P149. An artificial neural network approach to predicting mortality after emergency spine trauma surgery with near 90% accuracy

BACKGROUND CONTEXT Traumatic spine injury is a complicated injury to treat that carries a high risk of death. Current research has elucidated on patient characteristics that increase the risk of mortality during hospitalization after surgery. However, no definitive model has been presented to take into consideration the interaction of these characteristics to predict whether a given patient will die during hospitalization. PURPOSE In this study, one of the most advanced machine learning algorithm, Artificial Neural Network (ANN), is studied to determine its ability to predict postoperative mortality during hospitalization. STUDY DESIGN/SETTING The study was an analysis of National Inpatient Sample data. PATIENT SAMPLE The National Inpatient Sample (NIS) was queried between 2005 and 2014 to select patients who had been admitted to treat emergency trauma related injuries to the spine. OUTCOME MEASURES Death during hospitalization. METHODS After patient selection, the neural network's predictive factors were chosen: age, APR DRG Mortality Risk Score, primary procedure, preoperative blood loss, history of anemia deficiency, arthritis, congestive heart failure, chronic lung disease, coagulopathy, hypothyroidism, chronic diabetes, or neurological deficiency as well as obesity, gender, race, and principal insurance provider. Through random selection, patient data was split between training and testing data. Training data was used to calibrate the ANN, while the testing data was used to determine the accuracy of the algorithm. Prediction accuracy and area under the Receiver Operating Curve (AUC) were calculated to determine the ANN's efficacy in predicting mortality during hospitalization after surgery. RESULTS In this study, 9,244 patients were analyzed, of which 9.07% experienced mortality during hospitalization after surgery. Among the sample patients, 39.6% were female, 67.6% were Caucasian, and the average age was 49.1 years. The ANN was able to predict 87.7% of the test cases with an AUC of 0.849. CONCLUSIONS The ANN presented here demonstrated incredible predictive capacity in which patients would experience would die during hospitalization. Consequently, the ANN shows incredibly promise as tool to assess mortality risk for patients admitted to the hospital for traumatic spine injury. Additional research must be conducted to determine how the ANN can be integrated into the electronic medical records for clinical usage. FDA DEVICE/DRUG STATUS This abstract does not discuss or include any applicable devices or drugs.