Scholarlink

<p>Fracture rehabilitation must adapt as pain, function and warning signs change across healing phases, but many digital tools provide static information and cannot coordinate exercise selection, symptom assessment, progress monitoring and safety escalation. To address this gap, we developed FractureAgent, a reasoning-and-action (ReAct)-style system that couples a domain-adapted Qwen3.5-9B backbone with five typed rehabilitation tools and a deterministic safety gate to provide phase-aware support. The model was adapted using quantized low-rank adaptation (QLoRA) supervised instruction tuning on 18,742 synthetic rehabilitation dialogues and tool-use traces derived from open-access clinical and patient-education sources. We evaluated FractureAgent using automated functional metrics, expert clinical ratings and 210 simulated-patient scenarios spanning six fracture types and three rehabilitation phases. It achieved a task completion rate of 91.4%, a mean clinical-appropriateness score of 4.21/5.00, pain-assessment concordance of 0.873, exercise-appropriateness of 0.896 and complication-detection sensitivity of 0.843, outperforming the evaluated baselines, including the unfine-tuned Qwen3.5-9B backbone. These findings indicate that FractureAgent can coordinate multiple rehabilitation tasks and provide adaptive support in simulated settings, addressing the technical limitations of static digital guidance; however, they do not establish clinical effectiveness or readiness for deployment.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

FractureAgent: a multi-tool LLM-based intelligent agent for personalized fracture rehabilitation management

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月28日

查看全文
  • Hang Cao,
  • Fangwei Hu,
  • Lin Xu,
  • Kun Guo,
  • Bingchuan Xue,
  • Ning Zhang,
  • Jinhao Sun,
  • Weijuan Gong,
  • Xiao Ouyang

摘要

Fracture rehabilitation must adapt as pain, function and warning signs change across healing phases, but many digital tools provide static information and cannot coordinate exercise selection, symptom assessment, progress monitoring and safety escalation. To address this gap, we developed FractureAgent, a reasoning-and-action (ReAct)-style system that couples a domain-adapted Qwen3.5-9B backbone with five typed rehabilitation tools and a deterministic safety gate to provide phase-aware support. The model was adapted using quantized low-rank adaptation (QLoRA) supervised instruction tuning on 18,742 synthetic rehabilitation dialogues and tool-use traces derived from open-access clinical and patient-education sources. We evaluated FractureAgent using automated functional metrics, expert clinical ratings and 210 simulated-patient scenarios spanning six fracture types and three rehabilitation phases. It achieved a task completion rate of 91.4%, a mean clinical-appropriateness score of 4.21/5.00, pain-assessment concordance of 0.873, exercise-appropriateness of 0.896 and complication-detection sensitivity of 0.843, outperforming the evaluated baselines, including the unfine-tuned Qwen3.5-9B backbone. These findings indicate that FractureAgent can coordinate multiple rehabilitation tasks and provide adaptive support in simulated settings, addressing the technical limitations of static digital guidance; however, they do not establish clinical effectiveness or readiness for deployment.