Predicting High Variable Direct Cost in Burn Care: A Comparison of Logistic and Random Forest.

INTRODUCTION: Burn care is uniquely resource-intensive, and early identification of patients at risk for high hospital-incurred costs may enable proactive care coordination and more efficient resource allocation. This study uses the actual hospital cost data, rather than indirect billing metrics, to develop early prediction models for high-cost burn admissions. METHODS: We performed a retrospective cohort study of 363 adult and pediatric burn patients admitted to a regional burn center between 2022 and 2023. High variable direct cost was defined as variable direct cost exceeding the 75th percentile of the study cohort. Two predictive models, logistic regression and Random forest, were developed using admission variables available early in the hospital course: age, total body surface area, inhalation injury, and burn mechanism. Model performance was evaluated using area under the receiver operating characteristic curve, calibration, and decision curve analysis (DCA) to assess clinical utility. RESULTS: High variable direct cost occurred in 24.8% of patients (90/363). Both models demonstrated similar discrimination (area under the receiver operating characteristic curve ≈ 0.74), with total body surface area as the most influential predictor. Logistic regression showed higher specificity (0.823) and more stable calibration, whereas the Random forest model demonstrated higher sensitivity (0.786) but greater calibration instability. On DCA, the Random forest model provided greater net benefit across clinically relevant threshold probabilities (0.05-0.25), indicating superior utility for identifying patients at risk for high costs. CONCLUSIONS: Admission-based prediction models can identify high-resource burn patients early in their hospitalization. Machine learning approaches, particularly Random forest modeling, may better support prospective resource planning, though calibration limitations warrant caution before use in patient-specific decision-making. Incorporating clinical utility metrics such as DCA helps align predictive performance with real-world operational decision-making in burn care.