Identification of acute brain dysfunction in pediatric sepsis using four definitions

Introduction: Studies of acute brain dysfunction (ABD) in sepsis are limited by imprecise clinical variables suggestive of neurologic/behavioral dysfunction. We previously validated a new computational phenotype of ABD indicative of clinician concern for an acute change in neurologic/behavioral function. We hypothesized that this computational phenotype would identify ABD with higher sensitivity and specificity than common ABD criteria, including Proulx, Goldstein and the Pediatric Organ Dysfunction Information Update Mandate (PODIUM). Methods: We performed a retrospective study of 527 randomly selected pediatric sepsis episodes in a research sepsis registry from a single institution between 2011 and 2019. The reference standard for ABD was established using chart review for diagnosis of acute neurologic or behavioral change during sepsis. The computational phenotype defined ABD as completion of either brain imaging (CT or MRI) or electroencephalogram (EEG) irrespective of results and/or new antipsychotic administration. Proulx and Goldstein criteria defined ABD as Glasgow Coma Scale (GCS) < 5 and ≤11, respectively. PODIUM defined ABD as total GCS ≤8, motor GCS ≤4, Cornell Assessment of Pediatric Delirium (CAPD) score ≥9, or seizures, attenuation or suppression on EEG. Missing data were assumed to be normal (ie, not ABD). We used the probability test of proportions to compare sensitivity and specificity of the four criteria to identify ABD against the reference standard. Results: The 527 sepsis patients has a median age of 6.6 years [1.6-13.9] and were 54% male. 64% had more than one complex chronic condition and 36% had more than one non-neurologic organ dysfunctions. The computational phenotype was more sensitive to identify ABD (83%, 95%CI 76-89%) compared to Proulx (22%, 16-29%; p< 0.05), Goldstein (40%, 33-48%; p< 0.05), and PODIUM (73%, 66-80%; p=0.03). Specificity to exclude ABD was significantly higher for the computational phenotype (93%, 95% CI 90-96%) compared to Goldstein (73%, 68-77%; p< 0.05) and PODIUM (68%, 63-80%; p< 0.05), but not Proulx (93%, 90-95%; p=1.00). Conclusions: The computational phenotype identified ABD in pediatric sepsis with a higher combination of sensitivity and specificity than other commonly used criteria that rely heavily on GCS.