Machine learning (ML) techniques have become essential tools in the identification and classification of mental health disorders. Despite significant advances, limited research has specifically addressed the improvement of diagnostic processes and the exploration of factors contributing to the rising incidence of social phobia, or social anxiety disorder (SAD). SAD is characterized by a persistent and intense fear of social interactions, often resulting in avoidance behaviors that impair the formation and maintenance of interpersonal relationships. Accurate estimation of anxiety levels is critical for identifying at-risk individuals and guiding personalized therapeutic interventions. This study analyzes a dataset that integrates sociodemographic characteristics, clinical symptoms, and specific dimensions related to fear of social interaction scored on a scale from 0 to 10. In addition, three machine learning algorithms are implemented to classify the level of anxiety, and the Boruta algorithm is used to extract the most representative characteristics associated with SAD, improving the diagnosis and identifying the most representative features of this condition. Furthermore, it is important to highlight the interest in contributing to the development of personalized medicine; therefore, in this research, the SHAP algorithm was applied to interpret the results generated by the implemented machine learning models. The best performance is shown by the logistic regression algorithm with an AUC of 0.8464, which was achieved using the six features that passed the Boruta test.