OBJECTIVES: Preliminary studies have generated utility scores from SF-36 data using linear regression models, providing variable results. Of particular importance is the ability of these models to overcome floor effects. The objective of this study was to determine if an improvement in the accuracy of this modelling could be achieved using neural networks. METHODS: Data on 12,268 subjects were abstracted from the Health Outcomes Data Repository (HODaR) in Cardiff, UK, and split into training, validation and test sets. A single layer, feed-forward neural network was constructed and trained using data from 6268 respondents. A validation set containing data from 3000 respondents was used to find the optimal network structure. For comparative purposes, a linear regression model was then fitted to the same data, and both the regression model and neural network were evaluated using the independent test set data containing 3,000 respondents. RESULTS: The following results related to the minimum and maximum utility scores, lower and upper quartiles, and median and mean values, respectively. Actual results from the survey were as follows: -0.48, 1.0, 0.58, 0.88, 0.72, and 0.67, respectively. The 'trained' neural network gave the following values -0.38, 1.0, 0.53, 0.87, 0.72, and 0.68, respectively; the multiple regression model gave 0.01, 1.0, 0.47, 0.85, 0.69, and 0.67, respectively. Correlations between actual and predicted utility were 0.82 for the neural network and 0.80 for the regression model. CONCLUSION: In early analysis, the modest improvement in correlation between actual and predicted utility scores obtained via the neural network was primarily due to this class of model being more reliable in mapping lower utility values (<0.5). The bimodal distribution of EQ5D data (UK scoring algorithm) complicates standard regression modelling, favouring the use of these more flexible, data driven neural networks. It is hypothesised that these latter methods are more appropriate.