Exploring the Capabilities of a Language Model-Only Approach for Depression Detection in Text Data

Depression is a prevalent and debilitating mental health condition that requires accurate and efficient detection for timely and effective treatment. In this study, we utilized the E-DAIC (Extended Distress Analysis Interview Corpus-Wizard-of-Oz) dataset, an extended version of the DAIC-WOZ dataset, which consists of semi-clinical interviews conducted by an animated virtual interviewer called Ellie, controlled by a human interviewer in another room. With 275 participants, the E-DAIC dataset represents a valuable resource for investigating depression detection methods. Our aim is to predict PHQ-8 scores through text analysis. Leveraging state-of-the-art speech processing, LLM-based text summarization, and a specialized depression detection module, we demonstrate the transformative potential of language data analysis in enhancing depression screening. By overcoming the limitations of manual feature extraction methods, our automated techniques provide a more efficient and effective means of evaluating depression. In our evaluation, we achieve robust accuracy on the development set of the E-DAIC dataset, with a Mean Absolute Error (MAE) of 3.65 in estimating PHQ-8 scores from recorded interviews. This remarkable performance highlights the efficacy of our approach in automatically predicting depression severity. Our research contributes to the growing evidence supporting the use of LLMs in mental health assessment, showcasing the role of innovative technologies in advancing patient care for depression.

I. Introduction

Depression is a widespread mental health condition with a significant global impact. It affects thoughts, behaviors, emotions, and overall well-being, impacting approximately 300 million individuals worldwide [1]. The subjectivity of the assessment process presents a significant challenge in providing effective treatment and care for depression, potentially leading to inaccurate evaluations [2]. Questionnaires, such as the Patient Health Questionnaire (PHQ) [3], heavily rely on patient responses, which may be compromised by the subjective nature of the questions. Efficient tools are crucial for practitioners who dedicate their valuable time to tasks like screening and seeking second opinions. However, relying solely on the PHQ as a screening tool can yield false positives or false negatives, hindering accurate results and challenging early depression diagnosis [2]. To address this challenge, researchers have explored behavioral indicators for automated depression detection and prediction [4], [5]. While cues like facial expressions and speech patterns have been investigated, text-based approaches leveraging advancements in natural language processing (NLP) and language models offer scalable and widely applicable depression screening. Enhancing language models allows for the automatic extraction of subtle features and patterns indicative of depression from text data.