Identifying Pediatric Sepsis in the Emergency Department

Pediatric sepsis is a major cause of morbidity and mortality, with timely identification and treatment being critical. The Emergency Department (ED) is a key catchment area for sepsis screening, but the non-specific nature of sepsis symptoms makes definitive diagnosis challenging. Leveraging existing data on physiologic and laboratory-based biomarkers can be used to determine risk and eventual diagnosis of sepsis. We present a method of compressing large-volume retrospective EHR data into a low-dimensional representation to determine the earliest time point after sepsis intervention in which sepsis diagnosis can be reliably confirmed.

Vital signs and laboratory values (SpO2, HR, RR, BP, GCS, temperature, weight, Na, K, Cl, Ca, PLT, HGB, WBC, HCT, MCV, MCH, MCHC, RDW, RBC, MPV, anion gap, pCO2, pH, lactate, bilirubin, ALT, AST, CRP, procalcitonin) were collected in a retrospective cohort of pediatric (ages 3 mo-18 yrs) ED patients (2012-2024, n=10,924) with suspected infection requiring presumptive sepsis intervention (IV antibiotics and fluid bolus), at times ranging from 0-10 hours after initial intervention. Sepsis was defined as a composite outcome consisting of ICD-10 diagnosis or dysfunction in ≥2 organ systems. Time after intervention and encoding dimensionality was optimized via a hold-out validation set, and predictive performance of the encoded features was compared with L1- and L2- regularized logistic regression models trained on the full feature set. Patient demographics, vitals, and laboratory results were used to build predictive models.

Model performance was optimized at 5 hours post-intervention (driven largely by decreased missingness) and with 5 encoding dimensions, after which point performance plateaued. This model yielded an AUROC of 0.818 (95% CI: 0.798-0.837) and AUPRC of 0.795 (0.775-0.816). This compared favorably to the best-performing fully featured model, with an AUROC and AUPRC of 0.789 (0.769-0.810) and 0.766 (0.744-0.788), indicating that compression did not impair the model’s predictive capability.

Leveraging existing data on vital signs and laboratory parameters offers a novel way towards pediatric sepsis risk prediction. This approach can be combined with incremental data such as point-of-care labs, cytokine panels, and transcriptomics to bring precision diagnostics to the patient’s bedside.