AI Tool Predicts Suicide Risk in Emergency Departments from Routine Data

Suicide-related behavior (SRB) risk can be difficult to identify in emergency rooms, particularly when patients do not disclose mental health concerns. Existing screening approaches in South African emergency rooms can also require substantial clinician time, adding pressure as staff triage patients while prioritizing urgent medical needs. To support earlier identification, researchers have developed an artificial intelligence-based approach for detecting SRB risk after emergency room admission.

The AI model was developed by researchers from the University of Johannesburg to assist emergency room staff rather than replace clinical judgment. It uses routine patient intake data and lightweight computing to estimate SRB risk after hospital admission. The approach is designed to work without adding new data-entry requirements in hospitals that already maintain electronic patient records.

The model was built with historical open-source patient data from the United States, using a 41,202-patient dataset extracted from MIMIC-IV. The Massachusetts Institute of Technology Laboratory for Computational Physiology developed MIMIC-IV in collaboration with Beth Israel Deaconess Medical Center. The dataset contains anonymous records from patients admitted to hospitals and intensive care units within 24 hours of an emergency room visit between 2008 and 2019.

The model used age, gender, admission characteristics, vital signs, and previous psychiatric diagnoses recorded in hospital systems. Its strongest reasons for flagging SRB risk were patient age, pain, and a prior mood disorder documented at the same hospital. Temperature, blood pressure, prior substance use, and anxiety diagnoses had smaller roles, while night and weekend admissions and prior self-harm notes were also included.

In the study, the model correctly predicted 866 true negatives per 1,000 emergency department visits. It missed 13 documented SRB-related outcomes within 30 days of admission. It flagged 121 patients per 1,000 visits for further clinician screening, including 23 true positives and 98 false positives. The findings were published in Scientific Reports.

Future versions are expected to use larger datasets, different algorithms, and hospital-specific risk thresholds. The current outputs include confusion matrices and SHAP (SHapley Additive exPlanations) plots, which can be difficult for clinicians to interpret. A user interface is being designed to make the model’s explanations more understandable for clinical teams.

“Some people do not realize they are at risk of SRB. People may not be aware of what is happening with them unless they speak to somebody. But it is not an easy topic to speak about. There is so much stigma. So how safe do you feel to tell the next person about what is happening with you?” said Nompumelelo Ndlovu, associate professor in the Department of Nursing within the Faculty of Health Sciences at the University of Johannesburg.

Related Links
University of Johannesburg