Abstract 212: Artificial Intelligence To Predict Late Onset Cardiomyopathy Among Childhood Cancer Survivors Using Electrocardiogram, Echocardiogram, And Clinical Data

Abstract

Objectives: Applying machine learning to predict the 10-year cardiomyopathy risk among adult survivors of childhood cancer. Methods: The St. Jude Lifetime Cohort Study (SJLIFE) is an ongoing study of adult survivors of childhood cancer with in-person clinical evaluations. ECG and ECHO data were obtained on participants who did not have cardiomyopathy (defined as left ventricular ejection fraction < 50% or absolute drop from baseline ≥ 10%) at initial SJLIFE evaluation and who had at least two evaluations with ECHO screening. Predictors including ECG-AI based risk (cardiomyopathy risk from our previously develop AI model using ECG-alone), ECHO parameters, age, race, sex, BMI, primary cancer diagnosis, heart rate, respiratory rate, systolic and diastolic blood pressures, smoking, diabetes, hypertriglyceridemia, hypertension, or hypercholesterolemia. Light gradient boosting machines (LGBM) were applied to predict risk for cardiomyopathy using training (60%) and validation (20%) data. The hold-out (20%) data was used to evaluate predictive performance as sensitivity, specificity, precision, negative predictive value (NPV) and AUC. Shapley Additive exPlanations method (SHAP) was used to uncover predictors’ importance. Findings: The study population included 1,046 survivors with mean age±std of 32.2±8.0. The median time between baseline and follow-up evaluations was 5.2 years (0.5-9.5). In the hold out set, a model using ECHO parameters-only yielded an AUC of 0.68 (CI: 0.55-0.81), ECG-AI based risk alone yielded a hold-out test AUC of 0.87 (CI: 0.79-0.95), and an AUC of 0.89 (CI: 0.82-0.96) was obtained on LGBM model when ECG, ECHO, and clinical risk factors were used together. The final model classified cardiomyopathy with 80% sensitivity, 82% specificity, precision of 32%, and NPV of 98%. SHAP results showed that ECG-AI based risk, LV Global Longitudinal Peak Strain, LV Diastolic Volume, LV systolic volume, and cumulative anthracycline dose (>200 mg/m 2 ) were associated with an increased risk, and LVOT Peak gradient was associated with decreased risk. Conclusion: Artificial intelligence-based analyses with ECG and ECHO images may aid in cardiomyopathy risk prediction for at-risk survivors of childhood cancer.