Explainable AI for Mental Health: Detecting Mental Illness from Social Media Using NLP and Machine Learning

Sidra Hameed¹Muhammad Nauman¹Nadeem Akhtar²Muhammad A B Fayyaz^3*Raheel Nawaz⁴

• ¹Islamia University, Bahawalpur, Pakistan
• ²University of Punjab, Lahore, Pakistan
• ³Manchester Metropolitan University, Manchester, United Kingdom
• ⁴staffordshire Universityu, Stoke-on-Trent, United Kingdom

Mental disorders are prevalent in modern society, causing significant personal and societal suffering due to various socioeconomic, clinical, and individual risk factors. Increasingly, users express their thoughts and emotions on social media platforms, which presents new opportunities to leverage user-generated content for addressing critical societal challenges, such as the early detection of mental disorders. This work specifically focuses on the early detection of depression, one of the most common mental disorders, using black box machine learning (ML) models, including Support Vector Machines (SVM), Random Forests (RF), Extreme Gradient Boosting (XGB), and Artificial Neural Networks (ANN). These models are combined with advanced Natural Language Processing (NLP) techniques such as TF-IDF, Latent Dirichlet Allocation (LDA), N-grams, Bag of Words (BoW), and GloVe embeddings to effectively capture linguistic and semantic features for depression detection. While these black box models achieve high accuracy, their lack of interpretability undermines trust and applicability in sensitive healthcare contexts. To mitigate this issue, we incorporate Explainable AI (XAI) methodologies, specifically LIME (Local Interpretable Model-agnostic Explanations), to enhance transparency and facilitate understanding of model predictions. The experimental results indicate that SVM outperformed RF, demonstrating a higher accuracy in detecting depression from the social media data. In contrast to the majority of prior depression detection studies that primarily optimise for classification accuracy, the present work accords equal importance to model interpretability. We systematically applied the Local Interpretable Model-Agnostic Explanations (LIME) technique, enabling a granular examination of the decision-making processes. This approach facilitated the identification of linguistic markers associated with depressive symptomatology that exhibit strong concordance with established findings in psychological research, thereby enhancing the transparency, interpretability, and potential clinical trustworthiness of the proposed models.