Abstract
Abstract
Introduction
Artificial intelligence has proven to be useful in a wide range of medical applications. The purpose of this study was to use artificial intelligence, through supervised machine learning, to predict sepsis in patients with burn injuries.
Methods
Burn-injured patients were identified from the 2010–2014 Nationwide Readmissions Database. Three machine learning classifiers --logistic regression, gradient boosted trees, and neural network-- were trained with different algorithms to predict the primary outcome of sepsis. The classifiers used categorical variables corresponding to: age, gender, TBSA percentage, burn degree, burn site, and burn mechanism. Classifier cross-validation was performed with ten groups including equal proportions of septic patients. Nine groups were used for training and one for validation. This process was repeated using each group for validation once. The receiver operating characteristic curves (ROC) were plotted for each validation and the mean areas under the curve (AUC) were calculated.
Results
There were 65,029 patients admitted for burns and the rate of sepsis was 2.8%. Logistic regression performed with an AUC of 0.876 ± 0.012 and an accuracy of 97.15%±0.04%. Neural network had an AUC of 0.860 ± 0.011 and an accuracy of 97.14%±0.10%. Gradient boosted trees performed with an AUC of 0.881 ± 0.010 and an accuracy of 97.19%±0.08%. The most important variables were TBSA ≥20% (57.32%), second degree (20.08%), third degree (4.99%), flame mechanism (2.89%), and age ≥65 (2.89%).
Conclusions
This study demonstrates the utility of artificial intelligence for the development of highly-accurate prediction models for sepsis in burn patients.
Applicability of Research to Practice
These models could be easily incorporated into future systems designed to identify and prevent septicemia in burn patients.