Research On Epilepsy State Recognition Algorithm Based On Physiological Signals

Objective: Epilepsy is a chronic brain disorder characterized by recurrent and unpredictable seizures caused by abnormal electrical activity in the neurons of the brain.Patients with epilepsy often have no warning signs before a seizure,and if left uncontrolled for a long time,it can lead to depression,cardiovascular disease,and even sudden death.It is of great significance to pay timely attention to the condition of epilepsy patients and intervene to improve their quality of life and reduce mortality.Electroencephalography(EEG)is a commonly used diagnostic technique for epilepsy.However,the EEG waveform of epilepsy patients is complex and varies greatly,and currently relies mainly on subjective judgments by examiners,which is inefficient.Therefore,automatic identification of epileptic states is crucial for clinical diagnosis and treatment.Although existing automatic detection methods for epilepsy have achieved certain results,they still face some problems.In the automatic classification algorithm of epileptic EEG signals,there are issues such as complex feature extraction processes,insufficient information extraction,and low accuracy of multiclassification.Moreover,analyzing and studying only EEG signals cannot fully reflect the various states of epilepsy patients,and there are certain limitations in current epilepsy automatic detection algorithms based on the fusion of multiple physiological signals,which still need to improve their accuracy.Methods and results: In this thesis,we propose an automatic detection algorithm for epilepsy based on EEG signals,ECG signals,and the fusion of the two signals,respectively.This algorithm aims to achieve the classification and identification of epileptic seizures,interictal periods,and normal states.(1)In this thesis,we designed a Residual Attention Module Neural Network(RAMNet)model for automatic classification of epileptic EEG signals.This algorithm first denoises and segments the signal to enable the network to more effectively extract detailed features.Then,according to the amplitude characteristics of the EEG signal in the timefrequency domain,the signal is transformed into a two-dimensional time-frequency image as the input of the model.Finally,inspired by the residual network,we incorporate the attention mechanism into each residual block to enhance the network’s ability to extract effective information.The innovation of this algorithm lies in directly analyzing the signal in the time-frequency domain,fully retaining the details of the signal,and overcoming the shortcomings of traditional machine learning classifiers in feature extraction.The attention mechanism enables the network to focus on key information and suppress unnecessary information,thereby improving the classification accuracy.Through clinical data verification,the classification accuracy reached 97.16%.(2)This thesis proposes an automatic epilepsy detection model based on electrocardiogram(ECG).First,the ECG signal is denoised,and then it is directly inputted into the constructed classification model.The model consists of Convolutional Neural Networks(CNN)and Long Short-Term Memory(LSTM)networks.The CNN is used to extract local key features of the ECG data,while the LSTM is used to extract temporal sequence features of the ECG data.Then,the feature data is processed by attention mechanism to extract key features,and finally inputted into a fully connected layer for classification recognition.Currently,there are few studies using ECG signals to identify epileptic states.The proposed CNN-LSTM fusion attention mechanism classification model in this paper can comprehensively extract features of ECG data and achieve automatic classification of ECG signals.Clinical data validation showed an accuracy of 96.59%.(3)This thesis proposes an epilepsy state recognition algorithm based on the fusion of electroencephalogram(EEG)and ECG signals.Firstly,the models in(1)and(2)are used for classification and recognition of EEG and ECG signals,respectively.Then,an optimal weight distribution method is adopted to perform decision-level fusion of the classification results.Compared with various decision-level fusion methods through experiments,the method proposed in this paper achieved the highest accuracy of 98.79%.The multimodal fusion method has significant advantages over using EEG or ECG alone for automatic epilepsy detection.The fusion result comprehensively considers the individual recognition results of the two signals,thereby improving the overall performance of the model.Conclusion: This thesis proposes three automatic classification algorithm models for physiological signals,which are applied to the automatic identification of epileptic patient states.The classification and recognition models of EEG and ECG were constructed respectively,and then the classification results of EEG and ECG were fused and analyzed to improve the recognition accuracy.The proposed models have important application value in clinical auxiliary diagnosis of epilepsy.They not only improve the work efficiency of clinical doctors,but also assist doctors in observing the patient’s condition in a timely manner and formulating more effective treatment plans to control the patient’s condition.This can to some extent alleviate the patient’s suffering and improve the patient’s quality of life.