IEEE Access (Jan 2024)
Toward Early Detection of Depression: Detecting Depression Symptoms in Arabic Tweets Using Pretrained Transformers
Abstract
The COVID-19 pandemic and its associated setbacks have significantly impacted human mental health. Depression of various intensities has resulted due to a wide variety of losses that people have experienced. However, unlike physical illness, mental illness is still underestimated by patients themselves and by society due to various factors, such as the societal stigma of visiting a psychotherapist and being diagnosed with a mental health disorder. On the other hand, general practitioners can recognize signs of depression using the Patient Health Questionnaire (PHQ-9), which is used as a screening test for depression. The PHQ-9 questionnaire comprises nine questions that correspond to the nine symptoms of depression outlined in the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5). In this paper, we aim to detect the nine depression symptoms stated by DSM-5 from Arabic tweets, as recognizing the type of depression symptom is crucial in diagnosing depression. We used AraBERT and MARBERT pretrained transformers to classify tweets with depression symptoms. We also performed data augmentation using ChatGPT to balance the training set. The model was applied to a dataset consisting of 1,290 samples labeled with nine different symptoms, in addition to a ‘normal’ class which was also generated using ChatGPT. This work used four performance metrics to evaluate the models’ performance, which are accuracy, precision, recall, and F1 scores. The AraBERT and MARBERT transformers have yielded promising results, achieving accuracy, precision, recall, and F1 scores of 99.3%, 99.1%, 98.8%, and 98.9%, respectively, using the AraBERT transformer. While using the MARBERT transformer achieved accuracy, recall, precision, and F1 scores of 98.3%, 97.9%, 98.2%, and 98%, respectively.
Keywords
