Random Forests — A Better Anatomical Injury Severity Characterization (National Trauma Databank 2011-2015 Study)
Radu Filipescu1, Colin Powers1, Han Yu2, David Rothstein1, Carroll M Harmon1, Brian Clemency3, Weidun A Guo4, Kathryn Bass1
1Oishei Children's Hospital, Buffalo, NY;2Roswell Park Comprehensive Cancer Center (Biostatistics and Bioinformatics), Buffalo, NY;3Department of Emergency Medicine, SUNY Jacobs School of Medicine, Buffalo, NY;4Department of Surgery, SUNY Jacobs School of Medicine, Buffalo, NY
BACKGROUND—Injury Severity Score (ISS) overestimates mortality in children for ISS<25. Methodologically, it considers the absolute value of the worst Abbreviated Injury Scale (AIS) in 3 body regions (BR) and it does not account for combination of injuries. We hypothesized a random forest (RF) model, using the AIS, body region (BR) injured and age, would predict mortality more accurately across all ages.
METHODS—Using NTDB 2011-2015 data 3 cohorts were devised: pediatric (≤18), adult and combined. AIS mapped to 9 BR were obtained by deconstructing the injury codes. For BR with more than one injury, the worst AIS was used. RF models were trained and validated separately on all 3 cohorts and their performance was characterized as a significant change in the area under the curve (AUC) compared to ISS. The predictive impact of each variable was expressed as mean decrease of Gini Index.
RESULTS—We included 644,007 children and 3,283490. Mortality was 0.9% in children and 2.8% in adults. RF model outperformed ISS in both children (AUC 0.9538 vs 0.9396, P<0.001) and adults (AUC 0.8795 vs 0.8451, P<0.001). RF model trained and validated on the entire cohort, yielded an improved AUC in children (AUC-0.9555, P<0.001) with similar performance in adults. Age, injuries to the head, thorax and abdomen were the strongest predictors for mortality in the combined cohort.
CONCLUSIONS—RF methodology allows for a better characterization of the anatomical injury severity as it adjusts for age and takes account of injury combination. RF method may be a useful tool in comparing outcomes in pediatric trauma patients treated at mixed and adult facilities.
Back to 2019 Abstracts