Decrypting Sepsis-Associated Brain Dysfunction: A Computational Phenotype That Can Hack It*

In ancient Greece, fever accompanied by loss of the rational mind was termed "phrenitis," an ominous condition with a high fatality rate. The etymology of phrenitis relates to inflammation, indicated by "itis," and also references the diaphragm, or "phren," which ancient Greeks' believed harbored intellect and emotion (1). Perhaps a reflection of the condition's mysterious underpinnings, which prevented supplanting by more precise nomenclature, phrenitis continued to be diagnosed by physicians until the 19th century, when it was eventually ousted by terms that persist to this day, including encephalitis, meningitis, and delirium (2). This obsolescence of phrenitis might be likened to cleaning out a garage and throwing out a broken bicycle, then replacing it with a new handlebar bell, a new handlebar basket, and another broken bicycle. Encephalitis and meningitis each have separate, standalone utility. Both are words that speak to logical, underlying mechanisms responsible for clinical manifestations. Infection-related delirium, on the other hand, is rooted in the word "delirare," a term that meant "to be out of one's furrow," which originated as a farming metaphor referring to a plow veering from its intended path (3). The word "delirium" is as connected to the biological basis of altered mental status as the diaphragm is to one's personality. The use of conceptual metaphors to label infection-related brain dysfunction represents a need to bridge a gap between a complex, poorly understood phenomenon and the general understanding required to develop durable management strategies. Bridging this gap is indeed important because the ambiguous biological basis of sepsis-related brain dysfunction is juxtaposed with the sharply defined adverse outcomes associated with its development. Approximately 20% of children with sepsis or septic shock experience some form of associated brain dysfunction (4). Among one large cohort of children with septic shock, any acute pathologic neurologic sign or event was significantly associated with death or a clinically substantial deterioration in quality of life 3 months later (5). However, varied names and definitions for sepsis-associated brain dysfunction complicate efforts to identify optimal treatment approaches. As Alcamo et al (6) note in this issue of Pediatric Critical Care Medicine, brain involvement in sepsis and septic shock "…goes by a variety of terms including sepsis-associated encephalopathy, sepsis-induced brain dysfunction, or sepsis-associated delirium…" Accordingly, a validated, well-performing definition applicable to electronic health records (EHRs) is needed to aid surveillance efforts and epidemiological assessments, with the ultimate intent of improving outcomes. This is not a new undertaking for Alcamo et al (6), who have spent years developing and validating a structured, composite, "computational phenotype of acute brain dysfunction (CPABD)." Akin to other epidemiological breakthroughs, like the simple questionnaires used to link smoking with lung cancer in the 1950s, CPABD is constructed with a straightforward methodology that belies its broad utility (7). Using structured data ubiquitous in modern EHR databases, CPABD assesses whether a patient had MRI or CT brain imaging, an electroencephalogram (EEG), or was treated with an antipsychotic medication during an inpatient encounter. Initially developed and assessed in solid organ transplant recipients, CPABD proved useful in that population to reproduce, on a greater scale, an association between higher tacrolimus levels and neurologic morbidities (8). Subsequently, CPABD was applied to a retrospective population of children with sepsis, and content, criterion, and construct validity were confirmed through thorough evaluation (9). CPABD has also been astutely tethered to informatics best practices in previous work, poising it for use in massive, harmonized, multicenter collections of EHR data that are now being assembled through efforts such as the National Institute of Health's Bridge 2 Artificial Intelligence (10). To further establish the soundness of CPABD, Alcamo et al (6) compare it to other definitions of sepsis-related neurologic dysfunction, including those created by Proulx et al (11), the International Pediatric Sepsis Consensus Conference (IPSCC) (12), and the Pediatric Organ Dysfunction Information Update Mandate (PODIUM) (13). Proulx et al (11) put forward the most severe definition of sepsis-related brain dysfunction, requiring pupils to be fixed and dilated or a total Glasgow Coma Scale score (GCS) less than 5. IPSCC requires an acute change in mental status and a total GCS less than or equal to 11 or a change in mental status accompanied by a drop in GCS of at least 3 points. The recently introduced PODIUM definition includes any of the four following findings: 1) an EEG with seizure attenuation or suppressed activity, 2) acute injury, 3) a total GCS less than or equal to 8, or a GCS motor score less than or equal to 4, 4) a Cornell assessment of pediatric delirium score greater than or equal to 9. The assessment measures used to compare the definitions of brain dysfunction are those guided by the Standards for Reporting Diagnostic Accuracy (STARD) guidelines (14). These include the familiar measures of sensitivity, specificity, negative predictive value (NPV), and positive predictive value (PPV). A few more esoteric measures are also reported, including pretest and posttest probability, false discovery rate, false omission rate, and prevalence threshold. Pretest probability is measured as the prevalence of adjudicated brain dysfunction identified through chart review labeled according to an a priori established reference standard. Posttest probability is calculated by multiplying the positive likelihood ratio, determined using sensitivity and specificity, with pretest probability. The false discovery rate is calculated as 1-PPV, which is the same as the ratio of false positives to all positive cases and translates to labeling an encounter as having acute brain dysfunction when it does not meet the reference standard. The false omission rate is 1-NPV, or the ratio of false negatives to all negatives, and translates to missing a case of acute brain dysfunction when the reference standard was met. The prevalence threshold is the prevalence at which PPV drops most precipitously. CPABD outperforms the other acute brain dysfunction definitions across all measures. Sensitivity is 0.83, specificity is 0.93, PPV 0.84, and NPV 0.93. The prevalence threshold is 0.22, which is close to the actual population prevalence of 0.3 determined using the reference standard definition. The CPABD posttest probability of 0.84 is substantially higher than all the other definitions, which range from 0.39 (IPSCC) to 0.57 (Proulx et al [11]). Collectively, these characteristics add further support to the use of CPABD as a reliable, well-performing definition that can be applied to modern data collections. As the authors appropriately note, the narrower definitions relying heavily on GCS are more prone to misclassification and ascertainment biases, a finding likely rooted in the original construct of the GCS. GCS was developed empirically to standardize assessments of altered consciousness in the setting of trauma and brain surgery. Since then, GCS has burrowed its way into other domains of critical illness largely driven by clinician familiarity with the tool, a natural tendency to move things that work in one space to other proximate arenas, and the score's relative elegance compared with more comprehensive neurologic assessments. In contrast, the performance strengths of CPABD appear to stem from its reliance on clinical action rather than patient assessment. For CPABD, clinicians are the ultimate arbiter of determining acute brain dysfunction, a feature that arguably bolsters the tool's performance; however, it must also be noted that the reference standard used in the assessment was defined by "clinical experts" (the authors themselves) and therefore may have catered, both intentionally and inadvertently, to the computational phenotype. This is especially relevant given the recent, milestone release of the international Phoenix consensus criteria for sepsis in children, which defines sepsis-related neurologic dysfunction as a GCS less than or equal to 10 (1 point) or fixed pupils bilaterally (2 points) and is not accounted for by Alcamo et al (6) and Schlapbach et al (15) Instead, the reference standard in the present work encompasses much milder phenotypes of brain dysfunction, including "headache not relieved by fluid resuscitation" and "agitation" to name only 2 of the approximately 11 clinical scenarios described by the authors. How this decision will influence the use of CPABD as the world decides how to use the new Phoenix criteria remains to be determined, though the reference standard devised by Alcamo et al (6) might be interpreted as at least one collective voice claiming that brain dysfunction compatible with sepsis occurs well before a patient becomes stuporous, comatose, or develops two nonreactive pupils. We are no longer entering into an era of "big data" in medicine but are now solidly within the boundaries of this widely touted epoch of time. As mortality rates for pediatric sepsis continue to decline in the developed world, thanks in part to robust quality improvement initiatives also propelled by modern information systems, developing new breakthroughs that preserve the neurodevelopmental potential of critically ill children feels less dramatically ambitious and more realistically achievable in the coming years. Characterizing the scale of sepsis-related acute brain dysfunction is a necessary, fundamental step to deploy new treatments and interventions to help prevent both acute and long-lasting consequences of critical illness. Although the brain remains the body's most enigmatic organ, we have learned far more about it in the past 2 millennia than our current nomenclature for its afflictions might suggest. "Big data" tools like CPABD can be expected to help us hone in on meaningful patterns of illness that will lead to a better understanding of specific mechanisms and pathways underlying the varied reasons a plow might deviate from the furrow.