Prediction of depression and suicidality from social media activity using deep neural networks

Abstract

It's estimated that every 13 minutes, at least one American dies by suicide in the United States, making suicide the 2nd leading cause of death in the 15-34 age group (CDC, 2015). Suicide intervention is now one of the top priorities for public health agencies across the world, but unfortunately our screening efforts are still inadequate. In this study we propose a data-driven artificial intelligence approach towards more accurate and higher coverage screening for both suicidal ideation and depression. We trained deep neural networks on Facebook posts from 158 human subjects to distinguish depressed from non-depressed individuals and to distinguish suicidal from non-suicidal individuals, using their Beck Depression Inventory (BDI) and Columbia-Suicide Severity Rating Scale (C-SSRS) scores as the ground-truth labels. We achieved state-of-the-art accuracy, precision, and recall rates of 96%, 97%, and 94% respectively on the depression prediction task and accuracy, precision, and recall rates of 92%, 98%, and 83% respectively on the suicidality prediction task. In addition, the deep neural networks outperformed both a trained human psychiatrist as well as other standard machine learning-based classifiers on the same tasks. These results demonstrate that it is possible to achieve more accurate and timely screening for depression and suicidal ideation via deep learning, without the need to first seek out a physician or mental healthcare professional. This significantly lessens the time, effort, and money required to seek help, which is particularly important for low-income and marginalized communities with low access to adequate healthcare.