Machine Learning Models for Mortality Prediction in Burn Patients: Performance, Reliability, and Clinical Translation
Fangfang Xu1,2, Tong Wang2, Siyao Huang2
1The First Affiliated Hospital, Department of Burn, Plastic Surgery and Wound Repair, Hengyang Medical School, University of South China, Hengyang, Hunan, 421001, China.
Introduction:
Despite advances in contemporary burn care, the discriminative performance and clinical applicability of traditional prognostic scoring systems may be increasingly limited. Machine learning models have been applied to predict burn mortality; however, their overall predictive performance and methodological quality have not been systematically evaluated.
Methods:
This systematic review was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) statement. Discriminative performance was assessed using the C-statistic. Pooled C-statistics and corresponding 95% credible intervals for machine learning models were estimated within a Bayesian framework. Subgroup and sensitivity analyses were performed according to model type, including random forest, support vector machine, and logistic regression, as well as the degree of class imbalance. Risk of bias and publication bias were also assessed.
Results:
Twelve studies using diverse algorithms and data sources were included. The pooled C-statistic for the best-performing models was 0.96 (95% CrI: 0.93-0.98; 95% PI: 0.84-1.00), indicating excellent overall discrimination. Ensemble and more complex algorithms, such as random forest and support vector machine, outperformed single decision- tree models. Logistic regression also demonstrated stable performance. After excluding datasets with severe class imbalance, the overall findings remained largely unchanged. Funnel plot assessment and Egger's test did not indicate clear evidence of publication bias.
Discussion:
The findings indicate that ML models have considerable potential for predicting mortality in patients with burns. However, methodological limitations, including a high risk of bias, limited external validation, and insufficient model interpretability, may restrict their current clinical applicability.
Conclusion:
ML models demonstrate excellent discriminative performance for predicting mortality in patients with burns. Future high-quality multicenter studies with rigorous external validation, transparent reporting, and improved model interpretability are needed before these models can be routinely implemented in clinical practice.
Registration:
The protocol for this study was registered with PROSPERO (registration number: CRD420251153818).