Research On The Signal Representation And Classification Of Cardiotocography For Fetal Status Assessment

With the implementation of the "three-child" policy,not only the number of elderly mothers has increased,but also the risk of pregnancy complications for pregnant women has increased.It poses new challenges to the eugenics goal of reducing fetal mortality and birth defect rates.Cardiotocography(CTG),one of the routine fetal heart monitoring methods in obstetrics,is a key technology to ensure the health of the fetus until safe delivery.Due to the complex fetal heart variability,obstetricians have subjective differences in interpreting complex CTG signals,which easily lead to inappropriate clinical decisions.With the emergence of computer-aided systems,this situation has been alleviated,but there are still some outstanding problems,including: inaccurate interpretation,high false positive rate,and poor model interpretation,which greatly limit the development of fetal status assessment systems.In order to solve the above problems,this paper proposed an assessment idea based on multi-representation,multi-dimensional,and multi-model.On the basis of CTG morphological analysis,a closed-loop from automatic multi-representation analysis to fetal status classification was realized,which laid the foundation for building a fetal status assessment system with interpretability and high precision.The main contents of this paper are as follows:(1)Research on CTG signal morphological analysis algorithm.Aiming at the problem of inaccurate baseline estimation of existing morphological analysis algorithms,a novel algorithm based on segmentation network was proposed.To the best of our knowledge,this was the first study to apply a segmentation network model to baseline estimation.It achieved dynamic baseline fitting by combining with the proposed long-and short-term filter algorithm.The experimental results showed that: Compared with other 13 existing baseline estimation algorithms,the proposed algorithm was superior to the existing algorithms in all evaluation indicators.And its synthetic inconsistency index(i.e.,key indicator)was 13.73% smaller than that of the peer-best algorithm.It greatly improved the expert consistency of morphological analysis and laid a foundation for more accurate time-domain morphological features.(2)Research on fetal status assessment model based on feature engineering.On the basis of morphological analysis,the automatic extraction of multi-dimensional representation and their features were realized,including: time-domain representation and morphological features,frequency-domain representation and power features,time-frequency domain representation and energy features.Based on the above features,a variety of machine learning algorithms was introduced to realize the classification of fetal status.Aiming at the problem of poor model interpretability,the model interpretability algorithm was introduced.It revealed the contribution of different features to model decision-making,and provided a basis for quantitative evaluation of representations.The experimental results showed that: In the classification of fetal status,compared with the frequency-domain representation,the time-domain representation and the timefrequency-domain representation showed more significant statistical differences.Under the combination of time-frequency-domain and time-domain representations,the performance of the model was improved and the false positive rate was reduced.And its quality index reached 71.26%,which was 3.14% higher than that of the time-domain representation.(3)Research on fetal status assessment based on deep learning.Aiming at the problems of poor model interpretability and high false positive rate,an ensemble model combining traditional representation and network representation was proposed,which had both the interpretability of feature engineering and the high accuracy of deep learning.In network representation,a recurrent neural network based on temporal multi-scale awareness mechanism was proposed.It enabled the network to dynamically select different time and scales parameters according to the morphological nature of CTG signals.It realized the transparency of the time and scale perception process,and provided a feasible way for the interpretability of network models.The experimental results showed that: The performance of the proposed network was significantly better than other peer networks.Its quality index reached 72.13%,which was 3.83% higher than the peer-best network.The result of ensemble model based on the combined representation was significantly better than that of single representation,and its quality index reached 74.95%,which was 3.73% higher than that of time-domain representation,19.33% higher than that of time-frequency domain representation,and 5.92% higher than that of network representation.