Predicting left ventricular Ejection Fraction using Artificial Intelligence (AI) to save costs by using ECG over echocardiography

Grade:

8

Teacher:

Tamara Van Sickle

Abstract Background: Echocardiography remains the gold standard for assessing left ventricular ejection fraction (LVEF), but it is resource-intensive and less accessible in low-resource settings. This study hypothesized that an artificial intelligence (AI) model trained on electrocardiogram (ECG)-derived numerical features—specifically QRS duration, QT interval, and heart rate—can accurately predict LVEF, enabling a low-cost alternative for cardiac function assessment. Methods: The AI model was trained and validated using the EchoNext_metadata_100k dataset (PhysioNet), a comprehensive database of 100,000 de-identified patient samples from Boston representing diverse demographics. Each entry pairs ECG-derived features (age, gender, race, QRS duration, QT interval, heart rate) with corresponding echocardiographic LVEF values. The model was trained on 80,000 samples and validated on 20,000 unseen cases. Predictive performance was evaluated using mean squared error (MSE) and correlation analyses, stratified by ECG parameter categories. Results: For the QRS duration, predictive clustering of actual versus AI-predicted LVEF demonstrated high accuracy, with narrow QRS (<90 ms) yielding superior performance (88% accuracy; MSE 159.5) compared with broader complexes (>120 ms; 84% accuracy; MSE 256.1). Similar findings were observed for QTc duration, where lower values (<450 ms) corresponded to stronger LVEF prediction (MSE 159.1), whereas prolonged intervals resulted in greater dispersion (MSE 256.7). The ventricular rate exhibited a more uniform distribution pattern, showing weaker correlation with LVEF, consistent with its relative independence from depolarization and repolarization characteristics. Median ventricular rate was 82 bpm. Aggregate model accuracy across all parameters approached 85%, with optimal performance near normal EF values (50–60%). Accuracy declined to approximately 77% in cases of severe systolic dysfunction, indicating reduced generalization in extreme phenotypes. Conclusions: This study demonstrates that ECG-derived numerical features can be effectively leveraged by AI models to predict LVEF with clinically meaningful accuracy. The model captures physiologic relationships between electrical conduction (QRS, QTc) and mechanical performance (LVEF), reflecting the principle that electrical dyssynchrony underpins mechanical inefficiency. By predicting echocardiographic parameters from inexpensive, readily available ECG data, this approach has the potential to expand cardiac function screening, particularly in underserved populations lacking echocardiography access.

Project presentation

Research paper