基于特征筛选与机器学习的医疗保险报销比例预测研究

Study on health insurance reimbursement rate prediction by the combined method of feature selection and machine learning

目的对国家医疗保障疾病诊断相关分组中胸部大手术组的医疗保险报销比例进行数据驱动的辅助预测,为医疗保险经办机构及医疗机构精准有效地预测按病种分组医保支付情况提供参考。方法以四川省某大型三甲医院2020年的胸部大手术病例信息为样本,通过多元线性回归模型和基于特征筛选的机器学习改进方法,将全部数据集的70%作为训练数据集,30%作为测试数据集,对医保支出情况进行预测。结果随机森林、Logistic回归、支持向量机三种机器学习方法在筛选特征数量相同时,预测效果无统计学差异。预测效果最优的模型准确率为78.96%,敏感性为83.93%,特异性为71.27%,精确度为0.8188,AUC值为0.8414,Kappa值为0.6108。结论疾病诊断数量、手术操作数量及患者年龄对报销比例影响较大。治疗费、材料费、手术费及西药费为住院费用的主要方面。基于特征筛选的机器学习改进方法优于传统的统计线性模型,且选取合适的特征数量能够使模型在较高的效率下达到更好的预测效果。

Objective To perform data-driven,assisted prediction of health insurance reimbursement ratios for the major thoracic surgery group in CHS-DRG,in addition to providing an optional solution for health insurance providers and medical institutions to accurately and effectively predict the references of health insurance payments for the patient group.Methods Using the information on major thoracic surgery cases from a large tertiary hospital in Sichuan province in 2020 as a sample,70% of the total dataset was used as a training dataset and 30%as a test dataset.This data was used to predict health insurance spending through a multiple linear regression model and an improved machine learning method that is based on feature selection.Results When the number of filtered features was the same via three machine learning methods including random forest,logistic regression,and support vector machine,there was no significant difference in the prediction effectiveness.The model with the best prediction effect had an accuracy of 78.96%,sensitivity of 83.93%,specificity of 71.27%,precision of 0.8188,AUC value of 0.8414,and a Kappa value of 0.6108.Conclusion The basic characteristics such as the number of disease diagnoses and surgical operations,as well as the age of patients affect the reimbursement ratio.The cost of materials,drugs,and treatments has a greater impact on the reimbursement ratio.The combined method of feature selection and machine learning outperforms traditional statistical linear models.When dealing with a larger dataset that has many features,selecting the right number can enhance the prediction ability and efficiency of the model.