A lightweight deep neural network for personalized detecting ventricular arrhythmias from a single-lead ECG device.

Ventricular arrhythmia (VA) is a leading cause of sudden cardiac death. Detecting VA from electrocardiograms (ECGs) using deep learning techniques has potential to improve clinical outcomes. However, developing robust deep learning models for ECG analysis remains challenging due to: (1) inter-subject diversity among different individuals, and (2) intra-subject diversity within the same subject across different physiological state over time. In this study, we address these challenges by introducing enhancements in both the pre-training and fine-tuning stages. In the pre-training stage, we propose a novel approach combining model-agnostic meta-learning (MAML) with curriculum learning (CL) to effectively address inter-subject diversity. MAML efficiently transfer knowledge from large-scale datasets and enables rapid model adaptation to new individuals using limited records. Integrating CL further enhances the effectiveness of MAML by sequentially training models from simpler to more complex tasks. For the fine-tuning stage, we propose an improved pre-fine-tuning strategy specifically designed to manage the intra-subject diversity. We evaluate our methods on three publicly available ECG datasets and one real-world clinical ECG dataset collected using a portable device. Our proposed method achieves ROC-AUC = 0.984 / F1 = 0.940 with only 10 beats per class per subject on the test set and also achieves ROC-AUC = 0.965 / F1 = 0.937 on a real-world clinical collected data. Experimental results demonstrate that our proposed approach outperforms existing comparative methods across all evaluation metrics, and have a tendency to address intra-subject diversity. Ablation studies confirm that the combination of MAML and CL leads to more uniform performance across individuals, and our enhanced pre-fine-tuning technique substantially improves model adaptation to individual-specific data.