Gradient Boosted Tree Approaches for Mapping European Organization for Research and Treatment of Cancer Quality of Life Questionnaire Core 30 Onto 5-Level Version of EQ-5D Index for Patients With Cancer
Gradient Boosted Tree Approaches for Mapping European Organization for Research and Treatment of Cancer Quality of Life Questionnaire Core 30 Onto 5-Level Version of EQ-5D Index for Patients With Cancer
复制标题
DOI：
10.1016/j.jval.2022.07.020
复制
发表时间：
2023-02-06
期刊：
影响因子：
4.5
通讯作者：
Shimozuma，Kojiro
中科院分区：
文献类型：
--
作者：
Hagiwara，Yasuhiro;Shiroiwa，Takeru;Shimozuma，Kojiro
本研究旨在将欧洲癌症研究与治疗组织生活质量问卷Core 30的直接和响应映射算法开发到基于梯度提升树（GBT）的5级EQ-5D指数，这是一种有前途的现代机器学习方法。方法我们使用癌症研究数据（903例患者的903个观察结果）的生活质量映射算法来训练gbt并测试其预测性能。在癌症生活质量映射算法研究中，纳入晚期实体肿瘤患者，同时评估欧洲癌症研究与治疗组织生活质量问卷Core 30和5级版EQ-5D。日本的值集用于直接映射，而日本和美国的值集用于响应映射。我们对训练数据集中（80%）的gbt进行交叉验证，并通过测试数据集中（20%）的均方根误差（RMSE）、平均绝对误差（MAE）和平均误差来测试预测性能。结果GBT方法在测试数据集中的RMSE和MAE比先前开发的基于回归的方法更大。测试数据集的平均误差在GBT方法中往往比以前开发的基于回归的方法要小。结论与回归方法相比，GBT方法对RMSE和MAE的预测性能没有提高。GBT方法的灵活性有可能分别减少健康状况不佳和健康状况良好的人的过度预测和低估。需要进一步的研究来确定机器学习方法在将非基于偏好的措施映射到健康效用方面的作用。
ObjectivesThis study aimed to develop direct and response mapping algorithms from the European Organization for Research and Treatment of Cancer Quality of Life Questionnaire Core 30 onto the 5-level version of EQ-5D index based on the gradient boosted tree (GBT), a promising modern machine learning method.MethodsWe used the Quality of Life Mapping Algorithm for Cancer study data (903 observations from 903 patients) for training GBTs and testing their predictive performance. In the Quality of Life Mapping Algorithm for Cancer study, patients with advanced solid tumor were enrolled, and the European Organization for Research and Treatment of Cancer Quality of Life Questionnaire Core 30 and 5-level version of EQ-5D were simultaneously evaluated. The Japanese value set was used for direct mapping, whereas the Japanese and US value sets were used for response mapping. We trained the GBTs in the training data set (80%) with cross-validation and tested the predictive performance measured by the root mean squared error (RMSE), mean absolute error (MAE), and mean error in the test data set (20%).ResultsThe RMSE and MAE in the test data set were larger in the GBT approaches than in the previously developed regression-based approaches. The mean error in the test data set tended to be smaller in the GBT approaches than in the previously developed regression-based approaches.ConclusionsThe predictive performances in the RMSE and MAE did not improve by the GBT approaches compared with regression approaches. The flexibility of the GBT approaches had the potential to reduce overprediction and underprediction in poor and good health, respectively. Further research is needed to establish the role of machine learning methods in mapping a nonpreference-based measure onto health utility.