Meta-analysis and Systematic Review of Diagnostic Performance of Machine Learning Algorithms on Skeletally Mature Wrist Fractures

Abstract

Background:

Wrist fractures are prevalent, and the diagnostic performance of artificial intelligence for their detection requires clarification. This study aims to determine the diagnostic accuracy of machine learning in detection of scaphoid fractures, distal radius fractures, and other types of wrist fractures on skeletally mature wrist radiographs through a meta-analysis while accounting for heterogeneity of studies using a random-effects model. In addition, this study will compare the results of machine learning to those of experts, including orthopedic surgeons, hand surgeons, and radiologists.

Methods:

Following Preferred Reported Items for Systematic Reviews and Meta-Analysis guidelines, study design, fracture type, radiograph details, artificial intelligence algorithm features, ground truth, and diagnostic performance were extracted. Pooled estimates and heterogeneity were analyzed using a random-effects model. Data were presented using metrics such as area under the curve (AUC). Area under the curve reflects the overall ability of an algorithm to discriminate fractured from non-fractured wrists across all thresholds, providing a broad measure of diagnostic accuracy.

Results:

A total of 48 algorithms were identified from literature that aimed at detecting wrist fractures. The pooled AUC of detection of scaphoid, distal radius, and unspecified wrist fractures by machine learning were 0.88, 0.97, and 0.93, respectively. The AUC of expert detection of wrist fractures was 0.88 for scaphoid fractures and 0.98 for distal radius fractures.

Conclusion:

The performance of machine learning in wrist fracture detection is highly comparable to that of experts.