The researchers analysed records from 1,368 adults aged 18 or older who presented to the emergency department after a suicide attempt between January 2021 and June 2023. Of these patients, 589, or 43.1 per cent, had a history of at least one earlier attempt. The team set out to identify which patients carried that history using eleven demographic and suicide-related variables that are routinely available in emergency settings, rather than relying on specialised assessments.

Eight algorithms were trained and compared: logistic regression, a linear support vector machine, a decision tree, a random forest, a gradient boosting machine, a neural network, eXtreme gradient boosting, and naive Bayes. Performance was assessed with repeated stratified 10-fold cross-validation, and the researchers evaluated discrimination, calibration, and classification performance, reported with 95 per cent confidence intervals.

Most models performed similarly. The area under the receiver operating characteristic curve, a standard measure of how well a model ranks patients correctly, ranged from 0.71 to 0.73 for the majority of algorithms, with substantial overlap in their confidence intervals. The decision tree lagged behind, with an AUC of 0.66. Calibration, measured by the Brier score, was acceptable at approximately 0.21 for most models but poorer for the decision tree at 0.27.

Across all eight approaches, two variables stood out. Psychiatric history and age were consistently among the most influential predictors, indicating that a handful of routinely recorded clinical details carry meaningful signal about a patient’s attempt history. The study did not report that the models predicted future suicide risk directly; rather, they inferred a known risk marker from information available at presentation.

A notable finding was what did not happen: the more complex algorithms did not show a clear performance advantage over simpler ones. In a dataset limited to eleven variables, gradient boosting, neural networks, and XGBoost did not outperform logistic regression or naive Bayes. The authors attribute this to the constrained feature set, which limits how much additional structure sophisticated models can exploit.

The findings come from a single centre, and the models were not validated on external patient populations, so their performance in other hospitals, health systems, or countries remains unknown. The authors themselves describe the work as exploratory. The retrospective design also means the models reflect patterns in one institution’s records over a two-and-a-half-year period rather than a prospective test in live clinical workflow.

Within those limits, the study offers a measured result: machine learning applied to everyday emergency department variables can identify patients with a history of prior suicide attempts with moderate ability, with psychiatric history and age doing much of the predictive work.

Subject of Research: Machine learning prediction of prior suicide attempt history among emergency department patients in South Korea

Article Title: Development of machine learning algorithms for identifying repeat suicide attempters among emergency department patients

Article References: Na, S., Lee, J., Oh, J., Yoo, K. H., Cho, Y., & Lee, S. H. (2026). Development of machine learning algorithms for identifying repeat suicide attempters among emergency department patients. BMC Psychiatry. https://doi.org/10.1186/s12888-026-08678-7

Image Credits: AI Generated

DOI: 10.1186/s12888-026-08678-7

Keywords: suicide risk, machine learning, emergency department, psychiatric history, predictive modelling, BMC Psychiatry, South Korea, cross-validation, risk screening, logistic regression, Development, machine