The Performance of DeepSeek R1 and Gemini 3 in Complex Medical Scenarios: Comparative Study

Abstract

Background: Generative artificial intelligence models, especially reasoning large language models (LLMs), are gaining adoption in health care for diagnostic decision support and medical education. DeepSeek R1 is a reasoning LLM that generates extended chain-of-thought explanations to make its decision-making process more explicit. Traditional medical benchmarks often lack complexity and authenticity, motivating the adoption of scenario-rich datasets, such as the Massive Multitask Language Understanding Pro (MMLU-Pro) professional medicine subset, which provides multispecialty clinical vignettes for reasoning-centric evaluation. Objective: The objective of this study is to assess the diagnostic accuracy, reasoning quality, reasoning transparency, and practical usability of DeepSeek R1 and Gemini 3 Pro across closed- and open-ended clinical scenarios, with the intention of guiding their prospective application in practical clinical education and training. This evaluation was conducted by analyzing 162 diverse medical scenarios (both closed- and open-ended) from the MMLU-Pro health subset. Methods: In a 2-phase, dual-model evaluation, DeepSeek R1 and Gemini 3 Pro were applied to 162 matched clinical vignettes from the MMLU-Pro professional medicine subset spanning 21 specialties. Closed-ended, multiple-choice, and open-ended prompts were constructed for the same scenarios, and model outputs were coded for accuracy, reasoning steps, and citation behavior; descriptive statistics and the McNemar test were used to compare performance across formats. Results: DeepSeek R1 achieved an accuracy of 86.4% (140/162 scenarios) on closed-ended tasks and 80.9% (131/162) on open-ended questions across 162 clinical scenarios, indicating modest attenuation of performance when answer cues were removed. Gemini 3 Pro demonstrated 90.7% (147/162) closed-ended and 88.9% (144/162) open-ended accuracy on the same scenarios, showing a similar pattern of decreased performance without answer options. Error analysis indicated that incorrect answers typically involved longer reasoning chains, suggesting overthinking. In a structured review of open-ended responses, DeepSeek R1 produced an average of 18.7 (range 0-52) references per case, with 5.2 unrelated references and 13.1 (range 3-67) reasoning steps, whereas Gemini 3 Pro averaged 22.5 (range 12-50) references, 1.9 (range 0-8) unrelated references, and 4.4 (range 1-10) reasoning steps per case. Conclusions: DeepSeek R1 demonstrated moderate-to-excellent accuracy and reasoning in evaluating both closed- and open-ended medical scenarios. In parallel, Gemini 3 Pro showed broadly comparable but distinct performance and reasoning patterns. While the closed-ended format may inflate accuracy due to cueing, the open-ended evaluation yielded richer insights into the fidelity of reasoning. Side-by-side evaluation of two large reasoning models highlights the importance of format, specialty, and citation behavior when considering clinical and educational use. Continued validation across a wider range of specialties and real-world contexts will enhance the model's trustworthiness for diagnostic and teaching applications.

Ähnliche Arbeiten