Screening for Suicidal Ideation with Text Messages

Suicide is a leading cause of death in the US, with suicide rates increasing annually. Passive screening of suicidal ideation is vital to provide referrals to at-risk individuals. We study to what degree smartphone-based communication, in particular, text messages, could be leveraged for passively screening for suicidal ideation. We analyze the screening ability of texts sent in different time periods prior to reported ideation, namely, texts from specific weeks only versus accumulative over several weeks. Our approach involves performing comprehensive feature engineering and identifying influential features to train machine learning models. With just the prior week of texts, we were able to predict the existence of suicidal ideation with AUC = 0.88, F1 = 0.84, accuracy = 0.81, sensitivity = 0.94, and specificity = 0.68. The most influential features include word frequencies of words in the car, clothing, affection, confusion, driving, real estate, and journalism categories. This research, demonstrating the potential of text messages to screen for suicidal ideation, will guide the development of screening technologies.