Development and external validation of age-specific predictive models for sepsis in critically ill neonates and children using multidimensional hematological parameters: a multicenter retrospective study

Sepsis remains a leading cause of morbidity and mortality in neonatal and pediatric intensive care units (NICU/PICU). Given the rapid clinical deterioration of these vulnerable populations, shifting from reactive diagnosis to proactive prediction of sepsis onset is crucial. Multidimensional hematological parameters may provide new insights for risk stratification. While routine complete blood count (CBC) is widely used, the predictive potential of high-dimensional hematological cell morphology features that extend beyond traditional CBC remains under-investigated in large-scale pediatric cohorts. This multicenter study utilized a large-scale dataset from the CALM2302 project, encompassing 17,066 critically ill neonates and children across 30 Chinese hospitals between October 2020 and October 2023. We developed population-specific CatBoost models for NICU and PICU to predict sepsis onset. To mitigate class imbalance, random undersampling was applied to the training sets. From a candidate pool of 1,924 parameters (37 routine CBC and 1,887 research features), optimized signatures of approximately 200 features were selected for each population. Model performance was evaluated by the area under the ROC curve (AUC) across internal and dual external validation cohorts. Furthermore, SHAP (SHapley Additive exPlanations) analysis was performed to ensure model interpretability and identify key biological predictors. Among the total cohort, 10,347 infants were categorized into the NICU cohort (sepsis incidence: 7.48%), while 6,719 children formed the PICU cohort (sepsis incidence: 10.92%). CatBoost was identified as the optimal architecture after benchmarking nine machine learning algorithms. The population-specific simplified 18-variable models demonstrated robust performance. In the NICU cohort, the model achieved an internal AUC of 0.846 and external AUCs of 0.742 and 0.827. For the PICU cohort, corresponding AUCs were 0.803, 0.717, and 0.811. These simplified models maintained highly comparable discriminative power to the full models (range of AUC attenuation: 0.005–0.027). SHAP analysis identified 18 key predictive features for each cohort, with only two shared features between the two populations and the remaining being cohort-specific. Notably, most of these 18 features were research parameters rather than traditional blood counts indices. Furthermore, most of these features remained statistically significant in multivariable analyses after adjustment for age, sex, and study sites. In this large multicenter cohort, we developed and validated simplified 18-variable models that accurately predict sepsis in both neonatal and pediatric populations. By capturing early hematological morphological alterations beyond traditional blood counts, these models offer a robust, interpretable, and clinically accessible solution for proactive sepsis management in intensive care units.