AKUT KORONER SENDROMLU HASTALARDA HASTANE İÇİ MORTALİTENİN MAKİNE ÖĞRENME YÖNTEMLERİ KULLANILARAK KESTİRİMİ

Abstract

Akturk, S., Prediction of In-Hospital Mortality in Patients with Acute Coronary Syndrome Using Machine Learning Methods, Hacettepe University Graduate School of Health Sciences, Biostatistics Program, Master's Thesis, Ankara, 2026. This study aimed to compare the performance of different machine learning algorithms in predicting in-hospital mortality in patients with acute coronary syndrome. The retrospective study included 657 patients, and the dataset was stratified according to mortality status, dividing it into 80% training (n=525) and 20% test (n=132) sets. Decision tree, random forest, XGBoost, logistic regression, Naive Bayes, support vector machines, and artificial neural networks methods were used to predict in-hospital mortality. Five-fold cross-validation, hyperparameter and Youden index-based threshold optimization were applied in model development; class imbalance was addressed with SMOTE. Variable significance was evaluated using a model-independent permutation-based approach. The median age of the patients was 62 years (IQR: 53–71), 72.8% were male, and 41.4% had STEMI, with an in-hospital mortality rate of 5.6%. Model performance varied depending on the applied modeling approach. In the baseline evaluation, logistic regression; following hyperparameter and threshold optimization, random forest; among the baseline models with SMOTE, the decision tree; and in the final stage, in which SMOTE, hyperparameter optimization, and threshold optimization were applied together, XGBoost were the models that stood out.In variable significance analyses, acute heart failure, acute kidney injury, leukocytes, and glucose were prominent variables. In conclusion, no single model consistently superior across all performance measures was identified; model performance varied depending on the modeling approach and evaluation criteria. The findings highlight the importance of evaluating model performance in clinical data with class imbalances using different criteria together.