Regulatory-Compliant EEG Stress Recognition Using CNN–Transformer Models with Feature Unlearning

This is the authors' abstract. We don't add key points for this paper.

Stress recognition based on EEG data can be useful for healthcare, workplace safety, and mental health monitoring, among other applications. However, most studies do not consider how sensitive EEG data can reveal users’ cognitive and affective information. Existing models also do not offer participant data erasure methods that address compliance with data protection regulations (e.g., GDPR). In this regard, we present an improved hybrid Convolutional Neural Network (CNN) and Transformer model with Selective Feature Unlearning. The CNN improves the model’s ability to classify EEG signals based on spatial features, while the Transformer model improves classification by capturing the long-range temporal features of EEG data. A major novelty of the proposed framework is the use of Machine Unlearning (MU) to remove participant-specific data without requiring full model retraining. The proposed model also demonstrates that the privacy-aware unlearning model achieves a higher classification accuracy of 96.59%, compared with the baseline model’s accuracy of 95.67%. This EEG-based stress recognition system preserves participant privacy without compromising classification performance. The proposed approach can also be adapted to other stress recognition systems and privacy-sensitive applications.