MeDiTate: A Deep Learning Framework for Mental Disorder State Detection in the Era of Social Media

This article presents MeDiTate, a hybrid deep learning framework proposed to detect mental health disorders by analyzing social media content. The Menta dataset is collected, consisting of user-labeled mental health issues across 13 different disorders. The proposed framework employs various natural language processing techniques for data preprocessing and integrates deep learning models, specifically convolutional neural network (CNN), long short-term memory (LSTM), bidirectional LSTM (BiLSTM), and transformer architectures. Ablation studies are conducted to find the top performing architecture. The proposed MeDiTate achieves an accuracy of 88.32%, an F1-score of 88.51%, precision of 88.61%, and recall of 88.45% on Menta dataset. The framework surpasses state-of-the-art methods on two publicly available datasets, namely Kim et al. (2020) and Low et al. (2020) datasets, by $\sim$2%. Additionally, we leverage the effectiveness of transfer learning from models trained on user-labeled datasets on the performance. The results demonstrate a noticeable improvement. Furthermore, we empirically investigate cross dataset evaluation to assess the framework’s generalizability. Comprehensive quantitative and qualitative analyses further validate the framework’s discriminatory performance. ^{1 1}

The code and dataset are available at https://github.com/NourElmadany/MeDiTate.