Artificial Intelligence-Enabled ECG Analysis to Predict Incident Heart Failure.

Area of Science:

• Cardiology
• Artificial Intelligence
• Predictive Analytics

Background:

• Heart failure (HF) prediction is crucial for preventive strategies.
• Existing proprietary models show limited or inconsistent accuracy.
• There is a need for generalizable, publicly available AI tools for HF risk assessment.

Purpose of the Study:

• To develop and validate a publicly available convolutional neural network (CNN) for predicting incident HF risk.
• To utilize 12-lead ECG waveforms for HF risk prediction.
• To create an AI model named ECG-to-HF (ECG2HF).

Main Methods:

• Developed ECG2HF using data from 94,636 patients at Massachusetts General Hospital (MGH).
• Validated ECG2HF in three independent test sets (MGH, BWH, BIDMC) involving 93,868 individuals aged 30-79 without prior HF.
• Identified 10-year HF events using NLP and quantified model performance via AUC and net reclassification.

Main Results:

• ECG2HF demonstrated strong 10-year HF prediction in all test sets with AUCs ranging from 0.84 to 0.86.
• The model showed improved discrimination and net reclassification compared to the Pooled Cohorts Equations.
• Performance was consistent across large healthcare samples from the northeastern US.

Conclusions:

• ECG2HF is a validated, publicly available AI tool for predicting future HF risk from 12-lead ECGs.
• The model exhibits favorable and consistent performance across diverse patient populations.
• ECG2HF can facilitate efficient risk stratification for targeted HF prevention.