EHG Signal Analysis And Premature Detection Based On Optimal Synthetic Sampling

Premature birth is one of the global public health problems,and even in developed countries,the incidence of preterm birth is still around 7%.Premature birth not only brings serious emotional and economic burden to the family,but also is an important part of social medical payment.Although perinatal and neonatal care has been greatly improved,premature infants still have a high incidence of morbidity and mortality.It is estimated that 80% of neonatal deaths are directly or indirectly related to preterm birth.Therefore,the diagnosis and prevention of preterm birth has been one of the hot topics in clinical medicine.Due to the lack of understanding of the origin of uterine contractions,conventional examination methods cannot provide reliable test results.Uterine contractions are closely related to the electrophysiological activity of muscle cells,which can be recorded by electrodes placed in the abdomen of a pregnant woman to form uterine EHG(Electrohesterography)signals.The application of machine learning method to analyze EHG signal is an important research direction of current methods for the diagnosis of preterm labor.However,due to the limitation of EHG data samples,the uneven distribution of samples among different classes(preterm and non-preterm diseases)and the unclear characteristics,the detection accuracy needs to be further improved.In view of these problems,1)In addition to extracting the characteristic attributes of uterine EHG signals in the conventional time domain and frequency domain space,this project also used the signal recording time to depict the dynamic evolution process of uterine from resting to periodic synchronous contraction during the entire pregnancy for the first time,so as to improve the ability of distinguishing different cases of characteristics.2)Analyze the existing synthesis sampling methods of minority samples,take the distinguishing ability of features among classes as the index,analyze the advantages and disadvantages of different sampling methods,and determine the optimal synthesis algorithm of minority samples.3)The activation function is introduced to describe the influence of the number of samples on the classification of samples,and on this basis,the feature weight is combined to define the feature effective discrimination ability index,which is used to determine the optimal synthetic sampling algorithm parameters.The above feature extraction and sample synthesis sampling methods were applied to the open source EHG signal database TPEHG to build data samples,train classifiers and conduct crossvalidation.The experimental results show that,compared with the conventional uniform 1:1 synthesis of a small number of classes,the prediction accuracy of various classifiers under the condition of not introducing the dynamic attribute characteristics has been significantly improved under the condition of the imbalance of the optimal sample,among which the comprehensive accuracy of SVC classifier is 83.33% and AUC is more than 0.9.Combined with dynamic force attributes,classification accuracy is further improved.In order to promote the application of the research results,a platform for emg EHG signal analysis and premature delivery diagnosis based on optimal synthetic sampling was established.The detection results are given by using the existing EHG data and the classifier after the training of its features to apply to the newly uploaded EHG data.A user feedback mechanism is introduced to take the detected samples as training data to further optimize the classifier.With the increase of user data,the system provides reliable detection results for premature delivery.