Five-year cardiomyopathy risk prediction in survivors of childhood cancer using electrocardiogram.

Abstract

10022 Background: Childhood cancer survivors (CCS) face an increased risk of developing cardiomyopathy during adult life due to late onset of treatment (e.g. anthracyclines and chest directed radiation) associated cardiotoxicity. Early identification of survivors at elevated risk using low cost and easily accessible data modalities can identify survivors in need of echocardiographic screening. Our goal is to utilize 12-lead electrocardiogram (ECG) develop and externally validate an artificial intelligence model (ECG-AI) for prediction of five-year risk for cardiomyopathy among CCS. Methods: We developed three deep learning models including a modified ResNet (a convolutional neural network architecture), an encoder-attention network, and a dual attention network to predict cardiomyopathy risk from 10 second 12-lead ECGs. We used the St Jude Lifetime Cohort Study (SJLIFE) for model building using a 60/20/20 patient-level split for training, validation, and holdout testing. We evaluated model performance for all cardiomyopathy grades (Common Terminology Criteria for Adverse Events) and specifically grade 3 (severe) cases. The final model was externally validated in the Dutch Childhood Cancer Survivor Study (DCCSS-LATER) cohort. For the DCCSS-LATER cohort, we evaluated accuracy of ECG-AI for the five-year cardiomyopathy risk prediction. Results: SJLIFE analytical cohort included 7,632 ECGs from 4,795 unique participants with no cardiomyopathy. 228 participants developed cardiomyopathy at least one year after index ECG date. Participants were 79% white, 11% Black, and 49% male with mean age at ECG of 33±10 years. In SJLIFE holdout, the encoder-attention model achieved the highest performance (area under the receiver operating characteristic curve (AUC) 0.75 for all grades and 0.84 for ≥grade 3 cases). The modified ResNet and dual attention models achieved AUCs of 0.69 and 0.72, respectively. DCCSS-LATER data included 749 ECGs with from 330 unique patients (48% male, age at ECG of 28±10 years). 22 patients developed cardiomyopathy at least one year after index ECG date. The encoder-attention model achieved an AUC of 0.74 for 5-year cardiomyopathy risk prediction. We note that the cardiomyopathy grading was not available in DCCSS-LATER, this study instead used a broader cardiomyopathy diagnosis information. Conclusions: ECG-AI analysis of standard 10 second 12-lead ECGs can identify childhood cancer survivors at risk for future cardiomyopathy with moderate to high accuracy depending on the cardiomyopathy severity. Future studies will focus on improving accuracy by incorporating clinical data such as B-type natriuretic peptides, left ventricular ejection fraction.