Interpretable dementia-related EEG recognition via neurophysiological Prior–Constrained dynamic graph learning and multi-expert fusion

Objectives Dementia-related electroencephalography (EEG) signals are characterized by strong nonstationarity, substantial inter-subject variability, and dynamically evolving inter-channel coupling, which makes models relying solely on fixed connectivity assumptions or a single representation view insufficient to obtain stable and interpretable discriminative evidence. This study aims to develop an accurate and interpretable window-level framework for dementia-related EEG recognition. Methods To address these challenges, this paper proposes an end-to-end window-level graph learning framework. EEG channels are modeled as graph nodes, and a weighted dynamic graph structure is learned within each time window, with neurophysiological prior constraints incorporated to suppress noise and spurious connections while enhancing physiological plausibility. Furthermore, multiple graph views are constructed and integrated through a multi-expert gated fusion mechanism, enabling adaptive aggregation of complementary representations to capture heterogeneous neurophysiological phenotypes across different disease stages and subtypes. The proposed method is evaluated independently on two datasets using a subject-wise data partition protocol. Results The experimental results demonstrate consistent advantages in Accuracy, F1-score, and AUC across the two datasets. Visualization analyses, including PR and ROC curves, representation separability, and confidence calibration, further indicate improved threshold robustness and prediction reliability. SHAP-based analyses show that the model assigns substantial importance to biomarkers associated with dementia-related EEG slowing, rhythmic abnormalities, spectral changes, and signal complexity, thereby providing traceable feature-level evidence for the model predictions. Conclusion These results suggest that the proposed framework provides an effective and interpretable modeling approach for dementia-related EEG recognition. Nevertheless, further external and prospective clinical validation is required before routine clinical implementation.