Emotion-dependent integration and segregation in EEG functional brain networks revealed by data-driven sparsity optimization

Functional connectivity analysis based on electroencephalogram (EEG) provides an effective window for understanding the network-level mechanisms of emotional processing. However, traditional brain network construction methods typically relied on empirical thresholds, making it difficult to objectively reveal true emotion-specific connectivity patterns. This study proposed a data-driven sparsity optimization framework aimed at objectively identifying emotion-discriminative EEG connectivity patterns across multiple frequency bands. Functional networks based on Pearson Correlation Coefficient were constructed across five representative frequency bands, with network sparsity systematically varied from 10% to 100%. Utilizing ensemble learning models, we determined the optimal sparsity level by maximizing emotion classification performance. Under the optimized sparsity conditions, we further examined the graph-theoretic properties of three emotional states: neutral, sad, and happy. The proposed framework achieved peak classification accuracies of 94.28 ± 1.51% and 94.44 ± 1.89% in the Beta and Gamma bands, respectively, significantly outperforming fully connected networks. Crucially, network topology analysis revealed distinct emotion-dependent organizational patterns: the happy state exhibited higher global efficiency, indicating enhanced large-scale integration of emotional information; while neutral emotional states exhibited higher local efficiency and clustering coefficients, reflecting more pronounced small-world organization. These findings showed that sparsity-optimized networks boosted emotion recognition and revealed key differences in integration and segregation across emotions. The proposed method provided a principled framework for studying emotion-related brain network organization and contributed to a deeper understanding of the neural mechanisms underlying emotional regulation.