Neural Network-based ECG Delineation

Abstract

ECG delineation is crucial for assessing drug-induced proarrhythmic risks, but still heavily depends on expert cardiologists despite existing automated methods. Our study proposes a deep learning model, E-DelUnet, to accurately locate key ECG fiducial points: Ponset, QRSonset, QRSoffset, Tpeak, and Tend. E-DelUnet is an adapted U-Net without skip connections and up-sampling, improving speed, performance, and reducing overfitting. It processes 1.2s single-lead ECGs and outputs binary masks marking fiducials. We trained the model on 2,054 ECGs from Verapamil and Quinidine studies using 5-fold cross-validation and tested it on 13,085 ECGs from other drug studies. Compared to traditional U-Net, wavelet-based methods, and residual networks, E-DelUnet showed the best performance, closely matching cardiologist measurements (from 3.86 ± 2.9 ms for the QRSonset to 6.77 ± 7.41 ms for the Tpeak). This model holds promise for improving ECG analysis in drug safety assessments and clinical workflows.