Abstract 4369215: Generalizability of AI-based Cardiomyopathy Risk Prediction among Childhood Cancer Survivors

Abstract

Background: Children whose cancer is treated with anthracycline chemotherapy and/or chest directed radiation (RT) are at risk for premature cardiovascular disease including cardiomyopathy. Echocardiography screening is recommended every 2 to 5 years, depending on the cumulative dose of cardiotoxic treatment and modalities received. We previously developed and validated an AI model using ECG as a sole input (ECG-AI) that can predict 5-year risk for cardiomyopathy with moderate accuracy. Goal: The goal of this study was to compare ECG-AI accuracy to a baseline clinical model and assess whether incorporation of clinical variables increase accuracy. Methods: The original ECG-AI model (Model 1) was an attention-based encoder-decoder deep neural network using 10 second 12-lead ECGs as an input to predict 5-year cardiomyopathy risk. It was trained and validated on 80% of data from the St Jude Lifetime Cohort Study (SJLIFE) then it was tested internally on 20% holdout of SJLIFE and externally on the Dutch Childhood Cancer Survivor Study (DCCSS-LATER) cohort. Both SJLIFE and DCCSS are prospective cohorts of five-year survivors of childhood cancer. Most participants had exposure to prior cardiotoxic treatments such as anthracycline chemotherapy and/or chest RT. We built two additional models on the 20% SJLIFE holdout data of the previous study and validated models in DCCSS-LATER including clinical variables available in both cohorts. Model 2 used stepwise logistic regression including variables listed in Table 1. Model 3 used stepwise logistic regression incorporating the ECG-AI (Model 1) outcomes and the clinical variables in Table 1. Results: SJLIFE holdout data included 1,515 ECGs from 959 participants and DCCSS-LATER included 667 ECGs from 330 participants (Table 1). Results obtained from the models are reported in Table 2. Model 1 (ECG-AI) outperforms Model 2 (clinical model) in SJLIFE derivation cohort while Model 2 is failing to generalize to DCCSS-LATER cohort. Combining ECG-AI results with other clinical data result only in a marginal increase in accuracy. Conclusions: ECG-AI provides generalizable cardiomyopathy risk prediction relying solely on electrocardiogram as an input while additional clinical data resulted in marginal increase in accuracy. Future studies are needed to incorporate more comprehensive clinical and genetic risk factors to obtain higher accuracy.