Abstract
Electroencephalography (EEG) measures electrical signals on the scalp and can give information about processes near the surface of the brain (cortex). The goal of our research was to create models that predict depressivity (mapping to personality in general, not just sickness) and to find potential biomarkers in EEG data. First, to provide our models with cleaner EEG data, we designed a novel single-channel physiology-based eye blink artefact removal method and a mains power noise removal method. Then, we assessed two main machine learning model types (classification- and regression-based) with a total of eighteen sub-types to predict the depressivity of participants. The models were generated by combining four signal processing techniques with a) three classification techniques, and b) three regression techniques. The experimental results showed that both types of models perform well in depressivity prediction and one regression-based model (Reg-FFT-LSBoost) showed a significant depressivity prediction performance, especially for female group. More importantly, we found that a specific EEG frequency band (the gamma band) made major contributions to depressivity prediction. Apart from that, the alpha and beta band may make modest contributions. Specific locations (T7, T8, and C3) made major contributions to depressivity prediction. Frontal locations may also have some influence. We also found that the combination of both eye states’ EEG data showed a better depressivity prediction ability. Compared to the eyes closed data, the EEG data obtained from the state of eyes open were more suitable for assessing depressivity. In brief, the outcomes of this research provided the possibilities for translating the EEG data for depressivity measure. Furthermore, there are possibilities to extend the research to apply to other mental disorders’ prediction, such as anxiety.