Department of Clinical Information Engineering, Division of Social Medicine, Graduate School of Medicine, the University of Tokyo, Bunkyo-ku, Tokyo, Japan.
Department of Clinical Information Engineering, Division of Social Medicine, Graduate School of Medicine, the University of Tokyo, Bunkyo-ku, Tokyo, Japan.
Int J Med Inform. 2018 Mar;111:90-99. doi: 10.1016/j.ijmedinf.2017.12.009. Epub 2017 Dec 30.
Since the launch of a nationwide general health check-up and instruction program in Japan in 2008, interest in strategies to improve implementation of the program based on predictive analytics has grown. We investigated the performance of prediction models developed to identify individuals classified as "requiring instruction" (high-risk) who were unlikely to participate in a health intervention program.
Data were obtained from one large health insurance union in Japan. The study population included individuals who underwent at least one general health check-up between 2008 and 2013 and were identified as "requiring instruction" in 2013. We developed three prediction models based on the gradient boosted trees (GBT), random forest (RF), and logistic regression (LR) algorithms using machine-learning techniques and compared the areas under the curve (AUC) of the developed models with those of two conventional methods The aim of the models was to identify at-risk individuals who were unlikely to participate in the instruction program in 2013 after being classified as requiring instruction at their general health check-up that year.
At first we performed the analysis using data without multiple imputation. The AUC values for the GBT, RF, and LR prediction models and conventional methods: 1, and 2 were 0.893 (95%CI: 0.882-0.905), 0.889 (95%CI: 0.877-0.901), 0.885 (95%CI: 0.872-0.897), 0.784 (95%CI: 0.767-0.800), and 0.757 (95%CI: 0.741-0.773), respectively. Subsequently, we performed the analysis using data after multiple imputation. The AUC values for the GBT, RF, and LR prediction models and conventional methods: 1, and 2 were 0.894 (95%CI: 0.882-0.906), 0.889 (95%CI: 0.887-0.901), 0.885 (95%CI: 0.872-0.898), 0.784 (95%CI: 0.767-0.800), and 0.757 (95%CI: 0.741-0.773), respectively. In both analyses, the GBT model showed the highest AUC among that of other models, and statistically significant difference were found in comparison with the LR model, conventional method 1, and conventional method 2.
The prediction models using machine-learning techniques outperformed existing conventional methods: for predicting participation in the instruction program among participants identified as "requiring instruction" (high-risk).
自 2008 年日本推出全国性的一般健康检查和指导计划以来，人们对基于预测分析的提高计划实施效果的策略越来越感兴趣。我们调查了为识别不太可能参加健康干预计划的“需要指导”（高危）个体而开发的预测模型的性能。
数据来自日本的一个大型健康保险联盟。研究人群包括 2008 年至 2013 年期间至少进行过一次一般健康检查且 2013 年被确定为“需要指导”的个体。我们使用机器学习技术基于梯度提升树（GBT）、随机森林（RF）和逻辑回归（LR）算法开发了三个预测模型，并比较了所开发模型的曲线下面积（AUC）与两种传统方法的 AUC。这些模型的目的是识别那些在当年的一般健康检查中被确定为需要指导但不太可能参加指导计划的高危个体。
首先，我们在没有多重插补的情况下进行了分析。GBT、RF 和 LR 预测模型和传统方法的 AUC 值分别为 0.893（95%CI：0.882-0.905）、0.889（95%CI：0.877-0.901）、0.885（95%CI：0.872-0.897）、0.784（95%CI：0.767-0.800）和 0.757（95%CI：0.741-0.773）。随后，我们在进行了多重插补的数据上进行了分析。GBT、RF 和 LR 预测模型和传统方法的 AUC 值分别为 0.894（95%CI：0.882-0.906）、0.889（95%CI：0.887-0.901）、0.885（95%CI：0.872-0.898）、0.784（95%CI：0.767-0.800）和 0.757（95%CI：0.741-0.773）。在这两种分析中，GBT 模型的 AUC 均高于其他模型，与 LR 模型、传统方法 1 和传统方法 2 相比，差异具有统计学意义。
基于机器学习技术的预测模型在预测“需要指导”（高危）参与者参加指导计划方面优于现有的传统方法。
