AI As Good As Expert Radiologists in Diagnosing Hip Fractures from X-Rays

Predictions indicate that worldwide, the number of hip fractures will soar to 6.3 million by 2050, with associated expenses expected to reach USD 131.5 billion annually. These injuries not only introduce significant morbidity and mortality risks, but also cause a 1-year mortality rate of roughly 25% to 30%. Given these concerning statistics, there exists an urgent need for advanced technology that can enhance managing the condition, leading to better patient outcomes and economic advantages for healthcare systems. With the advent of artificial intelligence (AI), clinical diagnostic and prognostic tools for hip fractures can harness powerful predictive models, although little is known about the performance and impact of these new algorithms. Now, new study findings suggest that AI has the potential to automate hip fracture diagnosis, although overly complex and uninterpretable AI models may not yield any significant benefits when it comes to predicting patient-specific outcomes compared to the traditional, interpretable models.

Researchers at the University of Toronto (Toronto, ON, Canada) assessed the effectiveness of AI-driven algorithms in detecting hip fractures on radiographs and predicting postoperative clinical results following hip surgery. To achieve this, a systematic review and meta-analysis of 39 studies were carried out to determine the accuracy of hip fracture diagnosis by both AI models and expert healthcare professionals. The study found that AI models' diagnostic accuracy was comparable to that of expert clinicians, with similar error rates. The diagnostic precision of AI relative to expert clinicians was measured using odds ratios (ORs) with 95% CIs. Additionally, the researchers compared the area under the curves for the prediction of postoperative outcomes between traditional statistical models (e.g., multivariable linear or logistic regression) and machine learning (ML) models.

Out of the 39 studies that met the criteria, 18 (46.2%) used AI models aimed at detecting hip fractures from plain radiographs while 21 (53.8%) used AI models for predicting patient outcomes post hip surgery. In total, the studies used 39,598 plain radiographs and 714,939 hip fractures to train, validate, and test the ML models specific to the diagnosis and postoperative outcome prediction. The most commonly predicted outcomes were mortality and the length of stay in the hospital. After a pooled data analysis, ML models had a diagnostic error OR of 0.79 (95% CI, 0.48-1.31; P = .36; I2 = 60%) for hip fracture radiographs compared to clinicians. For ML models, the mean (SD) sensitivity was 89.3% (8.5%), specificity was 87.5% (9.9%), and F1 score was 0.90 (0.06). The average area under the curve for mortality prediction was 0.84 with ML models as compared to 0.79 for alternative controls (P = .09).

According to the results of this systematic review and meta-analysis, AI appears very promising for aiding the diagnosis of hip fractures using radiographs. The performance of AI models in detecting hip fractures was found to be comparable to that of expert radiologists and surgeons. Nevertheless, the current AI implementations for outcome prediction did not appear to offer significant benefits over traditional multivariable predictive statistics.

Related Links:
University of Toronto