Development and temporal validation of a screening-based prediction model for suicidal ideation at first contact with university counseling services

Abstract Background Suicidal ideation among university students is a major public health concern. Yet counseling services still lack practical tools that can identify students at elevated risk during their first clinical contact. This study aimed to develop and temporally validate a screening‑based prediction model for suicidal ideation at the time of initial presentation to university counseling services. Methods This retrospective observational study analyzed routinely collected data from a Japanese university counseling services. Predictor variables were exclusively derived from annual mental health screening data collected before participants used counseling services. The variables included psychological distress, self‑esteem, lifestyle‑related factors, and demographic characteristics. The outcome was suicidal ideation assessed during the first counseling intake interview. We developed a multivariable logistic regression model using a development cohort of students who accessed counseling services between 2017 and 2019 and temporally validated the model in an independent cohort from 2021 to 2023. We evaluated the model performance using discrimination, measured by the area under the receiver operating characteristic curve (AUC), calibration plots, and decision curve analysis (DCA). A reference value threshold was identified using a Youden Index–based cutoff. Results The development and validation cohorts included 1,136 and 1,286 students, respectively. The prevalence of suicidal ideation at intake was comparable across cohorts (14.0% and 14.7%). The model achieved moderate discrimination in the development cohort (AUC = 0.71) and good discrimination in the validation cohort (AUC = 0.74), with acceptable calibration in both cohorts. Higher levels of screening‑based psychological distress, particularly feelings of worthlessness, female sex, and lower self-esteem increased the odds of suicidal ideation at intake. Conversely, call-based or early outreach referral routes were associated with lower odds. DCA showed that the model provided greater net benefit than strategies that intervened in all or no students across clinically plausible threshold probabilities, including the Youden Index‑based cutoff (approximately 0.13). Conclusions Routinely collected mental health screening data can predict suicidal ideation at the time of initial counseling contact with reasonable accuracy and potential clinical utility. Screening‑based prediction models may strengthen existing student support systems by enabling earlier identification of elevated risk and guiding proportionate intervention strategies.