<p>Ground motion identified by pulse-like (PL) features exhibits significant effects on the built environment, since these pulses retain a substantial portion of the seismic energy, hence imposing a substantial demand on structural systems. Identifying the PL ground motions continues to be challenging due to a lack of agreement on the frequency content for defining the velocity pulses and the available diverse approaches for classifying these types of motions In this context, the present study utilizes the empirical mode decomposition technique to segment the near-fault ground motions into several intrinsic mode functions (IMFs), without predefining the frequency threshold. These real-valued IMFs are then transformed into an analytical signal using the Hilbert Transform and several key parameters are extracted from each IMF. Further, several tree-based, non-tree-based, linear, and boosting algorithms were employed to identify the optimal IMF. The results indicate IMF 4 and 5 as the optimal mode functions for the classification of ground motion as PL and non-pulse-like. The logistic regression was found to outperform the other machine learning algorithms with an accuracy of 86% and the lowest misclassification cost for IMF4. Additionally, the model validation on unseen PL ground motions and comparison of the present pulse indicator with the existing indicator shows a good agreement of results with an accuracy of 80%, which confirms the robustness and reliability of the model.</p>