Removal of Race and White Blood Cell Count in an Updated Pediatric Pneumonia Severity Model Free

Children enrolled in the Etiology of Pneumonia in the Community study (2010–2012) formed the development cohort.⁶  The validation cohort included children enrolled in a single-center observational pilot (2014–2017), followed by a 2-center observational cohort (2017–2019).³  Children aged <18 years presenting to the ED with signs and symptoms of acute lower respiratory tract infection, radiographic evidence of pneumonia, and a provider-confirmed diagnosis of pneumonia were included, as previously reported.^2,3  All studies were approved by the institutional review board.

The original models included: (1) a full model of 20 predictors, (2) an expert model of 10 predictors based on expert opinion, and (3) an EHR model of 9 predictors available as coded EHR fields.²  Modeling used ordinal logistic regression to simultaneously estimate risk probabilities for (a) mild (ED discharge) or moderate (non-ICU admission), (b) severe (ICU admission), and (c) very severe (invasive mechanical ventilation and/or shock requiring vasoactive medications) disease. The EHR model predictors included sex, race, temperature, age, heart rate, respiratory rate, systolic blood pressure, WBC count, and ratio of PaO₂ to fraction of inspired oxygen (P/F ratio).^2,3  Race was categorized as non-Hispanic white, non-Hispanic Black, Hispanic, and other. In development and validation, all 3 models performed similarly, with the EHR model holding greatest promise for implementation.^2,3  Notably, race and WBC count were not considered important predictors in the expert model and were omitted.

In this study, we compared the original EHR model to a modified EHR (mEHR) model removing race and WBC count. The original development cohort was used to reestimate model parameters for the mEHR model. Predictive performance was then estimated in the development and validation cohorts using discrimination (c-statistic) and calibration plots, with recalibration of intercepts and slopes where appropriate.

The mEHR model demonstrated very good discrimination, with similar performance to the EHR model (c-statistic across models 0.77–0.80) in both developmental and validation cohorts (Table 1). Calibration was excellent (Fig 1).

Removing the problematic and weak predictors of race and WBC count from our previously validated pneumonia severity model retained usability and maintained prediction performance. The updated mEHR model represents a suitable alternative to the original EHR model, with important improvements in health equity and generalizability.

Race was considered in the original models because of known differences in invasive pneumococcal disease by race.⁷  However, race is a sociopolitical construct without a clear biologic basis and more closely aligns with social health determinants. Including race in risk prediction models may perpetuate biases.⁵  Exclusion of race in our model aligns with the retirement of the American Academy of Pediatrics’ urinary tract infection guidelines and replacement of race with fever duration in the University of Pittsburgh’s UTICalc Version 3.0.^4,8,9  Critical evaluation of similar severity models for racial bias may further improve prediction and health equity.

Though WBC count is a common inflammatory biomarker associated with a variety of infectious diseases, studies have demonstrated weak associations with pneumonia disease severity.^2,10  Further, because blood tests are not often performed in children with pneumonia, removal of WBC count better aligns with clinical practice, improving generalizability, and future clinical application.

Our study has limitations. Although an EHR-based strategy simplifies implementation, it risks constraining predictor selection. As a result, the mEHR model omits important predictors, such as chest indrawing. Our study was performed at tertiary care children’s hospitals and may lack generalizability.

The mEHR model removed 2 suboptimal predictors, but still performed similarly to the original EHR model. Application of the mEHR model is favored over the original EHR model for future risk stratification and decision-making in pediatric pneumonia.