New ML tool helps gauge severity of pediatric head injuries

A new machine learning model is better at flagging which children can avoid CT scanning after a head injury than currently used tools.

CT scanning is the standard for the fast diagnosis of traumatic brain injury (TBI). However, it’s expensive and exposes patients to a greater risk of developing cancer.

The current clinical decision support tool used to determine whether a child who suffers a head trauma needs a CT scan is the Pediatric Emergency Care Applied Research Network (PECARN) rule, which uses predictors such as altered mental status and signs of fracture. Children with very low risk of clinically important TBI (ciTBI) don’t need a CT scan—those in the high-risk category always receive a CT scan. Children in the moderate risk category may receive a CT scan, depending on additional factors.

The researchers, from Brown University and the Massachusetts Institute of Technology, created a novel machine learning decision tool using optimal classification trees (OCTs) to improve on the success of the PECARN rules. OCTs create an entire decision tree in one high-performing computing step, rather than creating each decision split in isolation, allowing for a more complex decision model.

Brain Image-CROP.jpg

The study authors used data on 42,412 children with head trauma but not a severely altered mental status from 25 hospital emergency departments in North American participating in PECARN.

The OCTs identified 33% more children younger than age two at very low risk and 32% fewer patients in the high-risk category than the PECARN tool. For older children, the OCTs identified 14% more children in the very low risk category and 8% fewer patients at high risk than the PECARN tool. The OCTs also correctly identified all patients with ciTBI.

”We presented OCTs that appear to have a better predictive accuracy than PECARN in that they have similar sensitivity but improved specificity. The improvement in specificity was statistically significant and sizable. We also believe that the improvement is clinically important and economically consequential,” the study authors wrote.

The study was published in JAMA Pediatrics.

“We surmise that large health systems that aim to optimize operations by capitalizing on better predictive performance would consider easy-to-use implementations of the OCTs in their [EHR] systems,” the authors said.

For reprint and licensing requests for this article, click here.