Evidence-Informed Occupational Therapy Decision Support Using Graph Retrieval-Augmented Generation

Occupational therapy clinical decision-making requires support that is both evidence-informed and traceable. To develop and evaluate a clinical decision-support system (CDSS) for occupational therapy using the Japanese Association of Occupational Therapists (JAOT) Case Report Corpus. A total of 3,023 cases were included, with 90% used as a reference case set and 10% as a test case set. A knowledge graph linking assessment findings, goals, and intervention plans was constructed from the reference case set. Using GPT-5-mini, intervention plans were generated from assessment findings under five conditions: no reference information, random reference cases, keyword-based retrieval of similar cases, embedding-based retrieval of similar cases, and GraphRAG. Generated intervention plans were evaluated using the Retrieval-Augmented Generation Assessment (RAGAS) Faithfulness metric, and models were compared using paired t -tests. GraphRAG showed significantly higher faithfulness than the other reference-based models. GraphRAG may provide case-based and traceable support for occupational therapy decision-making.