Study On Neonatal Seizure Detection Based On Multi-channel EEG Signals

Seizures represent the most prevalent indication of abnormal brain function in newborns,and their prompt detection holds immense significance for the well-being of infants.In clinical settings,seizures are identified by monitoring electroencephalography(EEG)signals,a process that is both time-consuming and susceptible to errors.The automatic detection of seizures faces challenges such as noise,high dimensionality,and imbalanced data,rendering conventional algorithms inadequate for newborn seizure detection.Moreover,the performance of most algorithms suffers from limited cross-patient detection capability,influenced by gestational age factors.To tackle these challenges,this thesis presents a seizure detection approach that leverages data preprocessing and feature selection techniques.The key contributions of this work are outlined as follows:1.In order to improve the efficacy of seizure detection,this thesis introduces a comprehensive data preprocessing framework known as ARS(Amplitude-RateSampling).ARS combines signal transformation,feature extraction,and data balancing into a unified methodology.It effectively transforms noisy and temporally imbalanced seizure signals into balanced and structured data that can be utilized by machine learning models.Experimental results demonstrate that the application of the ARS framework yields impressive outcomes,with a random forest model achieving an accuracy of 99.3% and a sensitivity of 0.944 on publicly available datasets.2.Within the ARS framework,this thesis integrates the generation method of a EEG(Amplitude Electroencephalography)signals and introduces the c EEG(Changed Electroencephalography)signal transformation method.During this transformation process,the c EEG method preserves the denoising benefits observed in a EEG signals while eliminating unnecessary steps like logarithmic amplitude scaling.As a result,this approach better highlights the waveform characteristics during seizure periods and interictal periods,thereby enhancing the efficacy of feature extraction.Comparative analysis demonstrates that c EEG signals outperform both EEG signals and a EEG signals in the context of seizure detection.3.In the ARS framework,this thesis introduces the Amp-Fre(AmplitudeFrequency)feature extraction method,which is based on the waveform characteristics of c EEG signals.This method directly captures the waveform characteristics of c EEG signals during seizure periods,effectively emphasizing the distinctions between seizure periods and interictal periods.Importantly,this approach is not influenced by the inherent amplitude-frequency characteristics of the signals,resulting in improved crosspatient seizure detection capability.4.Within the ARS framework,this thesis presents the ABSRU(Adaptive Boosting-SMOTE-Random Undersampling)fusion sampling algorithm.This algorithm incorporates the outcomes of Ada Boost(Adaptive Boosting)model training to perform selective fusion sampling on the data,effectively mitigating the problem of data imbalance.Furthermore,adjustable parameters are incorporated into the algorithm to accommodate diverse datasets.Experimental results substantiate that the ABSRU algorithm outperforms its prototype algorithm as well as other commonly employed sampling algorithms in terms of sampling effectiveness.5.To tackle the high-dimensional problem arising from multi-channel signals,this thesis proposes the R-F-EN(Relief-Elastic Net)feature selection algorithm,which combines the Relief algorithm and Elastic Net algorithm.In comparison to other fusion approaches,R-F-EN achieves more comprehensive and reasonable fusion by modifying the cost function.After analyzing the cost function modification process of the R-F-EN algorithm,this thesis replaces the previously stringent numerical constraints with relatively loose feature ranking constraints,resulting in the Rank-RF-EN(Rank-ReliefElastic Net)algorithm.Experimental results on the CHB-MIT dataset demonstrate that the Rank-RF-EN algorithm successfully eliminates 90.2% of redundant features.In conclusion,this thesis presents a novel approach for the automatic detection of neonatal seizures across different patients.The proposed solution involves data transformation,feature extraction,resampling,and feature selection techniques.The algorithm developed for analyzing EEG signals holds potential for applications in other scientific domains that rely on EEG signals for research purposes.Furthermore,the algorithms proposed for data balancing and feature selection can be extended to address similar challenges within the field.Overall,this thesis provides a valuable contribution to the field of neonatal seizure detection and offers opportunities for further advancements in related areas.