AI tool uses emergency triage data to predict suicide risk within 30 days

Researchers trained an AI system on routine admission information to flag people at higher risk of suicide-related behaviour soon after an emergency visit.

Editorial illustration — not from the study.

Emergency departments are a key point of contact for people at risk, but standard suicide screening can be hard to carry out consistently because it can be time-consuming and depends on what patients disclose. This study evaluated whether AI could make risk prediction more scalable by using routine triage data collected at admission.

The researchers compared five supervised learning approaches and assessed performance with multiple measures, including decision-focused metrics and interpretability tests. They report that short-term suicide-related behaviour risk can be predicted using structured triage information available at presentation, with the LightGBM model performing most favourably among those evaluated.

AI model predicts 30-day risk

The study developed and evaluated five AI models—logistic regression, random forest, XGBoost, LightGBM, and a multilayer perceptron—to predict suicide-related behaviour within 30 days using routinely collected structured ED triage data available at admission. On the independent test set, the LightGBM model showed AUROC 0.88, AUPRC 0.25, Recall 0.79, and an F2-score of 0.41. Model interpretation using SHAP pointed to prior psychiatric history, age, and physiological variables as important contributors. The authors also report that threshold optimisation and decision-curve analysis supported clinical utility, and they describe the approach as decision-support for prioritisation, follow-up planning, and early intervention.

Support for triage decisions

If accurate enough, an AI-assisted tool that relies on information already captured during triage could help emergency departments identify people who may need additional support after discharge. The study frames the goal as supporting clinical prioritisation and planning rather than replacing clinicians, and it evaluates how the model behaves for real-world decision use (through threshold and decision-curve analyses).

Model testing with key caveats

This is a modelling study that trains and evaluates AI models using patient-level train-validation-test splits, with performance assessed on an independent test set and confidence intervals estimated via bootstrapping. The abstract notes that external validation and prospective studies are still needed. Because the results are based on how the models performed in the study data and splits, the findings may not fully reflect performance in other emergency departments or routine settings.

AI food images matched real food on health and calories — but not appeal

People judged AI-generated food images similarly to real food for healthiness and calorie content, but found them less realistic and were less willing to eat them. The results suggest AI images may be useful for controlled food-perception studies, but they are not interchangeable with images of real food.

21 hours ago·3 views

Society & Economics

Preprint finds measurable patterns in how AI skills emerge and fade

A preprint reports measurable patterns for how capabilities emerge and are forgotten during language-model training. The study says emergence depends on several model parts becoming useful together, while forgetting depends on how widely and strongly a capability is stored.

2 days ago·5 views

Society & Economics

In nursing mice, brain cells link bone, fat, energy use and pain sensitivity

A study in nursing mice identified hypothalamic nerve cells that linked prolactin signaling with bone and fat metabolism, food demand and pain sensitivity. Reducing parts of this pathway lowered food demand and pain sensitivity, but the findings have not been established in humans.

3 days ago·12 views