EEG based depression detection by machine learning: Does inner or overt speech condition provide better biomarkers when using emotion words as experimental cues?

BACKGROUND

Objective diagnostic approaches need to be tested to enhance the efficacy of depression detection. Non-invasive EEG-based identification represents a promising area.

AIMS: The present EEG study addresses two central questions: 1) whether inner or overt speech condition result in higher diagnositc accuracy of depression detection; and 2) does the affective nature of the presented emotion words count in such diagnostic approach.

METHODS

A matched case-control sample consisting of 10 depressed subjects and 10 healthy controls was assessed. An EEG headcap containing 64 electrodes measured neural responses to experimental cues presented in the form of 15 different words that belonged to three emotional categories: neutral, positive, and negative. 120 experimental cues was presented for every participant, each containing an "inner speech" and an "overt speech" segment. An EEGNet neural network was utilized.

RESULTS

The highest diagnostic accuracy of the EEGNet model was observed in the case of the overt speech condition (i.e. 69.5%), while a an overall subject-wise accuracy of 80% was achieved by the model. Only a negligible difference in diagnostic accuracy could be found between aggregated emotion word categories, with the highest accuracy (i.e. 70.2%) associated with the presentation of positive emotion words. Model decision was primarily influenced by electrodes representing the regions of the left parietal, the left temporal lobe and the middle frontal areas.

CONCLUSIONS

While the generalizability of our results is limited by the small sample size and potentially uncontrolled confounders, depression was associated with sensitive and presumably network-like aspects of these brain areas, potentially implying a higher level of emotion regulation that increases primarily in open communication.