A staged approach using machine learning and uncertainty quantification to predict the risk of hip fracture.

Hip fractures present a significant healthcare challenge, especially within aging populations, where they are often caused by falls. These fractures lead to substantial morbidity and mortality, emphasizing the need for timely surgical intervention. Despite advancements in medical care, hip fractures impose a significant burden on individuals and healthcare systems. This paper focuses on the prediction of hip fracture risk in older and middle-aged adults, where falls and compromised bone quality are predominant factors. The study cohort included 547 patients, with 94 experiencing hip fracture. To assess the risk of hip fracture, clinical variables and clinical variables combined with hip DXA imaging features were evaluated as predictors, followed by a novel staged approach. Hip DXA imaging features included those extracted by convolutional neural networks (CNNs), shape measurements, and texture features. Two ensemble machine learning models were evaluated: Ensemble 1 (clinical variables only) and Ensemble 2 (clinical variables and imaging features) using the logistic regression as the base classifier and bootstrapping for ensemble learning. The staged approach was developed using uncertainty quantification from Ensemble 1 which was used to decide if hip DXA imaging features were necessary to improve prediction for each subject. Ensemble 2 exhibited the highest performance, achieving an Area Under the Curve (AUC) of 0.95, an accuracy of 0.92, a sensitivity of 0.81, and a specificity of 0.94. The staged model also performed well, with an AUC of 0.85, an accuracy of 0.86, a sensitivity of 0.56, and a specificity of 0.92, outperforming Ensemble 1, which had an AUC of 0.55, an accuracy of 0.73, a sensitivity of 0.20, and a specificity of 0.83. Furthermore, the staged model suggested that 54.49 % of patients did not require DXA scanning, effectively balancing accuracy and specificity, while offering a robust solution when DXA data acquisition is not feasible. Statistical tests confirmed significant differences between the models, highlighting the advantages of advanced modeling strategies. Our staged approach offers a cost-effective holistic view of patient health. It can identify individuals at risk of hip fracture with a high accuracy while reducing unnecessary DXA scans. This approach has great promise to guide the need for interventions to prevent hip fracture while reducing diagnostic cost and exposure to radiation.