Transient aware EEG representation for sleep apnea subtype classification using adaptive segmentation

Accurate sleep apnea subtype classification is important because obstructive, central, and mixed apnea reflect different physiological mechanisms and may require different clinical management. Although EEG captures cortical responses to respiratory disturbances and microarousal, most existing methods treat each 30 s epoch as homogeneous, potentially obscuring brief and localized patterns. This study proposes a transient aware framework for single-channel EEG analysis. It combines event-sensitive intra-epoch segmentation, complementary segment encoding, and structured aggregation, enabling the joint preservation of transient heterogeneity and global spectral context. Multiscale spectral divergence and gamma-band burst cues identify intra-epoch boundaries. Hermite decomposition and Fourier sub-band analysis describe the resulting variable-length segments. Distributional, temporal, and diversity-related statistics then produce a fixed-dimensional representation for classification. On the UCDDB dataset, subject-wise five-fold cross-validation yielded a macro-F1 of 0.8546 ± 0.0231 for four-class classification among Normal, OSA, CSA, and MSA, and an F1 of 0.8608 ± 0.0141 for binary Normal-versus-Apnea detection. In subject-disjoint ablation experiments, adding adaptive local modeling to fixed-epoch spectral features increased the F1 by 6.3 and 5.8 percentage points in the four-class and binary tasks, respectively. Experiments on MIT-BIH further supported the applicability of the representation. These results indicate that transient intra-epoch EEG structure provides complementary information for sleep apnea subtype classification.