AI model flags suicide risk in emergency rooms using routine admission data

Researchers at the University of Johannesburg have published an artificial intelligence approach designed to help emergency room staff identify patients at risk of suicide-related behaviour, using only the routine data already collected when people are admitted to hospital. The study, led by Mr Tsholofelo Mokheleli as part of his PhD research in the Department of Applied Information Systems within the UJ College of Business and Economics, appears in the journal Scientific Reports. The model is intended to assist clinicians rather than replace them, particularly during very busy periods at resource-constrained hospitals.

The motivation is rooted in the pressures facing South African emergency rooms. According to Mokheleli, late at night and over weekends, some South African hospital emergency rooms can see one staff member triaging 30 or more patients. In such conditions, an objective and consistent measure to help staff rapidly screen patients could be valuable for the early detection of suicide-related behaviour risk. “The AI model will flag that a patient is at risk and provide reasons. A health care professional still needs to take the final decision however, the clinician must do a severity scaling to determine if the person is really suicidal or not,” says Mokheleli.

Existing screening for suicide-related behaviour in South African emergency rooms is resource intensive, says Professor Nompumelelo Ndlovu of the Department of Nursing in UJ’s Faculty of Health Sciences, a co-author of the study and one of Mokheleli’s PhD co-supervisors. When anyone first arrives at an emergency room, the first priority is to stabilise them medically. Even for conscious patients, the standard mini-mental status examination for suicide-related behaviour risk can take 15 minutes and varies between patients, she says. A full mental status examination and history taking for one patient at high risk could consume a day or more of a mental health professional’s time at the hospital.

“Not everyone who goes into a medical institution has mental health issues. What we are trying to do, is to find those people who do not even say it out loud,” says Mokheleli. Professor Ndlovu adds that some people do not realise they are at risk. “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?” she says.

The model was built with historical open-source patient data from the USA, drawn from a single hospital via the MIMIC-IV resource. MIMIC-IV was developed by the Massachusetts Institute of Technology’s Laboratory for Computational Physiology in collaboration with Beth Israel Deaconess Medical Centre, and contains anonymous records from patients admitted to hospitals and intensive care units within 24 hours of visiting an emergency room in the USA between 2008 and 2019. The hospital in the dataset had a mainly White and Hispanic patient population, which the researchers note is completely different from South Africa. Finding a suitable digital patient admissions dataset in South Africa proved difficult, according to Professor Ndlovu.

The information fed into the model includes patient age and gender, admission characteristics, vital signs and previous psychiatric diagnoses, all routinely recorded in hospital systems. “It includes demographics, age, gender, vital signs, temperature, heart rate, respiratory rate, oxygen saturations, and systolic and diastolic blood pressure. From that information, we calculate physiological features like the patient’s shock index and pulse pressure,” says Mokheleli. The model also checks a patient’s prior history at the same hospital: if substance abuse, anxiety or depression was documented with ICD-9 or ICD-10 codes on a previous visit, the model takes that into account, along with night and weekend admissions and notes from medical staff about self-harm on previous visits.

In the study, the model identified patient age, pain, and a prior mood disorder recorded at the same hospital as the major reasons for flagging a patient for suicide-related behaviour risk. Vital signs such as temperature and blood pressure follow on that list, with prior substance-use and anxiety diagnoses playing smaller roles. The model’s current output takes the form of confusion matrices and SHAP (SHapley Additive exPlanations) plots, which the researchers acknowledge are hard to understand. “A SHAP plot shows how a model made a decision about one patient. We are busy designing a user interface that clinicians can interact with and get answers they can understand,” says Mokheleli.

Performance figures from the study, based on a 41,202-patient dataset, illustrate both the promise and the limits of the approach. Per 1000 emergency department visits, the model correctly predicted 866 patients as not at risk of a documented suicide-related behaviour outcome in the 30 days after admission to the emergency room, and incorrectly predicted or missed 13 cases as not at risk although a suicide-related outcome was documented within those 30 days. The model flagged 121 patients per 1000 visits as potentially at risk and requiring further screening by clinicians, a resource-intensive task. Of those 121 flagged visits, 23 had a documented suicide-related outcome recorded within the 30 days following admission, while 98 flagged cases did not have any suicide-related outcome documented with ICD-9 or ICD-10 codes in the same hospital’s records in the 30 days after admission.

The researchers note that a number of factors influence how many patients the model flags correctly, and how many truly at-risk patients it misses. These include the size of the dataset, the dynamics of the patient population at the hospital from which the data was selected, how few patient records in the training dataset recorded suicide-related behaviour risk, the value of the risk decision threshold, and the AI algorithm selected for analysis. Because the model is built on existing electronic patient admissions data, emergency room staff would not need to do additional data capture in future implementations, although in hospitals with paper-based records, the requirement for digitised patient records in a database would be a barrier.

Data governance was a major reason for building the model on routine admissions data, says co-author Professor Patrick Ndayizigamiye, Mokheleli’s PhD main supervisor and a visiting Associate Professor in the Centre of Applied Data Science at UJ. “The question is always ‘who owns the confidential patient data and where is it stored?’ Routine admission data belongs to the patient, and it is stored within the hospital’s systems already,” he says, contrasting this with asking a cloud storage vendor to hold the data on their servers, which generates many data governance issues. Future versions of the model will use bigger datasets, potentially different AI algorithms and adjusted techniques to improve performance, says Dr Tebogo Makaba, a study co-author and PhD co-supervisor, and the patient risk threshold will need to be adjusted for each hospital since patient populations vary. The current model did not include job loss, bereavement, or drug- and alcohol-related observations from medical staff, but these can be included in future versions, says Mokheleli.

Subject of Research: An AI model using routine hospital admission data to predict suicide-related behaviour risk in emergency departments

Article Title: Promising AI-based approach to early detection of suicide-related behavior at emergency rooms

Article References: Promising AI-based approach to early detection of suicide-related behavior at emergency rooms. Original publication (Bibliographic details unavailable.)

Image Credits: AI Generated

DOI: Not provided

Keywords: artificial intelligence, suicide-related behaviour, emergency departments, MIMIC-IV, risk screening, machine learning, patient triage, mental health, University of Johannesburg, Scientific Reports, hospital admissions data, data governance