Risk stratification of Asian patients after ST-elevation myocardial infarction using machine learning methods

Abstract Background Risk stratification in ST-elevation myocardial infarction (STEMI) that is population-specific is essential. Conventional risk stratification methods such Thrombolysis in Myocardial Infarction (TIMI) score is used to evaluate the risk associated with the acute coronary syndrome (ACS) which are derived from Western Caucasian cohort with a limited participant from the Asian region. In Malaysia, multi-ethnic developing country, patients presenting with STEMI are younger, have a much higher prevalence of diabetes, hypertension and renal failure, and present later to medical care than their western counterparts. Purpose We aim to investigate the predictors, predict mortality and develop a risk stratification tool for short and long term mortality in multi-ethnic STEMI patients using machine learning (ML) method. Methods We created three separate mortality prediction models using support vector machine (SVM) to identify predictors and predict mortality for in-hospital, 30-days and 1-year for STEMI patients. We used registry data from the National Cardiovascular Disease Database of 6299 patient's data for in-hospital, 3130 for 30-days and 2939 for 1-year for ML model development. Fifty parameters including demographics, cardiovascular risk, medications and clinical variables were utilised for training the models. The Area under the curve (AUC) was used as the primary performance evaluation metric. All models were validated against conventional method TIMI and tested using testing data. SVM variable importance method were used to select and rank important variables. We converted the final algorithm into an online tool with a database for continuous algorithm validation. We implemented the online calculator in selected hospitals for further testing using prospective patients data. Results The calculator is available at myheartstemi.uitm.edu.my. The calculator outperforms TIMI on testing data for in-hospital (15 predictors) (AUC=0.88 vs 0.81), 30 days (12 predictors) (AUC=0.90 vs 0.80) and 1-year (13 predictors) (AUC=0.84 vs 0.76). Common predictors for in-hospital, 30 days and 1-year mortality model identified in this study are; age, heart rate, Killip class, fasting blood glucose and diuretics. Invasive and less invasive treatments such as PCI pharmacotherapy drugs are also selected as important variables that improve mortality prediction. Our results also suggest that TIMI score underestimates patients risk of mortality. 90% of non-survival patients are classified as high risk (>30%) by the calculator compared 10–30% non-survival patients by TIMI. Conclusions In the multi-ethnicity population, patients with STEMI are better classified using ML method compared to the TIMI score. ML allows identification of distinct factors in unique ASIAN population for better mortality prediction. Availability of population-specific calculator and continuous testing and validation allows better risk stratification. Machine learning and TIMI performance Funding Acknowledgement Type of funding source: Public Institution(s). Main funding source(s): University of Malaya Grant