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

By Hagai Hamami, Yosef Solewicz, Daniel Zur, Yonatan Kleerekoper and Joachim A. Behar

URL: https://arxiv.org/abs/2506.11238

Premature Ventricular Contractions (PVCs) are among the most common abnormal heartbeats, yet they are notoriously difficult to detect automatically. Their appearance shifts across patients, lead configurations, sensor types, and recording environments. This variability has long prevented PVC detection algorithms from reaching the level of robustness required for real world clinical use.

Our lab is excited to introduce uPVC-Net, a deep learning model designed specifically to overcome this challenge and deliver reliable PVC detection across a wide range of ECG devices and populations.

Why Conventional PVC Detection Models Fall Short

Many existing models perform well on the dataset they were trained on but struggle when evaluated on new data. This happens because PVC morphology can change significantly depending on:

• lead placement

• recording hardware

• patient demographics

• noise and motion artifacts

A system that only learns from one source is unlikely to generalize well. To build a universal model, the training process must reflect the diversity of real world ECG data.

Our Approach: uPVC Net

1. Multi Source, Multi Lead Training

uPVC-Net is trained using data from multiple independent ECG datasets collected across different countries, devices, and populations. During training, one dataset is always held out and used only for evaluation. This forces the model to learn patterns that are stable across domains rather than overfitting to a single source. This training strategy proved essential for achieving strong generalization across completely unseen ECG distributions.

2. A Lightweight and Efficient Architecture

uPVC-Net is built around a Bidirectional GRU backbone. This architecture captures temporal information from the ECG segment while remaining compact and computationally efficient. It operates on an eight second ECG window that is converted into a time frequency representation, allowing the model to detect PVC related morphological features without needing a large or complex network. Because of its small size, uPVC-Net is well suited for deployment on wearable ECG patches and other resource limited devices.

Performance Across Diverse ECG Sources

uPVC-Net was evaluated across four large scale ECG datasets representing different countries, device types, and lead configurations. The model delivered consistently strong performance across all of them, demonstrating its ability to detect PVCs even when confronted with unfamiliar signal characteristics.

Importantly, the model excelled not only on classical Holter recordings but also on data from a modern single lead wearable device. This highlights its suitability for real time monitoring, long term patient assessment, and consumer grade cardiac sensing.

Comparison to Prior Methods

We compared uPVC-Net to a widely cited deep learning model for PVC detection. uPVC-Net achieved higher accuracy, sensitivity, specificity, and F1 score while using significantly fewer computational resources. The simplicity of the architecture is a key strength, reducing inference cost and enabling broader deployment.

Understanding Error Patterns

Our team carefully examined instances where the model produced false positives or false negatives. Many misclassifications were associated with beat types that resemble PVCs or are difficult even for human annotators to distinguish, such as fusion beats or ventricular escape beats. In several cases the reference labels were likely incorrect.

This analysis suggests future improvements may come from targeted augmentation of rare and ambiguous beat types, as well as continued refinement of training data quality.

Why This Work Matters

uPVC-Net demonstrates a path toward creating universal, generalizable AI models for physiological signal analysis. Instead of tailoring a model to a single dataset or device, we focused on building a system that performs reliably across populations, hardware, and lead configurations.

This capability is essential for real world deployment. A universal PVC detector can support:

• continuous cardiac monitoring via wearable ECG patches

• early identification of high PVC burden

• prevention of PVC induced cardiomyopathy

• large scale arrhythmia research

• integration into clinical workflows without device specific retraining

Our work shows that high performance and broad generalization can be achieved simultaneously with the right combination of training strategy and model design.

Looking Ahead

Future directions for uPVC-Net include exploring simpler or more efficient recurrent architectures, incorporating attention mechanisms, expanding the diversity of training datasets, and evaluating performance under varying noise levels. We are also interested in integrating foundation model pretraining techniques, which have shown promise in other ECG applications.

Most importantly, we aim to bring uPVC-Net closer to real world deployment in wearable and clinical systems where robust PVC detection can meaningfully improve patient care.