• Nehemiah T Liu, Julie A Rizzo, Beth A Shields, Maria L Serio-Melvin, Robert J Christy, and José Salinas.
• U.S. Army Institute of Surgical Research, JBSA Fort Sam Houston, Texas.
• J Burn Care Res. 2018 Aug 17; 39 (5): 661-669.

AbstractThe intrinsic relationship between fluid volume and open wound size (%) has not been previously examined. Therefore, we conducted this study to investigate whether open wound size can be predicted from fluid volume plus other significant factors over time and to evaluate how machine learning may perform in predicting open wound size. This retrospective study involved patients with at least 20% TBSA burned. Various predictive models were developed and compared using goodness-of-fit statistics (R2, error [mean absolute error (MAE), root mean squared error (RMSE)]). Bland-Altman analysis was also performed to determine bias. A total of 121 patients were included in the analysis. Median TBSA burned was 31% (interquartile range: 26-46%). Average crystalloid volumes were 4.0 ± 2.7 ml/kg/TBSA in the first 24 hours. There were 24 (20%) patients who died. Importantly, multivariate analysis identified seven independent predictors of open wound size. Also, machine learning analysis was able to stratify patients based on the 20th day after admission, ~40% TBSA burned, and fluid volumes. Models for predicting open wound size varied in performance (R2 = .79-.90, MAE = 3.97-7.52, RMSE = 7.11-10.69). Notably, a combined machine learning model using only four features (fluid volume, days since admission, TBSA burned, age) performed the best and was sufficient to predict open wound size, with >90% goodness of fit and <4% absolute error. Bland-Altman analysis showed that there were no biases in the models. Open wound size can be predicted reliably using machine learning and fluid volume, days since admission, TBSA burned, and age. Future work will be needed to validate the utility of this study's models in a clinical environment.

19 keywords...

hide…