Abstract
INTRODUCTION: Pneumonia is globally recognized as a significant disease burden, particularly among pediatric patients in intensive care units (ICU), where its etiology is complex and prognosis often poor. METHODS: Data were extracted from a pediatric-specific intensive care (PIC) database, selecting 795 pediatric pneumonia patients in ICUs (2010-2018). After applying rigorous inclusion/exclusion criteria, 543 cases formed the study cohort. We analyzed patient baseline information and 70 laboratory indicators to identify 25 prognosis-associated biomarkers. For prognostic model construction, we used stepwise regression to filter 28 variables, then Spearman and Pearson correlation analyses to identify an intersection of 14 key indicators from the top 20 features. Twelve machine learning algorithms underwent parameter tuning and combination, forming 113 model combinations for survival outcome prediction. RESULTS: The "Stepglm [both] + GBM" combination achieved the highest average accuracy (79.4%) in both training and testing sets. Twelve prognostic variables were identified: WBC Count, Glucose, Neutrophils Count, Cystatin C, Temperature (body), Sodium (Whole Blood), Cholesterol (Total), Absolute Lymphocyte Count, Urea, Lactate, and Bilirubin (Total). DISCUSSION: These 12 variables provide a dependable basis and novel insights for prognostic evaluation, supporting clinical diagnosis, treatment, and early intervention.