Authors
Changxiao Han, Minrui Fu, Jinqiao Zou, Bochen Peng, Guangyi Yang, Hongtao Li, Guangwei Liu, Xianzheng Bu, Liguo Zhu, Minshan Feng
Published in
Zhongguo gu shang = China journal of orthopaedics and traumatology. Volume 39. Issue 10. Pages 1005-11. Oct 25, 2026.
Abstract
To systematically compare the performance differences between deep learning models and fine-tuned large language models in the diagnosis and treatment of cervical spondylosis using traditional Chinese medicine(TCM), so as to provide empirical evidence for selecting the technical route of intelligent auxiliary diagnosis and treatment systems.
A total of 100 real outpatient cases of cervical spondylosis from January 2023 to December 2024 were enrolled as the validation dataset, including complete four-diagnosis information, syndrome differentiation and classification, and Chinese herbal prescriptions. There were 39 males and 61 females, aged from 32 to 74 years with a mean age of (46.25±13.71 ) years, with a disease course of (4.25±2.06) months. The distribution of cervical spondylosis types showed cervical type 9 patients, nerve root type 21 patients, vertebral artery type 29 patients, spinal cord type 21 patients, and mixed type 20 patients. Two previously developed models were selected for comparison:a deep learning model (based on transformer architecture with multi-label classification probability output) and a fine-tuned large language model (based on Qwen2.5-7B-Instruct, with continuous pre-training and LoRA-supervised fine-tuning). Quantitative and qualitative evaluations were used to assess the accuracy of diagnosis and syndrome differentiation. The Jaccard similarity index (JSI) was adopted to evaluate the similarity of herbal composition. Normalized mean square error (NMSE) was used to evaluate the deviation between model-predicted dosages and actual prescription dosages from senior famous traditional Chinese medicine practitioners.
The diagnostic classification accuracy of the deep learning model and the fine-tuned large model was 93.0% and 95.0%, respectively;the accuracy for main syndrome differentiation was 89.0% and 91.0%, respectively;and the accuracy for concomitant syndrome differentiation was 35.0% and 67.0%, respectively, with the fine-tuned large model outperforming the deep learning model. The deep learning model achieved a higher JSI(0.616±0.361) than the fine-tuned large model (0.529±0.157), with a statistically significant difference (P<0.05), though the distribution was uneven, with 46% of samples falling in the extremely high range(0.8 to 1.0). The fine-tuned large model showed a significantly lower NMSE(0.089±0.076) than the deep learning model (0.112±0.093), with a statistically significant difference(P<0.05). The comprehensive expert scores of both models exceeded 4 points (full score:5 points). The intraclass correlation coefficient among experts was 0.83 [95%CI(0.78, 0.87)]. In terms of individualization, the fine-tuned large model scored higher (3.92±0.80) than the deep learning model (3.16±0.91), with a statistically significant difference (P<0.001).
Deep learning models are efficient and stable in standardized tasks, while fine-tuned large language models demonstrate superior performance in complex syndrome identification, clinical reasoning, and individualized treatment. The appropriate model should be selected based on practical application scenarios, or hybrid architectures could be explored.
PMID:
42988397
Bibliographic data and abstract were imported from PubMed on 06 Nov 2026.
Read full publication at:
Please sign in
to see all details.
Advertisement
Stats
- Recommendations n/a n/a positive of 0 vote(s)
- Views 11
- Comments 0