Real-time, objective anxiety detection has long been a gap in mental health care, where diagnosis depends almost entirely on self-report. A system that can read anxious brain states directly from EEG signals — and potentially guide biofeedback interventions — would shift anxiety management from reactive to continuous and personalized, particularly in underserved settings without psychiatrist access.
This proof-of-concept study built a convolutional neural network (CNN) trained on the GAMEEMO public EEG dataset, achieving 95.72% accuracy in classifying anxious versus non-anxious brain states. Researchers then tested transferability by collecting a small independent dataset from seven participants aged 18–60. The protocol deliberately induced acute cognitive stress via a custom Stroop color-word interference task, then elicited relaxation through a structured 4-7-8 breathing sequence. After domain adaptation to account for the shift from benchmark to real-world data, the same CNN architecture achieved 86.58% accuracy — a meaningful but expected drop, reflecting the inherent noise of small, heterogeneous, real-world EEG recordings.
The 9-percentage-point performance gap between the controlled benchmark and the live experimental cohort is arguably the most instructive finding here. It quantifies the translation penalty common to neural signal classifiers when laboratory conditions give way to individual variability, electrode placement inconsistency, and small sample sizes. Seven participants represent a fraction of what would be needed to assess clinical viability — this remains firmly in the feasibility stage. The CNN-to-neurofeedback pipeline concept is not new; prior work has coupled EEG classifiers with alpha-wave uptraining for stress reduction. What this study adds is a structured, deployable workflow using relatively affordable consumer-grade EEG hardware. For longevity-focused adults, the more distant but genuinely compelling implication is chronic stress monitoring: sustained anxiety accelerates cortisol-driven cellular aging, and a wearable that flags anxiety state in real time could support earlier intervention. This finding is incremental and exploratory, but the methodological scaffolding it establishes is the necessary first step toward clinically meaningful neurofeedback tools.
