Algorithm: Pre-fine-tuning.

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.

室性心律失常（Ventricular Arrhythmia, VA）是心源性猝死的主要诱因。利用深度学习技术从心电图（Electrocardiogram, ECG）中检测室性心律失常，有望改善临床结局。然而，开发用于心电分析的鲁棒深度学习模型仍面临挑战，原因在于：（1）不同个体间的受试者间差异（inter-subject diversity）；（2）同一受试者在不同生理状态下随时间变化的受试者内差异（intra-subject diversity）。本研究通过在预训练与微调两个阶段均引入改进方案来应对上述挑战。在预训练阶段，我们提出一种结合模型不可知元学习（model-agnostic meta-learning, MAML）与课程学习（curriculum learning, CL）的创新方法，以有效解决受试者间差异问题。模型不可知元学习可高效从大规模数据集迁移知识，并能通过有限的心电记录实现模型对新个体的快速适配。整合课程学习可通过按从简单到复杂的顺序训练模型，进一步提升模型不可知元学习的训练效果。在微调阶段，我们提出一种专为管理受试者内差异设计的改进预微调策略。我们在三个公开心电数据集与一个采用便携设备采集的真实临床心电数据集上对所提方法进行了评估。在测试集上，仅使用每类每受试者10次心跳数据时，所提方法取得了ROC-AUC=0.984、F1=0.940的性能；在真实临床采集数据集上，其ROC-AUC达0.965、F1为0.937。实验结果表明，所提方法在所有评估指标上均优于现有对比方法，且更能有效应对受试者内差异。消融实验证实，模型不可知元学习与课程学习的结合可使模型在不同个体间的性能表现更为均匀，而我们改进的预微调技术可显著提升模型对个体特异性数据的适配能力。