OrthoTraceGLM : A Knowledge-Traced Large Language Model for Orthopedic Consultation

Xue, Tao and Li, Pinjie and Wang, Ziwei and Lan, Fengbo and Cai, Jinfen and Zhang, Tao (2026) OrthoTraceGLM : A Knowledge-Traced Large Language Model for Orthopedic Consultation. In: 2025 China Automation Congress (CAC) :. 2025 China Automation Congress (CAC) . IEEE, pp. 5080-5085. ISBN 9798331589684

Text (cac2025)
cac2025.pdf - Accepted Version
Download (1MB)

Abstract

This paper proposes an orthopedic consultation system enable remote, intelligent, and personalized treatment and rehabilitation. To address the inherent hallucinations and domain knowledge deficiency in the general LLMs, we design a specialized multi-agent collaborative model to generate reliable orthopedics-related responses along with source knowledge through implicit knowledge injection and explicit retrieval. Inspired by task decomposition, we propose that each LLM agent handles a subtask, rather than relying on a single entity to manage the entire task globally. In this framework, four individual small-scale LLMs are trained to complete medical record filling, knowledge retrieval, response generation, and typesetting tasks, separately, to obtain reliable responses underpinned by verifiable and traceable knowledge. To evaluate the performance quantitatively, we create a new scoring metric from safety, helpfulness, and smoothness aspects, and results demonstrate that OrthoTraceGLM outperforms GLM-4-9B-Chat in both proposed evaluation scores and corresponding knowledge tracing accuracy.