An Intelligent EEG State Transition Detection Method for Cardiac-Related Neurodynamic Alterations Based on Multilayer Perceptron and Maximum Mean Discrepancy

Abstract Background Electroencephalography (EEG) is a high–temporal–resolution technique for monitoring central nervous system activity and is of considerable importance in elucidating autonomic regulation and its interaction with the cardiovascular system. In patients with cardiovascular disease, autonomic imbalance preceding acute cardiac events is frequently accompanied by alterations in central neural states, which may manifest as characteristic reorganisation of EEG rhythms and state transitions. These EEG state transitions may reflect early central nervous system responses to increased cardiac load and autonomic dysregulation, thereby providing quantifiable neurophysiological signals relevant to the assessment of potential cardiogenic cerebral risk. Purpose The purpose of this study was to propose an intelligent method for accurately detecting EEG state transitions associated with autonomic dysfunction, with the aim of supporting the exploration of early identification and monitoring of cardiogenic cerebral risk in cardiovascular disease. Methods We propose an intelligent EEG state transition detection method integrating a multilayer perceptron (MLP) classifier with the maximum mean discrepancy (MMD) statistic. The MLP is employed to extract nonlinear neural dynamic features from local EEG segments, while MMD quantifies distributional shifts between distinct neural states associated with changes in autonomic–cardiovascular control. These complementary outputs are intelligently integrated through fusion clustering to enhance robustness and reduce false positives. Method performance was first assessed using Monte Carlo simulations under noisy and correlated conditions, and subsequently compared with classical kernel change detection (KCD) and RuLSIF methods using paired t-tests. The proposed method was then applied to real EEG datasets to evaluate its practical utility in monitoring heart-related neurodynamic changes. Results On real EEG data, the proposed method achieved a recognition accuracy of 98.9%, an F1 score of 0.8, a false positive rate of 0.6%, and a positioning accuracy of 0.82. Paired t-tests demonstrated that the proposed method significantly outperformed KCD and RuLSIF in both positioning accuracy and false positive rate. The corresponding Cohen’s d values for positioning accuracy were 2.077 and 2.25, and for false positive rate were 1.196 and 22.817, respectively. These findings indicate that the method can sensitively capture subtle EEG dynamic changes associated with autonomic dysfunction preceding cardiac events. Conclusions The proposed EEG state transition point recognition method, integrating a multilayer perceptron with maximum mean discrepancy, enables accurate identification of state transitions in complex EEG signals while effectively reducing false positives. Continuous monitoring of EEG state transition points using this approach may contribute to early warning for cardiogenic cerebral risk in patients with cardiovascular disease, and may support risk stratification and clinical decision-making by providing quantitative, real-time neurodynamic information.