpretation is time-consuming, dependent on clinical expertise, and prone to human error. for deployment on resource-constrained platforms. Although deep learning has advanced automated ECG analysis, many existing models abstract away the signal’s intrinsic temporal and morphological features, lack interpretability, and are computationally intensive for deployment on resource-constrained platforms, precisely the class of system-level challenges that demand efficient, edge-ready architectures. In this work, we propose two novel lightweight 1D convolutional neural networks, ArrhythmiNet V1 and V2, optimized for efficient, real-time arrhythmia classification on edge devices. Inspired by MobileNet’s depthwise separable convolutional design, these models maintain memory footprints of just 302.18 KB and 157.76 KB, respectively, while achieving mean classification accuracies of 0.991±0.002[12pt]{minimal} {amsmath} {wasysym} {amsfonts} {amssymb} {amsbsy} {mathrsfs} {upgreek} {}{-69pt} {document}0.991 ± 0.002{document} (V1) and 0.978±0.003[12pt]{minimal} {amsmath} {wasysym} {amsfonts} {amssymb} {amsbsy} {mathrsfs} {upgreek} {}{-69pt} {document}0.978 ± 0.003{document} (V2) across five independent runs on the MIT-BIH Arrhythmia Dataset, covering five classes: Normal Sinus Rhythm, Left Bundle Branch Block, Right Bundle Branch Block, Atrial Premature Contraction, and Premature Ventricular Contraction. Statistical comparison using McNemar’s test confirms that the performance advantage of V1 over V2 is significant (χ2=74.01[12pt]{minimal} {amsmath} {wasysym} {amsfonts} {amssymb} {amsbsy} {mathrsfs} {upgreek} {}{-69pt} {document}χ ² = 74.01{document}, p<0.001[12pt]{minimal} {amsmath} {wasysym} {amsfonts} {amssymb} {amsbsy} {mathrsfs} {upgreek} {}{-69pt} {document}$$p < 0.001$${document}). The architectures are specifically designed to preserve the spatial morphology and temporal dynamics of ECG signals for robust performance. Hence, placing both architectures firmly on the efficiency-accuracy frontier relevant to embedded and supercomputing-edge deployment. In order to ensure clinical transparency and relevance, we integrate Shapley Additive Explanations (SHAP) and Gradient-weighted Class Activation Mapping (Grad-CAM), enabling both local and global interpretability. These techniques highlight physiologically meaningful patterns such as the QRS complex and T-wave that contribute to the model’s predictions, addressing a deployment barrier increasingly recognized in clinical AI: a model that cannot explain its decision cannot be trusted in a safety-critical pipeline, regardless of benchmark accuracy. We also discuss performance-efficiency trade-offs and address current limitations related to dataset diversity and generalizability. Our findings demonstrate the feasibility of combining interpretability, predictive accuracy, and computational efficiency in practical, wearable, and embedded ECG monitoring systems.