Evaluation of the diagnostic accuracy of artificial intelligence technology in detecting proximal femoral fractures in real-world clinical practice

Objective: to study the diagnostic accuracy of artificial intelligence-based software (AIT-based software) for detecting fractures of the proximal femur. Materials and methods. The research design is the scientific work performed in designing a diagnostic study, described according to the STARD 2015 guidelines. The study included two stages: retrospective and prospective. During the retrospective stage, we used a dataset of 100 subjects to calculate the following diagnostic metrics: AUROC, sensitivity, specificity, and accuracy. The prospective stage involved monitoring the diagnostic quality of the AI application and evaluating routine radiography studies (n = 11,685). We calculated the agreement between the AI application and medical experts, as well as performing an integral clinical estimate. The study duration was seven months. Results. We studied software based on AIT to detect proximal femur fractures. In retrospective testing, the area under the receiver operating characteristic curve (AUROC) was 0.887 (95% confidence interval [CI], 0.83-0.95); the specificity was 1.0 (95% CI, 1.0-1.0); the accuracy was 0.88 (95% CI, 0.81-0.94); and the sensitivity score was 0.76 (95% CI, 0.64-0.88). In a clinical setting, the AI application demonstrated sufficient data processing speed and high diagnostic metrics. The median clinical assessment exceeded 91%, and the AUROC ranged from 0.85 to 0.91. Conclusion. The study showed that the AI service has high diagnostic accuracy, making it possible to apply appropriate developments for interpreting X-ray results of the hip joints (in the context of the diagnosis of traumatic injuries).