Large Language Model Performance on Multistep Clinical Cases: Comparative Study Across Question and Case Levels.

backgroundMost large language models (LLMs) have achieved passing scores on medical licensing examinations. However, most evaluations focus on single-question accuracy, overlooking performance on multistep patient management scenarios, such as making a diagnosis followed by a treatment plan. It is unclear if LLMs can maintain high performance on complete clinical cases.

objectiveThis study aimed to evaluate the performance of LLMs on multistep clinical cases and to investigate the impact of model size scaling and case complexity on performance stability.

methodsWe curated a dataset of 189 unique clinical cases comprising 473 individual questions from the Chinese National Medical Licensing Examination. The number of questions related to each case ranged from 2 to 4. A physician with more than 8 years of clinical experience annotated the cases, confirming that 82.5% (156/189) contained sequentially dependent questions. We tested 4 representative LLMs: DeepSeek-R1, GPT-4o, Gemini-2.5-Flash, and Qwen2.5. We measured 2 metrics: the question pass rate (QPR) and the case pass rate (CPR), where a case was considered correct only if the model answered all its questions correctly. The consistency gap was calculated as the difference between QPR and CPR. Additionally, we tested Qwen2.5 at different parameter sizes (3B, 7B, 14B, and 32B) to examine the effect of model size. We also calculated the expected CPR and used the McNemar test with Bonferroni correction to compare model performance.

resultsAll LLMs achieved a QPR exceeding 83%. However, the CPR was lower for all models. DeepSeek-R1 achieved the highest QPR at 89.9% (425/473), with a CPR of 79.9% (151/189), corresponding to the smallest consistency gap at 10%. In contrast, GPT-4o exhibited a QPR of 83.5% (395/473) and a CPR of 65.6% (124/189), resulting in the largest consistency gap at 17.9%. DeepSeek-R1 and Qwen2.5-32B significantly outperformed GPT-4o in both QPR and CPR (P<.05). Observed and expected CPR values were closely aligned across models. Increasing the model size from 3B to 32B reduced the consistency gap by approximately 50%. As the number of questions per case increased, the consistency gap tended to increase across all evaluated LLMs.

conclusionsCurrent LLMs exhibit high accuracy in answering individual questions, but their ability to correctly answer all questions within a complete clinical case is substantially lower. This study design evaluated independent question answering, with responses aggregated at the case level. The performance depends heavily on model scale and case complexity. LLMs should be used as support tools rather than independent decision-makers in medical education and clinical practice.