A Novel Proof-of-Concept Data Platform Integrating Physiologic, Laboratory, and Transcriptomic Biomarkers for Emergency Department–Based Pediatric Sepsis Screening

Timely identification and treatment of sepsis is critical to reducing morbidity and mortality. However, definitive diagnosis remains challenging in part due to the non-specific nature of symptoms. Cytokine markers and host transcriptomic profiles may potentially improve performance characteristics of pediatric sepsis prediction models. We present a method of integrating physiologic parameters, cytokine concentrations, and transcriptomic profiles with traditional laboratory values to better distinguish between pediatric patients who are suspected of sepsis in the ED.

Vital signs, hematology, and chemistry values from a cohort of 10,924 pediatric (3 mo-18 yrs) ED patients suspected of sepsis were used to pre-train an autoencoder model 5 hours post-intervention (fluid bolus or IV antibiotics). In this proof-of-concept analysis, we include data from 24 presumed sepsis patients (12 with sepsis and 12 ill-appearing, but non-septic). Additionally, 12 inflammatory cytokine markers and 16 differentially-expressed genomic transcripts were integrated via an L2 logistic regression ensemble model, with sepsis onset as the outcome of interest as confirmed by clinician review. Training and validation of the ensemble was done using a 100-fold bootstrap, with mean (95% CI) AUROC and AUPRC across folds used for evaluation.

The baseline model with only EHR data was moderately predictive of sepsis, with an AUROC (95%CI) of 0.84 (0.80-0.88) and an AUPRC 0.80 (0.74-0.85). The addition of inflammatory cytokines did not alter model performance, with an AUROC 0.79 (0.74-0.84) and AUPRC 0.80 (0.75-0.85). However, adding the expression levels of the 16 assayed genes increased performance to an AUROC of 0.90 (0.87-0.94) and AUPRC 0.86 (0.82-0.90).

This proof-of-concept analysis on a small cohort of presumed sepsis patients demonstrates one method by which different data modalities could be combined to improve performance characteristics of prediction models. Future work will build upon this framework as additional cytokine and transcriptomic samples are collected. One potential avenue of improving this framework is better understanding of the varying longitudinal time points of collection of specimens for measurement of cytokines and transcripts relative to pathophysiological presentation and syncing with EHR input parameters.