Personalized Deep Networks for Enhanced ECG Segmentation.

Electrocardiography (ECG) plays a vital role in the diagnosis of cardiovascular diseases by analyzing the electrical activity of the heart. ECG semantic segmentation is a subfield focused on sample-wise delineation of ECG waveforms by assigning a physiological label to each time sample, enabling explicit estimation of clinically meaningful onset and offset boundaries. Recent advancements in deep learning have significantly improved ECG classification accuracy; however, the same has not yet been observed in automatic ECG segmentation. Existing models often lack explainability and adaptability to patient-specific variations, thereby reducing their generalizability. This study proposes a personalized deep neural network approach for enhanced ECG processing. This method incorporates convolutional neural networks (CNNs) and bidirectional long short-term memory (BiLSTM) networks, incorporating an attention mechanism to refine segmentation accuracy. A novel loss function is introduced to ensure smoother temporal transitions and better classification accuracy. The model was evaluated using the QT Database, demonstrating substantial improvements in P-wave and QRS delineation and in T-wave offset localization segmentation when fine-tuned for individual patients when fine-tuned and evaluated on held-out data from the same patient, demonstrating the benefit of intra-patient adaptation. Our results indicate that personalization improves delineation accuracy for challenging waveforms (notably P and T waves), supporting the potential of deep learning to better capture patient-specific morphology and providing a stronger basis for waveform-level, clinically interpretable ECG analysis.