Objectives
Suicide risk assessments currently rely on subjective clinical judgement, lacking objective measures. This study aimed to evaluate the association between biomarkers and suicide risk and to explore their predictive potential using machine learning (ML).
Methods
We analysed data from 2785 first-admission psychiatric inpatients across three institutions, including 103 biomarkers and 13 demographic and clinical variables. Suicide risk was assessed 1 week after admission using the Nurses’ Global Assessment of Suicide Risk.
Results
A total of 2785 patients met the inclusion criteria, from which 978 were selected via propensity score matching to minimise confounding from demographic factors, treatment differences and symptom severity. Multivariate random effects logistic regression identified nine biomarkers associated with elevated suicide risk and six with potential protective effects. Time-trend analyses further revealed that nine biomarkers showed significant changes following risk escalation. Integrating biomarkers with demographic data, treatment information and psychological scale scores substantially improved ML model performance, achieving an area under the receiver operating characteristic curve of 0.808 in the external testing cohort. The inclusion of biomarkers significantly enhanced predictive accuracy.
Conclusion
This study highlights the potential of biomarkers with ML to predict future risk, offering objective assessments and supporting early interventions.
