EGDN-KL: Dynamic graph–deviation network for EEG anomaly detection

Electroencephalography (EEG) anomalies are sustained or transient deviations – such as epileptiform spikes, abnormal rhythmic bursts, or pathologically elevated beta power – from the statistical regularities observed in healthy cortical rhythms. We introduce EGDN–KL, a structure-learning graph neural model that couples an explicit Kullback–Leibler divergence term with attention-guided LSTM forecasting to pinpoint such deviations. Evaluated on two public datasets, EGDN–KL achieves Recall 89.4%, Specificity 97.7%, F1 88.3%, and Accuracy 87.3% using its optimal configuration (11 channels, Top-K=10). These results surpass Graph Deviation Networks (GDN), EGDN-Naïve, ResCNN, BiLSTM-Attention and GCN baselines, whose best published accuracies range from 58.2% to 82.4%. By capturing both temporal dynamics and inter-channel connectivity, EGDN–KL delivers state-of-the-art anomaly detection while localizing the neural generators that underpin pathological activity.