On the basis of the per capita treatment cost per month of five common diseases, i.e. diabetes, intestinal polyp, hyperthyroidism, eutocia and cerebral infarction, in a 3-A-grade hospital in Chongqing from January, 2012 to December, 2014, we used BP neural network, generalized regression neural network (GRNN), grey system GM (1, 1) and non-linear regression analysis to predict the change of per capita treatment costs per month of these five diseases from January 2015 to August 2015. And the accuracy of these four models was judged by comparing the prediction results with real data. The results show that the minimum coefficients of determination (R²) of the four models are 0.278, 0.565, 0.048 and 0.097, respectively, while their maximum coefficients of determination(R²) are 0.826, 0.901, 0.600 and 0.747, respectively. The minimum prediction errors of the four models are 9.845%, 3.507%, 5.897% and 3.642%, respectively, while their maximum prediction errors are 15.450%, 13.940%, 30.518% and 17.204%, respectively. Compared with the other three models, the GRNN model can predict the cost of diseases more accurately.