ECG Machine Learning
Machine learning for electrocardiograms—supervised baselines, self-supervised contrastive pretraining, and transformer models for cardiac phenotypes, diagnostic support, and clinically reliable prediction with limited labels.
The electrocardiogram is a routine clinical test, but it also contains latent information that extends well beyond conventional interpretation. In this project, we develop machine learning methods that use ECGs to predict clinically important cardiac phenotypes, support diagnostic decision-making, and uncover signal patterns that are difficult to capture with handcrafted features alone. Our work treats ECG analysis not just as a classification problem, but as a broader challenge in building models that are clinically meaningful, robust across populations, and reliable enough for real-world use.
Technically, we combine several modeling directions. We study strong supervised baselines using open-source architectures, develop self-supervised contrastive pretraining to learn transferable ECG representations from large unlabeled cohorts, and design transformer-based models that better handle multi-scale temporal structure and inter-lead dependencies. These include grouped-lead encoders, hierarchical transformers, and dynamic feature aggregation across receptive fields. Across all of these directions, the central goal is to build ECG models that are accurate with limited labels, adaptable across tasks, and more informative than standard task-specific pipelines.
