Retrieval-augmented clinical decision support for structured hip-joint disease assessment

Hip-joint disease assessment requires integration of symptoms, imaging findings, staging criteria, and management context. Clinician-facing decision-support systems may help structure this reasoning but require defined clinical evaluation.

We conducted a retrospective, case-based validation of a hip-joint decision-support system integrating a local knowledge base, retrieval-augmented generation, and a multi-agent reasoning workflow. Seventy-four cases across five disease categories, stratified as easy, moderate, or complex, were evaluated. Validation diagnosis labels were established by an independent expert panel. Fifteen physicians (5 consultants, 10 residents/fellows) generated 1,110 evaluations. Endpoints included diagnostic accuracy, physician-rated clinical domains, confidence, decision time, and perceived acceptability.

Overall accuracy was 85.1% (95% CI, 76.8-93.4) for Hip-Agent, 94.6% (92.3-96.9) for consultants, and 73.5% (70.3-76.7) for residents/fellows. Hip-Agent accuracy was 94.4 and 92.0% in easy and moderate cases but 46.2% in complex cases. Inter-observer reliability for clinical-domain ratings was good to excellent (ICC 0.79-0.84). Perceived acceptability of the output format was high [mean 4.60 (0.49)].

The system demonstrated feasible retrospective performance for hip-joint case evaluation, with performance varying by complexity. Low accuracy in complex cases warrants caution. This study did not evaluate human-AI interaction; prospective workflow evaluation is needed to assess whether the system supports clinicians' judgment in practice.