JMIR Publications

ABSTRACT

Background:

Suicide risk fluctuates over hours and days, yet standard risk assessments depend on infrequent healthcare contacts and willingness to self-disclose. Digital phenotyping of phone data offers a unique opportunity to understand the ecological dynamics of risk; however, systematically analyzing even a small amount of phone data of deceased individuals has remained a challenge.

Objective:

This study aimed to characterize changes in suicide risk-related language in phone data during the year prior to suicide death, and whether these changes differ across outbound chats, internet searches, and personal notes.

Methods:

This retrospective cohort study used de-identified phone data from Stop Soldier Suicide's Black Box Project, donated by families of 41 individuals who died by suicide between 2014 and 2024 (93% male; 95% with military service). We analyzed outbound chats (n=38), internet searches (n=24), and personal notes (n=12), comprising 1,877 person-data-type-weeks and 1.7 million words during the year prior to death. Text was matched to 48 risk factors of the Suicide Risk Lexicon. A multilevel logistic regression estimated the weekly probability of mentioning any suicidal concept, and multilevel negative binomial models estimated weekly match rates for each risk factor by data type.

Results:

Almost all individuals (87.8%) used language related to suicidal concepts during their final year. The predicted probability of suicidal concepts in chats increased from 14.2% (95% CI, 9.0%-21.6%) at 52 weeks to 38.5% (95% CI, 28.6%-49.5%) at 1 week before death, while the probability in personal notes declined. Eleven of 125 risk factor and data type combinations showed significant time trends after false discovery rate correction; several followed non-monotonic trajectories, and no risk factor was significant in more than one data type. In chats, emotional pain & psychache, loneliness & isolation, and physical health issues & disability language increased over time; depressed mood language increased until 20 weeks before death and then declined, alcohol use language declined throughout, and passive suicidal ideation followed a U-shaped pattern. Internet search trends peaked 13 to 26 weeks before death. All the trends were driven by a few individuals.

Conclusions:

Phone data contain detectable signals of suicide risk during the final year of life. Suicidal concepts may increase in external chats; however, other risk factors fluctuate non-monotonically, consistent with the fluid vulnerability theory of suicide, and they differ by internal and external facing data types. These findings could support gatekeeper training for recognizing high-risk language, digital outreach where suicide-related internet searches occur, and data-type-specific and person-specific just-in-time adaptive interventions, pending comparison data from living individuals.