GCTRNet: global context-gated temporal refinement for short-window ECG-based premature ventricular contraction detection

Background Premature ventricular contractions (PVCs) are one of the most prevalent ventricular arrhythmias and therefore require a method for reliably identifying PVCs from short ECG strips, through an automated system capable of identifying PVCs and evaluating the burden of PVCs. While technology continues to advance, automatic identification of PVCs from very short ECG recordings remains problematic due to significant variability in PVC morphology across various patients as a result of the limited amount of information available from ECG data. Methods To overcome this difficulty, the Global Context-Gated Temporal Refinement (GCTR) module has been created which enhances bidirectional long short-term memory (BLSTM) representations by incorporating segment-level contextual information using context gating and residual learning. The GCTR module was integrated into a convolutional network architecture termed GCTRNet. GCTRNet was evaluated on the MIT-BIH Arrhythmia Database using a patient-independent partition. The frozen model was further evaluated on a predefined subset of CPSC2021 comprising 210 recordings from 105 patients. Results The frozen seed-2024 model of GCTRNet, when tested on MITDB data, reached an accuracy of 99.65%, a sensitivity of 96.82% and a specificity of 99.91%. The same frozen model achieved performance on the CPSC2021 subset with 97.83% accuracy, 84.75% sensitivity, and 98.11% specificity and an AUPRC of 80.49%. In addition, GCTRNet outperformed matched comparison models based on BLSTM architecture in many metrics of evaluation, thus supporting the potential of the temporal refinement method proposed. Conclusions GCTRNet provides an effective temporal refinement approach for patient-independent short-window PVC detection and shows potential for cross-database application.