The use of text-based responses to improve our understanding and prediction of suicide risk

Perceived Burdensomeness and Thwarted Belongingness in Text Messages of Suicide Attempt Survivors

OBJECTIVE

Perceived burdensomeness and thwarted belongingness are considered interpersonal risk factors for suicide. Examining these themes in personal text messages may help identify proximal suicide risk.

METHOD

Twenty-six suicide attempt survivors provided personal text messages and reported dates for past periods characterized by positive mood, depressed mood, suicidal ideation (with no attempt), or the two-week period leading up to suicide attempt(s). Texts were then classified into the applicable period based on matching dates. Texts (N = 194,083; including n = 86,705 outgoing texts) were coded for perceived burdensomeness and thwarted belongingness by masked trained raters. Multilevel models were fit to examine whether the target themes (combined into one overall interpersonal risk variable due to low base rate) were more prevalent in texts sent during higher risk episodes (e.g., suicide attempt vs. depressed mood episodes).

RESULTS

0.57% of outgoing texts contained either target theme. As hypothesized, logistic models showed participants were more likely to send texts containing the target themes during suicide attempt episodes relative to suicidal ideation (with no attempt) episodes, depressed mood episodes, and positive mood episodes, and during suicidal ideation (with no attempt) episodes relative to positive mood episodes. All contrasts were robust to post-hoc correction except for suicide attempt episodes vs. ideation (with no attempt) episodes. No other significant pairwise differences for episode type emerged.

CONCLUSIONS

Despite the small sample size and low base rate of target themes in the texts, perceived burdensomeness and thwarted belongingness were associated with intra-individual suicide risk severity in personal text messages.

Acceptability of Personal Sensing Among People With Alcohol Use Disorder: Observational Study

BACKGROUND

Personal sensing may improve digital therapeutics for mental health care by facilitating early screening, symptom monitoring, risk prediction, and personalized adaptive interventions. However, further development and the use of personal sensing requires a better understanding of its acceptability to people targeted for these applications.

OBJECTIVE

We aimed to assess the acceptability of active and passive personal sensing methods in a sample of people with moderate to severe alcohol use disorder using both behavioral and self-report measures. This sample was recruited as part of a larger grant-funded project to develop a machine learning algorithm to predict lapses.

METHODS

Participants (N=154; n=77, 50% female; mean age 41, SD 11.9 years; n=134, 87% White and n=150, 97% non-Hispanic) in early recovery (1-8 weeks of abstinence) were recruited to participate in a 3-month longitudinal study. Participants were modestly compensated for engaging with active (eg, ecological momentary assessment [EMA], audio check-in, and sleep quality) and passive (eg, geolocation, cellular communication logs, and SMS text message content) sensing methods that were selected to tap into constructs from the Relapse Prevention model by Marlatt. We assessed 3 behavioral indicators of acceptability: participants' choices about their participation in the study at various stages in the procedure, their choice to opt in to provide data for each sensing method, and their adherence to a subset of the active methods (EMA and audio check-in). We also assessed 3 self-report measures of acceptability (interference, dislike, and willingness to use for 1 year) for each method.

RESULTS

Of the 192 eligible individuals screened, 191 consented to personal sensing. Most of these individuals (169/191, 88.5%) also returned 1 week later to formally enroll, and 154 participated through the first month follow-up visit. All participants in our analysis sample opted in to provide data for EMA, sleep quality, geolocation, and cellular communication logs. Out of 154 participants, 1 (0.6%) did not provide SMS text message content and 3 (1.9%) did not provide any audio check-ins. The average adherence rate for the 4 times daily EMA was .80. The adherence rate for the daily audio check-in was .54. Aggregate participant ratings indicated that all personal sensing methods were significantly more acceptable (all P<.001) compared with neutral across subjective measures of interference, dislike, and willingness to use for 1 year. Participants did not significantly differ in their dislike of active methods compared with passive methods (P=.23). However, participants reported a higher willingness to use passive (vs active) methods for 1 year (P=.04).

CONCLUSIONS

These results suggest that active and passive sensing methods are acceptable for people with alcohol use disorder over a longer period than has previously been assessed. Important individual differences were observed across people and methods, indicating opportunities for future improvement.

Identifying suicide ideation in mental health application posts: A random forest algorithm

The growing use of digitized mental health applications requires new reliable early screening tools to identify user suicide risk. We used a lexicon-based random forest machine learning algorithm to predict suicide ideation scores from 714 online community text posts from December 2019 to April 2020. We validated predicted scores against expert-rated suicide ideation scores. The algorithm-predicted scores offered high validity and a low error rate and correctly identified 95% of expert-rated high-risk suicide ideation posts. Our findings highlight a potential new method to detect suicidal ideation of digital mental health application users.