Abstract
INTRODUCTION: The neuroimaging biomarkers for anxiety diagnosis remain poorly understood. This study aims to identify cortical regions whose hemodynamic patterns, derived from resting-state functional near-infrared spectroscopy (fNIRS) combined with machine learning, may serve as potential biomarkers to assist in the assessment of anxiety cases. METHODS: The final sample comprised ninety participants in emerging adulthood: 50 anxiety patients and 40 healthy controls (HC). After demographic data collection and Generalized Anxiety Disorder-7 (GAD-7) administration, all participants underwent a 1-minute baseline measurement and a 5-minute resting-state fNIRS recording. Following a stratified 70:30 train-test split, all feature selection procedures were performed using the training set. Model tuning and stability were assessed by five-fold stratified cross-validation within the training set, and final performance was evaluated on the independent test set. RESULTS: The machine-learning GBDT model utilized 13 distinct features and exhibited enhanced efficacy in discriminating between patients with anxiety and healthy controls compared with the RF classification model typically employed for anxiety identification (AUC: 0.900 vs. 0.832, sensitivity = 0.921, and specificity = 0.709). SHAP-based interpretability analysis revealed that oxyhemoglobin (HbO) fluctuations in the prefrontal regions, particularly the dorsolateral/middle frontal/orbital middle frontal gyrus and sensorimotor-visual areas (precentral/middle occipital gyrus), emerged as potential predictors. CONCLUSION: The study provides a new perspective for the development of anxiety diagnosis tools and contributes candidate biomarkers for anxiety disorder identification and intervention.