Evaluating the predictive accuracy of ChatGPT in risk stratification for chest pain in the emergency department

Abstract

Chest pain is a frequent cause of emergency department (ED) visits, yet accurately assessing the risk of major adverse cardiac events (MACE) remains challenging. This study evaluated the potential of ChatGPT 4.0 as a clinical decision-support tool for predicting MACE in patients with chest pain. We conducted a prospective observational study at the Rovereto Hospital ED from March to August 2024, analyzing 178 patients. ChatGPT received patient data in three sequential phases: initial clinical history and ECG, first troponin test, and second troponin test. Its predictive performance improved with additional data, with the area under the receiver operating characteristic curve (AUROC) increasing from 0.699 in the first phase to 0.776 after the second troponin test (p=0.039). However, ChatGPT misclassified several MACE cases as non-urgent, raising concerns about sensitivity and risk stratification. While ChatGPT demonstrated potential in identifying MACE cases, its current performance does not support routine ED implementation. Further refinements in AI-based models are needed to improve real-time risk assessment and ensure safer, more reliable clinical integration.

Ähnliche Arbeiten