LAUSR

Abstract Funding Acknowledgements Type of funding sources: None. Introduction Electrocardiography (ECG) can be easily obtained at a low cost and includes voltage and time interval representing heart conditions. We hypothesized that artificial intelligence (AI) detects a subtle abnormality in 12-lead ECG and may predict individual mortality. Methods Among 502,411 population in UK Biobank, 42,096 individuals had 12-lead ECG from 2013 to 2022. Among population with available ECG, 4,512 individuals were enrolled in this study adjusting the following inclusion criteria; age under 60 years, sinus rhythm, PR interval 120~200ms, QTc interval 350~460ms, and QRS duration 70~100ms. We developed and tested convolutional neural network (CNN) model to predict all cause death, cardiovascular (CV) death, or sudden cardiac arrest (SCA). The study population were divided into train (80%), validation (10%), and test (20%) set. Results Among 4,512 patients with median 3.7 years [IQR; 2.7-5.1] of follow-up, the rate of all-cause mortality was 11.6% (524). In overall study population, median age was 55.5 years and proportion of male sex was 42.2%. The patients with all-cause death were older (p<0.001) and had more comorbidities (p<0.001). In the train set, CNN model showed 0.93 in AUC for predicting all-cause death. In the test set, CNN model showed consistent good performance power (AUC 0.90) for all-cause death. In subgroup analysis, 102 of 4153 (2.46%) and 57 of 4065 (1.40%) patients experienced CV death and SCA, respectively. The performance power in test set were 0.90 in AUC for CV death and 0.87 in AUC for SCA. Conclusions AI detects and predicts future all-cause death, CV death, and SCA in median of 2.6 years by analyzing standard 12-lead ECG in generally looking normal sinus rhythm. Performance power of CNN model