# Applying machine-learning and deep-learning to predict depression from brain MRI and identify depression-related brain biology

## TL;DR

This study uses machine learning and deep learning to predict depression from brain MRI scans and identify brain regions linked to the condition.

## Contribution

The study introduces brain-based predictors of depression using machine learning and deep learning techniques on grey-matter MRI data.

## Key findings

- The BLUP predictor showed a significant association with MDD status (AUC = 0.57) in the UK Biobank.
- The BLUP predictor partially overlapped with a polygenic score but also captured a unique signal (combined AUC = 0.66).
- The morphometricity of MDD was estimated at 0.061, indicating limited potential for grey-matter-based predictors.

## Abstract

The accuracy of grey-matter predictors of depression has remained limited. In this study, brain-based predictors of major depressive disorder (MDD) were trained using machine-learning (Best Linear Unbiased Predictors [BLUP]) and deep-learning (ResNet3D) techniques applied to high-dimensional (voxel-wise) grey-matter structure extracted from T1-weighted structural MRI. The training sample comprised 987 MDD cases and 3934 controls from the UK Biobank. Predictors were evaluated in an independent sub-cohort of 483 MDD cases and 1939 controls from the UK Biobank and replicated in a clinical cohort (DEP-ARREST CLIN) of 64 cases and 32 controls. In the UK Biobank, the BLUP predictor showed a significant association with MDD status (AUC = 0.57; OR = 1.28 [1.15-1.43]; p-value = 1.1×10-5), which was confirmed in both males and females. By partitioning the BLUP predictor by brain regions of interest (ROI), we found nominal significance supporting the contribution of previously identified MDD-related ROIs (e.g. hippocampus and amygdala), though none passed multiple testing correction. The BLUP predictor overlapped partially with a polygenic score (PGS) of major depression (AUC = 0.65) but also captured a nominally significant signal that was not captured by the genetic score (combined AUC = 0.66, p-value = 0.024 when compared to PGS alone). No association passed multiple testing correction in the DEP-ARREST CLIN cohort, likely due to the small sample size. In contrast, the deep-learning predictor was not associated with MDD after multiple testing corrections. We estimated the morphometricity of MDD to be 0.061, implying limited potential of a brain-based predictor based on grey-matter structure (maximal AUC = 0.64). While the modest AUC values reiterate the challenge of developing brain-based MDD predictors for clinical applications, our predictors inform future research to explore brain-based relationships between MDD and comorbidities.