Thuật toán thông minh: Bước tiến mới trong việc phát hiện nguy cơ tự sát ở trẻ em tại phòng cấp cứu

• Notably, nearly 80% of adolescents who died by suicide had accessed healthcare services in the year prior, and 50% of them had visited the emergency department
• However, current risk identification methods still have significant gaps.

A new cohort study published in JAMA Network Open has opened a promising new direction: Using risk assessment algorithms based on clinical data has shown superior results compared to traditional screening methods.

Study scale and methodology

The study was conducted on a large cohort of 19,653 adolescents (aged 10 to 18) who visited the emergency department of a pediatric hospital in the Northeastern United States

The researchers compared two approaches:

• Direct screening: Using standardized questionnaires such as Ask Suicide-Screening Questions (ASQ) and the Columbia–Brief Suicide Severity Rating Scale (C-BSSRS)
• Risk algorithm: A machine learning model trained on patients’ electronic health records (EHR) over a 4-year period, including demographic data, suicide history, and diagnostic codes (ICD-10)
• Patients were followed for a period ranging from 6 months to 2.5 years to determine whether they engaged in suicidal behavior after the initial visit

Breakthrough results: AI algorithms “see” what humans miss
The analysis showed that the risk assessment algorithm outperformed direct screening methods across most metrics:
Detection capability: The algorithm correctly identified up to 127% more patients who later exhibited suicidal intent compared to screening methods (detecting 125 cases versus 55 cases by screening)

Sensitivity: The algorithm achieved a sensitivity of 50.7%, while traditional screening reached only 36.5%

Predictive value: The algorithm had a higher positive predictive value (PPV), meaning it was better at accurately predicting individuals who would actually engage in suicidal behavior

Notably, when combining both methods, clinicians were able to correctly identify up to 61.8% of individuals who later exhibited suicidal behavior

Why is the algorithm more effective?

The advantage of the algorithm lies in its ability to analyze dense historical data streams. The study found that patients identified by the algorithm often had more frequent healthcare visits and more complex diagnostic profiles

While screening questionnaires provide only a “snapshot in time” and can be affected by patients concealing their emotions, the algorithm connects data points from the past—such as depression, anxiety disorders, or emotion-related symptoms—to generate more accurate alerts

Practical implications and future directions

The application of these algorithms not only helps hospitals meet national patient safety standards (such as National Patient Safety Goal 15.01.01 by the Joint Commission) but also addresses healthcare resource challenges

In the context of limited mental health resources, these algorithms enable healthcare providers to focus interventions on high-risk patient groups, thereby reducing administrative burden and improving prevention effectiveness

Although there are still some limitations, such as reliance on ICD-10 coding and the single-site nature of the study, this clearly represents an important step forward in applying artificial intelligence to protect the lives of the younger generation

Read the full article on JAMA here.