Integrating Slow Neural Oscillations and Physiological Burden for Trait Anxiety Prediction

Jungyoun Janice Min, Jiong Chen, Jingxuan Bao, Shu Yang, Yize Zhao, Li Shen, Duy Duong-Tran

Keywords: Temporal Feature Extraction, Resting-State fMRI, Multimodal, Allostatic Load markers, Trait Anxiety, Graph Attention Network

TL;DR: Preserving slow-4/slow-5 temporal dynamics with multimodal graph attention yields stable and interpretable brain–body markers for trait anxiety

Abstract: Effective modeling of health outcomes from biomedical time series requires methods that capture both temporal and frequency dynamics. Trait anxiety, a stable disposition characterized by heightened anticipatory stress across contexts, manifests through both neural and systemic physiological burden, yet existing approaches rarely integrate these modalities. We present a graph-attention framework that models brain functional dynamics over structural connectivity and integrates them with allostatic-load-related blood biomarkers via cross-modal attention. In 120 young adults from the LEMON dataset, we systematically evaluated four feature extraction strategies and demonstrated that preserving temporal order in anxiety-relevant slow-4/slow-5 oscillations (0.01–0.073 Hz) was critical for stable prediction, while temporal order-discarding approaches consistently underperformed. Multimodal integration provided additional gains over brain-only models. Model interpretability analyses revealed that limbic and visual networks, along with metabolic and immune markers (creatinine, glucose, C-reactive protein), served as the most informative features. Our results show that temporal dynamics in neural oscillations are essential for modeling psychiatric vulnerability and establish a framework for integrating brain–body signals into interpretable digital biomarkers for mental health.

Submission Number: 64