Seizure Prediction Based On Wavelet Packet Features And Blending Classifier

Epilepsy is a brain disease caused by the abnormal discharge of brain neurons,which is sudden and repetitive.When the patient is sick,symptoms such as syncope and involuntary tremor may occur,and in severe cases,there may even be life-threatening.Therefore,accurate identification of the state before the onset of epilepsy can enable patients to take effective preventive measures in advance to avoid secondary injuries caused by epileptic seizures.The EEG signal records the discharge of neurons and is an important tool for studying the characteristics of epileptic seizures.The EEG signals of epilepsy patients have great differences in the characteristics of seizures,pre-onset and pre-seizure presentations.The study focused on seizure detection based on EEG signals,and there are few studies specifically for pre-onset prediction.Compared with seizure detection,accurate pre-prediction can help patients to take effective protective measures in advance and reduce secondary injuries.There are two main defects in the current study:(1)The pre-onset period is generally defined as about 1 hour,which is too large for the pre-seizure prediction.Premature pre-onset predictions can cause anxiety and adverse effects,and too late pre-expectations can lead to patients not having enough time to take protective measures.(2)For individual patients,the duration of the attack period is much less than that of the intermittent period and the pre-seizure period,which forms a typical problem of uneven set classification;however,the existing research results do not consider the imbalance of data distribution.In view of the above defects,this paper firstly divides the hour before the attack more subtly,and proposes a pre-seizure prediction model based on the nonlinear characteristics of wavelet packet and random forest.The validity of the model is demonstrated by experiments.Secondly,the research is carried out.The seizure detection and pre-seizure prediction methods based on the unbalanced set classification algorithm were optimized from two aspects of data set and classification algorithm respectively,and the unbalanced set classification model of KM-SMOTE+Blending was proposed.The specific research contents are as follows:1.A pre-seizure prediction model based on the nonlinear characteristics of EEG signal wavelet packet and random forest is proposed.Firstly,the upper part of seizure is divided equally,and then the wavelet packet nonlinearity,the Shannon entropy,the logarithmic energy entropy and the norm entropy are extracted on the basis of wavelet packet decomposition.Finally,the features are classified by random forest.To achieve the classification of epilepsy status.Experiments show that the average classification accuracy of the pre-seizure prediction model based on the EEG signal wavelet packet nonlinear feature combined with random forest on the CHB-MIT dataset is 84.82.A method for detecting pre-seizure and episodes based on the unbalanced set classification algorithm is proposed.In the long-term continuous EEG detection,seizures are sudden and short-lived,which constitutes a typical problem of unbalanced set classification.To classify this unbalanced EEG data,the KM-SMOTE algorithm is first used to equalize the unbalanced EEG dataset,and then the Blending algorithm is used to fuse multiple types of random forest classification.The experimental results show that the recognition model based on the KM-SMOTE and Blending fusion algorithms can effectively improve the predictive effect of pre-seizure state.