AI analysis for ejection fraction estimation from 12-lead ECG.

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Informatics

Background:

• Heart failure (HF) is a major cause of cardiovascular mortality, with rising prevalence.
• Ejection fraction (EF) measurement is vital for HF diagnosis and management.
• Echocardiography is the gold standard but limited by cost and accessibility, unlike electrocardiography (ECG).

Purpose of the Study:

• To explore the potential of 12-lead ECG signals for estimating EF using machine learning (ML) and deep learning (DL) models.
• To evaluate AI model performance for EF estimation in the underrepresented rural Appalachian population.
• To assess the impact of diverse demographics on AI fairness and accuracy in cardiovascular health.

Main Methods:

• Utilized a 12-lead ECG dataset of 55,500 patients from West Virginia.
• Applied a range of AI algorithms, including Random Forest and Transformer-based DL models, for EF estimation.
• Analyzed model performance using various thresholds, single vs. multi-lead ECG signals, and conducted interpretability analysis.

Main Results:

• Deep learning algorithms achieved the highest performance, with an Area Under the Receiver Operating Characteristic Curve (AUROC) of approximately 0.86 for EF estimation from 12-lead ECG.
• Individual ECG leads were insufficient for accurate EF estimation.
• Specific combinations of ECG leads significantly enhanced classification performance.

Conclusions:

• DL models show significant promise for estimating EF from 12-lead ECG, providing a scalable solution for HF monitoring.
• AI model performance in diverse populations, such as rural Appalachia, is critical for equitable healthcare.
• Optimizing lead combinations in ECG analysis can improve AI-driven EF estimation accuracy.