Journal of Brain Science and Mental Health
A Biomedical Computing Framework for Stress Severity As-sessment Using EEG Features and Ontology-Guided Trans-former Models
Abstract
Mohammed Mubasheeruddin Ahmed, Mohammed Tajuddin, Abdul Mateen Ahmed, Syed Raziuddin and Mudassar Ali Syed
Background and Objective: Accurate stress severity assessment is critical in biomedical informatics, as delayed detection may negatively impact mental health and decision-making. Existing approaches often rely on either textual or physiological signals, limiting robustness. This study proposes a multimodal framework integrating transformer-based language modeling, ontology- guided reasoning, and electroencephalography (EEG) features.
Methods: The proposed system combines contextual embeddings from a RoBERTa-based transformer, ontology-derived semantic features, and EEG biomarkers including alpha-band power, beta-band power, and frontal asym-metry. A weighted decision-fusion strategy integrates these modalities. The framework was evaluated using 10,000 annotated multilingual posts and EEG recordings from 200 participants. Performance metrics include precision, recall, F1-score, and false negative rate.
Results: The multimodal model achieved a precision of 0.88, recall of 0.90, and F1-score of 0.89, with a reduced false negative rate of 14%. The integration of ontology reasoning and EEG features improved sensitivity to moderate and severe stress cases.
Conclusions: Combining contextual, semantic, and physiological evidence enhances stress detection reliability. The proposed framework provides an in-terpretable and robust approach for biomedical stress assessment.