Diagnostic performance of artificial intelligence in traumatic femur fracture: A systematic review and meta-analysis.
Authors
Affiliations (12)
Affiliations (12)
- Mayo Clinic Alix School of Medicine, Rochester, MN. Electronic address: https://twitter.com/AudreyBankes.
- Department of Surgery, Mayo Clinic, Rochester, MN. Electronic address: https://twitter.com/SergMNavarro.
- Division of Health Care Delivery Research, Kern Center for the Science of Health Care Delivery, Mayo Clinic, Rochester, MN.
- Mayo Clinic Alix School of Medicine, Rochester, MN.
- Mayo Clinic Alix School of Medicine, Rochester, MN. Electronic address: https://twitter.com/Mgish24.
- Mayo Clinic Libraries, Rochester, MN.
- Department of Surgery, Mayo Clinic, Rochester, MN.
- Department of Radiology, Mayo Clinic, Rochester, MN. Electronic address: https://twitter.com/drashishcool.
- Department of Orthopedic Surgery, Mayo Clinic, Rochester, MN. Electronic address: https://twitter.com/CodyWylesMD.
- Department of Orthopedic Surgery, University of Minnesota, Minneapolis, MN.
- Department of Orthopedic Surgery, Mayo Clinic, Rochester, MN. Electronic address: https://twitter.com/KrystinHiddenMD.
- Department of Surgery, Mayo Clinic, Rochester, MN. Electronic address: [email protected].
Abstract
Artificial intelligence applications are expanding in medicine, yet the efficacy of artificial intelligence for diagnostic imaging in traumatic femur fractures is unverified. The purpose of this systematic review and meta-analysis is to determine the diagnostic accuracy of artificial intelligence at identifying acute traumatic femur fractures. A comprehensive search of 9 databases deployed on May 29, 2025 identified studies related to artificial intelligence applications in acute trauma. Using Covidence and a 2-reviewer method, a systematic review identified 37 studies that met inclusion criteria. Bivariate mixed-effects model was used to pool accuracy measures (area under the receiver operating characteristic curve, specificity, and sensitivity) across the studies. The 37 studies included 95,148 femur images. Pooled artificial intelligence specificity, sensitivity, and area under the receiver operating characteristic curve were 0.94 (95% confidence interval, 0.92-0.97), 0.94 (0.92-0.96), and 0.98 (0.97-0.99), respectively. Humans identified traumatic femur fractures with specificity, sensitivity, and area under the receiver operating characteristic curve of 0.93 (0.90-0.96), 0.91 (0.88-0.94), and 0.97 (0.95-0.98), respectively. When expert (orthopedic surgeons and radiologists) and nonexpert (trainees and physicians in other specialties) human readers were compared, experts demonstrated an area under the receiver operating characteristic curve of 0.989 (0.98-0.99) unaided and 0.99 (0.98-1.0) with artificial intelligence augmentation. Nonexpert readers performed with an area under the receiver operating characteristic curve of 0.95 (0.91-0.97) unaided and 0.97 (0.95-0.98) augmented. Current evidence suggests that artificial intelligence can identify traumatic femur fractures with diagnostic performance close to human experts. In addition, artificial intelligence-assisted interpretation improved diagnostic accuracy of nonexperts. While limitations must be considered, these results suggest that artificial intelligence is proficient at identifying acute traumatic femur fractures and has potential for augmenting nonexpert humans.