Research On Amplitude-integrated EEG Image Reconstruction And Classification Based On Ensemble SVM

Amplitude-Integrated EEG is an effective long-term monitoring technique for neonatal brain function.Due to its low cost,simple operation and long-term monitoring,aEEG has been increasingly used in the monitoring of clinically ill neonates in neonatal intensive care unit(NICU).Developing an effective method for automatic interpretation of aEEG tracings is very important.It not only liberates doctors from identifying brain disorders,making them focus on solving these brain diseases,but also has a far-reaching significance on the use and promotion of aEEG.The main work of this study is to design an automated aEEG signal interpretation algorithm.This algorithm learns models from a large number of data,and then utilize the model to predict the specific label of the coming aEEG samples.This paper first proposed a signal reconstruction method which can convert the aEEG signal to an amplitude frequency contour map,an image reflecting the local amplitudes changes more effectively than the original aEEG signal.In feature extraction,this study extracted four types of features including image features,linear features,histogram features and complexity features.Among them,the LBP operator is introduced into the analysis of aEEG amplitude frequency contour maps to get its image features,which can characterize the local amplitude variation of aEEG signals effectively;The lower border serves the discriminative features in the manual interpretation of aEEG but there is no clear definition about it.This paper not only redefined the low border of aEEG,but also introduced a scoring system to quantify it;In addition,auto-permutation entropy was used to describe nonlinear characteristics and complexity of aEEG signals.Finally,an SVM-based ensemble method called Hybrid-SVM was proposed to implement the automatic classification of aEEG signals.To prove the effectiveness and accuracy of our method,several experiments were conducted on 276 aEEG cases(217 normal cases and 59 abnormal cases).The experimental results show that the image features can effectively characterize the differences between normal aEEG signals and abnormal ones,and improve the performance of different classifiers in varying degrees.Compared with the original support vector machine and other ensemble methods,the overall performance of our proposed method is the best,with recognition accuracy achieving nearly 95.68%.This ensemble method with image features might be helpful for clinical detection of neonatal brain disorders in NICUs through classification of aEEGs.