PCT-Anchored Machine Learning for Pre-Culture Identification of Gram-Negative Sepsis in Children: A Four-Site Study

Procalcitonin (PCT) has moderate accuracy for bacteremia detection but is infrequently used in pediatric emergency departments, partly due to poor sensitivity when applied as a standalone threshold and guideline recommendations against its isolated use. We evaluated whether a machine learning (ML) approach that contextualizes PCT with additional clinical features has the potential to improve detection of gram-negative bacteremic organ dysfunction in children prior to culture results.

We conducted a retrospective analysis of 431 pediatric encounters across four sites where blood culture and PCT were co-ordered. Blood culture results were classified as gram-negative bloodstream infection (BSI), gram-positive BSI, contaminant, or negative using a priority-based NLP algorithm applied to free-text result strings. The primary outcome was gram-negative BSI with concurrent organ dysfunction ascertained by Phoenix-8 criteria excluding the immunologic domain. We evaluated PCT alone, a four-feature adult sepsis-aligned benchmarking model, and candidate multi-feature combinations identified through systematic univariate screening, using Random Forest (RF) and penalized logistic regression (LASSO) as co-primary algorithms with nested repeated stratified cross-validation and permutation importance analysis. Model discrimination was compared using paired fold-level AUROC testing and calibration was assessed using Platt scaling.

Among 431 encounters with a numeric PCT result, 20 (4.6%) met the primary outcome of gram-negative BSI with Phoenix-8 organ dysfunction. PCT alone achieved AUROC 0.762. A four-feature adult sepsis-aligned model achieved AUROC 0.862; respiratory rate and systolic blood pressure contributed negligibly. Systematic screening identified platelet count and creatinine as optimal co-features; the three-feature model (PCT, platelet count, creatinine) achieved AUROC 0.884 (BCa 95% CI 0.871-0.982), consistent across RF and LASSO (0.874), with Brier score 0.038 below the null model (0.044), and positive likelihood ratio 12.56 (number needed to assess of 2.6).

PCT alone showed limited performance. Embedding PCT within a 3-feature ML model improved discrimination over PCT alone (delta AUROC 0.122, p<0.001), with consistent performance across RF and LASSO. Unlike adult models, pediatric prediction relied on renal dysfunction (creatinine) rather than hemodynamics. These findings are hypothesis-generating and require prospective validation in adequately powered cohorts before clinical implementation.