| To Register SMDM Homepage | |
|
|
|
||||
PURPOSE: To compare the forecasting accuracy of Bayesian networks to other standard methods for determining the chance of employment one year after traumatic brain injury (TBI)
METHODS: This study compared four predictive models: (1) an independent Bayesian network that assumes independence of all predictor nodes from each other, (2) a complex Bayesian network where dependencies are allowed between predictor nodes, (3) a logistic regression model, and (4) a classification and regression tree (CART).
Data on demographic profile, pre-injury work information, injury severity, and neuropsychological test scores at one month post-injury collected from three longitudinal studies on TBI conducted at the University of Washington from 1980 to 1987 was analyzed. Data for 337 workers who were followed up for one year was incorporated.
The two Bayesian networks were constructed using Netica. Structure was determined manually while the parameters were learned from data. Logistic regression was performed using Stata. CART analysis was performed using DTREG.
Each model was trained and tested using a 10-fold cross-validation procedure. Probabilities of being employed at one year post-TBI were predicted for each subject in each model with probabilities of 0.5 or higher indicating employment.
Sensitivity, specificity, and overall classification rates were calculated and a receiver operating characteristic analysis was done for each model.
RESULTS: The complex Bayesian network (CBN) had the highest sensitivity, specificity, and overall classification rate at 88%, 72%, and 80% respectively. It also had the best performance in the ROC analysis with an area under the curve of 0.86.
The independent belief network had a sensitivity of 78%, specificity of 67%, and overall classification rate of 72%. The logistic regression model had a sensitivity of 74%, specificity of 70%, and overall classification rate of 72%. The CART model had a sensitivity of 75%, specificity of 72%, and overall classification rate of 73%. The areas under the ROC curve for these models were 0.79 for the independent belief network, 0.80 for logistic regression, and 0.77 for the CART. Statistical tests showed the CBN performed significantly better than the other approaches.
CONCLUSION: The results demonstrate the accuracy of a data-driven complex Bayesian network in determining chance of employment one year after traumatic brain injury. Automated decision support systems should consider this approach when making forecasting judgments.
See more of Poster Session - Public Health; Methodological Advances
See more of The 26th Annual Meeting of the Society for Medical Decision Making (October 17-20, 2004)