Efficient Two-Stage Neural Network Architecture for ECG Classification and Real-Time Monitoring

This paper proposes a monitoring system with on-device processing capabilities, using a two-stage machine learning (ML) approach for anomaly detection and multiclassification. Our approach involves an initial stage featuring a multilayer perception (MLP), followed by a convolutional neural network (CNN) combined with a long short-term memory (LSTM) neural network in the second stage. This design affords flexibility in deploying each stage separately: the first stage operates on-device with minimal computational requirements and high accuracy, while the subsequent stage, which is more complex in nature, operates on edge devices. Our model demonstrates superior computational efficiency compared to baseline approaches, with reduced numbers of multiplications, additions, and parameters. Specifically, our proposed MLP model has over 60 % fewer parameters compared to the baseline model, and the overall computational workload is reduced by more than 70%. Evaluation using the MIT-BIH Arrhythmia Dataset shows a competitive performance against state-of-the-art methods. Thus, the proposed model is suitable for applications in wireless remote health monitoring systems, enabling real-time anomaly detection and classification with minimal computational overhead.