Effectiveness of AI in detecting hip fractures, predicting postop outcomes

Editor's Note

This Canadian meta-analysis finds that artificial intelligence (AI) has the potential to automate hip fracture diagnoses; however, complicated models may not provide benefit over traditional patient-specific postoperative outcomes predictions.

Of 39 studies included in the analysis, 18 used AI models to diagnose hip fractures on plain radiographs and 21 used AI models to predict postoperative outcomes.

Among the findings:

• Mortality and length of hospital stay were the most predicted outcomes.
• Compared with clinicians, the odds ratio (OR) for diagnostic error of AI models was 0.79 for hip fracture radiographs.
• For the AI models, the mean sensitivity was 89.3%, specificity was 87.5%, and the F1 score was 0.90.
• The mean areas under the curve for mortality prediction were 0.84 with AI, compared with 0.79 for alternative controls.

The researchers concluded that the potential applications of AI to help with diagnosis from hip radiographs are promising; however, current uses in outcomes prediction do not provide substantial benefit over traditional multivariable predictive statistics.