Enhancing prediction of cancer therapy-related cardiomyopathy from surveillance echocardiograms: A Children's Oncology Group (COG) report

10012 Background: Many childhood cancer patients receive cardiotoxic therapies and need surveillance for therapy-related cardiomyopathy (CM). (Echo)cardiography is used to screen for cardiac dysfunction but has relatively poor discriminatory ability to predict who will develop CM. We sought to identify combinations of echo parameters that can better predict which patients subsequently develop CM. Methods: Longitudinal echos (obtained per routine care) of patients diagnosed with cancer < 21y who subsequently met CM criteria (left ventricular [LV] ejection fraction ≤50% / fractional shortening ≤28%, ≥2 times, with ≥1 time after completion of cancer therapy) plus clinical data were collated from COG sites. Echos and data from patients without CM were also collated. All echos were centrally remeasured for 42 routinely reported parameters in blinded fashion. We applied least absolute shrinkage and selection operator (LASSO) to fit logistic models to identify the most influential predictors for CM development within 2 and 5y of echo, plus age and sex. Data were randomly split into training (85%) and test (15%) sets with 10-fold cross validation. Prediction accuracy was calculated using area under the ROC curve (AUC), based on the model that provided the minimum mean cross-validated error. Results: Echos from 88 CM cases (248 echos predating CM) and 126 non-cases (518 echos) were available. Patients (n = 214) were diagnosed at mean age 7.7±5.2y with mean 9.5±4.0y follow-up. Mean doxorubicin equivalent doses for cases and non-cases were 350±188 and 272±199 mg/m 2 , respectively. For 2y CM prediction, models achieved training AUC 0.85 (95%CI 0.78-0.92; 261 echos) and test AUC 0.74 (95% CI 0.74-0.90; 46 echos). Factors selected included age, 2-dimensional (2D) measurements of LV geometry (LV systolic and end-systolic dimensions, and posterior wall thickness), systolic function (M-mode fractional shortening), diastolic function (mitral inflow E wave, septal E’, septal A’), and a measurement of combined systolic and diastolic function (myocardial performance index). For 5y CM prediction, models achieved training AUC 0.90 (95%CI 0.85-0.95; 175 echos) and test AUC 0.89 (95%CI 0.78-0.99; 30 echos). Age, LV end-systolic dimension (2D and M-mode), M-mode LV posterior wall thickness, 2D wall thickness-dimension ratio, mitral inflow E, septal E’, septal A’, and the myocardial performance index were selected. At both times 2D end-systolic dimension was the most influential parameter. Inclusion of anthracycline and chest radiotherapy dose did not meaningfully improve the AUCs. Conclusions: Prediction models that incorporate conventional echo data may be able to accurately identify childhood cancer patients at high risk of developing CM 2-5y prior to CM diagnosis. This may provide a window of opportunity to introduce interventions that may arrest or slow CM progression.