Session: 07-10-02: Medical Robotics, Rehabilitation, and Surgery

Paper Number: 191787

LLM-Based Lower-Limb Rehabilitation Device Recommendation With Multi-Turn Dialogue and Safety-Aware Mechanism 

Lower-limb rehabilitation recommendations carry critical clinical implications, yet large-scale implementation is increasingly challenging due to population aging and the uneven distribution of professional therapists, which leaves a substantial proportion of patients without timely and personalized guidance. Despite rapid progress in large language models (LLMs), mainstream LLM-based assistive tools remain unreliable in safety-critical medical interactions: they frequently over-infer from incomplete or ambiguous patient narratives, fail to maintain consistency across multiple dialogue turns, and lack appropriate mechanisms to intercept high-risk conditions, which could compromise patient safety in real-world clinical settings.To address scalability and safety challenges in rehabilitation recommendation, this study presents a computational framework for multi-turn triage and structured rehabilitation device guidance. The system processes sequential patient–clinician interactions and employs a JSON Schema–based constrained decoding algorithm to transform natural language reports into structured symptom representations, including affected anatomical regions, primary symptoms, duration, severity, and relevant risk indicators. Extracted data are further refined using LoRA-fine-tuned large language models, which incrementally update the semantic state with each dialogue turn, ensuring that the most recent input supersedes prior information. A field-level rule validation module enforces consistency by applying keyword presence checks, synonym mapping, approximate term detection, and negation analysis, thereby calibrating confidence scores and classifying fields as confirmed, requiring clarification, or unknown.At the recommendation stage, the system implements a hierarchical computational pipeline comprising a safety gate, constrained retrieval, and interpretable generation modules. The safety gate conducts binary evaluations of high-risk indicators and interrupts standard recommendation workflows when triggered, ensuring institutional safety protocols are followed. For cases passing the safety gate, the constrained retrieval module maps structured features to a rehabilitation configuration library under primary constraints—affected anatomical region, rehabilitation objective, and training intensity—producing a validated subset of candidate interventions. Dense vector retrieval using BGE-M3 performs semantic ranking via normalized cosine similarity, identifying the most relevant rehabilitation configurations. Finally, the interpretable generation module translates these ranked configurations into clear and structured guidance suitable for implementation in rehabilitation devices.The system is evaluated using two complementary datasets: real clinical cases and parametrically constructed cases derived from actual pathological manifestations, emphasizing boundary conditions, rare severe presentations, and complex comorbid scenarios. Comparative experiments against purely LLM-based and rule-based approaches demonstrate that the proposed framework consistently outperforms both baselines in recommendation effectiveness, safety control, and reliability. The results indicate that the framework achieves a superior balance between accurate guidance and risk mitigation, thereby enhancing clinical applicability, ensuring safe implementation, and providing a robust foundation for large-scale deployment of lower-limb rehabilitation recommendations.

Presenting Author: Zhenhao Liu Hefei University of Technology

Presenting Author Biography: Zhenhao Liu, born in 2002, is currently a master's student at the School of Mechanical Engineering, Hefei University of Technology, China. He received his bachelor's degree from Hefei University of Technology, China, in 2025.His research interests include rehabilitation robotics, robot mechanism design, and generative design.

Authors:

Zhenhao Liu Hefei University of Technology
Jianxing Liang Hefei University of Technology
Yuan Li Hefei University of Technology
Xinyu Cheng Hefei University of Technology）
Ping Zhao Hefei University of Technology