A Lightweight and Robust Parallel CNN–LSTM Network with Random Crop Augmentation for Single-Lead Raw ECG Arrhythmia Classification

Abstract

Cardiovascular diseases remain one of the leading causes of mortality worldwide, creating a strong need for continuous and reliable electrocardiogram (ECG) monitoring systems. Although recent deep learning methods have achieved high arrhythmia classification accuracy, many of them are computationally demanding and less suitable for wearable or edge-oriented applications. In addition, single-lead ECG recordings may exhibit temporal misalignment, segment variability, and noise, which can reduce model robustness in practical monitoring scenarios. To address these challenges, this study proposes a lightweight parallel CNN-LSTM framework for raw single-lead ECG arrhythmia classification, together with a Random Crop augmentation strategy designed to improve robustness to temporal variability. The proposed method was evaluated on the MIT-BIH Arrhythmia Database and the INCART Database under an intra-patient 5-fold cross-validation protocol. Three classification settings were considered: MIT-BIH NSVFQ, MIT-BIH NLRAV, and INCART NSV. The proposed parallel CNN-LSTM consistently outperformed CNN-only and LSTM-only baselines across all tasks. It achieved 99.34% accuracy and 95.59% macro F1-score on MIT-BIH NSVFQ, 99.58% accuracy and 98.63% macro F1-score on MIT-BIH NLRAV, and 99.74% accuracy and 97.22% macro F1-score on INCART NSV. Additional ablation results showed that Random Crop generally improved performance when combined with CNN-based and parallel feature learning, while its effect on LSTM-only modeling was more limited and task-dependent. To further assess practical applicability, input-length sensitivity, class-wise confusion-matrix analysis, and noise robustness experiments were also performed. The results showed that the selected input lengths of 340 samples for MIT-BIH and 240 samples for INCART provide effective heartbeat-centered context, and that the proposed framework remains highly stable under mild-to-moderate Gaussian and EMG-like noise. Grad-CAM visualizations further indicated that the model primarily attends to clinically meaningful waveform regions, especially around the QRS complex and class-dependent morphological transitions. In addition to predictive performance, computational efficiency was analyzed in terms of parameter count, FLOPs, and inference latency. The proposed model contains 2,070,917 parameters for the MIT-BIH configuration and 1,865,603 parameters for the INCART configuration, while maintaining real-time inference capability with CPU runtime below 1 ms per beat in both settings. These findings suggest that the proposed framework provides an effective and computationally practical solution for intra-patient single-lead ECG arrhythmia classification in wearable monitoring scenarios.