Mental health disorders remain a public health issue that poses critical challenges globally and there remains a need for effective methods of diagnosis, so that appropriate intervention and treatment can be undertaken in a timely manner. The present study outlines a methodological advancement using Variational Autoencoders (VAEs) for data augmentation in classifying mental disorders. The VAEs learn latent representations from the original dataset and generate synthetic samples, which can be used for data augmentation to mitigate the widely recognized problem of data imbalance seen in clinical datasets. The augmented dataset created using VAEs was also used to train and evaluate the performance of three of the most popular machine learning classifiers available - Random Forest, XGBoost and Linear Support Vector Machine. Findings indicate that VAE-generated samples can provide a degree of increase in performance of the model in terms of the varied metrics. With respect to precision, the very best model was XGBoost, whereas Random Forest and Linear SVM persistently achieving strong performance under the metrics of accuracy, recall, and F1-score. The findings suggest VAE-based data augmentation as valid option for improving model robustness and generalization in mental health diagnostics. It is anticipated that the methodological contribution will improve predictive performance, more fairly and robustly diagnostic tools in order to foster more personalized and data-driven mental health care.