ARCHIVES

Electrocardiogram (ECG) analysis is a key step in many clinical tasks, because the ECG is the most commonly used non-invasive examination for assessing cardiac condition. State-of-the-art algorithms rely on digital signal processing (DSP) techniques, which require considerable effort to re-tune rules for new morphologies. In contrast, deep learning (DL) algorithms, especially for classification, are attracting growing attention in academia and industry. This study proposes a three-stage hierarchical deep learning framework. In Stage 1, a one-dimensional U-Net performs coarse segmentation of the ECG signal to separate background from regions of cardiac activity. In Stage 2, an attention-fusion U-Net takes the probability map produced in Stage 1 to localize and classify the P wave, QRS complex and T wave more precisely in time. In Stage 3, complete heartbeat segments are extracted from the Stage 2 segmentation result and a beat-level classification model is trained to classify beats, following the AAMI EC57 standard, into normal beats (N) and three classes of abnormal rhythm (S, V and Q). The Stage 1 and Stage 2 segmentation models achieved Macro F1 scores of 0.965 and 0.999, respectively, indicating very high waveform localization accuracy. Combined with soft max post-processing, the final four-class beat classifier reached a Macro F1 of 0.9779 and an overall accuracy of 0.9970 on the test set, and remained stable on the minority classes (V and Q).