BERT based Deep Learning Approach for Classification of Mental Health Disorders

The rise of deep learning models for classification of mental disorders proven promising results. The transformer based CNN models like BERT is fairly accurate in processing neuro images, EEG signals, social media texts, clinical records and patient reviews from cancer treatment sites. This paper presents BERT based model trained with mental health data taken from Kaggle containing around 35k responses and 19 mental features. The major findings of this research highlights the accurate classification of mental states over traditional diagnostic methods. The train and test losses are collected for 100 epochs for the two BERT baselines. The strong correlation between the features relating to optimism & diagnosed-mapped (0.57) and concentration feature with sexual activity (0.63) shows that individuals with mental disorders tend to overthink more than the normal persons. The proposed model achieves 89% accuracy in classifying mental health in terms of bipolar disorders type I, II and depression.