P29 Variable Selection and Prediction for Longitudinal Data Using Bayesian Transfer Learning

The analysis of contemporary longitudinal health-science data can involve high-dimensional measurements of time-course gene expression data or brain images collected during different scanning sessions which are collected on a small number of patients. The estimation of such high-dimensional longitudinal models with limited sample sizes in a target study population poses substantial challenges and can lead to highly variable parameter estimates, poorly calibrated and unstable predictions, and low power in the identification of pivotal covariates that predict the outcomes. In many cases, multiple additional source datasets from related, but not necessarily identical, populations are available. Some of these sources datasets can be substantially larger than the target longitudinal dataset. In such instances, it becomes natural to borrow information from source datasets with similar covariates-outcome relations to improve the precision of the parameter estimates in the target data.
Transfer learning seeks to adeptly borrowing information from different source data cohorts for different data settings. We propose a novel Bayesian transfer learning model for longitudinal data using Bayesian mixed effects models. To enhance the accuracy of the parameter estimates and enable data-adaptive information borrowing, we leverage Bayesian mixture models for the discrepancies between fixed effects regression coefficients in source and target studies and for the means of the observation-specific time-trajectories and covariate effects. We define our Bayesian mixture models with the aim of minimizing the transfer of information from external source studies than would introduce large bias (i.e. these with large discrepancies). Through extensive simulation studies and real-data applications show that, compared to several alternative data analysis approaches, our Bayesian transfer learning model improves precision of the parameter estimates in the target study substantially and reduces the risk of bias when one or several source data sets are generated under population specific longitudinal models that are substantially different from the target mixed-effects longitudinal model.

Digital Health (e.g., Big Data, ML, AI)

ASA Biopharmaceutical Section Regulatory-Industry Statistics Workshop 2024