Classification of Major Depressive Disorder Using EEG-Derived Brain, Ocular and Muscular Independent Components: A Preliminary Study

Introduction

Electroencephalography (EEG) is widely employed for the automatic detection of Major Depressive Disorder (MDD), with most approaches focusing primarily on neural activity while discarding non-brain-derived components as artefacts. However, these non-neural sources may contain clinically relevant information, as they reflect the output of underlying neurophysiological processes.

Material and methods

In this study, we propose an EEG analysis framework that leverages components traditionally labeled as artefacts. Independent Component Analysis (ICA) was applied to a publicly available dataset comprising 34 subjects with depressive symptoms and 30 healthy controls to extract three categories of independent components (ICs): posterior brain ICs (Pb ICs), ocular ICs (Eye ICs), and electromyographic ICs (EMG ICs).

Features were extracted using the Wavelet Scattering Transform (WST), and classification was performed using a Long Short-Term Memory (LSTM) network. Model performance was evaluated across multiple values of the invariance scale (IS) parameter using a subject-level holdout split.

Results

The results indicate that classification performance was strongly dependent on the invariance scale (IS) parameter. Ocular ICs consistently provided higher discriminative performance across several configurations, whereas EMG ICs showed limited sensitivity. In contrast, posterior brain ICs did not exhibit meaningful discriminative performance, suggesting that reliable class-separating patterns could not be learned under the current experimental setup.

Conclusions

These findings suggest that non-neural EEG components, particularly ocular activity extracted via ICA, may contain useful information for MDD classification. However, given the use of a single subject-level split and the relatively small dataset size, the results should be considered preliminary. Further studies using more robust validation strategies and larger datasets are required to confirm the generalizability of this approach.