Research On Early Diagnosis Model Of Intrahepatic Cholestasis Of Pregnancy

Intrahepatic cholestasis of pregnancy(ICP)is an idiopathic disease that seriously endangers maternal and child health during pregnancy and its incidence is as high as 12%,which can cause a series of adverse pregnancy outcomes.As the etiology of ICP is not clear,early diagnosis,timely intervention and reasonable treatment are still the most effective measures for ICP.However,the current clinical diagnosis and treatment of ICP are mainly based on the screening of biochemical indicators such as bile acid.Its sensitivity and specificity are low,and ICP cannot be detected in time.Therefore,it is of great significance to promote maternal and child health if ICP can be detected in the early pregnancy and reasonable interventions can be made.Clinical trials have confirmed that the three major biomarkers,GCA,ACOX1 and palmitoyl carnitine,are closely related to the occurrence of ICP and can be used for the early diagnosis of ICP.Based on this finding,this paper divides the early diagnosis of ICP into two steps: biomarkers prediction and ICP diagnosis,then biomarker prediction models and ICP diagnosis models are constructed by machine learning methods.Firstly,a multivariate time series similarity measurement method is proposed for screening ICP sample collections with similar biomarker trends.Then,an ICP biomarker prediction model is proposed to achieve accurate prediction of ICP biomarkers in mid and late pregnancy.Finally,an ICP diagnostic model based on predictive data is constructed.The specific work is as follows:(1)The content of ICP biomarkers in different gestational weeks are in the form of multiple time series.In order to improve the accuracy of the prediction model,it is necessary to screen out a collection of samples with similar patterns of marker changes.Aiming at the problem that the existing similarity measurement methods cannot effectively extract the feature patterns of multivariate time series and perform similarity measurement,a multi-dimensional time series similarity measurement method,MSDWDTW,based on multi-dimensional segmentation and dynamic weighted dynamic time warping distance is proposed.First,multi-dimensional segmentation is performed on the multivariate time series,and the slope,mean value and time span of the segments are extracted as the segmentation feature representation.The correlation between variables and the shape and value characteristics of the sequence are retained while reducing the dimension.Aiming at the problem of malformed matching caused by dynamic time warping distance,a dynamic weighted dynamic time warping distance measurement method is proposed.This method assigns weight to each sequence point and reduces its weight adaptively according to the number of sequence point matching during the solution process.The experiments show that the KNN algorithm based on the MS-DWDTW method can achieve high search accuracy on data sets of different sizes.The clustering results of the K-means method based on MS-DWDTW also show that it can also get better measurement results on the ICP data set.(2)An ICP biomarker prediction model based on dual-LSTM and ARIMA is proposed.The prediction accuracy of the LSTM network model constructed directly on the original sequence is not high.Therefore,based on the idea of multi-dimensional segmentation,this paper extracts the trend feature of the sequence to obtain the trend feature sequence,and constructs the LSTM network prediction model of the sequence trend feature.According to the prediction error,the CRITIC weight assignment method is used to combine the sequence trend prediction results with the sequence value prediction results to obtain the preliminary prediction value based on the LSTM network.The ARIMA model is used to fit the prediction residual of the LSTM network,and then the model is used to predict the prediction error of the LSTM network,which is used to modify the prediction value of the LSTM.The experimental results show that compared with the single LSTM network,the trend sequence prediction network can reduce the overall prediction error to a certain extent,and ARIMA model can well modify the prediction results of the dual-LSTM network.Compared with the existing prediction methods,the proposed model can also obtain the smallest prediction error,which has great advantages in the prediction of ICP markers.(3)An improved cost-sensitive AdaBoost integration method is proposed.The traditional AdaBoost algorithm does not consider that the cost of diagnosing a patient as a healthy person is greater than the cost of diagnosing a healthy person as a patient,that is cost insensitive.The AdaCost algorithm,which is cost sensitive,needs to provide an uncertain cost factor of misclassification.In this paper,AdaBoost integration algorithm is improved,the recall rate of the base classifier to the patient is introduced into the objective function of the classifier integration,and the integration weight is determined based on the weighted error rate and the weighted recall rate of the base classifier,so that the base classifier with low classification error rate and high recall rate has higher weight coefficient,which improves the recognition ability of the diagnosis model to ICP patients.Experiments show that compared with other diagnosis methods,the method proposed in this paper does not improve the accuracy of diagnosis,but the diagnostic error rate of ICP patients is significantly reduced,which can meet the needs of ICP diagnosis.