Journal article
AN ARTIFICIAL NEURAL NETWORK AS A MODEL FOR PREDICTION OF SURVIVAL IN TRAUMA PATIENTS: EXTERNAL VALIDATION WITH A DISSONANT DATA SET
Critical care medicine, Vol.27(Supplement), p.A153
12/1999
DOI: 10.1097/00003246-199912001-00438
Abstract
Objective: We had previously demonstrated that an Artificial Neural Network (ANN) based on standard pre-hospital and admission variables derived from a regional area trauma database exceeded TRISS in its ability to predict trauma deaths. We now wish to extend this to data sets outside the original population. Patient Population: Development set: 10600 patients admitted to 23 hospitals comprising a seven county suburban/rural trauma region adjacent to a major metropolitan area. The data was generated as part of the New York State (NYS) trauma registry. Study period was from 1/93 through 12/96 (7936 patients 1993/94/95; 2673 patients 1996). Testing set: 3060 patients from a Midwest Level I Trauma Center. Study period was from 7/96 through 12/98 Methods: A standard feed-forward back propagation neural network was developed using a modified input set that included variables common to both datasets: GCS, systolic blood pressure, respiratory rate, age, sex, Injury Ecode, and Injury Severity Score (ISS). The network had a single layer of 40 hidden nodes. Development of the model was done on the NYS regional 1993/94/95 data, then tested on the NYS regional 1996 data, and the Midwest Hospital data. Comparison was made against TRISS using ROC area under the curve [ROC(Az)], and Lemeshow-Hosmer C-statistic (L/H C statistic). Results: The ANN again showed good discrimination when tested against the NYS regional data (ROC(AZ) = 0.91). When tested against the Midwest data discrimination was still adequate and exceeded TRISS: ROC(Az) was 0.86 for the ANN, 0.70 for TRISS. Conclusions: An ANN for predicting trauma probability of survival showed the ability to extend its application outside its development data set. It showed adaptability to predict for a dissonant data set, and exceeded TRISS in this respect. This model deserves future development and we are undertaking these studies.
Details
- Title: Subtitle
- AN ARTIFICIAL NEURAL NETWORK AS A MODEL FOR PREDICTION OF SURVIVAL IN TRAUMA PATIENTS: EXTERNAL VALIDATION WITH A DISSONANT DATA SET
- Creators
- Stephen M DiRussoGerald P KealeyThomas H SullivanLucy WibbenmeyerLori MorganSusan SchmidtMichael BurrJohn A Savino
- Resource Type
- Journal article
- Publication Details
- Critical care medicine, Vol.27(Supplement), p.A153
- DOI
- 10.1097/00003246-199912001-00438
- ISSN
- 0090-3493
- eISSN
- 1530-0293
- Language
- English
- Date published
- 12/1999
- Academic Unit
- Surgery; Injury Prevention Research Center
- Record Identifier
- 9984321870602771
Metrics
15 Record Views