부산대학교 도서관

Introduction and Hypothesis: Despite successful tetralogy of Fallot repair (rTOF) in childhood, advancing age is associated with escalating risk of morbidity and mortality. Published risk scores for prediction of major adverse cardiovascular events (MACE) typically incorporate invasive and/or specialized measures. We hypothesized that an artificial intelligence (AI) model using an array of parameters available in routine clinical practice could successfully predict MACE in rTOF.Methods: Adults with rTOF were identified using institutional databases (the prospective observational Comprehensive Outcomes Registry Late After Tetralogy of Fallot Repair [CORRELATE] and the retrospective Toronto Outcomes Registry of Congenital Heart disease [TORCH]). A random forest AI model for prediction of MACE was trained on CORRELATE using repeated random sub-sampling validation. External validation was achieved using non-overlapping retrospective data from TORCH. The MACE composite outcome included all-cause mortality, resuscitated sudden death, sustained ventricular tachycardia (>30 seconds) or heart failure (hospital admission>24 hours).Results: In total, 804 subjects were studied and MACE was observed in 73 (9%) subjects. Patients were either included in the training dataset (n=312, 59% male, median age 32 years [IQR 23-43], median follow-up 6 years [IQR 4-7]) or the testing dataset (n=492, 58% male, median age 26 years [IQR 20-37], median follow up 10 years [IQR 6-10]). Variables incorporated into the AI model are shown (Figure 1). Predictive capacity of the AI models were similar in testing (AUC 0.81, CI 0.75-0.86) and training (AUC 0.88, CI 0.71-0.99) datasets with excellent receiver operating characteristics (Figure 1).Conclusions: An AI model based on routinely used and widely available clinical and imaging variables could successfully predict MACE in rTOF. Further study is required to determine the value of AI for risk management in rTOF.