Depression Detection in Text Using Long Short-Term Memory-Based Neural Structured Learning

Abstract

These days, depression is a common illness worldwide which varies from usual mood fluctuations to challenges in everyday life. When depressive symptoms stay for long in the form of moderate or high intensity, it may result in becoming a serious health condition such as moving towards suicide. The advancement of machine learning can help technologies to build smart technologies to detect depression from different sources of data such as text. This work proposes a robust approach to detect depression from text messages using Long Short-Term Memory (LSTM)-based Neural Structured Learning (NSL). First of all, a text dataset is collected from the queries submitted in a Norwegian youth forum and then efficient features are obtained based on pre-defined handcrafted robust features developed by focusing on the symptoms of depression that were defined by medical practitioners and psychologists, rather than applying typical word frequency-based features (e.g., conventional term frequency-inverse document frequency (TF-IDF) and onehot) where the frequency of the words in the texts are focused but not the importance of the words. After that, LSTM-based NSL approach is applied as deep learning method to train the features discriminating depression and non-depression. The trained LSTM-based NSL is then utilized later to detect depression in the testing text messages. Besides, to explain the machine learning model's decision, a popular explainable Artificial Intelligence (XAI) algorithm is applied, which is Local Interpretable ModelAgnostic Explanations (LIME). The proposed approach shows the superiority against the traditional approaches on the dataset consists of Norwegian text, achieving the mean accuracy of 99%. Though it has been applied on Norwegian dataset, but the proposed concept can however be applied on other datasets as well, using the translated features.