AI-based Hip Fracture Classification from Pelvic X-ray Images Using a Hybrid PEFT-ViT and XGBoost Model

Accurate radiographic classification of hip fractures into femoral neck (intracapsular) versus intertrochanteric (extracapsular) types is essential for surgical planning, yet most AI studies focus on binary detection using large datasets from high-resource regions, with no prior work from the Middle East. This study proposes a Hybrid PEFT-ViT and XGBoost framework for three-class classification of hip radiographs (Normal, Intertrochanteric fracture, Neck of femur fracture), using 768 X-ray images collected from King Abdullah University Hospital (KAUH) in Northern Jordan. A LoRA-fine-tuned ViT-Base/16 backbone extracted fused class-token and patch-token features, which were classified using XGBoost. On an independent test set of 77 radiographs, the proposed model achieved an accuracy of 88.31% (95% CI 79.3-93.7%), a sensitivity of 87.75%, a specificity of 94.28%, and an F1-score of 87.75%, the highest among the evaluated CNN-based (MobileNetV2, DenseNet121) and transformer-based (ViT, DeiT, Swin) baselines. These results show that clinically meaningful hip fracture classification can be achieved using a single-center dataset, providing a practical framework for resource-limited settings, although external validation is still required.