PO-03-207 BUILDING A PREDICTION MODEL FOR SUDDEN CARDIAC DEATH IN A GENERAL POPULATION IN TAIWAN USING THE MACHINE LEARNING TECHNIQUE

Abstract

Predicting the risk of sudden cardiac death (SCD) is of paramount importance in preventive medicine. However, it remains a major challenge to prevent SCD in the general population. This study aimed to develop a prediction model for sudden cardiac deaths in the general population using the community-based cohort in Taiwan. Chin-Shan Community Cardiovascular Cohort (CCCC) enrolled participants ≥ 35 years of the age since 1990-1991, and participants were followed up until 2005. Participants without 12-lead ECG, echocardiography, and carotid artery duplex sonography data were excluded from this study. A total of 2193 subjects in the CCCC were analyzed (Cohort 1: whole CCCC). Among Cohort 1, 2105 participants without a prior history of CAD and heart failure (HF) with reduced ejection fraction (HFrEF: left ventricular [LVEF] < 35%) were also studied (Cohort 2). For Cohort 1 & Cohort 2, we randomly selected 55% subjects to be the training dataset, 30% as the validation dataset, and 15% as the test dataset. Feature importance in machine learning (ML) was applied to assign a score to input features based on how important they are in predicting SCD events. Four ML algorithms were compared, including: (1) XGboost; (2) random forests; (3) logistic regression; and (4) DNN (deep neural network). The receiver operating characteristic curves and the area under curves (AUC) were used to summarize the prediction performance. The cumulative incidence of SCD was 1.50% in Cohort 1. The AUC of the CCCC-SCD-Score (previously published by our research team) in predicting SCD risks was 0.888. Using the ML analysis, we established novel prediction models to predict SCD events. For Cohort 1 and Cohort 2, the AUC is within the range of 0.81-1.00 (using the algorithms of XGboost and random forests algorithms (Figure). The scores of feature importance for variable factors in Cohort 1 and Cohort 2 are shown in the Figure. For Cohort 1, coronary artery disease (CAD) diagnosed by electrocardiography (ECG) or history was the most significant determinant of SCD among all factors; For Cohort 2, CCCC-SCD scores established by our research team was the most significant determinant of SCD among all factors. Using the ML analysis, we established excellent prediction models for predicting SCD events of SCD via XGboost and random forests. Careful evaluation and management for subjects with risk factors of SCD may be useful in preventive medicine.