CE-452775-2 MARS-HCM: MULTI-MODAL DEEP LEARNING METHOD FOR VENTRICULAR ARRHYTHMIA (VA) RISK STRATIFICATION IN HYPERTROPHIC CARDIOMYOPATHY (HCM) PATIENTS

Abstract

Current clinical guidelines by the AHA/ACC and ESC remain inadequate for accurately identifying HCM patients at risk for sudden cardiac death (SCD) by VA. Although risk factors for SCD in HCM patients have been established previously, including demographics, lab tests, late gadolinium-enhanced magnetic resonance imaging (LGE-MRI) features, it is challenging to integrate these distinct data types in a predictive model. Whether deep learning can integrate clinical covariates and LGE-MRI images in a multi-modal predictor that stratifies VA risk in HCM patients better than current clinical guidelines has not been explored. To develop a multi-modal deep learning approach that stratifies the risk of VA (sustained VT or VF) in HCM patients using raw LGE-MRI images and clinical covariates. In a single-center retrospective study, LGE-MRI images and electronic health record (EHR) data (demographics, clinical risk factors, echocardiogram features, LGE-MRI features) were acquired from 819 HCM patients. We developed a deep learning model with two input branches for images and EHR. In the image branch, the left ventricle (LV) was automatically segmented from LGE-MRI images, and the LV with raw image intensities was fed into a 3D ResNet. We adopted a feedforward network for the EHR branch to learn patterns from clinical covariates. In the output, the model was trained to synergistically fuse knowledge from both branches for VA prediction. In a 5-fold cross-validation, the multi-modal model using both LGE-MRI and EHR identified HCM patients with VA with an area under the ROC curve (AUC) of 0.83 (Figure), outperforming single-modal models using either LGE-MRI (AUC=0.80) or EHR (AUC=0.79). As a comparison, AHA/ACC guidelines had an AUC of 0.65 with a sensitivity of 0.81 and specificity of 0.44; ESC guidelines had an AUC of 0.69 with a sensitivity of 0.94 and specificity of 0.22. While preserving the same level of sensitivity as AHA/ACC guideline, the multi-modal model achieved higher specificity, 0.66. The VA risk score by the multi-modal model also showed a higher correlation coefficient (r=0.16) with other future adverse events (ICD shock, death, etc.) than risk scores by AHA/ACC (r=0.06) and ESC (r=0.01); see Figure. Our multi-modal deep learning significantly improved the prediction of VA in HCM patients compared with single-modal models and current clinical guidelines. Our model has the potential to improve the precision of VA risk stratification in HCM patients.