Machine Learning Models for Predicting In-Hospital Mortality in Burn Patients.

Abstract

Aim: To develop and evaluate predictive models for in-hospital mortality in burn patients using machine learning (ML) techniques.

Methods: A retrospective cohort study was conducted using data from burn patients admitted to Ankara Bilkent City Hospital Burn Treatment Center between 2015 and 2020. Key variables including age, gender, total body surface area burned, burn depth, burn type, inhalation injury, inflammatory markers and inflammatory indexes were collected. Seven ML models-Logistic Regression, Random Forest, Support Vector Machine, Decision Tree, K-Nearest Neighbors, Naive Bayes, and Gradient Boosting-were trained and evaluated.

Results: The cohort included 218 patients (mean age 42.5 ± 18.5 years; 69.7% male, 30.3% female), with an in-hospital mortality rate of 18.8% (n = 41). Logistic Regression had the best performance (accuracy: 88.6%, Receiver Operating Characteristic (ROC)-Area Under Curve (AUC): 0.906), while Random Forest achieved the highest accuracy (90.9%) and recall (97.2%). K-Nearest Neighbors excelled in recall (99.0%), Gradient Boosting balanced precision and recall (91.6% each, ROC-AUC: 0.744), and Support Vector Machine showed moderate results (accuracy: 84.0%, ROC-AUC: 0.864).

Conclusions: ML models, particularly Logistic Regression and Random Forest, demonstrated strong predictive capabilities for mortality in burn patients. This study supports the potential for ML in burn care, offering a data-driven approach for personalized prognosis and clinical decision-making. Further multicenter validation is recommended.