Classification of mental disorders is a crucial task in psychiatric diagnosis and treatment planning. Traditional diagnostic methods often rely on subjective assessments and predefined symptom categories, which can lead to variability and inaccuracy in diagnoses. To address these limitations, deep generative models (DGMs) have emerged as powerful tools for modelling complex data distributions and uncovering latent representations in mental health data. This paper presents a comprehensive approach to classifying mental disorders using DGMs, focusing on how these models can capture the underlying structure of mental health conditions based on large-scale datasets, such as neuroimaging data, clinical reports, and behavioural assessments. We explore several types of DGMs, including Variational Autoencoders (VAEs), Generative Adversarial Networks (GANs), Gaussian Generative Models (GGMs) and Autoregressive Models, to learn meaningful representations of mental disorders. These models are trained on datasets with labelled mental health diagnoses to generate latent feature spaces that reflect the underlying psychopathology. The learned representations are then used to classify mental disorders, providing a more data-driven, objective approach to diagnosis. Our findings indicate that DGMs can achieve high classification accuracy, particularly in distinguishing between closely related disorders, such as different mood or anxiety disorders. Moreover, DGMs offer the ability to generate synthetic data that can be used to augment training datasets, addressing issues related to data scarcity. The generative aspect of these models also provides potential insights into the etiology of mental disorders by identifying patterns in the data that correlate with specific diagnoses. This study underscores the potential of deep generative models to transform mental health diagnosis by offering a more nuanced, data-driven framework.