Abstract 4370585: AI-based prediction of mortality in patients with ventricular tachycardia

Background: While ventricular tachycardia (VT) is frequently treated with anti-arrhythmic drugs (AADs), little is known about the impact of temporal patterns in medication exposure on clinical outcomes in patients with VT. We developed a Transformer-based framework that integrates longitudinal AAD exposure with baseline demographics to estimate risk of mortality in patients with VT. Methods: The dataset comprises a pure ventricular-tachycardia (VT) cohort of 10,929 adults who were put on one or more anti-arrhythmic drugs (AAD) between 2010 and 2024, of whom 1,311 died within the observation period. The dataset was divided into training and testing (70:30) using patient-based stratified splitting. We analyzed input features including baseline covariates (age, BMI, sex, race, ethnicity) and time varying prescription records ( "drug ON" and "drug OFF" from the EHR). A Transformer architecture with dimension=64, heads=4 and layers=3 was trained using these step functions to simultaneously predict a) all-cause mortality and b) Cox log-hazard. Baseline covariates were one-hot encoded and added to the classification model after the Transformer encoder. Harrell's concordance-index over 5 years prior to death was used to create a temporal risk score. Gradient-based SHAP was computed on a balanced 200-patient test subset to obtain drug importances. Results: The classifier predicted patients who ultimately died ( median time = 570 days from VT diagnosis; IQR: 1392 days) with AUROC 0.86 (Fig A). Optimum Youden index-based thresholding yielded Sensitivity = 0.95, Specificity = 0.67, F1-Score =0.65 for classifying all-cause mortality. Only three drugs contributed to mortality prediction (carvedilol, amiodarone and metoprolol, cumulative importance ~ 85%, Fig B) with slight variations in relative importance over time up to the point of death. Conclusion: An event duration-aware Transformer was able to predict mortality in patients with VT within a clinically relevant time window from baseline variables and time-varying medication use, in a large registry. The model was also able to extract the relative importance of medications up to the fatal event. This approach may have relevance for personalized therapy including the institution of advanced therapies.