Lauffenburger Julie C, Franklin Jessica M, Krumme Alexis A, Shrank William H, Brennan Troyen A, Matlin Olga S, Spettell Claire M, Brill Gregory, Choudhry Niteesh K
*Division of Pharmacoepidemiology and Pharmacoeconomics †The Center for Healthcare Delivery Sciences, Department of Medicine, Brigham and Women's Hospital and Harvard Medical School, Boston, MA ‡CVS Health, Woonsocket, RI §Data Science, Aetna, Hartford, CT.
Med Care. 2017 Jan;55(1):64-73. doi: 10.1097/MLR.0000000000000623.
With rising health spending, predicting costs is essential to identify patients for interventions. Many of the existing approaches have moderate predictive ability, which may result, in part, from not considering potentially meaningful changes in spending over time. Group-based trajectory modeling could be used to classify patients into dynamic long-term spending patterns.
To classify patients by their spending patterns over a 1-year period and to assess the ability of models to predict patients in the highest spending trajectory and the top 5% of annual spending using prior-year predictors.
We identified all fully insured adult members enrolled in a large US nationwide insurer and used medical and prescription data from 2009 to 2011.
Group-based trajectory modeling was used to classify patients by their spending patterns over a 1-year period. We assessed the predictive ability of models that categorized patients in the top fifth percentile of annual spending and in the highest spending trajectory, using logistic regression and split-sample validation. Models were estimated using investigator-specified variables and a proprietary risk-adjustment method.
Among 998,651 patients, in the best-performing model, prediction was strong for patients in the highest trajectory group (C-statistic: 0.86; R: 0.47). The C-statistic of being in the top fifth percentile of spending in the best-performing model was 0.82 (R: 0.26). Approaches using nonproprietary investigator-specified methods performed almost as well as other risk-adjustment methods (C-statistic: 0.81 vs. 0.82).
Trajectory modeling may be a useful way to predict costly patients that could be implementable by payers to improve cost-containment efforts.
随着医疗支出的增加，预测成本对于确定干预对象至关重要。许多现有方法的预测能力一般，这可能部分是由于未考虑支出随时间的潜在有意义变化。基于群体的轨迹建模可用于将患者分类为动态长期支出模式。
根据患者1年期间的支出模式对其进行分类，并使用前一年的预测指标评估模型预测处于最高支出轨迹和年度支出前5%患者的能力。
我们确定了美国一家大型全国性保险公司的所有全额参保成年成员，并使用了2009年至2011年的医疗和处方数据。
基于群体的轨迹建模用于根据患者1年期间的支出模式对其进行分类。我们使用逻辑回归和拆分样本验证评估了将患者分类为年度支出前第五百分位数和最高支出轨迹的模型的预测能力。模型使用研究者指定的变量和专有的风险调整方法进行估计。
在998,651名患者中，在表现最佳的模型中，对最高轨迹组患者的预测很强（C统计量：0.86；R：0.47）。表现最佳的模型中处于支出前第五百分位数的C统计量为0.82（R：0.26）。使用非专有的研究者指定方法的方法表现几乎与其他风险调整方法一样好（C统计量：0.81对0.82）。
轨迹建模可能是预测高成本患者的一种有用方法，付款人可以采用这种方法来改进成本控制措施。
