To identify an efficient and accurate feature selection algorithm for filtering seismic intensity indicators, the performance of four common feature selection algorithms, MIC, ReliefF, XGBoost and Lasso, was compared and analyzed. Based on the incremental dynamic analysis results of single-degree-of-freedom structures and the ground motion features, the feature selection regression model was established, the ground motion features was sorted and screened according to the Euclidean distance, the performance of the feature selection algorithm was evaluated according to the screening results, and the least squares regression model was established based on the incremental dynamic analysis results of the 2-storey, 4-storey, 8-storey and 12-storey reinforced concrete frame structures, and the standard deviation change of residual was used to measure the prediction ability of ground motion intensity measure selected by different feature selection algorithms for structural collapse. The results show that the accuracy of the ground motion features screened by the Lasso regression algorithm is 31% higher than that of other algorithms when used for structural collapse prediction. The results can be used as a feature selection algorithm reference for the selection of ground motion intensity measures in the uncertainty analysis of ground motion in the structural vulnerability analysis under the performance-based earthquake engineering (PBEE) framework, and can also be used as an effective feature selection algorithm reference for the selection of ground motion intensity measure s suitable for structural collapse prediction.