Sparse EEG neurodynamics of hypoglycemia in type 1 diabetes with personal euglycemic baseline adaptation

Introduction Hypoglycemia alters cerebral electrophysiological activity, providing a physiological basis for EEG-based detection. However, translation toward subject-independent monitoring remains constrained by uncertain minimum input requirements, insufficiently characterized stability of sparse representations and their physiological associations, and inter-participant baseline variability. Methods This study benchmarked 24 models spanning engineered-feature machine learning, feature-domain deep learning, and raw-waveform deep learning under leave-one-subject-out validation. Channel–Feature Sparsification was applied using a predefined performance-preservation criterion. Sparse representations were evaluated for selection stability, cross-paradigm concordance, and physiological association. Personal Euglycemic Baseline Adaptation (PEBA) was subsequently applied using a short personal euglycemic reference without participant-specific model retraining. Results LDA achieved the highest observed balanced accuracy within the engineered-feature paradigm, whereas Sinc-SSNET performed best in both deep-learning paradigms. Sparsification retained the F7 log-theta descriptor for LDA, the F7 channel row for feature-domain Sinc-SSNET, and two channels for raw-waveform Sinc-SSNET. F7 was retained across all three paradigms, and its log-theta descriptor increased during hypoglycemia in 30 of 34 participants (88.2%). PEBA improved balanced accuracy from 0.71 to 0.77 for LDA, from 0.68 to 0.83 for feature-domain Sinc-SSNET, and from 0.71 to 0.78 for raw-waveform Sinc-SSNET. Compared with the corresponding full-input models, reduced configurations decreased parameter counts by 88.51%–99.88%, parameter storage requirements by 87.20%–99.88%, and multiply–accumulate operations by 94.72%–99.91%. Discussion These findings demonstrate the preservation of hypoglycemia-related EEG information under a predefined sparsification criterion, while personal euglycemic normalization improves cross-participant discrimination. Physiological consistency and reduced computational requirements support further evaluation of compact EEG-based hypoglycemia detection. Independent cohorts, continuous glycemic-transition recordings, and native sparse EEG acquisition are required before prospective application.