AI-Based Multi-Class Mental Health and Stress Level Detection using Ensemble Machine Learning
- Version
- Download 12
- File Size 654.76 KB
- File Count 1
- Create Date 19 March 2026
- Last Updated 19 March 2026
AI-Based Multi-Class Mental Health and Stress Level Detection using Ensemble Machine Learning
Dr. Y. Mohammed Iqbal1, N. Mohamed Irfan2, Dr. S. Peerbasha3, Dr. M. Mohamed Surputheen4, Dr.
M. Rajakumar5
Department of Computer Science, Jamal Mohamed College, Affiliated to Bharathidasan University,
Tiruchirappalli, Tamil Nadu, India
ABSTRACT-The way we use social media has changed a lot over the years; now, it is a primary means of expressing emotions and relating to others. Because of this, social media has created an entirely new form of digital content that can be analyzed for health purposes. However, manual analysis of large volumes of unstructured text data is impractical and subject to errors. Therefore, this paper presents an end-to-end Machine Learning framework capable of automatically detecting multiple mental health conditions, including Anxiety, Depression, Stress,Bipolar Disorder, Suicidal Behaviour, and Personality Disorders. The framework is built on the idea that multiple types of mental health states exist and should be classified separately. Whereas most current classification models can only classify text as "Stress" or "No Stress", our proposed model is equally capable of classifying text into seven different types of psychological states, allowing for a more detailed analysis with respect to the socio-cultural context of individuals living in Tamil Nadu. To provide greater robustness and decrease the variance of the predictions generated by our proposed model, we developed a Weighted Soft-Voting Ensemble Framework (Fusion Model) which combines Logistic Regression (for interpretability) with Random Forest (fornon-linear prediction). We employed a dataset that contained 52,681 records to develop the models, which made use of advanced Natural Language Processing (NLP) techniques such as Lemmatization and TF-IDF vectorization for effective feature extraction. Testing results revealed that the proposed Fusion Model had a betteroverall accuracy of 77.17% in comparison with both types of standalone models when evaluated by means of 80-20 split test methodology outlined in this study.
Keywords: Mental Health Analytics, Ensemble Learning, Fusion Model, Natural Language Processing (NLP), TF-IDF, Soft Voting, Multi-Class Classification.
Download