Authors: Prof. Dr. Murat Ersel , Assoc. Prof. Dr. Oktay Yildiz , Research Asistant Sumeyra Kus Ordu , Dr. Mehmet Ragip Aktemur

Abstract

Introduction

Machine learning (ML) models analyze various risk predictors to accurately estimate patients' risk by understanding nonlinear relationships and interactions among multiple variables. By many trials ML models  tested for prediction of mortality and MACE risk for chest pain patients within 30 days.

Aims & Objectives

In this study, we aimed to compare the prediction accuracies of HEART and TMACS scores and various ML methods for estimating the risk of mortality and MACE in patients presenting to the ED with chest pain.

Material & Methods

The total number was 888, of consecutive patients aged 18 and older, presented to the ED of Ege University Hospital, Izmir, Türkiye, with chest pain or symptoms indicative of ischemia, applying the exclusion criteria, 328 patients were excluded, resulting in a final sample size of 560 patients.

The data of 560 patients were used for both training and testing through K-Fold Cross Validation. For evaluation of the dataset, several ML methods were applied, including K-Nearest Neighbor (K-NN), Naive Bayes, Logistic Regression (LR), Support Vector Machine (SVM), and Decision Trees.

The performance of models developed using these five methods to calculate 30-day mortality and MACE risks.  

Results

The highest AUC value for prediction of 30-day MACE belonged to the LR ML model with (0,860), and followed by SVM (0,840), which are higher than HEART score AUC value (0,827) and similar to T-MACS scores (0,851).

Even for 30-day mortality prediction LR has the highest AUC value (0,950) (HEART: 0,929, T-MACS: 0,875) While HEART, K-NN and SCM have 100% recall for mortality, LR has a recall of 0,998, with its balanced structure and highest AUC value, LR seems to be most preferable ML algorithm by clinical use to predict 30-day mortality.

Conclusion

Our LR model showed better AUC values for 30-days mortality than HEART scores, and SVM model found to be better than all models and scoring systems to predict the 30-days MACE. This  suggest that these models could potentially be integrated into electronic patient records to better assess the risk of mortality and MACE. 

Keywords: Machine Learning, MACE, chest pain, emergency department