Automatic Depression Assessment using Deep Learning Techniques
Depression assessment , Transfer learning , Natural language processing , Deep learning , Clinical decision support , Multi-task learning , Speech processing , Model calibration
Early recognition and treatment of depression can avert escalation of the mental disorder and alleviate suffering for the patients and their families. To assist mental health care providers with the early recognition and assessment of depression, research to develop a standardized, accessible, and non-invasive technique has garnered considerable attention. This work investigates the use of deep learning techniques for the task of automatic depression assessment. We propose two novel solutions to address current limitations faced by machine learning-powered depression assessment systems. Our first solution utilizes multi-task learning to improve upon calibration of a prediction model. Calibration is an important trait in developing trustworthy models for high-stake applications. Our second solution leverages recent advances in pretrained large language models and parameter-efficient tuning techniques to effectively learn with limited data. Using speech and spoken text extracted from clinical interviews, our resultant models achieve new state-of-the-art performance results on a benchmark depression dataset against strong baselines and previously published methods. Results from this work show promise in applying deep learning models to assist care providers with depression assessment.