A Prioritization Model for Suicidality Risk Assessment

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8124–8137 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 8124 A Prioritization Model for Suicidality Risk Assessment Han-Chin Shing Computer Science University of Maryland College Park, MD shing@cs.umd.edu Philip Resnik Linguistic/UMIACS University of Maryland College Park, MD resnik@umd.edu Douglas W. Oard iSchool/UMIACS University of Maryland College Park, MD oard@umd.edu Abstract We reframe suicide risk assessment from so- cial media as a ranking problem whose goal is maximizing detection of severely at-risk indi- viduals given the time available. Building on measures developed for resource-bounded doc- ument retrieval, we introduce a well founded evaluation paradigm, and demonstrate using an expert-annotated test collection that mean- ingful improvements over plausible cascade model baselines can be achieved using an ap- proach that jointly ranks individuals and their social media posts. 1 Introduction Mental illness is one of the most significant prob- lems in healthcare: in economic terms alone, by 2030 mental illness worldwide is projected to cost more than cardiovascular disease, and more than cancer, chronic respiratory diseases, and diabetes combined (Bloom et al., 2012). Suicide takes a terrible toll: in 2016 it became the second leading cause of death in the U.S. among those aged 10-34, fourth among those aged 35-54 (Hedegaard et al., 2018). Prevalence statistics suggest that roughly 141 of the 3,283 people who attended ACL 2019 have since had serious thoughts of suicide, 42 have made a plan, and 19 have actually made attempts.1 The good news is that NLP and machine learn- ing are showing strong promise for impact in men- tal health, just as they are having large impacts everywhere else. Traditional methods for predict- ing suicidal thoughts and behaviors have failed to make progress for fifty years (Franklin et al., 2017), but with the advent of machine learning ap- proaches (Linthicum et al., 2019), including text analysis methods for psychology (Chung and Pen- nebaker, 2007) and the rise of research on mental 1Approximately: ACL is international, but these figures use prevalence statistics for U.S. adults (SAMHSA, 2019). health using social media (Choudhury, 2013), algo- rithmic classification has reached the point where it can now dramatically outstrip performance of prior, more traditional prediction methods (Linthicum et al., 2019; Coppersmith et al., 2018). Further progress is on the way as the community shows in- creasing awareness and enthusiasm in this problem space (e.g., Milne et al., 2016; Losada et al., 2020; Zirikly et al., 2019). The bad news is that moving these methods from the lab into practice will create a major new chal- lenge: identifying larger numbers of people who may require clinical assessment and intervention will increase stress on a severely resource-limited mental health ecosystem that cannot easily scale up.2 This motivates a reformulation of the techno- logical problem from classification to prioritization of individuals who might be at risk, for clinicians or other suitably trained staff as downstream users. Perhaps the most basic way to do prioritization is with a single priority …

Content truncated to 500 words