In this paper, we propose a feedback algorithm based on the B-spline models of channel responses(CRs) variation across subchannels. We determine the coefficients of the B-splines based on the least-squares fitting principle. The coefficients of the B-splines are fed back to the transmitter for the CRs reconstruction. In the feedback process, we also apply efficient quantization algorithms based on the Gaussian quantization. The simulation results verify that the proposed B-spline model has smaller coefficients variances, and this leads to smaller meansquares error than the Polynomial model  when using fewer number of feedback bits. The simulation results also show that the proposed algorithm can attain the upper bound of channel capacity for the precoded MIMO-OFDM system in the frequency-selective channel.