Comparison of DeepSeek-V3.2 and ChatGPT-5.2 in the outpatient diagnosis and treatment of prostate cancer: a multicenter evaluation of application value and multilingual adaptability

Le résumé fourni par la source

To conduct a multicenter evaluation comparing the diagnostic and therapeutic reasoning performance of two large language models (LLMs), DeepSeek-V3.2 and ChatGPT-5.2, across key prostate cancer (PCa) clinical scenarios, and to assess their multilingual adaptability in Chinese and English clinical environments. In this retrospective, multicenter study, 480 cases of pathologically confirmed PCa were enrolled from three tertiary hospitals between January 2023 and September 2025. Cases were stratified into four categories (n = 120 each): localized, treatment-naïve PCa (cT1-T2N0M0); post-radical treatment follow-up (after surgery or radiotherapy); metastatic hormone-sensitive PCa (mHSPC); and metastatic castration-resistant PCa (mCRPC). A centralized multidisciplinary team (MDT) developed unified consensus-based management plans for individual cases based on current guidelines. De-identified case data were input into both LLMs using standardized prompts in Chinese and English. Outputs were evaluated by three blinded clinicians across five domains via a validated 5-point Likert scale: diagnostic accuracy, treatment consistency, follow-up rationality, complication warning completeness, and readability. A comprehensive weighted score was calculated. In the Chinese language environment, DeepSeek-V3.2 achieved a significantly higher comprehensive score than ChatGPT-5.2 (4.40 ± 0.51 vs. 4.25 ± 0.47, P < 0.001). In the English environment, ChatGPT-5.2 outperformed DeepSeek-V3.2 (4.33 ± 0.38 vs. 4.20 ± 0.49, P < 0.001). Both LLMs scored highest in diagnostic accuracy and readability, and lowest in complication warning. Performance declined with advancing disease stage, with the lowest scores observed in mCRPC. The MDT maintained consistently high performance across all stages. Both LLMs demonstrate competent assistance potential for PCa outpatient management but do not surpass expert MDT judgment. Performance is stage-dependent, with significant challenges in complex metastatic settings. The models exhibit distinct linguistic strengths, suggesting that optimal tool selection may depend on the clinical language context. LLMs should be integrated as adjunctive tools within a clinician-led, MDT-based framework.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.