A 12-Lead ECG Delineation Algorithm based on a Quantized CNN-BiLSTM Auto-encoder with 1-12 Mapping

Abstract

12-lead electrocardiogram (ECG) delineation is a critical step in diagnosing of various heart diseases. Current practices for 12-lead ECG delineation typically involve processing each of the 12 leads separately using a network, which is computationally expensive. To solve this issue, 1-12 mapping strategy is proposed to directly map one lead network predictions to other leads and then fine-tune boundaries. CNN-BiLSTM autoencoder architecture is employed to model the sequential dependencies of ECG signal. Besides, data augmentation and mixed losses are utilized to enhance the robustness of the network. Evaluated on QTDB and LUDB, the delineation results for 12-lead ECG achieve a Se of 97%, 99%, and 98%, DS of 95.3%, 96.2%, and 94.4% for P-wave, QRS complex, and T-wave respectively. At last, quantization-aware training is employed to convert float32 model to int8 one with only about a 2% drop of accuracy.