基于未来引导学习的心电图心律失常分类研究

Research on ECG Arrhythmia Classification Based on Future-Guided Learning

摘要: 可穿戴设备在心电图实时监测中对计算资源和推理延迟要求极高，通常只能截取极短的局部信号片段进行分析。然而，短窗口轻量级模型因缺失心跳前后的全局上下文信息，在心律失常分类中面临性能瓶颈。为解决极低延迟要求与局部信息缺失之间的矛盾，本文提出了一种基于未来引导学习的心律失常分类方法。该方法构建了非对称的教师–学生网络架构，在训练阶段，教师模型利用包含心跳前后非对称扩展的宽视窗捕捉代偿间歇等长程演变机制；随后通过跨时域知识蒸馏，将结构化的暗知识传递给仅输入窄视窗的学生模型。此外，针对数据集类别极度不平衡问题，设计了基于有效样本数的类别平衡交叉熵损失函数以稳定训练过程。在MIT-BIH心律失常数据库上的实验结果表明，该框架在不增加学生模型推理阶段计算成本的前提下，将宏平均F1分数由基准模型的94.12%显著提升至95.60%。尤其是对于形态复杂且极易混淆的室上性异常心拍和融合波，其F1分数分别实现了3.57%和2.15%的绝对提升。

Abstract: Wearable devices for real-time electrocardiogram (ECG) monitoring impose stringent requirements on computational resources and inference latency, typically necessitating the analysis of extremely short, localized signal segments. However, lightweight models operating on such short windows suffer from a performance bottleneck in arrhythmia classification due to the lack of global contextual information before and after the heartbeat. To resolve the conflict between the extremely low latency requirement and the absence of local information, this paper proposes an arrhythmia classification method based on future-guided learning. This method constructs an asymmetric teacher-student network architecture. During the training phase, the teacher model utilizes a wide window that asymmetrically extends before and after the heartbeat to capture long-term evolution mechanisms such as compensatory intervals. Subsequently, through cross-temporal knowledge distillation, it transfers structured dark knowledge to the student model, which only receives a narrow window as input. Furthermore, to address the severe class imbalance in the dataset, a class-balanced cross-entropy loss function based on the effective number of samples is designed to stabilize the training process. Experimental results on the MIT-BIH Arrhythmia Database demonstrate that this framework significantly increases the macro-averaged F1-score from 94.12% (baseline model) to 95.60% without adding computational cost during the student model’s inference phase. Notably, for morphologically complex and easily confusable supraventricular ectopic beats and fusion beats, absolute improvements of 3.57% and 2.15% in F1-score were achieved, respectively.