Predictive Adherence Modeling (PAM) Study

Stanford University, Stanford CA

Investigators

Linked publications, trials & patents

Abstract

A major problem for both clinicians and patients is patient adherence. In the field of sleep medicine, patients with obstructive sleep apnea (OSA) have variable adherence to the gold standard treatment for this condition: continuous positive airway pressure (CPAP) therapy. The proposed Predictive Adherence Modeling (PAM) Study will use two large OSA datasets [the NHLBI-supported Apnea Positive Pressure Long-term Efficacy Study (APPLES) and the AHRQ-supported Comparative Outcomes Management with Electronic Data Technology (COMET) Study] and three NIDA-supported datasets, to accomplish three specific aims: (1) To construct a general, calibration-based approach for deriving prognostic definitions of adherence. The goal is to develop this approach by using adherence to continuous positive airway pressure for patients with obstructive sleep apnea as a testbed. (2) To develop a predictive model for adherence. Continuous measures of adherence (e.g., mean hours of adherence per night), will be used so that the outcome is kept at full resolution and highest information content, which maximizes opportunities for predictive models to distinguish among patients of differing behaviors. Adherence will also be operationalized as a multivariate outcome and predictive-modeling methods for multivariate outcomes will be used, in addition to modern regularized methods that will allow sifting through extensive lists of candidate predictors. The project will include methods that are specially designed to explore predictive interactions, such as regression trees, and we will allow for nonlinear predictors through use of various spline basis expansions, tree-based methods, and neural net technology. Ensemble methods will be employed, such as boosting, wherein many different regression models are fit and then combined to capitalize on their collective ability to predict outcome, and there will be correction for overfitting through use of validation techniques. Using these methods will allow the team to identify predictive models that are more robust, in that predictive performance will be sustained in other data sets. Further, the preceding techniques will be combined in order to construct models that optimize prediction of adherence. Finally, existing statistical methodology will be extended and adapted to the specific problem of adherence prediction, developing new statistical technology as needed. (3) To build a suite of statistical tools that will facilitate development of predictive models of adherence in any field of medicine. The plan is to develop a suite of statistical tools that will facilitate development of predictive models of adherence in any field of medicine, which will include three essential elements: (a) A description of the statistical methods contained within the suite in language accessible to non-statistician medical professionals. (b) A user-friendly package of code will be provided for the suite of statistical tools. This code will be provided in two languages, SAS® and the freeware R. (c) The code will include a number of visualization tools to facilitate interpretation and utilization of predictive models by clinical practitioners.