Automated EEG-Based Classification of Nonclinical Depressive States via the Integration of Automatic Speech Recognition and a Pretrained Language Model

The early detection of depressive states is crucial for the effective prevention of clinical depression. To this end, a previous method enabled the classification of nonclinical depressive states using electroencephalogram (EEG) data recorded during a daily activity, such as listening to news, for continuous daily monitoring. However, this method required the manual annotation of word onset times and emotional valence, making its real-time application impractical. Furthermore, manual feature extraction from EEG responses cannot adequately address variability caused by differences in news content and participant characteristics such as age. To overcome these limitations, this study integrates automatic speech recognition to extract word onset times and a pretrained language model-based sentiment analysis to classify the emotional valence of the news contents, thereby enabling automated annotation. In addition, a convolutional neural network-based end-to-end classification framework is proposed to account for variability in EEG responses. In practical evaluation involving 186 participants aged 22-77, including 44 individuals who self-reported a depressive state, the proposed method outperformed the previous manual method. These findings demonstrate the feasibility of classifying nonclinical depressive states using automated EEG data analysis collected during passive listening.