基于偏好多目标优化特征选择的骨质疏松预测

Osteoporosis prediction based on preference multi-objective optimization feature selection

为快速筛查糖尿病患者群体中患有骨质疏松的高风险个体,基于决策者偏好多目标优化算法建立糖尿病患者的骨质疏松预诊断模型.以糖尿病患者的临床数据为驱动,随机森林作为基础分类器,非支配排序遗传算法-Ⅱ作为优化搜索引擎进行特征选择,搜寻帕累托最优特征方案解集.基于决策者偏好需求开发替换分数计算器,以决定帕累托最优解集中的首选特征方案.与具有34个变量的全模型相比,最佳模型的准确率提高了5.13%,优于对比模型,同时使用的特征数减少了91.18%.对最佳模型进行的特征重要性排序分析,增强了模型的可解释性,并为骨质疏松临床诊断提供参考.基于提出方法可建立低成本、紧凑有效的骨质疏松预诊断工具,为糖尿病患者的骨骼健康管理提供支持.

For the purpose of rapidly screening individuals at high risk of suffering from osteoporosis in the diabetic population,an osteoporosis pre-diagnostic model of diabetic patients has been developed based on a multi-objective optimization algorithm with decision maker preferences.The search for the Pareto optimal feature solution set is driven by clinical data,with random forest as the base classifier and with non-dominated sorting genetic algorithm-Ⅱas the optimization search engine for feature selection.A replacement score calculator is developed based on decision maker preference requirements to determine the preferred feature solution in the Pareto optimal solution set.Compared with the full model with 34 variables,the best model improves the accuracy by 5.13%while using 91.18%fewer features,and outperforms the machine learning models compared.The feature importance ranking analysis of the best model enhanced the interpretability of the model and provides a reference for clinical diagnosis of osteoporosis.Finally,a low-cost,compact and effective pre-diagnostic tool for osteoporosis based on the proposed method can be established to support the management of diabetic patients′bone health.