MSc Thesis Defense: Mansur Kiraz, EXPLAINABLE ACS PREDICTION AT THE EMERGENCY DEPARTMENT FROM ROUTINE TESTS AND VITALS

EXPLAINABLE ACS PREDICTION AT THE EMERGENCY DEPARTMENT FROM ROUTINE TESTS AND VITALS

Mansur Kiraz
Data Science, MSc. Thesis, 2025

Thesis Jury

Assoc. Prof. Öznur TAŞTAN (Thesis Supervisor)

Asst. Prof. Onur VAROL

 Asst. Prof. Gülden OLGUN

Date & Time: December 19th, 2025 – 3:30 PM

Place: FENS G029

Keywords: Acute Coronary Syndrome, Machine Learning, Explainable AI, NSTEMI, Risk Stratification

Abstract

Acute coronary syndrome (ACS) is a major cause of mortality worldwide. Rapid diagnosis in emergency departments is challenging yet critical. Decisions rely on ECG interpretation, serial troponin testing, and scoring systems such as HEART, TIMI, and GRACE, which incorporate subjective clinical judgment and patient-reported history---information not always available in high-volume settings. This challenge is most acute for NSTEMI, which lacks clear early ECG findings, leading to missed diagnoses and unnecessary admissions. This thesis develops an explainable machine learning framework for NSTEMI risk assessment using only objective data collected at emergency department arrival. We analyzed approximately 5,000 patients with suspected ACS from Şişli Hamidiye Etfal Training and Research Hospital. The dataset includes vital signs, laboratory tests, ECG parameters, and expert-verified outcome labels. Models were trained to distinguish high- from low-risk patients without requiring patient history or serial measurements, comparing clinical-only models against ECG-augmented variants. The clinical-only LightGBM model achieved AUC-ROC of 0.91, compared to 0.81 for troponin threshold and 0.60 for GRACE score on the same cohort. Adding ECG features provided only marginal improvement, indicating that routinely available vital signs and laboratory values capture most predictive signal. Explainable AI analyses (SHAP, counterfactuals) identified troponin as the dominant predictor, with age, renal function, and hemodynamic markers providing secondary stratification. These findings demonstrate that an explainable, objective-data-only model can deliver transparent NSTEMI risk assessment without serial testing or subjective judgment.