Machine Learning Approach for Prediction of Bone Mineral Density and Fragility Fracture in Osteoporosis

Osteoporosis is a prevailing bone disease, which weakens the bone and is one of the major factors of disability, especially in elderly persons. In this thesis, we developed various machine learning models to predict fracture risk of osteoporosis. These mod- els were built to base their predictions on genotype and phenotype data of patients. We performed two dierent types of analysis: fracture risk prediction (a classica- tion model) and bone mineral density (BMD) prediction (a regression model). For fracture risk prediction we implemented four dierent algorithms: logistic regression, random forest, gradient boosting, and multi-layer perceptron (MLP) based on dier- ent risk factors identied. We performed our experiments using 307 and 1103 Single Nucleotide Polymorphism (SNPs) with data from 5133 patients. For both 307 and 1103 SNPs the performance of MLP was the best with area under curve (AUC) of 0.970 and 0.981 respectively. Logistic regression had the worst performance among four models with AUC of 0.816 and 0.904. For BMD prediction we implemented linear regression, random forest, gradient boosting and MLP and as a performance metric we plotted mean squared error (MSE) versus number of iterations for both train and test set of data. The random forest performed the best in both cases with MSE of 0.004 and linear regression was the worst with MSE of 0.104 in the test data for both sets of SNPs