TY - JOUR
T1 - An artificial neural network model for prediction of survival in trauma patients
T2 - Co-morbid conditions do not improve model performance
AU - Dirusso, S.
AU - Sullivan, T.
AU - Kamath, R.
AU - Holly, C.
AU - Cuff, S.
AU - Siegel, B.
AU - Savino, John
PY - 1999
Y1 - 1999
N2 - Overview: We have previously developed an Artificial Neural Network (ANN) for prediction of survival in trauma patients at a regional level 1 trauma center. We have validated that model and extended it to include 23 hospitals in that center's trauma region. The model has good accuracy and calibration. The NY State trauma registry also includes co-morbid conditions, which are not included in the original ANN model. In at least one patient population these were shown to relate to survival. We now test whether the addition of these variables will improve performance of the ANN model. Patient Population: 2552 patients admitted to a Regional level 1 trauma center which is part of a 23 hospital, seven county suburban/rural trauma region adjacent to a major metropolitan area. The data were generated as part of the New York State (NYS) trauma registry. Study period was from January 1995 through December of 1996 (1160 patients 1995; 1392 patients 1996). The NYS trauma registry definition of co-morbid conditions was used: cardiac disease, hypertension, pulmonary disease, liver disease, diabetes, and renal failure. Methods: A standard feed-forward back propagation neural network was developed using GCS, systolic blood pressure, heart rate, respiratory rate, hematocrit, age, intubation status, Injury E-code, and Injury Severity Score (ISS) as input variables. The model was also tested with the addition of the co-morbid conditions as addition variables. The network had a single layer of "hidden" nodes. The model was developed using the 1995 dataset and tested on the 1996 dataset. The ANN model was tested against TRISSRTS and ISS using ROC area under the curve [ROC(Az)], Lemeshow-Hosmer C-statistte (L/C statistic), and calibration curves. Results: The original ANN continues to show good clustering of the data and excellent separation of non-survivors and survivors (ROC(Az) 0.922 ANN, 0.848 TRISS (RTS), and 0.766 for ISS]. However the addition of the co-morbid conditions to the model did not improve it's performance: ROC(Az) ANN with co-morbid conditions: 0.920. Conclusions: An ANN developed for trauma patients using pre-hospital and admission data continues to give good prediction of survival. However the addition of the co-morbid conditions to the model did not improve its accuracy. The current input variables appear to be the most accurate set using the NYS trauma registry data. We are, however continuing to explore other network architectures to see if they can improve model performance.
AB - Overview: We have previously developed an Artificial Neural Network (ANN) for prediction of survival in trauma patients at a regional level 1 trauma center. We have validated that model and extended it to include 23 hospitals in that center's trauma region. The model has good accuracy and calibration. The NY State trauma registry also includes co-morbid conditions, which are not included in the original ANN model. In at least one patient population these were shown to relate to survival. We now test whether the addition of these variables will improve performance of the ANN model. Patient Population: 2552 patients admitted to a Regional level 1 trauma center which is part of a 23 hospital, seven county suburban/rural trauma region adjacent to a major metropolitan area. The data were generated as part of the New York State (NYS) trauma registry. Study period was from January 1995 through December of 1996 (1160 patients 1995; 1392 patients 1996). The NYS trauma registry definition of co-morbid conditions was used: cardiac disease, hypertension, pulmonary disease, liver disease, diabetes, and renal failure. Methods: A standard feed-forward back propagation neural network was developed using GCS, systolic blood pressure, heart rate, respiratory rate, hematocrit, age, intubation status, Injury E-code, and Injury Severity Score (ISS) as input variables. The model was also tested with the addition of the co-morbid conditions as addition variables. The network had a single layer of "hidden" nodes. The model was developed using the 1995 dataset and tested on the 1996 dataset. The ANN model was tested against TRISSRTS and ISS using ROC area under the curve [ROC(Az)], Lemeshow-Hosmer C-statistte (L/C statistic), and calibration curves. Results: The original ANN continues to show good clustering of the data and excellent separation of non-survivors and survivors (ROC(Az) 0.922 ANN, 0.848 TRISS (RTS), and 0.766 for ISS]. However the addition of the co-morbid conditions to the model did not improve it's performance: ROC(Az) ANN with co-morbid conditions: 0.920. Conclusions: An ANN developed for trauma patients using pre-hospital and admission data continues to give good prediction of survival. However the addition of the co-morbid conditions to the model did not improve its accuracy. The current input variables appear to be the most accurate set using the NYS trauma registry data. We are, however continuing to explore other network architectures to see if they can improve model performance.
UR - http://www.scopus.com/inward/record.url?scp=33750808421&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=33750808421&partnerID=8YFLogxK
M3 - Article
SN - 0090-3493
VL - 27
SP - A179
JO - Critical care medicine
JF - Critical care medicine
IS - 1 SUPPL.
ER -