DataSheet1_Identification of target groups and individuals for adherence interventions using tree-based prediction models.PDF

Background: In chronically ill patients, medication adherence during implementation can be crucial for treatment success and can decrease health costs. In some populations, regression models do not show this relationship. We aim to estimate subgroup-specific and personalized effects to identify target groups for interventions.Methods: We defined three cohorts of patients with type 1 diabetes (n = 12,713), type 2 diabetes (n = 85,162) and hyperlipidemia (n = 117,485) from German claims data between 2012 and 2015. We estimated the association of adherence during implementation in the first year (proportion of days covered) and mean total costs in the three following years, controlled for sex, age, Charlson’s Comorbidity Index, initial total costs, severity of the disease and surrogates for health behavior. We fitted three different types of models on training data: 1) linear regression models for the overall conditional associations between adherence and costs, 2) model-based trees to identify subgroups of patients with heterogeneous adherence effects, and 3) model-based random forests to estimate personalized adherence effects. To assess the performance of the latter, we conditionally re-estimated the personalized effects using test data, the fixed structure of the forests, and fixed effect estimates of the remaining covariates.Results: 1) our simple linear regression model estimated a positive adherence effect, that is an increase in total costs of 10.73 Euro per PDC-point and year for diabetes type 1, 3.92 Euro for diabetes type 2 and 1.92 Euro for hyperlipidemia (all p ≤ 0.001). 2) The model-based tree detected subgroups with negative estimated adherence effects for diabetes type 2 (-1.69 Euro, 24.4% of cohort) and hyperlipidemia (-0.11 Euro, 36.1% and -5.50 Euro, 5.3%). 3) Our model-based random forest estimated personalized adherence effects with a significant proportion (4.2%–24.1%) of negative effects (up to -8.31 Euro). The precision of these estimates was high for diabetes type 2 and hyperlipidemia patients.Discussion: Our approach shows that tree-based models can identify patients with different adherence effects and the precision of personalized effects is measurable. Identified patients can form target groups for adherence-promotion interventions. The method can also be applied to other outcomes such as hospitalization risk to maximize positive health effects of an intervention.

背景：在慢性病患者中，实施过程中的药物依从性对于治疗的成功至关重要，并且可以降低医疗成本。在某些人群中，回归模型并未显示出这种关系。我们的目标是估计亚组和个性化的影响，以识别干预措施的目标群体。
方法：我们根据2012年至2015年德国索赔数据，定义了三个队列，包括12,713名1型糖尿病患者、85,162名2型糖尿病患者和117,485名高脂血症患者。我们估计了第一年内（覆盖天数比例）的依从性与随后的三年平均总成本之间的关联，同时控制了性别、年龄、Charlson合并症指数、初始总成本、疾病严重程度和健康行为替代指标。在训练数据上拟合了三种不同类型的模型：1）线性回归模型，用于估计依从性与成本之间的总体条件关联；2）基于模型的树模型，以识别具有异质依从性效应的患者亚组；3）基于模型的随机森林，以估计个性化的依从性效应。为了评估后者的性能，我们使用测试数据、森林的固定结构和剩余协变量的固定效应估计条件性地重新估计了个性化的效应。
结果：1）我们的简单线性回归模型估计了正的依从性效应，即1型糖尿病的总成本每增加10.73欧元/依从性点/年，2型糖尿病为3.92欧元，高脂血症为1.92欧元（所有p ≤ 0.001）。2）基于模型的树模型检测到2型糖尿病（-1.69欧元，队列的24.4%）和高脂血症（-0.11欧元，36.1%和-5.50欧元，5.3%）具有负估计依从性效应的亚组。3）我们的基于模型的随机森林估计了个性化的依从性效应，其中显著比例（4.2%–24.1%）的效应为负（高达-8.31欧元）。对于2型糖尿病和高脂血症患者，这些估计的精度很高。
讨论：我们的方法表明，基于树的模型可以识别具有不同依从性效应的患者，并且可以衡量个性化效应的精度。这些识别出的患者可以形成依从性促进干预措施的目标群体。此方法还可应用于其他结果，如住院风险，以最大化干预措施的健康效益。