Machine Learning Advances in Predictive Pediatrics 

In recent years, health informaticians have developed tools that assist caregivers in navigating today’s fast-paced, information-loaded hospitals. By analyzing hundreds of thousands of data points taken from electronic health records (EHR), these new tools aim to discover patterns and connections important to the course of care which might otherwise be imperceptible or liable to pass under the radar of caregivers. 

These devices not only have applications for improving patient outcomes, but their benefits impact the entire field of healthcare, from reining in hospital budgets to bettering the lives of hospital staff, pointing to the range of exciting opportunities open to health informatics professionals entering the field today.

Amarachi Erondu and Emma Hegermiller, 2020 graduates of the Master of Science in Biomedical Informatics (MScBMI), assessed for their capstone project the efficacy of four machine learning models in predicting pediatric patient decline. Extending work begun by their scientific advisor, Anoop Mayampurath, PhD, a research assistant professor in pediatrics at UChicago, they worked with models that aggregate patient EHR data to generate predictive early warning scores for hospitalized children.

Hegermiller, a research technician in immunology at UChicago, saw the project as an ideal way to gain experience using machine learning tools on clinical data to improve patient care. For Erondu, a medical student at the Pritzker School of Medicine who hopes to maintain a research focus in the future, the project was a great way to gain confidence in practical data analysis skills.

Setting a New Metric for Health Assessment

Since its introduction at UChicago Medicine in 2015, the predictive clinical score tool eCART has effectively alerted caregivers to potential cardiac arrest in hospitalized adult patients. Its success has spurred efforts to develop similar tools for other conditions, like respiratory failure, acute kidney injury, and sepsis. 

In children, a tool used for predicting deterioration, BedsidePEWS, has gained less traction at hospitals. Because BedsidePEWS includes subjective assessments of a patient’s clinical status gathered by nurses and other care staff, the tool tends to lack consistency across hospitals. Requiring manual inputs of data has led to poor compliance as well.

One aim of pCART, the new tool being developed at the University of Chicago, is to outperform BedsidePEWS. To accomplish this, the model utilizes EHR data indicative of a patient’s physiological condition and advanced machine learning methods to generate a score that predicts the risk of a patient’s transfer to the ICU in the next twelve hours. 

For their study, Erondu and Hegermiller drew on eleven years of data from Comer Children’s Hospital at the University of Chicago. The results of their work have been accepted as an abstract at the 2021 Virtual Critical Care Congress by the Society of Critical Care Medicine.

“We took Anoop’s logistic regression model to the next level,” Hegermiller says. “We added lab data and used an advanced machine learning method. We determined that the most important factors were things like heart rate, respiratory rate, and whether the patient had recently arrived on the ward.”

The actions that a high-risk score will trigger are still under discussion, notes Hegermiller. “You can imagine they might include bringing a specialized care team in to analyze the patient. Or maybe the patient gets automatically transferred to the pediatric intensive care unit.”

Erondu, who is familiar with the setting of Comer Children’s Hospital from her experience as a medical student, envisions a variety of possible scenarios in which the score might be useful. As an additional metric, it could be an important piece of information when assessing the overall condition of the patient.

“I’m picturing myself as a third-year medical student. I go in and the patient doesn’t look so great and the nurse feels that way too. If the pCART score also indicates a potentially worrisome situation, that gives me an extra number to report to the clinical team that might alert the attending physician to monitor the patient a little more closely.”

Exciting Prospects for pCART’s Future

Having centered their Capstone Project around fine-tuning the machine learning model, Erondu and Hegermiller note that there are now additional adjustments that might improve the model’s accuracy further. 

One addition may involve focusing on data trends. “Not just what their blood pressure was at this point in time, but also whether their blood pressure suddenly plummeted,” Erondu explains. “If there’s an acute change in those metrics, then you’d be worried that something’s going on.”

Also important will be to ensure that the model’s results are not specific to the data set or the clinical setting in which the model was trained and built. For that, the next order of business involves running the model on data acquired from another hospital.

For now, the project has been passed on to another capstone team in the MScBMI program. Their task will be to pursue these new avenues, while also tweaking the machine learning model’s overall performance. Erondu and Hegermiller are eager to see where it leads.

“I’ll be really excited to see where it goes in the end and whether this might be a whole new score that’s used at Comer,” says Erondu. “We already have the eCART score on the adult’s side, so it would be amazing if this became a standard tool in our EHR for children.”

The Graham School will not be admitting new students to the Master of Science in Biomedical Informatics (MScBMI) in Autumn 2024. The University will take this opportunity to consider future programming in the Biological Sciences Division (BSD). Please see the BSD website for more information about their offerings.