Unsupervised deep learning of electrocardiograms enables scalable human disease profiling.

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Informatics

Background:

• The 12-lead electrocardiogram (ECG) is a widely accessible and inexpensive diagnostic tool.
• The full potential of ECG for detecting diverse human diseases remains largely unexplored.

Purpose of the Study:

• To develop and evaluate a deep learning model for identifying associations between ECG data and a broad spectrum of diseases.
• To assess the efficacy of ECG-derived latent space representations in disease detection.

Main Methods:

• Development of a deep learning denoising autoencoder for ECG analysis.
• Systematic evaluation of associations between ECG encodings and approximately 1,600 Phecode-based diseases across three independent datasets.
• Meta-analysis of results to identify robust disease associations.

Main Results:

• The latent space ECG model identified significant associations with 645 prevalent and 606 incident diseases (Phecodes).
• Enriched associations were observed in circulatory, respiratory, and endocrine/metabolic disease categories.
• The model demonstrated superior disease discrimination compared to traditional ECG intervals and demographic factors (age, sex, race).

Conclusions:

• Deep learning analysis of ECG data can detect a wide range of human diseases beyond cardiac conditions.
• ECG latent space models offer a powerful, non-invasive approach for disease detection and individual profiling.
• This technology has the potential to expand the clinical utility of the ECG.