Phenotyping in Pediatric Traumatic Brain Injury

Phenotypes developed using machine learning are widely used in industry to target marketing activities. In medicine, phenotypes are the “measurable biological, behavioral, and clinical markers of a condition or disease.”(1) In critical care research, phenotypes may represent disease categories such as septic shock subtypes(2), treatment pathways and clinical events(3), or classes of patient outcomes.(4) The heterogeneity of traumatic brain injury (TBI) has been a barrier to successful trials of new interventions.(5) High-quality outcome phenotypes are an important part of understanding that heterogeneity, but pediatric TBI currently lacks them.

Outcome phenotypes based on data available at hospital admission or shortly thereafter might be useful in several ways. First, they might inform pre-clinical and translational investigation into mechanisms and biomarkers of injury and potential new treatments. Second, they could be used to identify patients for high-risk interventions. Third, they could inform our conversations with parents of injured children.

In this issue of Pediatric Critical Care Medicine (6), Rosario and colleagues attempt to identify outcome phenotypes using the Cool Kids phase 3 hypothermia randomized trial cohort.(7) The trial was conducted at 15 centers in the U.S., New Zealand, and Australia, but most of the 77 patients were enrolled at 3 large U.S. centers. All patients had acute severe TBI (post-resuscitation total Glasgow Coma Scale [GCS] less than 9), GCS-Motor score less than 6, an intracranial pressure (ICP) monitor, and an abnormal head CT. Important exclusion criteria included known or suspected inflicted injury, a total GCS of 3 with bilateral non-reactive pupils, hypotension for more than 10 minutes, hypoxia for more than 30 minutes, or a severe non-head injury (another Abbreviated Injury Scale [AIS] body region score greater than 3).

Patient, injury, and imaging characteristics were examined for association with two outcomes: mortality and the Glasgow Outcome Scale Extended Pediatric Revision (GOS-E-Peds)(8) at 3 months after injury. Variable selection was performed using pre-screening with bivariable testing followed by forward stepwise selection. Multivariable logistic (mortality) and proportional odds (trichotomized GOS-E-Peds) regression models were then built. The authors found that higher severity of spinal injury (spine AIS body region score) and midline shift on CT were associated with mortality. Subarachnoid hemorrhage on CT, any hypoxia, and the number of non-reactive pupils were associated with worse GOS-E-Peds score. No differences were found between the normothermic and hypothermic groups from the original Cool Kids trial.

Strengths of the current manuscript include the well-characterized prospective trial cohort and the gold-standard outcome measures collected. Detailed reads of CT images by multiple study radiologists and variables that are difficult to obtain without extensive chart abstraction (e.g. duration of hypoxia and hypotension, seizure occurrence) were particularly valuable. Unfortunately, the authors found that the dataset had an insufficient sample size to independently develop and validate novel outcome phenotypes.

How does pediatric TBI research move forward? Faced with similar challenges, adult TBI investigators combined several trial datasets into a large, diverse sample and then developed and validated new prognostic models for outcome phenotypes.(4) The recently developed U.S. Federal Interagency Traumatic Brain Injury Research (FITBIR) informatics system (https://fitbir.nih.gov/) is intended to facilitate TBI dataset combination and re-use. The high-quality Cool Kids trial data would be a valuable addition to FITBIR. The Approaches and Decisions in Acute Pediatric TBI (ADAPT) study has recently prospectively enrolled 1,000 children with severe TBI and ICP monitors at 50 international centers.(9) The ADAPT data, all of which will ultimately be submitted to FITBIR, are expected to be a rich source for pediatric TBI phenotype development. In order to generate a representative sample, ADAPT purposefully enrolled patient types excluded from many previous studies (e.g. those with inflicted injury or major non-head injuries).

What methodologic advances might we leverage? First, a new set of data-driven phenotypes is effectively a novel classification system. Recent guidelines(10) provide best-practices for classification and prediction model development and reporting. These best-practices include training and testing datasets, unbiased and reproducible variable selection, and internal validation using resampling methods such as cross-validation and bootstrapping. Machine learning techniques such as k-means clustering(11) and deep neural networks(12) may enhance our ability to identify novel critical care phenotypes. Combinations of clinical, biological, and image data as inputs to machine learning models may generate valuable new hypotheses such as those forthcoming from the adult-focused Transforming Research and Clinical Knowledge in Traumatic Brain Injury (TRACK-TBI) initiative.(13)

Although not an explicit goal of the current study, computable phenotypes algorithmically derived from clinical data may be particularly useful.(14) Because they can be built into an electronic health record (EHR), computable phenotypes have the potential for research, clinical, and quality improvement impact. Natural language processing(15) could be leveraged to extract difficult-to-obtain but potentially powerful features from unstructured text in the EHR (e.g., “75 mph” or “3 story fall to concrete”). Once developed, we should share our computable phenotypes using resources such as the Phenotype KnowledgeBase (https://phekb.org/).

Overall, phenotyping is an active area of both informatics and critical care research. The current study adds to the literature identifying risk factors for poor outcome after pediatric TBI and moves the field forward in an important direction.