A comparison of neural network models for the prediction of the cost of care for acute coronary syndrome patients.

Acute coronary syndromes have remained the focus of many clinical economic studies due to the increasing prevalence of the disease and the tightening of cost controls. An accurate descriptive cost model for this population would be a valuable tool for clinical researchers. With such a model, the relative importance of different factors upon the total cost of care could be determined through computer simulation. This study explored the use of different neural network architectures in creating a descriptive cost model. This was a difficult problem in that the costs span 3 orders of magnitude but the output variable of the neural network must be restricted to the range 0-1. Models that used logarithmic transformations and multiple modular networks were created and analyzed. It was found that the model with a single network and logarithmic transformation performed significantly better than other more complicated networks.