Self-Trained Model for ECG Complex Delineation

TL;DR

This work tackles robust ECG fiducial-point delineation under limited labeled data by introducing the ISP 12-lead annotated dataset and a self-training framework that leverages unlabeled data. The ECG-CODE CNN processes 12-lead signals converted to mel spectrograms to predict P, QRS, and T wave onset/offset across intervals, trained with a composite loss and post-processing. A novel self-training pipeline uses pseudolabels from unlabeled PTB-XL, selected via a delineation score $= |0.5 - \hat{c}|$, and retrains from scratch before fine-tuning on ISP, yielding improved delineation, especially for P and T waves. Experiments across ISP and LUDB demonstrate improved generalization and highlight the value of unlabeled data for enhancing ECG delineation performance.

Abstract

Electrocardiogram (ECG) delineation plays a crucial role in assisting cardiologists with accurate diagnoses. Prior research studies have explored various methods, including the application of deep learning techniques, to achieve precise delineation. However, existing approaches face limitations primarily related to dataset size and robustness. In this paper, we introduce a dataset for ECG delineation and propose a novel self-trained method aimed at leveraging a vast amount of unlabeled ECG data. Our approach involves the pseudolabeling of unlabeled data using a neural network trained on our dataset. Subsequently, we train the model on the newly labeled samples to enhance the quality of delineation. We conduct experiments demonstrating that our dataset is a valuable resource for training robust models and that our proposed self-trained method improves the prediction quality of ECG delineation.

Figures (3)

• Figure 1: PQRST segment of an ECG signal.
• Figure 2: The architecture of ECG-CODE network.
• Figure 3: Example of ECG signal delineated by ECG-CODE trained from scratch of pseudolabeled PTB-XL and fine-tuned on ISP.