Study-Level Diagnostic Performance of Artificial Intelligence for Fracture Detection on Musculoskeletal Radiographs: A Retrospective Multi-Reader Study

Fracture detection on musculoskeletal radiographs remains challenging, particularly when findings are subtle or visible on only one projection. This retrospective two-centre multi-reader diagnostic performance study evaluated an artificial intelligence system at the level of the complete radiographic examination and compared its performance with seven anonymized readers. Radiographic examinations from AGEL Šternberk Hospital and AGEL Vítkovice Hospital, acquired between 15 December 2025 and 25 January 2026, were retrospectively analysed. The primary endpoint was study-level fracture positivity, defined as at least one positive image within an examination; the same aggregation rule was applied to AI and reader assessments. Examinations of the spine, ribs, and hemithorax were excluded because these anatomies were outside the supported scope. The cohort comprised 526 examinations, including 101 fracture-positive and 425 fracture-negative studies. Sensitivity was 0.911 (95% CI, 0.839–0.952), specificity 0.899 (0.866–0.924), PPV 0.681 (0.599–0.754), NPV 0.977 (0.957–0.988), and accuracy 0.901 (0.873–0.924). Mean reader sensitivity, specificity, PPV, NPV, and accuracy were 0.853, 0.853, 0.598, 0.961, and 0.853, respectively. The AI showed higher point estimates than the mean of the seven readers across all principal diagnostic metrics. These findings support prospective evaluation of AI as a second-reader or workflow-support tool for fracture detection, particularly in examinations containing subtle or easily overlooked findings.