Residual Spatio-Temporal Attention Based Prototypical Network for Rare Arrhythmia Classification

Arrhythmia is a common cardiovascular disease that requires early detection and treatment to improve prognosis. Electrocardiogram (ECG) is an important tool for diagnosing and monitoring heart health. However, existing ECG diagnostic methods suffer from incomplete capture of spatio-temporal features and poor recognition ability for rare categories. In this paper, we introduce few-shot learning for ECG signals and propose a Residual Spatio-Temporal Attention based Prototypical Network (RSTA-ProtoNet) for rare arrhythmia classification. In the model, a spatio-temporal attention residual network is constructed as the backbone network. This network uses interleaved temporal and spatial attention encoders for extracting spatio-temporal features of ECG. Meanwhile, we build a few-shot learning framework based on prototypical networks to classify rare arrhythmia classes. This meta-training framework can learn useful features for classifying rare categories even with extremely limited samples of rare diseases. We evaluate our method on a large public ECG dataset, and the N-way K-shot experimental results demonstrate that RSTA-ProtoNet outperforms the state-of-the-art approaches in rare arrhythmia classification.