With the widespread implementation of electronic health records (EHR), large health care datasets are available. Given these rich data sources, it is now possible to tailor treatments based on the patients’ characteristics and preferences. Leveraging data from multiple sources can further enable us to effectively use the data and provide better health care services. The task is challenging as different data sources may have different characteristics such as population composition; ignoring such differences will make the personalized treatment rule (PTR) developed for one site not generalizable to another site. In particular, we lack methods to systematically incorporate multi-source data for personalized treatment recommendations.
To fill these methodological gaps, we will build on the existing PCORI award for evaluating the University of Wisconsin (UW) complex case management (CCM) program, which is implemented for patients with two different health insurers: Medicare (mostly older or disabled patients) and Commercial (mostly younger patients).
We have estimated a PTR for Medicare patients that is currently being implemented in the UW electronic health record to support nurse and physician decisions to recommend patients for enrollment into CCM. However, the sample size for the Commercial population is not large enough for stable estimation. With this proposal, we hope to borrow strength from the Medicare population to develop a PTR for the Commercial population (i.e., the target population).
We aim to develop a generalizable PTR for CCM enrollment for the target population when various levels of source and target population information are available. In particular, for Aim 1, we will develop meta-analysis framework for fitting a linear PTR, which requires outcome, treatment, as well as covariate information for both populations. For Aim 2, we propose a framework to weight samples in the source population to overcome the problem of covariate shift, i.e. the distributions of covariates in the source population and target population are different. For Aim 3, we will develop criteria based on large margin classifier for selecting patients in the target population to apply the rule developed from the source population when we have individual level information on the target population. In addition, for all aims, our method is robust as it will directly model the optimal PTR, which could bypass the need to build a prediction model for the outcome.
The research described here will build innovative statistical methods to produce generalizable personalized treatment recommendations for various populations. Such rules can be incorporated into a health system’s EHR to support better treatment decision making and ultimately to provide better care for patients with complex conditions.
*All proposed projects, including requested budgets and project periods, are approved subject to a programmatic and budget review by PCORI staff and the negotiation of a formal award contract.