Accurate prediction of hip fracture by machine learning approach — ASN Events
  1. Schedule
  2. Proffered papers - Clinical II
  3. Accurate prediction of hip fracture by machine learning approach

Accurate prediction of hip fracture by machine learning approach (#77)

Thao P. Ho-Le 1 , Jackie R. Center 2 3 , John A. Eisman 2 3 4 , Hung T. Nguyen 1 , Tuan V. Nguyen 1 2 3 4 5
  1. Centre for Health Technologies, Falcuty of Engineering and Information Technology, University of Technology, Sydney, Australia, Sydney, New South Wales, Australia
  2. Bone Biology Division, Garvan Institute of Medical Research, Sydney, NSW, Australia
  3. St Vincent Clinical School, University of NSW, Sydney, New South Wales, Australia
  4. School of Public Health and Community Medicine, UNSW, Sydney, New South Wales, Australia
  5. School of Medicine, Notre Dame University , Sydney, New South Wales, Australia

Hip fracture is a serious public heath problem among post-menopausal women with osteoporosis. Existing models for predicting hip fracture unrealistically assume that risk factors exert additive effects on fracture risk, and their area under the receiver operating characteristic (AUC) curves ranged between 0.70 to 0.85. Here, we sought to develop new predictive models using machine learning approach to account for complex interactions between risk factors.

The study was part of the Dubbo Osteoporosis Epidemiology Study that involved 1167 women aged 60 years and above. The women had been followed up for up to 20 years, and during the period, the incidence of new hip fractures was ascertained. We "trained" a series of models with different machine learning algorithms, including Artificial Neural Network (ANN), Support Vector Machine (SVM), and K-nearest neighbours (KNN) to predict 10-year hip fracture in one cohort. The models were then validated in another cohort. The data for training and validation included age, bone mineral density (BMD), clinical factors, and lifestyle factors which were ascertained prior to the fracture event.

Results of validation showed that each of the three machine learning algorithms yielded better predictive accuracy than the Cox's regression model. Although KNN and SVM had the same sensitivity as ANN (80.6%), the specificity and accuracy for KNN and SVM were lower than ANN. Using ANN, the accuracy of model I (which included only lumbar spine and femoral neck BMD) and model II (which included non-BMD factors) was 81% and 84%, respectively. When both BMD and non-BMD factors were combined (Model III), the accuracy increased to 87%. The AUC for model III was 0.94.

These findings indicate that machine learning is able to predict hip fracture more accurately than any existing statistical models, and that machine learning predicted fracture risk can help stratify individuals for clinical management.