Abstract Systemic inflammation after pediatric burn injury frequently causes fever, complicating early recognition of infectious complications. Improved risk-stratification may help identify patients at risk for adverse clinical events during hospitalization. This study aimed to develop and validate a machine learning (ML)–based model using a Random Forest (RF) algorithm to predict fever and related adverse outcomes in hospitalized pediatric burn patients. We conducted a retrospective analysis of 595 pediatric burn patients admitted to a tertiary center between 2012 and 2022. Extracted data included demographics, burn characteristics, clinical interventions, laboratory values, and outcomes. RF models were trained to predict three key endpoints: fever (38.5°C), transfer to pediatric intensive care unit (PICU), and need for surgical intervention. To address missing data and class imbalance, we employed multiple imputation techniques and generated synthetic data through bootstrap sampling to improve model robustness. The patient cohort had a mean age of 4.27 (range: 0.2-18.1) years and an average total body surface area (TBSA) of 5.49 (range: 0.3-45.0). The RF models demonstrated high predictive accuracy, with F1-scores of 0.81±0.037 (fever), 0.88±0.091 (PICU transfer), and 0.81±0.027 (surgery). Area Under the Curve (AUC) values were 0.96, 0.97, and 0.95, respectively. Feature importance analysis identified younger age, lower body weight, female sex, and head and neck burn location as key predictors. These ML-based RF models demonstrate strong potential for early risk-stratification of fever and high-risk trajectories in hospitalized pediatric burn patients, guiding monitoring intensity, diagnostic vigilance, and resource planning. Prospective evaluation is needed to determine whether model-informed workflows improve outcomes.
Har‐Shai et al. (Thu,) studied this question.