Abstract:
|
Modern large-scale observational psychiatric studies collect data in a plethora of modalities, including questionnaires, structured clinical interviews, life histories, and many biological variables, including, e.g., structural and functional brain imaging, genetics, inflammatory measures. An important goal of such studies is to obtain a biological foundation for psychiatric diagnoses that are predictive of outcomes and/or response to specific treatments. However, a major difficulty in analyzing data from these studies is reducing dimensionality via revealing latent structures that inform about relationships across modalities, while simultaneously accounting for "batch" effects and method variance within modalities of measurement. Here, we present a Bayesian multi-level model that uncovers both shared and idiosyncratic factors within blocks (data modalities). We demonstrate that this methodology is effective in uncovering latent structure and predicting clinical outcomes in the T-1000 data, a large-scale of psychiatric disorders collecting data in scores of domains, including structural and functional imaging.
|