ChatGPT-4o assists emergency physicians in enhancing diagnostic accuracy for fever of unknown origin: retrospective analysis

OBJECTIVE: To evaluate ChatGPT-4o’s diagnostic accuracy for fever of unknown origin (FUO) compared to emergency physicians and assess its utility as an adjunctive diagnostic tool. METHODS: This retrospective analysis included 60 adult patients presenting to the emergency department (ED) with FUO (fever ≥ 38.3℃ for ≥ 3 weeks without diagnosis after initial evaluation). Only patients with a confirmed final discharge diagnosis were included; FUO cases remaining undiagnosed at discharge were excluded. ChatGPT-4o and emergency medicine (EM) physicians independently generated preliminary diagnoses and comprehensive diagnoses. Diagnostic accuracy was compared against final discharge diagnoses. EM physicians subsequently revised their diagnoses after reviewing ChatGPT-4o’s output. Statistical analysis employed Welch’s ANOVA. RESULTS: ChatGPT-4o significantly outperformed EM residents in both preliminary (70.0% vs. 46.11%; 95% CI, 57.6%-79.9% vs. 34.4%-59.3%, P = 0.008) and comprehensive diagnoses (75.0% vs. 55.0%, 95% CI, 62.9%-84.1% vs. 42.5%-66.9%, P = 0.002). While numerically higher than EM specialists in both preliminary (70.0% vs. 57.78%) and comprehensive diagnoses (75.0% vs. 68.89%), these differences did not demonstrate consistent statistical superiority. Incorporating ChatGPT-4o’s suggestions significantly improved accuracy for both EM residents (preliminary: 67.78% vs. 46.11%, 95% CI, 55.8%-78.6% vs. 34.4%-59.3%, P = 0.002; comprehensive: 78.33% vs. 55.0%, 95% CI, 66.8%-86.6% vs. 42.5%-66.9%, P = 0.01) and specialists (preliminary: 70.22% vs. 57.78%, P = 0.023), though the improvement in specialists’ comprehensive diagnoses remained non-significant (80.89% vs. 66.67%, P = 0.076). Stratified analysis showed that ChatGPT-4o significantly improved diagnostic accuracy across major etiologic categories, including infectious (76.0% vs. 59.4%, P = 0.003) and cancer-related causes (72.2% vs. 50.0%, P < 0.001). CONCLUSIONS: ChatGPT-4o demonstrates potential to augment FUO diagnosis, particularly aiding less experienced clinicians. While this study highlights AI’s complementary value, prospective trials are needed to validate its impact on clinical efficiency.