Automatic detection of psychological distress indicators and severity assessment in crisis hotline conversations

Psychological health disorders pose a growing threat to society. Disorders such as Depression, Post-Traumatic Stress Disorder (PTSD), and mild Traumatic Brain Injury (mTBI), are often under-diagnosed and under-treated. Crisis hotlines are often the last resort for people who, from the lack of proper treatment, are considering suicide or intend to harm themselves or others. This paper describes a system that automatically analyzes online crisis hotline chats to (1) extract fine-grained distress indicators that map to Diagnostic and Statistical Manual of Mental Disorders (DSM) IV codes, and to (2) perform triage classification based on the severity of distress. For distress detection, we present several approaches which leverage annotator rationales and dialogue structure to improve classification performance, demonstrating significant gains over a state-of-the-art approach from literature. For triage classification, we demonstrate early detection capability for the most severe triage code. We evaluate our work on a large corpus of chats from the U.S. Department of Veterans Affairs' online Crisis Hotline.