Comparative feasibility of reasoning and non-reasoning large language models for gynecologic cancer emergency care

Recent advancements in large language models (LLMs) have led to increasing exploration of their clinical applications, particularly in enhancing decision-making and patient management. This study aimed to evaluate the feasibility of LLMs, specifically the non-reasoning Generative Pretrained Transformer (GPT)-4o and reasoning o3-mini-high, in supporting emergency care of patients with gynecologic cancer.

In this retrospective, single-center study, 15 real-world emergency cases in gynecologic oncology were selected. Two gynecologic oncology fellows, two obstetrics and gynecology residents, GPT-4o, and o3-mini-high assessed each case through four steps: providing differential diagnoses and suggesting relevant examinations; interpreting findings, establishing diagnoses, and proposing treatment; prescribing medical orders; and generating patient education materials (LLMs only). Responses were scored for relevance and speed. Paired tests with bootstrapping were used to estimate mean differences (MDs) and 95% confidence intervals (CIs).

Both LLMs completed the tasks significantly faster than physicians, with an average time reduction of approximately 300 s per model (P < 0.001). GPT-4o achieved higher total scores than physicians (MD, 3.55; 95% CI, 2.98-4.10; P < 0.001) and maintained superiority when speed metrics were excluded (MD, 1.27; 95% CI, 0.80-1.79; P < 0.001). o3-mini-high also outperformed physicians in total scores (MD, 3.05; 95% CI, 1.98-3.88; P < 0.001), but not with speed metrics. Satisfaction scores for LLM-generated management were 1.9/2.0 for GPT-4o and 1.8/2.0 for o3-mini-high.

Both GPT-4o and o3-mini-high are feasible tools for emergency care of patients with gynecologic cancer. GPT-4o may provide advantages, reflecting the pattern-based structure of emergency care in this domain. LLM selection should be based on the domain-specific medical knowledge required rather than on the reasoning status or model version. Further prospective multicenter studies are warranted to confirm our findings as well as clinical effectiveness of LLMs in gynecologic cancer emergency care.