Using effective connectivity-based predictive modeling to predict MDD scale scores from multisite rs-fMRI data

Background

Major depressive disorder (MDD) is a severe mental illness, and the Hamilton Depression Rating Scale (HAMD) is commonly used to quantify its severity. Our aim is to develop a predictive model for MDD symptoms using machine learning techniques based on effective connectivity (EC) from resting-state functional magnetic resonance imaging (rs-fMRI).

New method

We obtained large-scale rs-fMRI data and HAMD scores from the multi-site REST-meta-MDD dataset. Average time series were extracted using different atlases. Brain EC features were computed using Granger causality analysis based on symbolic path coefficients, and a machine learning model based on EC was constructed to predict HAMD scores. Finally, the most predictive features were identified and visualized.

Results

Experimental results indicate that different brain atlases significantly impact predictive performance, with the Dosenbach atlas performing best. EC-based models outperformed functional connectivity, achieving the best predictive accuracy (r = 0.81, p < 0.001, Root Mean Squared Error=3.55). Among various machine learning methods, support vector regression demonstrated superior performance.

Comparison with existing methods

Current phenotype score prediction primarily relies on FC, which cannot indicate the direction of information flow within brain networks. Our method is based on EC, which contains more comprehensive brain network information and has been validated on large-scale multi-site data.

Conclusions

Brain network connectivity features effectively predict HAMD scores in MDD patients. The identified EC feature network may serve as a biomarker for predicting symptom severity. Our work may provide clinically significant insights for the early diagnosis of MDD, thereby facilitating the development of personalized diagnostic tools and therapeutic interventions.