Decoding 2.3 million ECGs: interpretable deep learning for advancing cardiovascular diagnosis and mortality risk...

Area of Science:

• Cardiology
• Medical Artificial Intelligence
• Computational Biology

Background:

• Electrocardiograms (ECGs) are crucial for cardiovascular disease evaluation.
• The application of artificial intelligence (AI) in interpreting ECGs for novel clinical insights remains underexplored.
• AI holds potential for enhancing diagnostic accuracy and risk stratification using ECG data.

Purpose of the Study:

• To develop and validate a deep-learning model for fine-grained ECG interpretation.
• To advance cardiovascular diagnosis, mortality risk stratification, and identify new clinical information from ECGs.
• To explore AI's capability in gender identification and hypertension screening using ECGs.

Main Methods:

• A large dataset of 2,322,513 ECGs from 1,558,772 patients was utilized.
• A deep-learning model was developed for interpretable diagnosis, gender identification, and hypertension screening.
• Model performance was evaluated using Area Under the Receiver Operating Characteristic Curve (AUC) and hazard ratios (HR).

Main Results:

• The AI model achieved high AUC scores for diagnostic tasks (up to 0.998).
• ECG-predicted results identified high mortality risks for sinus tachycardia and atrial fibrillation.
• The V1 ECG lead was found crucial for hypertension screening and mortality risk stratification in hypertensive patients.

Conclusions:

• The AI model demonstrated cardiologist-level accuracy in cardiac diagnosis and mortality risk stratification using ECGs alone.
• The model has the potential to facilitate clinical knowledge discovery for gender and hypertension detection.
• This AI approach advances cardiovascular diagnostics and risk assessment beyond traditional methods.