From Bag-of-Words to Transformers: A Deep Dive into the Participation in the eRisk Early Risk Detection of Depression Tasks with Classical and New Approaches

Abstract

AbstractWorld Health Organization (WHO) describes depression as the leading contributor to global disability and names it a major reason for suicide. Regardless of the prevalence and seriousness of depression, only about 50% of all affected individuals are under professional treatment. Pandemics such as COVID-19 have shown that a global crisis with drastic changes in living conditions can lead to increased symptoms of anxiety and depression, making the demand for early risk detection more relevant than ever. This chapter describes the models and results submitted by the FHDO Biomedical Computer Science Group (BCSG) to the eRisk tasks of 2017 and 2018 on Early Detection of Signs of Depression. After those challenges, some extended models have been developed to see if the results could be improved. These tests include the use of language models, which will be described in this chapter as well. Besides using hand-crafted features based on linguistic metadata, there were also attempts to use emotions and sentiment lexica for the early risk detection of depression. For experiments with machine learning models, a new word embedding was trained on a large corpus of the same domain as the described task and is evaluated. Finally, different prediction thresholds and ensembles of the developed models are used to explore the possible improvements, and the proposed alternative early detection metrics (\(F_{latency}\) and ERDE\(_{o}^\%\)) are evaluated.