Abstract 4353134: Explainable Deep Learning Predicts Future Adverse Outcomes in Non-Ischemic Cardiomyopathy From Multi-domain Digital Health Data

Background: Despite recognized phenotypic heterogeneity of idiopathic non-ischemic cardiomyopathy (NICM), management remains generalized and dominantly guided by left ventricular (LV) ejection fraction (EF) and New York Heart Association class. Deep learning (DL) models can integrate complex, multimodal phenomics data to support individualized prognostication; however, are considered black box models that limit clinical adoption. We developed an explainable DL model, DeepPhenome-NICM, for patient-specific prediction of major adverse cardiac events (MACE) in NICM leveraging multi-domain phenomics data captured at time of cardiovascular MR (CMR) imaging. Methods: 1,142 patients with CMR confirmed diagnosis of NICM were identified from the Cardiovascular Imaging Registry of Calgary, defined as LV EF <50% in the absence of any identifiable ischaemic or non-ischaemic aetiology. All patients underwent baseline health questionnaires with standardized reporting at the time of CMR imaging and were followed for a minimum of 6 months for the composite outcome of all-cause mortality, survived cardiac arrest, ventricular tachycardia, or heart failure hospitalization. A total of 50 routinely captured variables were included in a final trained DL survival model (DeepPhenome-NICM), inclusive of patient-reported, CMR-derived, and electronic health record-derived variables. Data were split into training (80%) and test (20%) sets. Model performance was assessed on the test set. Shapley values, a measure of additive feature contribution to model prediction, were estimated to deliver model explainability. Results: Baseline characteristics of the study population are reported in Table 1. Over a median follow-up of 3.8 years, 210 patients (18.4%) experienced MACE. Using the hold-out test set, the DeepPhenome-NICM model achieved a mean time-dependent AUC of 0.83 (95% CI 0.75-0.89) with a 1- and 5-year AUC of 0.87 (0.79-0.93) and 0.82 (0.73-0.89), respectively. Stratification of patients by the median predicted patient-specific risk score yielded significant discrimination of event-free survival, with the high-risk group experiencing a 4.3-fold increased risk (HR; 95% CI 2.1-8.9; p<0.001; Figure 1) . Figure 2 shows the respective influence of top predictors on model prediction. Conclusions: DeepPhenome-NICM is a multimodal DL model that identifies high risk patients with NICM at time of CMR using a composite phenomics based approach. External validation of this model is planned.