Predicting the risk of in-hospital Asian NSTEMI patients using stacked ensemble learning

Abstract

Abstract Background NSTEMI affects 60–75% of patients with MI and has a variety of clinical manifestations and prognoses. Ensemble learning (EL) utilizes the machine learning (ML) models on a classification task to make more accurate predictions than any single ML model. It has also been demonstrated to be superior to conventional risk scores in predicting mortality risk. However, no studies on the application of EL to ASIAN NSTEMI patients have been published. Purpose To identify factors associated with Asian in-hospital NSTEMI patients using ML and to develop a stacking EL-based NSTEMI risk score that is specifically tailored to the Asian population. Methods From 2006 to 2016, 10464 data were collected from the Malaysian National Cardiovascular Disease Database (NCVD-ACS) registry on hospitalized multiethnic Asian patients admitted with NSTEMI. This study utilized 39 variables, including demographic, cardiovascular risk, medication, and clinical variables. For stacked EL model development, four algorithms, Support Vector Machine (SVM), Random Forest (RF), eXtreme Gradient Boosting (XGB), and Naive Bayes (NB), are used as base learners, and Generalized Linear Model (GLM) is used as the meta learner. Significant variables were chosen and ranked using ML variable importance with backward elimination. As a predictive performance metric, the area under the curve (AUC) was used. Using a validation dataset, algorithms were validated against the TIMI for NSTEMI, and the net reclassification index (NRI) was calculated. Results The stacked EL models developed using SVM feature selection (13 predictors, AUC = 0.823, CI: 0.779 - 0.867) outperforms the conventional risk score, TIMI (AUC = 0.563, CI: 0.515 - 0.610). RF contributes the most to the EL model predictions, followed by NB, XGB and SVM (Figure 1). Common predictors between SVM feature selection and TIMI are; age, hypertension and ECG-type depression. SVM feature selection also identified gender, MI history, chronic lung disease, heart rate, systolic blood pressure, Killip class, fasting blood glucose, LMWH, beta-blockers and angiotensin II receptor blocker as predictors that improve mortality prediction in ASIAN NSTEMI patients. Our findings indicate that the TIMI score underestimates the mortality risk of NSTEMI patients. Stacked EL model using selected predictors classified 47.4% of nonsurvivors as high risk (risk probabilities > 50%), whereas TIMI (score > 5) classified only 3.3% of nonsurvivors as high risk (Figure 2). The NRI of NSTEMI patients using the EL model (SVM selected var) resulted in an NRI of 79.8% with a p-value < 0.00001 when compared to the TIMI for NSTEMI risk score. Conclusions The stacked EL algorithm classifies ASIAN NSTEMI patients more accurately than the single ML model and TIMI score. A precise prognosis is useful for selecting the appropriate level of care and pharmacological or invasive treatment for patients with Asian NSTEMI.Relative influence of each ML modelsStacked EL performance on validation set