Document resource
Introduction Mass casualty scenarios, such as natural disasters, terrorist attacks and war, presents extreme challenges to triage and effective evacuation of injured patients. Similarly, for polytraumatized patients in scenarios such as the recent Covid 19 pandemic where the healthcare system is overwhelmed, effective triage is imperative to maximize resource utilization, and ultimately secure optimal treatment of patients given the available resources. We wanted to leverage machine learning techniques to estimate the need for ICU care and length of stay (LOS) for seriously injured adult patients (18 yrs or older, NISS > 15).Methods Based on data from the national Swedish trauma registry (SweTrau) from 2015 to 2019, using only data available at the time of initial presentation we developed a decision support tool for triage and planning. Missing data was handled with the missForest algorithm, and we performed feature selection using the Boruta algorithm to identify the 10 parameters with the highest significance for the outcome. Both algorithms are based on Random Forest (RF) modelling and are effective on datasets with mixed categorical and continuous data. Also, they are robust to outliers and can handle nonlinear relationships between variables. We then trained a generalized linear model (GLM), a RF model and an eXtreme Gradient Boosting (XGB) model for both outcomes separately using tenfold repeated cross validation on the data from 2015 to 2018 (n=6706). The summarized training runs were used as validation sets and the last calendar year of data from 2019 was used as a temporal internal-external test set (n=2350). As the range of days in the training set was too large to estimate individual days (1–376 days), we created categories of 1 - 2 days, 2 - 10 days and 10 days or more, rather than attempting to predict a unique day for ease of use. Results were compared using ROC plots and calibration curves.Results The XGB models were the best performers for both outcomes with AUC of 0.69, 0.64, 0.71 for the three-time intervals od admission respectively and 0.85 for ICU care needs. Both models were well calibrated. Finally, we created an online application that can be used in a hand-held device such as a phone or a tablet and that can be developed as a stand-alone app to be used without the need for internet connectivity.Conclusion While this work is based on casualties of regular peace-time trauma care, it could also be applied as a decision support tool during extreme events, and a similar tool could easily be developed if data from other scenarios were available.Abstract A20 Figure 1ROC and calibration curve for estimation of whether patients will need ICU care based on 10 easily available parameters from the initial presentation using eXtreme gradient boosting (XGB) modellingDisclosures No relevant disclosures.