Scholarlink

<p>Suicide remains a major public health challenge, with emergency departments (EDs) representing a critical point of contact for individuals at risk. However, conventional suicide screening approaches are often time-intensive, disclosure-dependent, and difficult to implement consistently in high-acuity settings. This study developed and evaluated an artificial intelligence (AI)-based model for predicting suicide-related behaviour (SRB) within 30&#xa0;days of ED presentation using routinely collected structured triage data available at admission. Five supervised learning models, logistic regression, random forest, extreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM), and a multilayer perceptron (MLP), were developed and evaluated using patient-level train-validation-test splits. Performance was assessed using a multidimensional framework including area under the receiver operating characteristic curve (AUROC), area under the precision-recall curve (AUPRC), Recall, F2-score, Matthews Correlation Coefficient (MCC), Balanced Accuracy, threshold optimisation, decision-curve analysis, bootstrap confidence intervals, and SHapley Additive exPlanations (SHAP). Among the evaluated models, LightGBM showed the most favourable balance between overall discrimination, minority-class discrimination, sensitivity, and decision stability, achieving an AUROC of 0.88, an AUPRC of 0.25, a Recall of 0.79, and an F2-score of 0.41 on the independent test set. Threshold optimisation and decision-curve analysis further supported the model’s clinical utility, while SHAP analyses identified prior psychiatric history, age, and physiological variables as important contributors to risk prediction. These findings demonstrate that short-term SRB risk can be predicted using structured ED triage information available at presentation. Rather than replacing clinicians, the proposed model aims to provide transparent, scalable decision support to support prioritisation, follow-up planning, and early intervention. Although external validation and prospective studies remain necessary, the study provides a deployment-oriented framework for integrating AI-assisted suicide risk prediction into emergency care workflows.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Artificial intelligence approach for predicting suicide-related behaviour in emergency departments

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月2日

查看全文
  • Tsholofelo Mokheleli,
  • Tebogo Makaba,
  • Patrick Ndayizigamiye,
  • Nompumelelo Ndlovu

摘要

Suicide remains a major public health challenge, with emergency departments (EDs) representing a critical point of contact for individuals at risk. However, conventional suicide screening approaches are often time-intensive, disclosure-dependent, and difficult to implement consistently in high-acuity settings. This study developed and evaluated an artificial intelligence (AI)-based model for predicting suicide-related behaviour (SRB) within 30 days of ED presentation using routinely collected structured triage data available at admission. Five supervised learning models, logistic regression, random forest, extreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM), and a multilayer perceptron (MLP), were developed and evaluated using patient-level train-validation-test splits. Performance was assessed using a multidimensional framework including area under the receiver operating characteristic curve (AUROC), area under the precision-recall curve (AUPRC), Recall, F2-score, Matthews Correlation Coefficient (MCC), Balanced Accuracy, threshold optimisation, decision-curve analysis, bootstrap confidence intervals, and SHapley Additive exPlanations (SHAP). Among the evaluated models, LightGBM showed the most favourable balance between overall discrimination, minority-class discrimination, sensitivity, and decision stability, achieving an AUROC of 0.88, an AUPRC of 0.25, a Recall of 0.79, and an F2-score of 0.41 on the independent test set. Threshold optimisation and decision-curve analysis further supported the model’s clinical utility, while SHAP analyses identified prior psychiatric history, age, and physiological variables as important contributors to risk prediction. These findings demonstrate that short-term SRB risk can be predicted using structured ED triage information available at presentation. Rather than replacing clinicians, the proposed model aims to provide transparent, scalable decision support to support prioritisation, follow-up planning, and early intervention. Although external validation and prospective studies remain necessary, the study provides a deployment-oriented framework for integrating AI-assisted suicide risk prediction into emergency care workflows.