Simple models vs. deep learning in detecting low ejection fraction from the electrocardiogram.

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• Deep learning models show high accuracy in detecting left ventricular systolic dysfunction (LVSD) from electrocardiogram (ECG) waveforms.
• However, deep learning models present challenges in clinical interpretability and broad deployment.

Purpose of the Study:

• To evaluate if simpler models using standard ECG measurements can achieve comparable accuracy to deep learning models in detecting LVSD.
• To assess the clinical utility and portability of simpler ECG-based models.

Main Methods:

• Trained and validated various models on a large dataset (40,994) of matched 12-lead ECGs and echocardiograms from Stanford University Medical Center.
• Included external validation datasets from Columbia Medical Center and UK Biobank.
• Compared performance using area under the receiver operator characteristic curve (AUC), including random forest, logistic regression, and deep learning models.

Main Results:

• A random forest model with 555 measurements achieved an AUC of 0.92, comparable to a deep learning model's AUC of 0.94.
• A logistic regression model using five measurements demonstrated high performance (AUC 0.86), outperforming NT-proBNP.
• Simpler models exhibited greater portability across external validation sites.

Conclusions:

• Simple electrocardiographic models offer a valuable alternative to deep learning for LVSD detection.
• These models achieve high accuracy while being significantly easier to implement and interpret in clinical settings.