HEAL-MinD: hybrid contextual and semantic embeddings with attention and linguistic features for mental health detection
TL;DR
A novel deep learning model referred to as HEAL-MinD to detect mental health based on online user posts is presented and achieves an F-score and accuracy of up to 95% and 96%, respectively, across all three datasets.
Abstract
The occurrence of events, such as data breaches, loss of sensitive information, and inappropriate use of an individual’s personal information often adversely impacts the mental health of users. Insightful analysis of such online complaint-based posts provides valuable insights into the mental health challenges faced by users. In this article, we present a novel deep learning model referred to as HEAL-MinD to detect mental health based on online user posts. The model begins with an input layer that processes the input text followed by the embedding layer which employs two types of embeddings to capture contextual and semantic information. A convolution gated and attention (CGAT) layer is then introduced which is one of the most important components of HEAL-MinD. In addition, a rich set of hand-crafted auxiliary features is incorporated into the model to develop it more accurate and informative. Finally, a dense layer performs the final classification, and the output layer produces the prediction by categorizing each input as either mental health or non-mental health. We conduct extensive experimental evaluations on three newly developed datasets to demonstrate the effectiveness of the HEAL-MinD model. The proposed model outperforms five existing studies and 12 baseline models. It achieves an F-score and accuracy of up to 95% and 96%, respectively, across all three datasets. The required source code to implement the HEAL-MinD model is available at GitHub: https://github.com/vishxl/HEAL-MinD .