Clinical Electrophysiology -> SCA Risk Assessment: -> Clinical Risk Assessment Techniques D-PO04 - Poster Session IV (ID 15) Poster

D-PO04-216 - Computational Heart And Artificial Intelligence (CHAI): A Novel Methodology For Ventricular Tachycardia (VT) Risk Stratification In Patients With Cardiac Sarcoidosis (CS) (ID 1238)

Disclosure
 J.K. Shade: Nothing relevant to disclose.

Abstract

Background: While VT in CS is associated with the amounts and locations of active granulomatous inflammation and diffuse fibrosis in the myocardium, these characteristics are insufficient for clinical VT risk stratification. Here we propose a novel VT risk stratification methodology, the CHAI approach, which uses artificial intelligence (AI) to build a VT risk predictor from assessments of arrhythmogenicity in virtual-heart models as well as imaging and clinical biomarkers.
Objective: To determine if the CHAI approach provides superior VT risk assessment.
Methods: In a retrospective study of 40 CS patients, personalized virtual heart models were generated based on fusion of MRI and PET and simulations were performed in each model. 34 patients were randomly selected to train and cross-validate a support vector machine with 3 types of inputs: characteristics of arrhythmogenicity in the CS virtual heart, imaging features characterizing the amount and heterogeneity of CS remodeling, and baseline clinical data. The remaining patients were used for testing.
Results: CHAI predicted VT with validation AUC=0.85 (Fig.1A), sensitivity=83%, and specificity=82%. The CHAI validation accuracy was greater than the accuracies of virtual heart simulations (Sim) alone (p=0.05), MRI (p=0.002), PET (p=0.01), and LVEF (p=0.02, Fig.1B). CHAI and Sim achieved the highest test accuracies and had statistically significant odds ratios (Fig.1C).
Conclusion: The CHAI approach provided superior risk stratification to clinical metrics and achieved strong generalizability even with a small training data set. This is the first use of AI and virtual heart modeling together for VT risk prediction.
Collapse