Deep learning for improved prediction of late-onset cardiomyopathy among childhood cancer survivors: A report from the St. Jude Lifetime Cohort (SJLIFE).

Abstract

10545 Background: Early identification of survivors at high risk for treatment-induced cardiomyopathy may allow for prevention and/or early intervention. We utilized deep learning methods using COG guideline-recommended baseline electrocardiography (ECG) to improve prediction of future cardiomyopathy. Methods: SJLIFE is a cohort of 5-year clinically assessed childhood cancer survivors including baseline ECG measurements. Development of cardiomyopathy was identified from clinical and echocardiographic measurement using CTCAE criteria (grade 3-4). We applied deep learning approaches to ECG, treatment exposure and demographic data obtained at baseline SJLIFE assessment. We trained a cascaded model combining a 12-layer 1D convolutional neural network to extract features from waveform ECG signals with a 2-layer dense neural network to embed features from other phenotypic data in tabular format to determine if use of deep learning with ECG data could improve prediction of cardiomyopathy. Results: Among 1,218 subjects (median age 31.7 years, range 18.4-66.4) without cardiomyopathy at baseline evaluation, 616 (51%) were male, 1,041 (85%) white, 157 (13%) African American and 792 (65%) were survivors of lymphoma/leukemia. Follow-up averaged 5 (0.5 to 9) years from baseline examination. Mean chest radiation dose was 1350 cGy (range 0 to 6,200 cGy) and mean cumulative anthracycline dose was 191 mg/m2 (range o to 734 mg/m2). A total of 114 (9.4%) survivors developed cardiomyopathy after baseline. A cascaded deep learning model built on a training set (N = 974 participants) classified cardiomyopathy in the test set (N = 244 participants) using both clinical and ECG data with a sensitivity of 70%, specificity of 73%, and AUC of 0.74 (95% CI 0.63-0.85), compared to a model using clinical data alone (sensitivity 61%, specificity 62%, and AUC 0.67, 95% CI 0.56-0.79). In subgroup analyses, models predicting cardiomyopathy within 0-4 years following baseline had a sensitivity, specificity, and AUC of 77%, 78%, and 0.78 (0.65-0.91), respectively. When predicting cardiomyopathy 5-9 years following baseline, model performance dropped to a sensitivity, specificity, and AUC of 70%, 70%, and 0.68 (0.50-0.87), respectively. Conclusions: Deep learning using ECG at baseline evaluation significantly improved prediction of cardiomyopathy in childhood cancer survivors at high risk for cardiomyopathy. Future directions will incorporate deep learning approaches to echocardiography to further improve prediction.