A0729
Title: A non-parametric integrative Bayesian approach for variable selection and prediction
Authors: Thierry Chekouo - University of Minnesota (United States) [presenting]
Abstract: Linear models may not adequately capture complex, nonlinear associations between outcomes and features. Moreover, with advancements in technology, data from multiple platforms are now collected on the same individuals, necessitating methods that effectively integrate multi-platform data. To address these limitations, a nonparametric Bayesian variable selection approach that employs Gaussian process priors is proposed to flexibly model the response surface within a model-based data integration framework. The novelties of the method lie in integrating multi-view data using multiple kernels in a Bayesian framework, allowing for simultaneous variable selection for each data type and accurate prediction. The proposed model measures the importance of each data view in predicting clinical outcomes while performing view-specific variable selection. A key feature of the method is that it can also be utilized in accelerated failure time (AFT) models when dealing with censored time-to-event outcomes. Several simulation studies are presented where we demonstrate the capability of the approach to detect significant variables across various data platforms and to predict outcomes effectively.
