Integrating Slow Neural Oscillations and Physiological Burden for Trait Anxiety Prediction

Effective modeling of health outcomes from biomedical time series requires methods that capture both temporal and frequency dynamics. Trait anxiety, a transdiagnostic risk factor, manifests in neural activity and systemic physiology. We present a multimodal graph-attention framework that integrates resting-state fMRI time series with structural connectivity and allostatic-load biomarkers via cross-modal attention. Using 120 participants from the LEMON dataset, the model achieved modest but stable predictive accuracy. Within the brain branch, we systematically compared feature extraction strategies and found that preserving temporal order in slow-4/slow-5 oscillations was essential for prediction, while approaches discarding temporal structure consistently underperformed. Interpretability analyses highlighted limbic–visual circuits and metabolic–immune markers as reproducible contributors. These findings demonstrate that capturing temporal dynamics is critical in health time-series modeling, and show how multimodal graph-attention can provide both predictive value and interpretable digital biomarkers for anxiety vulnerability.