The present study investigates the best seismic parameters for modeling the dynamic response of various non-linear structural systems by comparing different Machine Learning (ML) algorithms. A total of 400 synthetic excitations were generated and analyzed against 23 seismic parameters. These signals were used in a step-by-step numerical analysis to calculate the dynamic responses of 1000 single-degree-of-freedom (SDOF) systems with varying mechanical properties. The data obtained from these responses were processed using 20 ML algorithms, including linear regression, tree, support vector machine (SVM), boosted and bagged trees, and artificial neural network (ANN). Each ML algorithm used a single seismic parameter as input to determine the most predictive parameters for modeling structural responses, defining the high predictive seismic parameters (HPSP) set. To validate the obtained results, the most effective model predictions have been compared with the results of the parametric step-by-step analyses performed for a new group of natural ground motions. The findings demonstrate that with a properly calibrated training phase, considering the specific site hazard and selecting seismic parameters from the HPSP set, the ML model can accurately estimate seismic responses whit a significantly reduced computational effort. This study underscores the potential of integrating ML techniques into the performance-based seismic design approach.