Researchers Develop Powerful Machine Learning Models to Enhance Suicide-Risk Prediction Among Children (July 21, 2023)

Machine Learning Models

An innovative study conducted by experts at UCLA Health has thrown light on the limitations of current approaches for detecting youngsters who are at danger of self-harming themselves. This comes at a time when there is a growing worry over the mental health of young people. The research also presents a potentially useful answer in the form of machine learning models, which have the potential to dramatically enhance the accuracy of suicide risk prediction among children.


Machine Learning Models

Machine Learning Models to Enhance Suicide-Risk Prediction: The Gap in Current Risk-Prediction Methods

As the youth mental health crisis deepens, mental health providers are striving to identify children who may be at risk of suicide or self-harm in order to intervene before it’s too late. However, the current systems often fall short in accurately capturing this vital information. Health systems frequently lack a comprehensive understanding of the children seeking help for self-injurious thoughts or behaviors, leading to incomplete data that hampers the accuracy of risk-prediction models.

Dr. Juliet Edgcomb, lead author of the study and associate director of UCLA’s Mental Health Informatics and Data Science (MINDS) Hub, highlights the key issue: “Our ability to anticipate which children may have suicidal thoughts or behaviors in the future is not great – a key reason is our field jumped to prediction rather than pausing to figure out if we are actually systematically detecting everyone who is coming in for suicide-related care. We sought to understand if we can first get better at detection.”

Machine Learning Models

The Shortcomings of Current Coding Methods

Many existing risk-prediction models rely on diagnostic codes, specifically the International Classification of Diseases, version 10 (ICD-10), used by healthcare providers to categorize care provided. However, this approach may overlook children who have self-injurious thoughts or behaviors but have been coded for underlying mental health issues like depression or anxiety. Another commonly used method is analyzing the “chief complaint” given by the patient at the start of a healthcare visit. Nevertheless, children might not always disclose their suicidal thoughts and behaviors during their initial emergency department visit.

Machine Learning Models

Unveiling the Discrepancies

To delve deeper, experts meticulously reviewed clinical notes from 600 emergency department visits involving children aged 10 to 17 within a large health system. The findings were eye-opening. The ICD-10 codes missed a significant 29% of children with self-injurious thoughts or behaviors, while the chief complaint approach missed a staggering 54% of such cases. Even when combining both methods, around 22% of these vulnerable patients remained undetected.

Furthermore, certain demographics appeared to be disproportionately missed. The screening methods displayed a tendency to overlook male children compared to females, as well as preteens when compared to teenagers. Disturbingly, there was a clear indication that Black and Latino youth were at a higher risk of being overlooked, potentially resulting in skewed risk prediction models.

Machine Learning Models

The Power of Machine Learning

In response to these shortcomings, the researchers turned to machine learning as a potential solution. They designed three distinct machine learning models to determine whether an automated system could enhance the detection of children with self-injurious thoughts or behaviors.

The most comprehensive model incorporated a rich array of 84 data points from a patient’s electronic record. This encompassed details like previous medical care, medications, demographic information, and even neighborhood disadvantage indicators. The second model employed all mental health diagnostic codes, going beyond just suicide-related codes, and the third model factored in additional indicators such as medications and lab test results.


All three machine learning models outperformed the traditional ICD-10 codes and chief complaint method in identifying children at risk. Notably, there was no significant disparity in performance among the machine learning models, implying that even moderately sophisticated models could notably enhance risk-flagging capabilities.

Dr. Edgcomb emphasizes, “Adding more information helps, but you don’t necessarily need a bells-and-whistles approach to get better detection.”

Machine Learning Models

Embracing Sensitivity in Screening

Despite the fact that the machine learning models displayed increased sensitivity, it is essential to keep in mind that they also indicated some patients who were not in imminent danger of harming themselves. Dr. Edgcomb contends that the benefits much exceed the potential drawbacks, stating that “Depending on the situation, it may be better to have some false positives and have a medical records analyst double-check those charts that screen positive, than to miss many children entirely.”

machine learning models

A Glimpse into the Future

The continued study of Dr. Edgcomb will continue to focus on developing more accurate models for predicting the risk of suicide among young people. Notably, another objective of the research is to fill a vacuum in the existing literature by developing risk prediction models for children of elementary school age, which have been noticeably absent up until this point.

machine learning models

Published on July 21, 2023, in JMIR Mental Health, this study signifies a crucial step towards a more accurate and comprehensive approach to identifying and supporting children at risk of self-harm.


  1. Why is accurate suicide-risk prediction among children so important? Accurate prediction helps mental health providers intervene earlier, potentially saving lives and offering timely support.
  2. How do the machine learning models enhance risk prediction? The models incorporate a wide range of data points, offering a more holistic view of a child’s situation, leading to better detection.
  3. Are there downsides to using more sensitive screening tools? While there may be some false positives, the benefits of early detection and intervention outweigh potential drawbacks.
  4. What groups are more likely to be overlooked by current risk-prediction methods? Certain demographics, such as male children, preteens, and Black/Latino youth, have a higher likelihood of being missed.
  5. What’s next for this research? The study aims to further refine risk prediction models, especially for elementary-school age children, and bridge existing gaps in the field.





Leave a Comment

Your email address will not be published. Required fields are marked *

Share via
Copy link