From data to decisions: Predicting inpatient burn mortality with advanced classification models.

Area of Science:

• Medical Informatics
• Data Mining in Healthcare
• Burn Injury Research

Background:

• Burn injuries represent a significant global health burden.
• Accurate mortality prediction models are crucial for effective burn patient management and treatment decisions.
• Ensuring the long-term reliability of predictive models is essential for safe clinical application.

Purpose of the Study:

• To develop a high-performing machine learning model for predicting mortality in burn patients.
• To rigorously evaluate the temporal stability and clinical reliability of the developed model.
• To identify key predictors of mortality in burn injuries.

Main Methods:

• A retrospective cohort study involving 651 burn patients and 93 predictive features.
• Development and comparison of five tree-based models, including Gradient Boosted Trees (GBT).
• Model evaluation using 10-fold cross-validation and temporal validation on distinct patient cohorts.

Main Results:

• The GBT model achieved high performance with 93.1% accuracy and an AUC of 0.966 during cross-validation.
• Key predictors identified include the Abbreviated Burn Severity Index (ABSI), Total Burn Surface Area (TBSA), and percentage of third-degree burns.
• Temporal validation showed good model stability (AUC 0.948), but a decrease in sensitivity from 78.1% to 68.3% was noted.

Conclusions:

• Tree-based models, especially GBT, are effective for predicting burn mortality.
• Despite strong temporal stability, performance shifts (e.g., reduced sensitivity) necessitate continuous monitoring and recalibration.
• A robust governance framework is critical for maintaining the safety and efficacy of clinical predictive models over time.