Artificial Intelligence in Pelvic Fracture Diagnosis and Outcome Prediction: A Systematic Review and Meta-analysis

Abstract

ObjectiveTo synthesize the performance of AI applications for detecting pelvic fractures, classifying severity, and predicting clinical outcomes relative to clinicians. Patients and MethodsThe study was designed as a systematic review and meta-analysis (PROSPERO CRD420251141768).Ovid Embase, Ovid MEDLINE, PubMed, Scopus, and Cochrane CENTRAL were searched for articles published from database inception to September 11, 2025.Studies were included if they evaluated AI models for pelvic ring fractures in adults using pelvic radiographs.Case series, reviews, and abstracts without full data were excluded.Summary level data were independently extracted using a standardized template.Diagnostic metrics were pooled using a random-effects model.Outcomes included pooled sensitivity, specificity, area under the receiver operating characteristic curve (AUC), and accuracy. ResultsFourteen studies were included.Thirteen evaluated radiographic fracture detection or classification (n=31,166 radiographs) and one evaluated outcome prediction.AI demonstrated high pooled performance: accuracy 0.96 (95% CI 0.91 -0.98;I 2 =93.3),AUC 0.94 (95% CI 0.89 -0.97;I 2 =97.9%), sensitivity 0.90 (95% CI 0.84 -0.94; 2 =0.42), and specificity 0.93 (95% CI 0.85 -0.97; 2 =1.30).In three studies directly comparing AI with clinicians, AI models showed comparable or marginally superior performance.One study on clinical outcomes reported strong performance for predicting hemodynamic instability (AUC 0.92) and mortality (AUC 0.90). ConclusionAI algorithms show promise as supportive tools for pelvic fracture detection, achieving diagnostic performance comparable to expert clinicians.However, included studies exhibit substantial heterogeneity, selection bias, and limited external validation.Large-scale, prospective validation is necessary before widespread clinical adoption.

Ähnliche Arbeiten