Controlling Confounds in Biomarker-Based Clinical Prediction Models

In recent years, brain-based clinical prediction models have shown great promise in identifying biomarkers for psychiatric and neurological disorders. These models, which often rely on neuroimaging data, are increasingly used to predict clinical outcomes and guide personalized treatment decisions. However, their reliability is compromised by confounding variables — i.e. factors such as demographic characteristics or baseline clinical measures — that influence both brain data and the outcomes of interest. Without proper adjustment, confounds can lead to misleading conclusions, casting doubt on the validity of brain-behavior associations.

Research Project

This project (funded by the IMF Münster) seeks to systematically evaluate existing methods for adjusting confounding variables in brain-based prediction models. Building on this evaluation, we will develop and propose novel approaches that offer more robust solutions to confound adjustment. Central to this project is the development of a new toolbox based on PHOTONAI, an established machine learning framework within the Medical Faculty of Münster, designed specifically to address confounding issues. This toolbox will provide a comprehensive suite of tools for confound correction, optimizing model accuracy, and facilitating the development of clinically useful predictive models.