The effects of uncertainty are becoming increasingly important in structural dynamic analysis. Though there are many analytical methods, the parameter-sampled method is still one of the most effective ways to quantify response uncertainties in engineering practice. However, one severe limitation is that repeated runs of numerical realization on high fidelity simulations are extremely costly. To address this issue, a data-driven surrogate approach is presented to accelerate the uncertainty propagation of dynamical systems with frequency response function (FRF) output. A comparison study of a half-car model of vehicles is made based on a standard Monte Carlo simulation using finite element models. It is shown that the proposed method enjoys the computational advantages of predicting complete FRFs without degradation of accuracy.