An Evaluation of Energy-efficient AI Models for Edge Deployment in 12-lead ECG Classification

In recent years, artificial intelligence (AI) techniques, particularly deep learning, have shown significant promise in modern tele-medicine. Deep learning models, particularly Convolutional Neural Networks (CNNs), have been trained on large datasets to identify patterns and anomalies within 12-lead electrocardiograms (ECGs), which are crucial for diagnosing various cardiac conditions. In this study, we conducted a comparative analysis to identify and evaluate models that can be effectively deployed in rural healthcare settings, where energy efficiency, low computational requirements, and ease of use are critical factors. We compare the performance of two well-known sequential data processing model architectures: time-series RetNet and Mamba. The best model from this comparative study achieves a notable precision of over 95% in a 10 fold cross-validation for the classification of nine types of cardiac arrhythmias (CA) in the 2018 PhysioNet/CinC Challenge (CPSC 2018) dataset.