By Joseph J. Zorc, James M. Chamberlain, Lalit Bajaj
JAMA Pediatrics, May 13, 2019
In this issue of JAMA Pediatrics, Bertsimas et al1 describe a novel machine-learning approach to derive a revised version of the head injury prediction rule developed by the Pediatric Emergency Care Applied Research Network (PECARN). The PECARN rule was derived and validated using a prospectively collected data set of more than 42 000 patients to classify which children with head injury are at very low risk of clinically significant intracranial abnormalities.2 The ultimate goal of such a decision rule is to reduce unnecessary computed tomographic imaging and associated radiation. Bertsimas et al1 analyzed a public use data set from the PECARN study using a technique called optimal classification trees. The revised rule has improved specificity and predictive value, identifying 33% more children younger than 2 years, and 14% more children 2 years or older as having a very low risk for intracranial injury compared with the PECARN rule, without missing any additional cases of intracranial injury. Although this is good use of the public use data sets now required for federally funded research, interpreting machine-learning techniques may be challenging for clinicians to understand and apply as the techniques become increasingly complex. Although we live in an era of precision medicine, with the ability to tailor personalized recommendations, it is also an era emphasizing shared decision making between clinicians and patients. It may be difficult for clinicians to counsel patients about the implications of a rule that is perceived as a black box or ghost in the machine, which may provide recommendations for unclear reasons.