Evaluation of ChatGPT’s Potential for Predicting Mortality Among Patients with Sepsis in the Emergency Department at Warinchamrap Hospital, Ubon Ratchathani Province

Abstract

Background: Sepsis is a medical emergency associated with high mortality. Early prognostication is important for treatment planning. However, evidence regarding the potential of ChatGPT for mortality prediction among patients with sepsis in real-world clinical practice in Thailand remains limited.

Objective: To evaluate the potential of ChatGPT in predicting 7-day and 28-day mortality among patients with sepsis presenting to the emergency department.

Methods: This prospective study included 272 patients with sepsis presenting to the emergency department at Warinchamrap Hospital, Ubon Ratchathani Province, between January and May 2026. First-day clinical data were entered into ChatGPT (GPT-4) using a structured prompt to predict 7-day and 28-day mortality. Prediction outputs were converted into binary outcomes and compared with observed outcomes.

Results: Among 272 patients, 12 (4.4%) died within 7 days, while a total of 26 patients (9.6%) had died by 28 days. Patients who died were significantly older, had more comorbidities, and had greater disease severity. They had a higher prevalence of septic shock, lower systolic blood pressure, higher respiratory rate, lower oxygen saturation, more frequent vasopressor use, and greater intravenous fluid administration within the first hour than survivors (p<0.05). ChatGPT predicted 7-day mortality with a sensitivity of 83.3%, specificity of 53.5%, PPV of 7.6%, NPV of 98.6%, and an AUROC of 0.684 (95% CI: 0.57–0.79). For 28-day mortality, the sensitivity was 73.1%, specificity was 56.1%, PPV was 15.0%, NPV was 95.2%, and the AUROC was 0.646 (95% CI: 0.55–0.73).

Conclusions: ChatGPT demonstrated the ability to discriminate between survivors and non-survivors, with an AUROC of 0.684 for predicting 7-day mortality and 0.646 for predicting 28-day mortality.