The effect of well-known burn-related features on machine learning algorithms in burn patients' mortality prediction

Abstract

Background: Burns is one of the most common traumas worldwide. Severely injured burn patients have an increased risk for mortality and morbidity. This study aimed to evaluate well-known risk factors for burn mortality and comparison of six machine learn-ing (ML) Algorithms' predictive performances.

Methods: The medical records of patients who had burn injuries treated at Izmir Bozyaka Training and Research Hospital's Burn Treatment Center were examined retrospectively. Patients' demographics such as age and gender, total burned surface area (TBSA), Inhalation injury (II), full-thickness burns (FTBSA), and burn types (BT) were recorded and used as input features in ML models. Pa-tients were analyzed under two groups: Survivors and Non-Survivors. Six ML algorithms, including k-Nearest Neighbor, Decision Tree, Random Forest, Support Vector Machine, Multi-Layer Perceptron, and AdaBoost (AB), were used for predicting mortality. Several different input feature combinations were evaluated for each algorithm.

Results: The number of eligible patients was 363. All six parameters (TBSA, Gender, FTBSA, II, Age, BT) that were included in ML algorithms showed a significant difference (p<0.001). The results show that AB algorithm using all input features had the best predic-tion performance with an accuracy of 90% and an area under the curve of 92%.

Conclusion: ML algorithms showed strong predictive performance in burn mortality. The development of an ML algorithm with the right input features could be useful in the clinical practice. Further investigations are needed on this topic.

Figure 1

Comparison of k-NN (a), DT (b), RF (c), SVM (d), MLP (e), and AB (f), respectively. Red circles indicate burn injury patients, blue circles indicate non-burn injury patients, and black circles indicate unclassified patients.

Figure 2

Comparison of ML algorithms’ ROC curves with different combinations. False-positive rates (1–specificity) and true-positive rates (sensitivity) are reported on the x- and y-axis, respectively. (a) ROC curves with all features (b) ROC curves only without Age (c) ROC curves only without TBSA (d) ROC curves without Age and TBSA (e) ROC curves only with Age and TBSA. TBSA: Total Burned Surface Area.

Figure 3

Performance metrics of algorithms based on accuracy, sensitivity, specificity, PPV, and AUC.