Analysis of the neural mechanisms of social anxiety based on EEG features and machine learning and construction of a diagnostic model

Social anxiety is a common psychological problem, and its accurate diagnosis and investigation of underlying neurophysiological mechanisms are of significant importance. This study aims to explore the neuroelectrophysiological characteristics and diagnostic value of social anxiety by integrating event-related potentials (ERP), event-related spectral perturbation (ERSP), and machine learning methods. A total of 128 participants with varying levels of social anxiety completed an emotional face Go/No-Go task, during which electroencephalographic (EEG) data were collected for ERP and time-frequency analyses. The results revealed that individuals with social anxiety exhibited abnormalities throughout the entire process, from early perception to late regulation: during the early processing stage, signs of hypervigilance (enhanced N1), deficits in conflict monitoring (abnormal N2 latency), and biases in emotional evaluation (flattened P2) were observed, accompanied by increased synchronization in the delta/theta frequency bands. In the late regulation stage, there were signs of disorganized classification functions (reversed P3), ineffective emotional maintenance (absence of LPP regulation), and excessive recruitment of compensatory attentional resources (enhanced alpha/beta desynchronization). These EEG indicators were found to be broadly correlated with symptoms of social anxiety. The support vector machine (SVM) model constructed from the selected features demonstrated excellent performance on the original dataset (accuracy 88.462 %, AUC 90.196 %) and maintained good generalization performance on an independent validation dataset (n = 38), achieving an accuracy of 68.421 % and an AUC of 80.000 %. This study provides new insights for the objective diagnosis of early social anxiety and the exploration of its neurophysiological mechanisms, while also indicating directions for the development of targeted intervention strategies. Future research should further expand the sample size, enhance ecological validity, and explore the potential for individualized prediction and neural modulation.