Artificial Intelligence-Based Assessment of Hip Fracture Detection from Radiographic Images: Diagnostic Accuracy Compared with that of Orthopedic Surgeons and Radiologists

Abstract

Purpose: Hip fracture is a major global public health concern and one of the leading causes of morbidity and mortality among older adults. Diagnostic inaccuracies often result in delayed treatment and poor outcomes. Artificial intelligence (AI) has shown promise in fracture detection, but studies did not fully reflect real-world clinical practice. We aimed to evaluate the feasibility and capacity of a YOLOv8-based AI model to detect hip fractures from anteroposterior pelvic radiographs as accurately as orthopedic surgeons and radiologists.

Methods: A total of 345 anonymized radiographs were used, comprising 45 images for physician comparison and 300 for extended testing. Various clinicians reviewed 45 images, evenly distributed among normal, femoral neck, and intertrochanteric fractures. Diagnostic accuracy, sensitivity, specificity, and error types were analyzed. The AI model was trained by simulating real-world hospital conditions.

Results: AI achieved an overall accuracy of 0.94, with 0.92 sensitivity and 0.91 specificity, comparable to radiologists and orthopedic surgeons and superior to physicians. Model performance remained stable when tested on the larger dataset (p > 0.05). Most errors occurred in minimally displaced femoral neck fractures, though accuracy for this group improved with larger test data. Mean processing time was 1.9–2.3 seconds per image.

Conclusions: The YOLOv8-based AI system demonstrated expert-level diagnostic performance and high processing efficiency without requiring advanced hardware. Our findings highlight its applicability in hospitals. Although occasional misclassifications and mislocalizations occurred, the model shows promise as a clinical decision-support tool for improving diagnostic confidence, reducing delays, and enhancing patient safety.