Predicting risk of mental health deterioration using multimodal data from the UK biobank

Introduction
Prospective prediction of mental health risk is critical for early intervention to reduce the burden of depression, anxiety, and cognitive disorders.
Objectives
We aimed to investigate both the associations and genetically informed relationships of biological markers and lifestyle factors with mental health outcomes, and to develop predictive models for future symptom worsening using UK Biobank data.
Methods
We analyzed multimodal data from 157,733 participants, including 2,911 plasma proteins, 2,126 brain imaging-derived phenotypes (IDPs), 77 lifestyle factors, and 103 mental health or cognitive questionnaires. We performed polygenic risk score-based phenome-wide association studies (PRS-PheWAS) and one-sample Mendelian randomization to assess phenotypic associations and genetically informed effects, followed by mediation analyses. Stacked machine learning models were developed to predict worsening depression, anxiety, and cognitive decline using biomarkers, questionnaire or lifestyle factors.
Results
We identified 220 plasma proteins and 138 IDPs associated with at least one symptom score. Among these, GDF15 and HGF were found to be associated with depression and cognitive outcomes, with IDPs from 11 brain regions acting as potential statistical mediators in these associations. Models based on plasma proteins or IDPs achieved Areas Under the Curve (AUCs) ranging from 0.803 to 0.817 for predicting worsening depression, anxiety and cognitive decline. In comparison, models based on questionnaires achieved AUCs ranging from 0.859 to 0.916.
Conclusion
Questionnaire and lifestyle data-based models have shown better predictive performance for mental health outcomes, supporting their use in scalable risk stratification. Biomarker-based models highlighted candidate proteins (e.g., GDF15 and HGF) and brain structural features potentially related to mental health outcomes, which warrant further investigation.