Detailed Rubric Prompts Improve LLM Evaluation of Clinical Reasoning Reports in Physical Therapy Education

【Background】 Rubric-based evaluation is recommended for objective assessment of clinical reasoning reports in physical therapy education. Recently,Large Language Model (LLM) are expected to support text evaluation.However, abstract rubrics often overlook the complex and implicit reasoning of physical therapists. 【Purpose】 This study examined whether detailed rubric prompts improve LLM evaluation of clinical reasoning reports. 【Methods】 Clinical reasoning reports were collected from 20 physical therapists with 1–5 years of experience using a simulated femoral neck fracture case. Faculty member A, Faculty member B, and an LLM (ChatGPT-5.5) independently scored each report using a 15-item rubric (maximum score: 15 points). Spearman’s rank correlation analysis was used to examine relationships among the evaluators.First, the LLM evaluated the reports using the original abstract rubric items. Next, more detailed prompts were provided to the LLM. These prompts included predefined sub-items and criteria that had been developed before this study by faculty members specializing in clinical reasoning education. 【Results】 Faculty members A and B showed a strong positive correlation (ρ = 0.915, p < 0.001), whereas correlations between the faculty members and the LLM were moderate (ρ = 0.525 and 0.48).The LLM tended to assign lower scores to rubric items related to ICF relationships and prognosis prediction. After providing more detailed prompts, strong positive correlations were found between the faculty members and the LLM (Faculty member A: ρ = 0.788, p < 0.001; Faculty member B: ρ = 0.727, p = 0.002). 【Conclusion】 LLMs had difficulty evaluating clinical reasoning using abstract rubric items. However, correlations with faculty evaluations improved when rubric prompts designed for educational purposes were used. 【Clinical Relevance】 In LLM-based evaluation of clinical reasoning reports, rubric prompts should be designed not only as scoring criteria but as structured representations of clinical reasoning.