Predicting mortality of Malaysian patients with acute coronary syndrome (ACS) subtypes using machine learning and deep learning approaches / Muhammad Firdaus Aziz

Muhammad Firdaus , Aziz (2022) Predicting mortality of Malaysian patients with acute coronary syndrome (ACS) subtypes using machine learning and deep learning approaches / Muhammad Firdaus Aziz. PhD thesis, Universiti Malaya.

	PDF (The Candidate's Agreement) Restricted to Repository staff only Download (250Kb)
	PDF (Thesis PhD) Download (4051Kb)

Abstract

The conventional risk score for predicting short- and long-term mortality following Acute Coronary Syndrome (ACS) is typically not population-specific and does not accommodate for Asian patients. The purpose of this study is to use machine learning (ML) and deep learning (DL) algorithms to predict and identify variables linked to short and long-term mortality in Asian STEMI and NSTEMI/UA patients and to compare these results to a conventional risk score. Model development for STEMI: in-hospital (6299 patients), 30-days (3130 patients), and 1-year (2939 patients) and NSTEMI/UA: in-hospital (4771 patients), 30-days (2402 patients), and 1-year (2304 patients) datasets was done using the National Cardiovascular Disease Database (NCVD) Malaysia registry of a multi-ethnic, heterogeneous Asian ACS population. 50 variables were considered for STEMI and 39 for NSTEMI/UA. ML algorithms were used to examine significant variables utilising feature selection methods. The ML feature selection approach was then used to develop ML and DL models using all and selected variables, which were then compared to the Thrombolysis in Myocardial Infarction (TIMI) score. For STEMI patients, the best ML model, a Support Vector Machine (SVM) classifier with sequential backward elimination (SBE) selected variables, produced AUC values of 0.88 for in-hospital, 0.90 for 30 days, and 0.84 for 1 year, while the best model for NSTEMI/UA patients produced AUC values of 0.85 for in-hospital, 0.87 for 30 days, and 0.80 for 1-year mortality prediction. The same variables were then used to create the best DL model for STEMI (AUC 0.96 in-hospital, 0.93 for 30 days, and 0.90 for 1-year mortality prediction) and NSTEMI/UA (AUC 0.97 in-hospital, 0.91 for 30 days, and 0.88 for 1-year mortality prediction). TIMI risk score reported lower performance for STEMI (In-hospital: AUC=0.81, 30 days: AUC=0.80 and 1-year: AUC=0.76) and NSTEMI/UA patients (In-hospital: AUC=0.42, 30 days: AUC=0.49 and 1- year: AUC=0.42) as compared to ML and DL algorithms. Age, heart rate, Killip class, fasting blood glucose, and diuretics were found to be the common variables across the three time points in the STEMI dataset, whereas age, heart rate, Killip class, and intake of Low-molecular-weight heparin (LMWH) were found to be the common variables in the NSTEMI/UA dataset. When compared to the TIMI risk score, both ML and DL were better at classifying ACS patients in a multi-ethnic population. ML enables the identification of distinct variables in Asian populations to improve mortality prediction. In the future, continuous testing and validation will enable improved risk classification, possibly modifying management and results.

Actions (For repository staff only : Login required)