uPVC-Net: A Universal Premature Ventricular Contraction Detection Deep Learning Algorithm

TL;DR

This work tackles the challenge of reliably detecting premature ventricular contractions (PVCs) from single-lead ECGs across diverse sensors and populations. The authors introduce uPVC-Net, a universal detector built on a Bi-GRU backbone that processes 8-second Mel-spectrogram segments and is trained with a multi-source, multi-lead strategy to bolster out-of-distribution generalization. Evaluated on four independent datasets totaling 8.3 million beats, uPVC-Net achieves high AUROC values (97.8%–99.1%), with wearables reaching 99.1%, demonstrating robust performance across lead configurations and devices. The study also provides ablation, error analyses, and a benchmark comparison, highlighting practical implications for continuous PVC monitoring in wearables and real-world clinical workflows.

Abstract

Introduction: Premature Ventricular Contractions (PVCs) are common cardiac arrhythmias originating from the ventricles. Accurate detection remains challenging due to variability in electrocardiogram (ECG) waveforms caused by differences in lead placement, recording conditions, and population demographics. Methods: We developed uPVC-Net, a universal deep learning model to detect PVCs from any single-lead ECG recordings. The model is developed on four independent ECG datasets comprising a total of 8.3 million beats collected from Holter monitors and a modern wearable ECG patch. uPVC-Net employs a custom architecture and a multi-source, multi-lead training strategy. For each experiment, one dataset is held out to evaluate out-of-distribution (OOD) generalization. Results: uPVC-Net achieved an AUC between 97.8% and 99.1% on the held-out datasets. Notably, performance on wearable single-lead ECG data reached an AUC of 99.1%. Conclusion: uPVC-Net exhibits strong generalization across diverse lead configurations and populations, highlighting its potential for robust, real-world clinical deployment.

Figures (5)

• Figure 1: Research overview. (a) ECG recordings from both traditional Holter monitors and modern patch-based ECG sensors are included in the study. In total, four independent datasets from diverse geographical regions are utilized, comprising 8.3 million beats. (b) To train uPVC-Net, an 8-second ECG strip centered around a reference annotation serves as input to a deep neural network after being subjected to a series of transformations. A multi-source, multi-lead training strategy is employed to ensure robust model performance. (c) Out-of-distribution generalization performance of uPVC-Net is evaluated across all included datasets and compared against the benchmark.
• Figure 2: Receiver operating characteristic (ROC) curves for each channel in each dataset.
• Figure 3: Six 8-second ECG segments centered around a FN beat (marked by a red $V$) with the lowest predicted PVC probability by uPVC-Net, among all test samples in the first channel of the MIT-BIH dataset. (a) Nonsustained ventricular tachycardia. (b) Fusion beat. (c) Artifact. (d) Ventricular escape beat. (e + f) Bigeminy with LBBB.
• Figure 4: Training strategy. uPVC-Net was successively retrained using either $n$ training examples randomly drawn from a single source domain, or a total of $n$ training examples randomly drawn from three source domains. OOD-GP is then evaluated over both channels of MIT-BIH. For the single source training strategy the median AUROC over models trained on one dataset among Icentia11k, INCART or CPSC2021 is reported.
• Figure 5: Ablation study. uPVC-Net was successively retrained using the same multi-source multi-lead training strategy, while first omitting the band-pass filter, then also omitting the Bi-GRU layer. OOD-GP is displayed for the left-out dataset.