Machine learning model predicts pediatric adverse events

A new EHR-based machine learning model was effective in continually assessing hospitalized children and predicting their risk of experiencing critical events, according to new research.

Published in JAMA Network Open, the study describes the development of a novel machine learning model called the pediatric Critical Event Risk Evaluation and Scoring Tool (pCREST) that assesses critical event risk across emergency departments, wards and intensive care units. The study authors noted that risk stratification for the pediatric population tends to be fragmented, with each hospital unit using different approaches. Thus, the researchers aimed to create a model to assess pediatric patient risk across their hospital stay.

They conducted a retrospective study of 135,621 pediatric admissions to the ward and the ICU across three tertiary care centers in Illinois and Wisconsin. They identified critical events among the patient population, defined as invasive mechanical ventilation, vasoactive drug administration or in-hospital mortality within 12 hours of a vital sign or laboratory result observation.

The researchers used EHR data, including patient age, hospital unit, vital signs, laboratory results and prior comorbidities, to derive a regression-based model, an extreme gradient-boosted machine (XGB) model and two deep learning models. 

The XGB model, called pCREST, was the best-performing model, outperforming the other models in terms of discrimination and other clinically relevant metrics. Additionally, the model performed as well as or better than models trained for a specific hospital unit. Meanwhile, the deep learning models did not exhibit improved performance.

"Our model could be used to monitor a child's health seamlessly throughout their hospital journey, facilitating early recognition of deterioration and timely intervention," the researchers concluded.  

This is the latest example of machine learning models used in pediatric risk prediction.

A study published in 2023 showed that a machine learning model that used the deterioration risk index was able to predict hospitalized children at risk for pediatric deterioration earlier than other approaches. The study detailed how Nationwide Children's Hospital researchers developed and deployed the model, finding that it exhibited greater sensitivity and more precise alerting than the existing situational awareness program.

Further, the pilot study reported a 77% reduction in deterioration events during the first 18 months of the model's use compared to the situational awareness program.

However, machine learning tools are prone to bias, meaning that the algorithms can produce systemically prejudiced results due to incorrect assumptions. Healthcare AI developers and users must mitigate this algorithmic bias as it can perpetuate care disparities.

Anuja Vaidya has covered the healthcare industry since 2012. She currently covers the virtual healthcare landscape, including telehealth, remote patient monitoring and digital therapeutics.