Predicting Survival From Large Echocardiography and Electronic Health Record Datasets: Optimization With Machine...

Area of Science:

• Cardiology
• Medical Informatics
• Machine Learning

Background:

• Echocardiography is crucial for predicting patient survival, typically relying on ejection fraction (EF) and comorbidities.
• Electronic health records and additional echocardiographic measurements may offer enhanced predictive insights beyond current clinical standards.

Purpose of the Study:

• To develop and evaluate machine learning models for more accurate survival prediction post-echocardiography.
• To compare the predictive performance of nonlinear machine learning models against traditional logistic regression and clinical scoring systems.

Main Methods:

• A large dataset of 171,510 patients and 331,317 echocardiograms was analyzed.
• Machine learning models (including random forest) were trained using clinical variables, EF, and 57 echocardiographic measurements.
• Model performance was assessed using the area under the curve (AUC) across various survival durations and validated through cross-validation.

Main Results:

• Machine learning models demonstrated superior prediction accuracy (AUC > 0.82) compared to clinical risk scores (AUC = 0.61–0.79).
• Nonlinear random forest models outperformed logistic regression, with the model incorporating all echocardiographic data achieving the highest accuracy.
• A small subset of 10 variables, including 6 echocardiographic measures, achieved 96% of maximum predictive accuracy; tricuspid regurgitation velocity was a key predictor.

Conclusions:

• Machine learning effectively leverages diverse data inputs for superior survival prediction after echocardiography.
• The study highlights the potential of machine learning to refine prognostic accuracy in cardiology by integrating comprehensive echocardiographic and clinical data.