Transformers in depression detection from semi-structured psychological Interviews

Abstract

The expansive adoption of Transformer models across the Machine Learning landscape is undeniable, and health is not an exception. This study undertakes a rigorous exploration of the efficacy of these novel architectures in discerning depression indicators from semi-structured psychological interviews. A key focus of this study is the extrapolation of the pre-training knowledge inherent in these models, and the comparison with traditional state-of-the-art Machine Learning models. In doing so, the thesis proposes a comprehensive framework designed to facilitate objective comparison. The study extends its inquiry into the differential performance of text and speech modalities, in isolation and combination, within the context of depression detection. Moreover, this research delves into the importance of topical relevance in the detection process, culminating in an evaluative discussion of crucial themes integral to accurate depression detection. Ultimately, this thesis contributes to the deepening understanding of the complex interplay between Transformer models, modality use, and topic importance in the realm of depression detection.