Aiming at the problem of design of the RBF neural network structure, an optimal algorithm based on variance significance in output sensitivity is proposed. Firstly, it is tested whether the variance in output sensitivity for the different patterns is significantly different from zero. If the variance in output sensitivities is significantly different from zero or not, the hidden units corresponding can be inserted or pruned. Then, the gradient descent method for the parameter adjusting ensures the fitting precision of the network. Finally, the proposed optimal algorithm is applied to the simulation experiment. Simulation results show that the proposed optimal algorithm can adjust network structure adaptive and possess good approximation and generalization ability.