Using State Data to Predict a Single Institution Mortality for Patients That Fall

Background: Falls are the most common cause of injury-related death for patients older than 45. We hypothesized that a machine learning algorithm developed from state-level registry data could make accurate outcome predictions at a level 1 trauma hospital. Methods: Data for all patients admitted for fall injury during 2009 – 2019 in the state of Pennsylvania were derived from the state trauma registry. Thirteen variables that were immediately available upon patient arrival were used for prediction modeling. Data for the test institution were withheld from model creation. Algorithms assessed included logistic regression (LR), random forest (RF), and extreme gradient boost (XGB). Model discrimination for mortality was assessed with area under the curve (AUC) for each algorithm at our level 1 trauma center. Results: 180,284 patients met inclusion criteria. The mean age was 69 years ± 18.5 years with a mortality rate of 4.0%. The AUC for predicting mortality in patients that fall for LR, RF, and XGB were 0.797, 0.876, and 0.880, respectively. The variables which contributed to the prediction in descending order of importance for XGB were respiratory rate, pulse, systolic blood pressure, ethnicity, weight, sex, age, temperature, Glasgow Coma Scale (GCS) eye, race, GCS voice, GCS motor, and blood alcohol level. Conclusions: An extreme gradient boost model developed using state-wide trauma data can accurately predict mortality after fall at a single center within the state. This machine learning model can be implemented by local trauma systems within the state of Pennsylvania to identify patients injured by fall that require greater attention, transfer to a higher level of care, and higher resource allocation.