Organizations: Hebei Provincial Hospital of Traditional Chinese Medicine, Shijiazhuang, China · Beijing Wenge Technology Co., Ltd., Beijing, China · Institute of Automation, Chinese Academy of Sciences, Beijing, China · Tianjin University, Tianjin, China
The clinical burden of spleen-stomach disorders is substantial. While large language models (LLMs) offer new potential for medical applications, they face three major challenges in the context of integrative Chinese and Western medicine (ICWM): a lack of high-quality data, the absence of models capable of effectively integrating the reasoning logic of traditional Chinese medicine (TCM) syndrome differentiation with that of Western medical (WM) disease diagnosis, and the shortage of a standardized evaluation benchmark. To address these interrelated challenges, we propose DongYuan, an ICWM spleen-stomach diagnostic framework. Specifically, three ICWM datasets (SSDF-Syndrome, SSDF-Dialogue, and SSDF-PD) were curated to fill the gap in high-quality data for spleen-stomach disorders. We then developed SSDF-Core, a core diagnostic LLM that acquires robust ICWM reasoning capabilities through a two-stage training regimen of supervised fine-tuning. tuning (SFT) and direct preference optimization (DPO), and complemented it with SSDF-Navigator, a pluggable consultation navigation model designed to optimize clinical inquiry strategies. Additionally, we established SSDF-Bench, a comprehensive evaluation benchmark focused on ICWM diagnosis of spleen-stomach disorders. Experimental results demonstrate that SSDF-Core significantly outperforms 12 mainstream baselines on SSDF-Bench. DongYuan lays a solid methodological foundation and provides practical technical references for the future development of intelligent ICWM diagnostic systems.
Figures & tables
Fig. 1: Overview of DongYuan. The framework integrates three specialized datasets (SSDF‑Syndrome, SSDF‑Dialogue, SSDF‑PD), a core diagnostic LLM trained via a two‑stage pipeline (SFT → DPO), a pluggable consultation navigation model (SSDF‑Navigator), and a comprehensive evaluation benchmark (SSDF‑Bench). Application scenarios cover a diverse array of ICWM clinical and research-oriented tasks.
Fig. 2: An example from the SSDF-Syndrome dataset. Each sample is constructed from original medical records, presenting fully structured diagnostic information that explicitly integrates TCM syndrome differentiation, WM disease diagnosis, and other structured clinical information.
Fig. 3: An example from the SSDF‑Dialogue dataset. The dialogue follows a clinically coherent progression: starting from an initial symptom, the model sequentially inquires about severity fluctuation, temporal patterns, and associated impacts on daily activities.
Fig. 4: Overview of SSDF‑Navigator. The model uses a Transformer-based encoder-classifier architecture, with the symptom set from historical dialogue as input, to predict the next symptom. Training combines behavior cloning with offline reinforcement learning. During inference, it collaborates with SSDF‑Core via a candidate-based inquiry selection mechanism to guide multi‑turn proactive consultation.
Type
Model
ME
LLMS
TCM-BC
WM-BC
DDSSD
Total score
General LLMs
DeepSeek-V3.2 (w/)
0.81
0.89
0.8143
0.8358
0.5994
0.7398
Gemini-3 (w/)
0.51
0.70
0.8617
0.8158
0.6010
0.6969
GPT-5.2-2025-12-11 (w/)
0.9
0.82
0.5819
0.6794
0.6051
0.6663
GLM-4.5 (w/)
0.77
0.96
0.7803
0.7825
0.6063
0.7280
Qwen3-32B (w/)
0.75
0.97
0.7651
0.7069
0.5323
0.6793
TCM-Specific LLMs
Biancang-14B (w/o)
0.88
0.97
0.7507
0.7275
0.4542
0.6623
TABLE II: Performance Comparison Results on SSDF-Bench
Fig. 5: Performance Comparison Across Dimensions of the DDSSD Task.
Model
ME
LLMS
TCM-BC
WM-BC
DDSSD
MPC
Qwen3-14B (base)
0.73
0.97
0.7162
0.6847
0.5499
0.5702
SSDF-Core w/o DPO
0.85
0.99
0.7861
0.8275
0.6995
0.4663
SSDF-Core-DPO
0.89
0.97
0.8759
0.8231
0.7457
0.6066
TABLE III: Ablation Results of the two-stage training strategy on SSDF-Bench
Configuration / Demension
Recallcrit
Effdia
SSDF-Core w/o Navigator
0.6080 (357/653)
0.3444
SSDF-Core w/ Navigator
0.6876 (449/653)
0.3701
TABLE IV: Ablation results of SSDF-Navigator on the MPC task
Fig. 6: Effectiveness Evaluation of the Judge Model in SSDF-Bench.
Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen, China · Guangdong Provincial Laboratory of Traditional Chinese Medicine Hengqin, Guangdong, China
the Department of Computer Science, Johns Hopkins University, Baltimore, USA · the School of Pharmacy, Nanjing University of Chinese Medicine, Nanjing, China · the School of Information Engineering, Huzhou University, Huzhou, China +3
Department of Computer Science and Technology, Beijing Jiaotong University, Beijing, China · Tianjin Tasly Digital Chinese Medicine Technology Co., Ltd., Tianjin, China · Tasly Biopharmaceuticals Co., Ltd., Tianjin, China +11