Explaining Models of Mental Health via Clinically Grounded Auxiliary Tasks

Models of mental health based on natural language processing can uncover latent signals of mental health from language. Models that indicate whether an individual is depressed, or has other mental health conditions, can aid in diagnosis and treatment. A critical aspect of integration of these models into the clinical setting relies on explaining their behavior to domain experts. In the case of mental health diagnosis, clinicians already rely on an assessment framework to make these decisions; that framework can help a model generate meaningful explanations.In this work we propose to use PHQ-9 categories as an auxiliary task to explaining a social media based model of depression. We develop a multi-task learning framework that predicts both depression and PHQ-9 categories as auxiliary tasks. We compare the quality of explanations generated based on the depression task only, versus those that use the predicted PHQ-9 categories. We find that by relying on clinically meaningful auxiliary tasks, we produce more meaningful explanations.