Digital Health and Intelligent Medicine
PENG Dongge, CHEN Xiaojing, WAN Ziye, YANG Pengcheng, ZHANG Yang, LU Ning
Objective This study aims to evaluate the clinical value of the locally deployed DeepSeek R1 model, developed by DeepSeek Inc., in multi-task oncology scenarios. Specifically, this research attempts to systematically assess its real-world performance in assisted diagnosis, treatment pathway recommendation, clinical trial matching, and medical record quality control, and to verify its feasibility and potential for implementation in smart hospital development. Methods Real-world data were collected from the Oncology Department of the Chinese PLA General Hospital of Xinjiang Military Command between May and August 2024. A locally deployed framework incorporating retrieval-augmented generation technology was constructed, followed by multi-scenario performance evaluations. In the assisted diagnosis task, 30 patients with gastric cancer were enrolled to compare the diagnostic performance between the DeepSeek R1 model and junior physicians. In the treatment recommendation task, another 30 patients with gastric cancer were included to evaluate the model’s capacity, relative to ChatGPT-4, in generating suggestions for comorbidity management, nutritional intervention, and dynamic treatment plan adjustment. For the clinical trial matching and doctor-patient communication tasks, 10 patients with non-small cell lung cancer were assessed to measure the model’s accuracy in enrollment recommendation and its effectiveness in enhancing patient communication. In the medical record quality control task, 50 retrospective medical records were reviewed to evaluate the model’s sensitivity in detecting logical inconsistencies. Results In the diagnostic task, the DeepSeek R1 group demonstrated a significantly higher diagnostic accuracy (96.7% vs. 76.7%, P = 0.008), lower misdiagnosis rate (6.7% vs. 20.0%, P = 0.039), and shorter diagnostic time compared to the junior physician group. In treatment recommendation, the DeepSeek R1 group outperformed the ChatGPT-4 group in identifying comorbidities and generating personalized treatment suggestions (100.0% vs. 13.3%, P<0.001), assessing nutritional risk and providing intervention plans (86.7% vs. 6.7%, P<0.001), and dynamically adjusting treatment strategies (93.3% vs. 0, P<0.001). The model achieved a clinical trial matching accuracy of 90%. In the communication and quality control tasks, the shared decision-making questionnaire score was significantly higher in the DeepSeek R1 group than in the conventional communication group (38.7 ± 2.8 vs. 21.3 ± 3.5, P<0.001), and the model demonstrated 91.7% sensitivity in identifying tumor staging inconsistencies. Conclusions Through localized deployment and multimodal data integration, the DeepSeek R1 model demonstrates substantial potential in supporting various oncology-related decision-making tasks. Therefore, it provides a cost-effective technological pathway for advancing smart hospital development.