To address the challenge of decoding motor imagery-based electroencephalography (EEG) signals, a novel casual sparse optimization method was proposed, focusing on selecting casual discriminative features to improve decoding performance. Specifically, The Hilbert-Huang transform (HHT) extracted marginal spectral energy from EEG channels. And the Sample Reweighted Decorrelation Operator (SRDO) optimized features by weighting, eliminating spurious relationships among interfering features and discriminative features crucial for the decoding of motor imagery patterns. Particularly, the Affinity Propagation (AP) clustering algorithm identified potential relationships among feature spatial distributions. And the iterative Sparse-Group Lasso (iSGL) optimized features by considering features’ importance with intra- and inter-group to improve the decoding accuracy of motor imagery-based EEG signals. A 5-fold cross-validation experiment on a binary classification motor imagery dataset using the Support Vector Machine (SVM) showed an average accuracy of 92.30%, nearly 4% higher than the original features. Comparative experiments demonstrated the superiority of the proposed method, suggesting its effectiveness for advancing brain-computer interfaces.