Recognition Of Uterine Contractions In Electrohysterogram Signal With Convolutional Neural Network

It is contractions that maternal uterine muscle would have a regular electrical activity during the labor.It is the main driving force for delivery and an important feature of maternal labor.The electrohysterogram(EHG)is the result of uterine smooth muscle stimulation and contraction and can reflect the activity of the uterus.EHG and uterine contractions have a good correlation and synchronization,using the EHG to monitor the contraction of pregnant women which will reflect the normal maternity or not.Therefore,EHG has important value in pregnancy and delivery monitoring.A convolutional neuron network(CNN)is a network in deep learning that has a great advantage in processing images.This study uses CNN to identify EHG contractions.In order to obtain the onset of the contraction,the tocodynamometer(TOCO)and annotation contents of the Icelandic 16-electrode electrohysterogram database(EHG database)were used to determine the onset of contractions.The onset of TOCO recordings contraction was obtained from EHG database.It was defined as the point where the amplitude value changed the most,that was the derivative value of the point suddenly increased.The time difference(TD)of the uterine contraction(UC)onset was defined as the difference between the onsets of UC felt by a pregnant woman and determined from the TOCO signal.A total of 295 TDs from 78 recordings was obtained from a total of 33 participants.The results showed that 85.42% of TDs was within [-40,40] s.The overall mean±SD of TD was 3.04±28.02 s,indicating that there was no significant difference between the UC onset determined from TOCO and maternal perception(p>0.05).It was noticed that 61.5% recordings(48 out of 78 recordings)had all positive or negative TD for all the available UCs within a recording.Only seven recordings had equal number of positive and negative TD within a recording.This shows that TD has a strong towards one side within a recording.In addition,In relation to the gestational week,the TD decreased or even became negative with the increase of gestational week.When subjects approached delivery,the intensity of contractions increases significantly and the pregnancy is more likely to perceive contractions.The combination of TOCO records and maternal perception was thus used as a reference to identify the location of contraction.In order to identify the contractions and non-contractions of EHG signals,a convolutional neuron network(CNN)method in deep learning was applied to identify EHG signals.There were the following major steps.1)Filtering EHG,using 0.04-4Hz bandpass filtering to remove irrelevant frequency and DC components;2)Interdicting of the contractions and non-contractions of EHG.Obstetrics and gynecology clinicians marked contractions in the TOCO signal,and then combining the perception of pregnant women to intercept 45-second duration contractions and non-contractions EHG.3)Grouping dataset,contractions and non-contractions picture composed the datasets of a total 5560 pictures.4)Processing datasets,including unifing image size,conversion data format,and zero-mean of datasets.5)Training the Caffenet model with datasets,and adjusting the network parameters of the model to optimal settings.The 10-fold cross model test the average accuracy and loss rate were 95.43% and 10.48%,respectively.In order to optimize the layout of 8 and 4 detection electrodes in collecting EHG signals,the 4×4 electrode distributions were divided into 7 and 13 regions,respectively,corresponding to different physiological locations of the uterus.The EHG signals of the detection electrode positions corresponding to different layouts were assembled into a dataset.The CNN model was trained and a 5-fold cross test was performed.The accuracy,the receiver operating characteristic curve(ROC),the area under the curve(AUC),sensitivity and specificity were used to evaluate the model.Using the scoring method,the top 8 or 4 electrodes of the model with the highest evaluation parameter were added 1 point,cumulative score,and the layout of higher score as the optimal 8-electrode or 4-electrode layout.The results showed that the 4-2-2 structure was the optimal layout in the 8-electrode layout,which 4 electrodes were placed on the uterine fundus,2 electrodes on the uterus,and 2 electrodes on the cervix.In the 4-electrode layout,a 2-2 structure was used,with 2 electrodes placed on the right uterus of pregnant women and 2 electrodes placed on the right cervix was the optimal layout.Therefore,using the CNN model helps to determine the electrode layout of the acquired EHG signals.In order to verify the clinical application of the CNN model,a total of 20 labor EHG records were collected,including EHG signals,TOCO signals,and annotations.According to the TOCO signal and the uterine contractions perceived by pregnant women,384 EHG signals were obtained from contractions and non-contracted segments.Use the previously established CNN model to identify clinical EHG contractions and non-contractions.The accuracy of the uterine contraction recognition was 76.70%,the sensitivity was 82.69%,the specificity was 99.75%,and the AUC value was 0.75.Therefore,the model had a good performance identifing clinical EHG contractions and non-contractions.It also has a good application prospects in the obstetric clinic.