NLP Detection of Short-Term Suicide Risk in Adolescents Using Smartphone Language

A case series of five adolescents hospitalized for suicidal thoughts and behaviors was analyzed using natural-language-processing (NLP) on passively collected smartphone messages. All five participants showed increased suicide-related language; four had significantly sustained negative sentiment before admission. Peaks in suicidal language appeared 1–5 days prior to hospitalization, while negative sentiment rose over a longer period. Topic modeling flagged treatment, sleep, and school but often missed interpersonal conflict noted by clinicians.

Nature

Real-time suicide risk monitoring usually relies on clinical visits, leaving distress unobserved. Prior passive-sensing linked GPS and sleep to risk; this study applied an adolescent-specific suicidal-language lexicon and transformer embeddings, described as identity-protective and cost-effective.

Nature

All five participants showed increased suicide-related language; four had sustained negative sentiment. The acute window examined was ten days before admission. Topic models captured treatment, sleep, school but missed peer conflict; baseline spikes sometimes reflected distress without imminent risk.

Nature

The authors of the Nature study propose that NLP alerts could enable just-in-time digital interventions for youth lacking clinical access, improve alignment of observed behavior with self-report, and require privacy-preserving safeguards.

Nature

Critics note that models miss nuanced triggers like interpersonal conflict, generate false positives, and risk stigma. The observational design prevents causal claims, and the small sample limits power. Baseline language spikes may indicate distress without imminent risk, reducing specificity.

Nature

NLP signals matched clinician judgments of ideation in many cases but diverged on interpersonal triggers; baseline spikes sometimes reflected distress without predicting attempts.

Nature

• “In the current case study of five adolescents hospitalized for suicide risk, NLP methods showed sensitivity to risk periods, with all individuals experiencing increases in suicidal language, and 4 of 5 individuals showing sustained increases above the mean in negative sentiment prior to hospitalization.” — Study authors

Nature

• “Suicidal language increases often corresponded to clinician judgments of suicidal ideation, attempts, and help-seeking.” — Study authors

Nature

• “NLP detection of short-term suicide risk factors could facilitate digital treatment approaches and just-in-time interventions, which may be especially helpful for individuals without access to other clinical services [62, 63].” — Study authors

Nature

Future work should combine contextual language models with multimodal sensors, expand sample sizes, test time-lagged associations, develop ethical frameworks for privacy-preserving digital surveillance, and assess user acceptability.

Nature