Seeing the forest beyond the trees: Predicting survival in burn patients with machine learning

Area of Science:

• Medical Informatics
• Public Health
• Burn Care Research

Background:

• Burn injuries represent a significant public health concern requiring effective survival prediction.
• Identifying factors influencing survival is crucial for optimizing patient care and resource allocation.

Purpose of the Study:

• To identify patient- and hospital-level predictors of survival in burn patients.
• To apply machine learning techniques for robust predictive modeling in burn care.

Main Methods:

• Utilized Healthcare Cost and Utilization Project (HCUP) State Inpatient Databases (SID) for California, Florida, and New York.
• Integrated patient data with hospital data from the American Hospital Association (AHA) Annual Survey.
• Employed random forest and stochastic gradient boosting (SGB) machine learning models to identify survival predictors.

Main Results:

• Analyzed data from 31,350 burn patients across 670 hospitals.
• Stochastic gradient boosting (AUC 0.93) outperformed random forest (AUC 0.82) in predicting survival.
• Identified patient factors (e.g., age, absence of renal failure) and hospital factors (e.g., full-time residents, nursing staff) significantly associated with survival (p < 0.001).

Conclusions:

• Both patient-specific characteristics and hospital operational factors significantly predict burn patient survival.
• While patient factors are often uncontrollable, hospital-level factors offer actionable insights for improving treatment location decisions and resource management.