Predictors of suicide attempt in adolescents with a history of self-harm: a machine learning study combining psychosocial and actigraphy measures
Abstract
Youth suicide attempt is an important clinical and public health concern, yet prediction of suicide attempt among youth with a history of self-harm remains challenging. Disrupted states of arousal may play an important role in the etiology of suicide attempt but have rarely been measured objectively in prospective studies. In this study, we examined to what extent actigraphy-derived arousal markers and questionnaire-based psychosocial measures collected at age 14 could predict self-reported suicide attempt at age 17 in a population-based sample of 1,321 adolescents with a history of self-harm from a United Kingdom cohort. We conducted both inferential analyses (univariable and multivariable logistic regressions), and predictive (machine learning) modeling using nested cross-validation. Cannabis use, peer problems, depressive symptoms, higher BMI, lower self-esteem, and sex emerged as significant independent predictors of later reported suicide attempt. Machine learning models achieved moderate predictive performance (AUROC = 0.65–0.67), with relatively high sensitivity (0.71–0.81) but lower specificity (0.37–0.53). Peer problems and depressive symptoms were consistently among the most important predictors across models. Although some actigraphy-derived measures contributed to prediction in specific models, two-stage analyses suggested little incremental predictive value beyond questionnaire-based measures alone. These findings highlight the importance of cannabis use, interpersonal difficulties and depressive symptoms in predicting suicide attempt among adolescent who self-harmed and suggest a limited additional contribution of actigraphy-derived arousal markers for long-term prediction.
Related
- Dietary amino acid composition and elevated fasting glucose in children and adolescents from four Chinese provinces: a cross-sectional study combining regression and machine learning
- Beyond Fertility: Ethical and Psychosocial Dimensions of Fertility Preservation in Children and Adolescents
- Psychosocial predictors and justification patterns of traditional and alternative medicine in Peru
- Relative importance of psychosocial and organizational predictors of burnout among European researchers: a dominance analysis
- Professionals' perspectives on barriers to improving psychosocial care for adolescents with emotional and behavioural problems - a concept mapping case study.
BibTeX
@misc{silva2026predictors,
title = {{Predictors of suicide attempt in adolescents with a history of self-harm: a machine learning study combining psychosocial and actigraphy measures}},
author = {Martin Sellier Silva and Bhavya Patwa and Dylan James and Martin Lepage and Lena Palaniyappan and Massimiliano Orri},
howpublished = {PsyArXiv},
year = {2026},
month = sep,
doi = {10.31234/osf.io/qhbde_v1},
url = {https://osf.io/preprints/psyarxiv/qhbde_v1/},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Predictors of suicide attempt in adolescents with a history of self-harm: a machine learning study combining psychosocial and actigraphy measures
AU - Martin Sellier Silva
AU - Bhavya Patwa
AU - Dylan James
AU - Martin Lepage
AU - Lena Palaniyappan
AU - Massimiliano Orri
PY - 2026
DA - 2026/09/13
PB - PsyArXiv
DO - 10.31234/osf.io/qhbde_v1
UR - https://osf.io/preprints/psyarxiv/qhbde_v1/
AB - Youth suicide attempt is an important clinical and public health concern, yet prediction of suicide attempt among youth with a history of self-harm remains challenging. Disrupted states of arousal may play an important role in the etiology of suicide attempt but have rarely been measured objectively in prospective studies. In this study, we examined to what extent actigraphy-derived arousal markers and questionnaire-based psychosocial measures collected at age 14 could predict self-reported suicide attempt at age 17 in a population-based sample of 1,321 adolescents with a history of self-harm from a United Kingdom cohort. We conducted both inferential analyses (univariable and multivariable logistic regressions), and predictive (machine learning) modeling using nested cross-validation. Cannabis use, peer problems, depressive symptoms, higher BMI, lower self-esteem, and sex emerged as significant independent predictors of later reported suicide attempt. Machine learning models achieved moderate predictive performance (AUROC = 0.65–0.67), with relatively high sensitivity (0.71–0.81) but lower specificity (0.37–0.53). Peer problems and depressive symptoms were consistently among the most important predictors across models. Although some actigraphy-derived measures contributed to prediction in specific models, two-stage analyses suggested little incremental predictive value beyond questionnaire-based measures alone. These findings highlight the importance of cannabis use, interpersonal difficulties and depressive symptoms in predicting suicide attempt among adolescent who self-harmed and suggest a limited additional contribution of actigraphy-derived arousal markers for long-term prediction.
ER -