Abstract:Electroencephalogram (EEG) signals is significant in dealing with seizure detection. In order to achieve early warning of seizures, weak change information of EEG in the five frequency bands of δ, θ, α, β and γ waves and the unique advantages of the graph model were fully utilized. A method for weak abnormal change detection of EEG signal based on multi-frequency graph model was proposed. Firstly, dynamic modeling of graph models on the five frequency waves of the filtered EEG signal was performed in the proposed method, and the distance function was used to quantify the similarity scores between the graph models. Then comprehensive indicators were obtained by all the similarity scores with an adaptive weighed fusion algorithm. A common null hypothesis test was finally employed for detecting whether the state of EEG signals was abnormal. The public CHB-MIT scalp EEG database and the EEG database of the Department of Neurology of the Second Hospital of Shandong University were used to conduct experiments, and precision, recall and F-score was evaluated the detection performance of the proposed method. Compared with the benchmark methods, the experimental results show that the proposed method is superior to the benchmark method in terms of precision and F-score, and the recall is 100%, it is show that the proposed method is able to detect all potential abnormal changes in weak EEG signals, achieving outstanding superiority and broad application potential for the detection of changes in all seizure moments.