In this research effort a neural network approach was used as a method of extrapolating the presence of mercury in human blood from animal data. We also investigated the effect of different data representations (As-is, Category, Simple binary, Thermometer, and Flag) on the model performance. In addition, we used the Rough Sets methodology to identify the redundant independent variables and then examined the proposed extrapolation model performance for a reduced set of independent variables. Moreover, a quality measure was introduced that revealed that the proposed extrapolation model performed extremely well for the Thermometer data representation.