A Culturally Adapted AI Framework for Early Mental Health Disorder Detection Among Indian Students

Abstract

Mental Health Issues are now a global concern, contributing and capable of taking lives away from students. Especially in India every year more than 13,000 student's suicidal deaths due to Mental health issues and even with preventive measures of the government, improvements are nearly insufficient which stands with availability paradox in mental health cases (Availability doesn't prove the ability of solution). Though the Artificial Intelligence Systems are performing good in predicting and classifying the underlying issue, still it is highly proved with English text but countries with language diversity not only use proper English for communication which becomes a barrier for classification models. youngsters highly prefer Romanised text in social media for expedient and ubiquitous access seeks attention and increases its popularity. Standard English classification models are insufficient and have no special pattern recognition to find the nuances in Romanised text which cause predictions into random guesses, this degradation determines the unreliability of risk prediction in real world settings. We conducted an experiment to prove the inability of classification models with Romanised text, we used RoBERTa-Base and trained it with 1000 English mental health context of labels such as low stress, medium stress and high stress and tested against 200 test Tanglish(Romanized version of Tamil language) samples, there are no bench mark datasets available in Romanised text of any Indian language as well as there are no Neural Machine Translators or Language Service Providers hence we synthetically prepared Tanglish dataset of 200 samples for testing purpose. The results showed the accuracy of $38.50 \%$ and F1 of 28.42%. Again, we took RoBERTa-Base and trained it with synthetically made Tanglish dataset of 1000 samples and tested against same test samples. The results showed accuracy of $92.00 \%$ and F1 score of $91.91\%$ proving the inaccuracy of romanization effect and gap of NLP to fit with diverse cultural environment of India. The improvements made by 53.50% (138.96% relative improvement) proved our proposed system works if proper Romanised benchmark dataset is used and also the presented framework promises romanization regulations and proper escalation system to save lives in real world situations.