Detection and prediction of Future Mental disorder from Social Media Data using Machine Learning, Ensemble Learning, and Large Language Models.

Abstract

Social media platforms are used widely by all people to express their feelings, opinions, and emotional states. Billions of people worldwide use them daily to share what they think and feel in their posts. Amongst all social media available platforms, Facebook only contains around three billion personal accounts. In this work Reddit dataset is used to automatically detect mental illness from social media posts. This study is not only limited to early detection of already existing mental illness or disorder like depression and anxiety from social posts, but also and most importantly the study is extended to predict successfully potential mental illness that would happen in future. This study deploys Nineteen different models to study the capability of them in detecting and predicting mental disorders from social media posts. Some of the deployed models are classical machine learning classifiers, some are ensemble learning models, and the rest are large language models (LLMs). Six machine learning classifiers were used in this work for the automatic detection and prediction of mental illness and logistic regression proved to be the best amongst other classifiers in this task. Nine Ensemble methods were also used and examined. Amongst the Nine ensemble learning models VC2, Light GBM, Bagging estimator, and XGBoost proved to be superior in this task. Four large language models were also used and examined for the same task. RoBERTa and OpenAI GPT proved to outperform the rest of models in this task. All those models were built, trained, tested, and compared with previous work in literature to get the best possible results. The study covers the main four mental disorders which are ADHD, Anxiety, Bipolar, and Depression. The work proposed in this paper succeeded in outperforming the results in literature in terms of number of addressed mental disorders, number of models used and tested, and dataset size used to validate results. The proposed work also outperformed the only attempt in literature that addressed all mental disorders in results of detection and prediction noticeably. This work achieved the detection of already existing mental disorders F1-score of 0.80 from clinical data and of 0.52 from non-clinical data, and it achieved a prediction of future mental disorder F1-score of 0.43 from non-clinical data.