End-to-End Processing-on-Chip Wearable Ear EEG Device with Tiny Neural Network for Multilevel Stress Detection

The study proposes an innovative method for predicting stress levels that employs electroencephalography (EEG) data collected from behind the ear (BTE) and incorporates tiny neural networks (TNN). The system, which consists of an EEG processing device and a companion smartphone application for monitoring, is designed to: (1) collect EEG signals from behind the ear rather than the scalp, prioritizing comfort and reducing hair interference; (2) filter out irrelevant signals from the desired EEG; (3) extract relative power spectral density (PSD) features from EEG data using Fast Fourier Transform (FFT); and (4) classify stress levels using a proposed TNN. Importantly, all processing stages occur on the chip itself, rather than on external devices such as PCs or servers, providing fast and synchronized processing and preventing data loss during transmission. In addition, (5) a user-friendly smartphone app is being created for rapid and straightforward stress level monitoring. Experimental results show that the on-chip TNN is successful, with accuracy of 90.83%, precision of 90.87%, recall of 90.83%, and an F1-score of 0.9081, all verified using leave-one-subject-out cross-validation (LOOCV). This merging of BTE EEG analysis with on-chip TNNs offers the potential for enhancing mental stress detection and other health-related uses.