Inferential methods for functional data from wearable devices

点击翻译按钮获取中文摘要

Project Summary/Abstract This is a project to develop new statistical methods for comparing groups of subjects in terms of health outcomes that are assessed using data from wearable devices. Inexpensive wearable sensors for health monitoring are now capable of generating massive amounts of data collected longitudinally, up to months at a time. The project will develop inferential methods that can deal with the complexity of such data. A serious challenge is the presence of unmeasured time-dependent confounders (e.g., circadian and dietary patterns), making direct comparisons or borrowing strength across subjects untenable unless the studies are carried out in controlled experimental con- ditions. Generic data mining and machine learning tools have been widely used to provide predictions of health status from such data. However, such tools cannot be used for significance testing of covariate effects, which is necessary for designing precision medicine interventions, for example, without taking the inherent model selection or the presence of the unmeasured confounders into account. To overcome these difficulties, a systematic de- velopment of inferential methods for functional outcome data obtained from wearable devices will be carried out. There are three specific aims: 1) Develop metrics for functional outcome data from wearable devices, 2) Develop nonparametric estimation and testing methods for activity profiles and a screening method for predictors of activity profiles, 3) Implement the methods in an R package and carry out two case studies using accelerometer data. For Aim 1, the approach is to reduce the sensor data to occupation time profiles (e.g., as a function of activity level), and formulate the statistical modeling in terms of these profiles using survival and functional data analytic meth- ods. This will have a number of advantages, the principal one being that time-dependent confounders become less problematic because the effect of differences in temporal alignment across subjects is mitigated. In addition, survival analysis methods can be applied by viewing the occupation time as a time-to-event outcome indexed by activity level. For Aim 2, nonparametric methods will be used to compare and order occupation time distributions between groups of subjects that are specified in terms of baseline covariate levels or treatment groups. Further, a new method of post-selection inference based on marginal screening for function-on-scalar regression will be developed to identify and formally test whether covariates are significantly associated with activity profiles. Aim 3 will develop an R-package implementation, and as a test-bed for the proposed methods they will be applied to two Columbia-based clinical studies: to the study of physical activity in children enrolled in New York City Head Start, and to the study of experimental drugs for the treatment of mitochondrial depletion syndrome.