An NLP-Based System for the Screening of Clinical Depression رسالة ماجستير

Abstract

Clinical depression is a common mental disorder characterized by depressed mood and loss of interest/pleasure. Other symptoms include decreased energy, feelings of guilt or low self-worth, disturbed sleep or appetite, and poor concentration. According to the WHO, clinical depression affects over 75% of people worldwide. Usually, clinical depression screening is based on psychological evaluation through face-to-face interviews. Such conventional methods are time- consuming and error-prone. Previous research showed that it is possible to identify patients with clinical depression through the analysis of social media posts. However, analyses of social media content were neither stratified according to the various symptoms of clinical depression nor based on the exploitation of extermal resources of medical semantics. In addition, current screening solutions can be characterized by a loosely coupled nature, hence the low level of validity and reliability of these tools. In particular, the validity of these methods is not based on the mathematical structure of symptoms in a multidimensional construct. Accordingly, we developed a strongly coupled system that combines Natural Language Processing (NLP)% and medical knowledge resources to assist healthcare professionals in screening for clinical depression; deploying a reliable and efficient tool that can passively and automatically assist in identifying subjects with clinical depression symptoms based on their social media posts. Given the dire state of mental health services and resources, the proposed system is expected to address immediate clinical relevance as it integrates the Beck Depression Inventory IL (BD1-1I) with social media post-analysis for both English and Arabic languages.

For the English Language, we used the Cross-Language Evaluation Forum (CLEF) eRisk 2020

Jaiaset which is a global dataset that includes social media posts of depressed people with different severity levels, and posts of people who have signs of Pathological gambling with signs of self-hanming. Our goal in this context is to find whether social media posts reflect the presence over tne Arabic social media posts, as well as severity of clinical depression symptoms if the social media posts have been written using English Language. To do this, we use multiple semantic resources in the domain of psychiatry to map the content of social media posts to their corresponding symptoms of clinical depression in the BDI-II. We utilized the outcomes of this step to train the proposed system and develop a reliable and efficient screening procedure. As the produced results indicate, the utilization of proposed pipeline and integrating it with the Bidirectional Encoder Representations from Transformers (BERT) model have resulted in 87% accuracy rate, while integrating the proposed pipeline with other classifiers, such as Logistic regression, Naive Bays, XGBboost, Support Vector Machine (SVM( and Random Forest )RF) classifiers produced 55.8%, 50%, 47.4%, 57%% and 49%, respectively.

On the other hand, for the Arabic language, we have used (Arabic Sentiment Analysis 2021 Dataset) dataset lo develop an NLP-based pipeline to detect depressed subjects based on their social media posts that are written in Arabic. To evaluate the developed model, we have utilized multiple machine learning based classifiers using the Arabic dataset and achieved 81% accuracy rate using the SVM and 70% accuracy using the Logistic Regression (LR), NB and RE classifiers, respectively.