A Novel R-Peak Detection Model and SE-ResNet-Based PVC Recognition for 12-Lead ECGs.

The real-time and accurate recognition of premature ventricular contractions (PVC) in dynamic 12-lead ECGs poses a clinical challenge due to noise and variability. The accurate location of the QRS complex is crucial for efficient PVC heartbeat recognition. This study proposes a robust PVC recognition approach, combining a self-adaptive multi-detector fusion model for R-peak detection and a multi-parameter squeeze–excitation ResNet-based heartbeat classifier. The detection results of multiple detectors are weighted with coefficients, and decision fusion is performed through adaptive threshold comparison. Tested on the INCART arrhythmia and 2018 China Physiological Signal Challenge databases, the R-peak detection results exhibit that our proposed fusion model outperforms majority, mean, and median voting strategies, with sensitivity improvements of 0.33%, 0.78%, and 0.41% for INCART dataset and 0.28%, 0.61%, and 0.34% for 2018 Physiological Signal dataset. In addition, our model is also superior to the best single annotator used in this paper. Evaluation of the multi-parameter SE-ResNet classifier reveals increased F1 scores of PVC heartbeat recognition by 4.33%, 3.44%, 5.04%, 12.41%, and 1.56% using INCART dataset compared to CNN, Inception, MLP, AlexNet, and LSTM, respectively, and 2.51%, 1.8%, 2.32%, 12.54%, and 2.27% using 2018 Physiological Signal dataset. Finally, the Se and Sp metrics also show improvement on the two datasets using the SE-ResNet. [ABSTRACT FROM AUTHOR]