Jaclyn Schwartz Kirshenbaum

Adolescent suicide rates have risen over the past two decades, underscoring the need for improved risk detection strategies. Although natural language processing (NLP) tools are increasingly used to flag suicide-related content, little is known about how such approaches perform on adolescents’ smartphone communications. Addressing this gap, this study leverages passively-collected smartphone data to identify suicide-related language in adolescents’ keyboard usage via NLP. We developed a lexicon of suicide-related adolescent language and validated it with labeled data (N=171,468 text entries; e.g., messages, web searches), demonstrating higher performance in identifying suicide-related text than few-shot prediction with large language models (LLMs) and lexicons not designed for youth. Across two independent cohorts at elevated suicide risk (Ns=208 & 257; >6 million text entries), lifetime suicidal thoughts and behaviors (STB) and current suicidal ideation were associated with increased frequency of smartphone suicide-related language. Human coding indicated varied language, including authentic first-person current suicidal ideation (14.5%) and jokes or hyperbole (20.2%). Compared with the lexicon alone, human coding of suicide-related entries with first-person language showed stronger associations with STB history. However, an LLM showed limited performance in identifying whether suicide-related text indicated authentic, first-person, and current STB (F1=.45). These findings highlight that effective NLP-based tools for suicide prevention will require more nuanced and context-specific approaches to better distinguish suicidal intent.

Rising adolescent suicide rates underscore an urgent need for better detection of short-term risk in the days and hours leading up to attempts. Passive smartphone sensing of language offers a promising approach, yet its performance during vulnerable periods remains unclear. This case study examined five adolescents (3 male, 2 female) who were hospitalized for suicidal crises while enrolled in a smartphone sensing study. Participants contributed outgoing text entries over six months (M = 21,000/person), which were analyzed using natural language processing (NLP) to assess suicide-related content, sentiment, and topics (e.g., school, treatment). In addition, clinicians conducted qualitative reviews of the text entries to identify potential risk events. Results showed that 4 of 5 adolescents exhibited increased suicide-related language and negative sentiment during the 10 days prior to psychiatric hospitalization. Especially elevated suicide language was found within 5 days of hospitalization, while negative sentiment peaked between 5–10 days prior to hospitalization. These signals, however, also occurred outside of acute risk periods, highlighting the challenge of separating suicide risk from distress more generally. Clinical annotations revealed that suicidal thoughts and behaviors often co-occurred with NLP signals of suicide-related language, and topic models identified clinically relevant language related to substance use and psychiatric treatment. Clinical annotations of interpersonal conflict and school stressors were not identified by topic models. Discrepancies largely originated from the inability of NLP methods to infer context (e.g., text conversation history). Although smartphone language data showed low missingness and some sensitivity to acute crises, enhancing contextual analysis is essential for personalized risk detection.

Rising adolescent suicide rates underscore an urgent need for better detection of short-term risk in the days and hours leading up to attempts. Passive smartphone sensing of language offers a promising approach, yet its performance during vulnerable periods remains unclear. This case study examined five adolescents (3 male, 2 female) who were hospitalized for suicidal crises while enrolled in a smartphone sensing study. Participants contributed outgoing text entries over six months (M=21,000/person), which were analyzed using natural language processing (NLP) to assess suicide-related content, sentiment, and topics (e.g., school, treatment). In addition, clinicians conducted qualitative reviews of the text entries to identify potential risk events. Results showed that 4 of 5 adolescents exhibited increased suicide-related language and negative sentiment during the 10 days prior to psychiatric hospitalization. These signals, however, also occurred outside of acute risk periods, highlighting the challenge of separating suicide risk from distress more generally. Clinical annotations revealed that suicidal thoughts and behaviors often co-occurred with NLP signals of suicide-related language, and topic models identified clinically relevant language related to substance use and psychiatric treatment. Clinical annotations of interpersonal conflict and school stressors were not identified by topic models. Discrepancies largely originated from the inability of NLP methods to infer context (e.g., text conversation history). Although smartphone language data showed low missingness and some sensitivity to acute crises, enhancing contextual analysis is essential for personalized risk detection.

Adolescent suicide is a public health emergency with interpersonal factors playing a critical role in risk for suicidal thoughts and behaviors (STB). This study examined whether smartphone-based ecological momentary assessment (EMA) surveys capturing stress and social context could enhance the identification of youth at risk for STB. Adolescents ages 13-18-years-old (N=207, 166 female sex) reporting depressive, anxiety, and/or substance use disorders were recruited from psychiatric outpatient programs, emergency departments, medical research registries, and social media. Participants with STB history were oversampled; 66% reported current suicidal ideation and 27% had a past-year attempt. Assessments of suicidal events (attempts, emergency department visits, hospitalizations for suicide concerns) were completed at baseline, 1-, 3-, and 6-month follow-up assessments. EMA probed momentary stress severity, affect, and recent social context (time spent with peers, family, or alone) 4-7x/day. Compared to psychiatric controls (N=64), adolescents with STB history (N=143) were more likely to report spending time alone (aOR=1.75, 95% CI=[1.15, 2.39]), less time with family (aOR=0.59, 95% CI=[0.31, 0.91]), and higher stress (d=0.28). Further, EMA of stress (aOR=1.86, 95% CI=[1.11, 3.26]) and negative affect (aOR=2.32, 95% CI=[1.34, 4.36]) were prospectively associated with suicidal events above and beyond prior STB history. Broadly, results were similar between EMA measures of stress and negative affect, indicating such responses may reflect shared underlying processes. These findings underscore the importance of evaluating dynamic family and peer contact among adolescents at risk for STB and highlight EMA as a potential tool for assessing interpersonal exposures related to STB risk.

Background: Poor sleep health is pervasive and contributes to long-lasting physical and psychological problems. As traditional sleep measurement can be burdensome, testing scalable and accessible sleep measurements is important. Objective: The aim of this study was to test whether sleep features obtained through a smartphone app are comparable to other modes of sleep measurement (ie, daily diary and wearable actigraphy). Methods: Healthy college students (n=29, 18‐24 years old) with no prior diagnosis of a sleep disorder consented to downloading a smartphone app, the Effortless Assessment Research System (EARS) (Center for Digital Mental Health, University of Oregon; Ksana Health). For 1 week, the EARS app collected data continuously using the phone’s accelerometer, gyroscope, and exposure to light. Each morning, participants received a notification via EARS to complete a sleep diary, which asked participants what time they got into bed, fell asleep, woke up, and got out of bed. A random subset (n=13) of participants also consented to wear an ActiGraph wristwatch. All analyses examined bedtime (ie, time going to bed), risetime (ie, wake-up time), and time-in-bed (ie, duration between time-in-bed and out-of-bed) as 3 measures of interest. For all analyses, diary data were considered the reference measurement, and analyses were repeated with ActiGraph as the reference measurement. Results: On average, EARS showed a high mean true positive rate (86.6%) and low mean false positive rate (4%) based on diary-reported bedtime and risetime. Supplementary analyses comparing EARS to ActiGraph data showed a similar true positive rate (83.7%) and slightly higher false positive rate (8.5%). Although there were no significant differences in mean bedtime, risetime, and time-in-bed among sources (P≥.07), there was some misalignment. Compared to the diaries, EARS estimated bedtime to be later by an average of 20 minutes, risetime to be earlier by an average of 21.83 minutes, and time-in-bed to be shorter by an average of 41.82 minutes. Relative to ActiGraph estimates, EARS estimated bedtime to be earlier by an average of 21.87 minutes, but estimated risetime to be later by 2.5 minutes, and time-in-bed to be longer by 24.4 minutes. Day-to-day correlations showed that bedtimes, risetimes, and time-in-bed were positively correlated between diary and EARS (0.29≤r≤0.52, P≤.002). Similarly, day-to-day bedtimes and time-in-bed were positively correlated between ActiGraph and EARS (0.38≤r≤0.55, P≤.01), and while correlated in the expected direction, risetimes were not significantly associated between ActiGraph and EARS (r=0.29, P=.07). Conclusions: Smartphone-based sleep sensors show acceptable alignment with more established methods and may provide a feasible alternative to measuring daily sleep patterns in a scalable way. Future studies will require larger, diverse samples to corroborate findings of concordance among EARS, diary, and actigraphy data in other populations.

Intensive longitudinal research-including experience sampling and smartphone sensor monitoring-has potential for identifying proximal risk factors for psychopathology, including suicidal thoughts and behaviors (STB). Yet, missing data can complicate analysis and interpretation. This study aimed to address whether clinical and study design factors are associated with missing data and whether missingness predicts changes in symptom severity or STB. Adolescents ages 13- to 18 years old (N = 179) reporting depressive, anxiety, and/or substance use disorders were enrolled; 65% reported current suicidal ideation and 29% indicated a past-year attempt. Passively acquired smartphone sensor data (e.g., global positioning system, accelerometer, and keyboard inputs), daily mood surveys, and weekly suicidal ideation surveys were collected during the 6-month study period using the effortless assessment research system smartphone app. First, acquisition of passive smartphone sensor data (with data on ∼80% of days across the whole sample) was strongly associated with survey data acquisition on the same day (∼44% of days). Second, STB and psychiatric symptoms were largely not associated with missing data. Rather, temporal features (e.g., length of time in study, weekends, and summer) explained more missingness of survey and passive smartphone sensor data. Last, within-participant changes in missing data over time neither followed nor predicted subsequent change in suicidal ideation and psychiatric symptoms. Findings indicate that considering technical and study design factors impacting missingness is critical and highlight several factors that should be addressed to maximize the validity of clinical interpretations in intensive longitudinal research. (PsycInfo Database Record (c) 2024 APA, all rights reserved).