Appraising the Value of AI in Wrist Fracture Detection: A Clinical Review - 11/02/26

Doi : 10.1016/j.hansur.2026.102597

Aidin Gharavi ^a,⁎ , Sergio M. Navarro ^b,^c, Andy Tom ^a, Audrey Banks ^a, Matthew Gish ^a, Michele McGinnis MSIS ^d, Matthew D. Rich ^e

^a Mayo Clinic Alix School of Medicine, Rochester, MN, United States

^b Department of Surgery, Mayo Clinic, Rochester, MN, United States

^c Department of Surgery, University of Minnesota, Minneapolis, MN, United States

^d Mayo Clinic Libraries, Rochester, MN, United States

^e Division of Hand Surgery, Department of Orthopedic Surgery, University of California Irvine, Irvine, CA, United States

^⁎ Corresponding author at: Mayo Clinic Alix School of Medicine, 200 First Street SW Rochester, Minnesota 55905, United States. Mayo Clinic Alix School of Medicine 200 First Street SW Rochester Minnesota 55905 United States

Sous presse. Manuscrit accepté. Disponible en ligne depuis le Wednesday 11 February 2026
Cet article a été publié dans un numéro de la revue, cliquez ici pour y accéder

Abstract

Introduction

Artificial intelligence (AI) has become increasingly tested for improving fracture detection. This scoping review evaluates how AI can improve human performance in detecting traumatic wrist fractures.

Methods

A scoping review was conducted to identify studies comparing AI and human performance for wrist fracture detection. Nine resources were searched on 05/15/2024, including key databases Embase, MEDLINE, and SCOPUS. Studies which did not include AI being used as an assistant were excluded.

Results

Nine studies were included in the review, where AI models demonstrated sensitivity from 83% to 97.7%, specificity from 77% to 96%, negative predictive value (NPV) from 89% to 90%, and positive predictive value (PPV) from 83% to 92%. Human performance demonstrated sensitivity from 58% to 94%, specificity from 77% to 97%, a NPV from 62% to 87%, and a PPV from 74% to 93%. When AI was used as a diagnostic aid, combined performance generally improved across all studies. Sensitivity gains ranged from 3% to 19%, with greater improvements among less experienced readers. Specificity changed between -4% to 11% when AI was used as a diagnostic aid. NPV and PPV changed from 6% to 13% and –5% to 7% when AI was used as a diagnostic aid, respectively.

Conclusion

Models using AI improved sensitivity and negative predictive value across all readers, while some readers had a reduced specificity or positive predictive value when using AI. Overall, AI has demonstrated promising results in traumatic wrist fracture detection when used as an assistant to support diagnosis.

Le texte complet de cet article est disponible en PDF.

Keywords : Artificial Intelligence, Wrist Fracture, Diagnostic Accuracy, Deep Learning, Distal Radius Fracture