Journal Article

Predicting depression in healthy young adults: A machine learning approach using longitudinal neuroimaging data

Zhang, Haobo, Zhang, Ailing

2025

Overview

•Early identification of depressive symptoms may reduce associated societal costs.•Machine Learning selects predictors of subclinical depression from vast MRI data.•Generated models demonstrated varied but overall good performance for prediction.•Functional abnormalities in brain areas like the Orbital Gyrus predicted depression.•The best predictive model achieved 85 % accuracy and an AUC of 0.80 in healthy people. Accurate prediction of depressive symptoms in healthy individuals can enable early intervention and reduce both individual and societal costs. This study aimed to develop predictive models for depression in young adults using machine learning (ML) techniques and longitudinal data from the Beck Depression Inventory, structural MRI (sMRI), and resting-state functional MRI (rs-fMRI). Feature selection methods, including the least absolute shrinkage and selection operator (LASSO), Boruta, and VSURF, were applied to identify MRI features associated with depression. Support vector machine and random forest algorithms were then used to construct prediction models. Eight MRI features were identified as predictive of depression, including brain regions in the Orbital Gyrus, Superior Frontal Gyrus, Middle Frontal Gyrus, Parahippocampal Gyrus, Cingulate Gyrus, and Inferior Parietal Lobule. The overlaps and the differences between selected features and brain regions with significant between-group differences in t-tests suggest that ML provides a unique perspective on the neural changes associated with depression. Six pairs of prediction models demonstrated varying performance, with accuracies ranging from 0.68 to 0.85 and areas under the curve (AUC) ranging from 0.57 to 0.81. The best-performing model achieved an accuracy of 0.85 and an AUC of 0.80, highlighting the potential of combining sMRI and rs-fMRI features with ML for early depression detection while revealing the potential of overfitting in small-sample and high-dimensional settings. This study necessitates further research to (1) replicate findings in independent larger datasets to address potential overfitting and (2) utilize different advanced ML techniques and multimodal data fusion to improve model performance.

• Publisher: Elsevier Inc,Elsevier Limited,Elsevier
• Subject: Adult / Algorithms / Biomarkers / Brain / Brain - diagnostic imaging / Brain - physiopathology / Depression / Depression - diagnosis / Depression - diagnostic imaging / Depression - physiopathology / Female / Frontal gyrus / Functional magnetic resonance imaging / Humans / Learning algorithms / Longitudinal Studies / Machine Learning / Magnetic resonance imaging (mri) / Magnetic Resonance Imaging - methods / Male / Medical imaging / Mental depression / Mental disorders / Mental health / Neuroimaging / Neuroimaging - methods / Older people / Parahippocampal gyrus / Prediction models / Predictive modeling / Random forest / Support Vector Machine / Support vector machine (svm) / Teenagers / Young Adult / Young adults