LAKE BUENA VISTA, FLA. – An algorithm might predict whether patients with severe traumatic brain injury are recovering well or need interventions to preempt evolving intracranial hypertension.
“Valid predictive algorithms have the potential to revolutionize the care of patients with traumatic brain injury [TBI] and transform physiologic data from just a pure numeric value buried in a never-ending nursing flow sheet into a useful triage and decision-assist tool,” study author Dr. Brandon Bonds said at the annual scientific assembly of the Eastern Association for the Surgery of Trauma.
Patrice Wendling/Frontline Medical News
Dr. Brandon Bonds
A minimum of 10 hours of continuous data on vital signs (intracranial pressure, heart rate, systolic blood pressure, shock index, and mean arterial pressure) were used to predict intracranial pressure (ICP) values for a retrospective cohort of 132 adults with severe TBI, 97% of which was the result of blunt trauma. Even relatively brief episodes of elevated ICP have been shown to be associated with poor outcomes in TBI patients, while marked elevation of ICP may lead to herniation and death, said Dr. Bonds of the R. Adams Cowley Shock Trauma Center, University of Maryland, Baltimore.
At the trauma center, vital signs are automatically collected every 6 seconds, 24 hours a day, on all TBI patients. This granularity of data was used to map patterns in the patients’ physiology. The approach used a nearest neighbor regression (NNR) method: A model was constructed that predicts future numerical values for an individual based on comparisons to data from historical subjects.
The same mathematical principal is used by a variety of industries to predict likely responses. NetFlix, for example, uses a system similar to the NNR method to predict future television and movie picks based on prior selections, Dr. Bonds explained.
About 20 minutes of continuously collected, automated vital sign data were then used to test the algorithm on a per-patient basis. The algorithm was used to predict future ICP values at 5 minutes to 2 hours from that time. The predictions are made on a rolling basis, with patient data updates every 5 minutes.
The NNR model was good at predicting actual ICP at 5 minutes, with a bias of 0.02 (± 2 standard deviations of 4 mm Hg). As expected, agreement was somewhat lessened at 2 hours (± 2 standard deviations of 10 mm Hg), “but this may still represent a clinically significant value,” Dr. Bonds said.
The next step is a prospective study of the algorithm’s utility.
Dr. Bonds said that NNR research really isn’t all that alien to medicine. Think about the experienced emergency physician who can look out into the wait room and “tell the nurse to bring back [a certain patient] because he didn’t look good,” Dr. Bonds said. Such a physician uses “the minimum amount of data he has and compares that patient to the historic data set of the thousands of patients that he’s seen previously to identify a patient that’s not going to do well. What we’re trying to do with this model is take this subjective skill and turn it into an objective tool.”
In an interview, session comoderator Dr. David A. Hampton, M.Eng., of Oregon Health and Science University in Portland, commented that he could definitely see the NNR method eventually having utility in severe TBI.
Future work will need to address outliers in the data because the standard deviation of 4 mm Hg “is pretty big for ICP swings” and to determine whether multiple libraries of data will need to be created based upon the different types of patients who come in, he said.
The study was funded by the United States Air Force. Dr. Bonds and his coauthors reported no financial disclosures.