Evaluating artificial intelligence for accurate detection of hand and wrist fractures: a systematic review and meta-analysis

Abstract

<ns3:p> Background and Objectives Hand and wrist fractures are among the most frequently encountered injuries in emergency departments and are often misdiagnosed, particularly when interpreted by non-specialist clinicians. These diagnostic errors can lead to treatment delays and long-term complications. Artificial intelligence (AI), particularly deep learning algorithms, is emerging as a promising adjunct to improve diagnostic accuracy in radiographic fracture detection. This study aims to evaluate the effectiveness of Artificial Intelligence (AI) in detecting hand and wrist fractures compared to manual radiographic interpretation by clinicians. Materials and Methods A systematic review and meta-analysis were conducted to assess the diagnostic performance of AI models in detecting hand and wrist fractures compared to conventional radiographic interpretation by clinicians. A comprehensive search of PubMed, Google Scholar, Science Direct, and Wiley Online Library was performed. Eligible studies included those utilizing AI for fracture detection with sensitivity and specificity data. Pooled estimates were calculated using fixed- and random-effects models. Heterogeneity was assessed via I <ns3:sup>2</ns3:sup> statistics, and publication bias was examined using funnel plots and Egger’s test. Results Eighteen studies met inclusion criteria. The pooled sensitivity and specificity under the random-effects model were 0.910 and 0.912, respectively, indicating high diagnostic accuracy of AI models. However, substantial heterogeneity (I <ns3:sup>2</ns3:sup> = 99.09% for sensitivity; 96.43% for specificity) and publication bias were observed, likely due to variations in AI algorithms, sample sizes, and study designs. Conclusions Most AI models demonstrated good diagnostic accuracy, with high sensitivity and specificity scores (≥90%). However, some models fell short in sensitivity and specificity (≤90%), indicating performance variations across different AI models or algorithms. From a clinical perspective, AI models with lower sensitivity scores may fail to detect hand and wrist fractures, potentially delaying treatment, while those with lower specificity scores could lead to unnecessary interventions—treating hands and wrists that are not fractured. </ns3:p>

Ähnliche Arbeiten