基于Boosting算法的医疗费用预测——以鼻咽癌为例

Medical Expenses Prediction Based on Boosting Algorithms—Using Data of Nasopharyngeal Carcinoma (NPC)

摘要: 本文的数据来源于广东省某肿瘤医院，共计2064个鼻咽癌病案，我们对其进行数据挖掘，并预测病人的医疗费用。本文通过以下四步对数据进行研究。首先，我们选取了病人的年龄、性别、TNM诊断分期以及住院天数等特征为预测变量。然后，基于回归决策树算法(CART)建立费用预测模型。其后，分别使用两种Boosting算法，AdaBoost和Gradient Boosting对已有模型进行改进。接着，通过直观比照和回归评价指标，分析三种算法建立的预测模型的效果并进行比较，得到效果最好的DBRT (Gradient Boosting Decision Tree)预测模型，其预测准确率约为85%。最后，通过特征重要度和部分依赖关系图，解释基于Boosting算法的模型的现实意义，为医疗保险资源的分配和单个病例预期费用提供了参考。

Abstract: The data of this paper come from 2064 cases of NPC in a Cancer Hospital of Guangdong Province. We mine the data and predict the medical cost per patient. This paper studies the data through the following four steps. First, we select the characteristics of patients’ age, gender, TNM diagnosis stage and length of stay as the prediction variables. Then, we build the cost prediction model based on the regression decision tree algorithm (CART). Then, two boosting algorithms, AdaBoost and gradient boosting, are used to improve the existing model. Then, through the visual comparison and regression evaluation index, the effect of the prediction model established by the three algorithms is analyzed and compared, and the best DBRT (gradient boosting decision tree) prediction model is obtained, with the prediction accuracy of about 85%. Finally, the significance of the model based on boosting algorithm is explained through the feature importance and partial dependency graph, which provides a reference for the allocation of medical insurance resources and the expected cost of a single case.