Abstract 12780: Machine Learning Models Show That Global Longitudinal Strain is a Strong Predictor of Survival After Echocardiography That is Superior to Ejection Fraction

IntroductionGlobal longitudinal strain (GLS) is closely related to adverse clinical outcomes (such as mortality) and likely a more sensitive measure of cardiac function than left ventricular ejection fraction (LVEF). Recent work used machine learning to improve the prediction of mortality after echocardiography by combining common clinical variables (CV) with echocardiography-derived measures such as LVEF. We hypothesized that incorporation of GLS would further improve the ability of machine learning models to accurately predict all-cause mortality after echocardiography.MethodsWe randomly selected 1081 echocardiograms from 1051 patients who had clinically reported resting LVEF and either a recorded death within 5 years of echocardiography (469 studies) or a known living encounter 5 years after echocardiography (612 studies) in the Geisinger electronic health record (EHR). QLAB was used to calculate GLS from 4-chamber views. CV (age, sex, systolic and diastolic blood pressures, HDL, LDL, smoking and diabetes status) were extracted from the EHR. We predicted 1-year and 5-year all-cause mortality after echocardiography using a random forest classifier with multiple combinations of the following input featuresCV, LVEF and GLS. Model performance was evaluated using the mean area under the curve (AUC) across 10 cross-validation folds, and feature ranking was obtained.ResultsUsing GLS alone produced higher accuracy than using LVEF alone (Figure, p&lt0.01 for both survival durations). CV+GLS yielded the highest accuracies (AUC=0.85 and 0.8, respectively for 1 and 5-year mortality prediction), while adding LVEF did not improve the AUC. GLS was the most important variable for predicting mortality, while LVEF was on average the 7 most important variable.ConclusionsMachine learning models can utilize clinical variables and imaging-derived metrics such as global longitudinal strain to predict survival after echocardiography with high accuracy. Global longitudinal strain outperforms other variables, including LVEF, for predicting survival after echocardiography.