A deep learning phenome wide association study of the electrocardiogram

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Deep learning shows promise in electrocardiogram (ECG) disease detection.
• The full spectrum of detectable diseases and underlying ECG features remain unclear.

Purpose of the Study:

• To identify all cardiac and non-cardiac conditions detectable from ECG waveforms.
• To elucidate the specific ECG features contributing to disease classification.

Main Methods:

• Trained PheWASNet, a multi-task deep learning model, on large ECG and electronic health record datasets from two medical centers.
• The model was designed to detect 1243 distinct disease phenotypes from raw ECG data.

Main Results:

• Confirmed ECG detectability for chronic kidney disease (AUC=0.80), cirrhosis (AUC=0.80), and sepsis (AUC=0.84).
• Identified new detectable conditions: respiratory failure (AUC=0.86), neutropenia (AUC=0.83), and menstrual disorders (AUC=0.84).
• Found 35 of 37 non-cardiac conditions were detectable via models for just four diseases, indicating shared ECG effects.

Conclusions:

• Uncovered a broad range of ECG-detectable diseases, including novel phenotypes.
• Advanced the understanding of ECG features crucial for accurate disease detection.