Identifying Osteoporosis in Diabetic Individuals based on Multi-objective Optimization Method

Abstract:
To rapidly screen and long-term monitor high-risk individuals with osteoporosis in the diabetic population and reduce the incidence of osteoporotic fractures in patients, a cost-effective osteoporosis pre-diagnosis model was developed for diabetic patients by applying a hybrid multi-objective optimization method. With the objectives of maximizing the model prediction accuracy and minimizing the number of risk predictors, Random Forest was used as the basic evalua-tor and Non-dominated Sorting Genetic Algorithm-II was applied as the search engine to optimize feature solutions for finding the optimal feature solution. Based on the preference needs of decision-makers, a flexible replacement score calculator was developed to address the best-in-class problem existing in the Pareto optimal set. The pre-diagnostic model was developed based on real-world clinical data. The contribution of risk predictors in osteoporosis diagnosis decisions was also analysed using the importance ranking method. The results showed that compared to the full model trained with 34 features, the best model built by the proposed method arrived at 93.59% prediction accuracy with 55.88% fewer predictors, and outperformed other comparable models. The proposed method can be applied to develop cost-effective pre-diagnostic tools for the early monitoring and prevention of osteoporosis.