A systematic review of automated pre-processing, feature extraction and classification of cardiotocography

Database

Article search is done systematically using three databases such as (1) IEEE Xplore digital library, (2) Science Direct and (3) Web of Science (WOS). Article searches are based on the facilities provided by the three databases both in simple and complex searches by searching for various journals and conference articles in various fields of science such as engineering, computer science and medical. Therefore, the literature in the medical sciences and engineering is included in the search scope to provide broader understanding and views for researchers related to a variety of science majors.

Article selection procedure

The selection procedure of the article in this study was based on an intensive search for the relevant publications in this field of study using the following steps: (1) Scan titles and abstracts that are in accordance with the topic and exclude inappropriate and not related papers and to find any duplication of papers. (2) The next step is to read the full papers thoroughly.

Article search

The articles search was conducted using three databases namely IEEE, Science Direct (SD) and Web of Science (WoS) through the database search on March 17, 2018. In all three databases, searches were performed using queries (Cardiotocogram OR Cardiotocograph OR Cardiotocography) with the following keywords (“fetal heart rate” OR baseline OR “baseline variability” OR acceleration OR deceleration OR “uterine contractions”). Searches are added using this keyword (“FHR-monitoring” OR “feature extraction” OR classification OR diagnosis) to limit the search space to cardiotocography cases related to engineering and computer science. Search queries are shown in Fig. 1.

Article searches are added with advanced search options for each database used to select conference and journals articles, and excluded articles from each chapter of the book or other documents. Articles in journals and conference articles are chosen for consideration because articles in journals and conferences are most likely involved in the latest research and relevant to this study. The presentation of rule information used in conducting search requests is shown in Table 1.

Criteria for selected articles

Figure 1 shows the criteria for the selected article and all articles that meet the criteria are selected. Mapping of research space on feature extraction and cardiotocography classification becomes a general taxonomy and a rough overview of the four categories defined as initial targets. The division of this category is based on an intensive pre-survey of the selected literature resources. Any article that did not match the specified criteria as depicted in Fig. 1 was not selected. Criteria that are not included are as follows: (1) Papers not written in English. (2) Focus on fetal heart rate and uterine contraction extraction and classification but with manual analysis features.

The process of collecting articles

In managing the list of articles, the author uses the Excel application which is composed of the three databases used. The selected article is read in full by the authors and given several highlights and important comments, and the taxonomy that has been set at the beginning is refined in the classification of the articles in the listing. Each article is Summarize, tabulate, and all-important findings are elaborated in detail. The word and excel application are used to store related information such as lists of surveyed articles, source indexes, summary tables and descriptions, objectives, review sources, audiences, data sets used, criteria of evaluation, techniques of validation, etc. These data are presented in the supplementary materials as a comprehensive indication of the obtained results.

Article statistics information and search results article

The taxonomy used to review these research articles that focus on automatic feature extraction and cardiotocography classification is as depicted in Fig. 2. This taxonomy stated the entire development of various researches and modern applications. The classification stated four main stages. The method proposed is discussed in the first stage (stage 1), this includes all proposed method that presents steps and algorithms in the preprocessing stage, feature extraction and classification and their use in cardiotocography (20/50 articles). Development of system for automatic feature extraction along with cardiotocography classification is conducted in second stage (stage 2) (7/50 articles). Next, the third stage (stage 3) comprised of reviews and survey articles related to automatic feature extraction and cardiotocography classification (3/50 articles). Finally, evaluation as well as comparative studies in determining the most optimum technique related to extracting and classifying features and its comparisons is discussed in the last stage (stage 4) (20/50 articles).

Techniques and algorithms

This section discusses the methods and algorithms used in feature extraction and cardiotocography classification of all articles that have been included in this study. The general context of articles in this taxonomy is about achieving high performance or increasing the accuracy of cardiotocography. This first stage has 20 articles out of 50 and it is divided as follows:

• The first subclass {7/20 articles}: focuses on the classification using the new proposed method or modify conventional methods.

• The second subclass {4/20 articles}: focuses on preprocessing and classification methods.

• The third subclass {6/20 articles}: discusses feature extraction (features selection) and classification methods.

• The fourth subclass {3/20 articles}: discusses preprocessing, feature extraction and classification methods.

As mentioned earlier, the first segmented group of articles discussed classification. Firstly, Volterra based Neural Networks (VNN) using the Volterra series expansion are proposed in Haweel & Bangash (2013). Secondly, the scattering transform is proposed in Chudáček et al. (2014) as a new method for analyzing the variability of FHR in the intrapartum period. Thirdly, The CTG trail using a 3-level classification presented in Das, Roy & Saha (2015a) to classify CTG as Normal (N), Suspicious (S), and Pathological (P). Further, an ordinal classification approach as used in Georgoulas et al. (2017) based on Binary Decomposition for prediction, that used C.45 algorithm split the data and select a test that provided the best information gain, and used Synthetic Minority Oversampling Technique (SMOTE). Next, Permanasari & Nurlayli (2017) proposed a method that implemented in the decision tree (DT) to analyze the CTG data to determine Fetal Distress. Moreover, in clinical practice, the qualitative assessment of FHR records used fuzzy classification with the Weighted Fuzzy Scoring System (WFSS) technique as proposed in Czabański et al. (2013). Lastly, proposed a scheme that is based on adaptive neuro-fuzzy inference systems (ANFIS) was trained to predict the normal and the pathological state (Ocak & Ertunc, 2013).

Next, the second segmented group discussed preprocessing and classification. Firstly, in Georgoulas et al. (2014) Interpolate using Matlab implementation of Hermite Spline Interpolation proposed a classification approach used for missing data, the technique is based on a latent class analysis method that seeks to produce more objective labeling of training cases, and this is an important step in any classification problem. Secondly, as proposed in Gavrilis, Nikolakopoulos & Georgoulas (2015) Principal Component Analysis (PCA) used both as part of the one class classification process as well as for visualization, while Isomap, a nonlinear dimensionality method, is used to confirm the visualization results of the PCA and in this study, two simple one-class classifiers are tested: the nearest neighbor (NN) and the Parzen window classifier. Thirdly, preprocessing proposed in Jyothi, Hiwale & Bhat (2016) used a moving average filter to compensate for the noise and for contraction classification using K-Nearest Neighbor (KNN). Lastly, in preprocessing using Principal Component Analysis (PCA), the data utilized for training and testing using the Adaptive Boosting algorithm (AdaBoost) to obtain strong groupings for classifying unknown data and predict the state of the fetus proposed in Zhang & Zhao (2017).

Further, the third group discussed feature extraction and classification. Firstly, in Shah et al. (2015), Shah et al. proposed the use of correlation feature selection subset evaluation (CFS) with CFS calculated the correlation between features and classes for classification, Random Forest applied to classify data samples into classes and REPTree applied to produce decision trees, by calculating information acquisition using entropy and also using J48. Secondly, the K-means algorithm utilized to obtain the important features and for classification Support Vector Machine (SVM) was used as reported in Chamidah & Wasito (2015). Thirdly, as proposed in Ocak (2013), the Genetic Algorithm (GA) implemented to minimize the number of features and determine the ideal subset of features and further classify fetuses using Support Vector Machine (SVM). Fourthly, feature extraction based on Genetic Algorithm (GA) used as proposed in Xu et al. (2014) and as for classification linear regression, linear Support Vector Machine and kernel radial basis function (RBF) SVM were utilized. Moreover, as presented in Inbarani, Banu & Azar (2014), in order to identify the most important features, unsupervised swarm-based reduction techniques, hybrids of swarm intelligence and rough sets were employed and to check the level of errors generated from the reduce set by applying K-means clustering, and for classification, this study used single decision tree (DT), multilayer perceptron (MLP) neural network, probabilistic neural network (PNN) and random forest. Finally, as proposed in Kim, Yang & Lee (2017), it used fitMine as a new nonlinear dynamic model that reflected the relationship between signals of Fetal Heart Rate and Uterine Contraction by combining chaotic population model and unscented Kalman filter algorithm. In the last segmented group of articles, it discussed three components namely the preprocessing method, feature extraction, and classification. Firstly, as reported in Tomas et al. (2013) were used the principal component analysis (PCA) used as feature reduction, the correlation feature selection subset evaluation (CFS), CFS calculated the correlation between features and classes and three algorithms were used in this work: Random Forest, REPTree and the Linear Discriminant Analysis. Secondly, for preprocessing was used R-R intervals , while sequential feature selection algorithm was used to determine the informative subset of HRV features and Support Vector Machine was used to classify healthy fetuses and fetuses with adverse outcome presented in Warmerdam et al. (2016). Lastly, (Spilka et al., 2017) proposed cubic spline interpolation for preprocessing, as for feature extraction this study used FIGO enhanced and automated features (FI), Spectral Energy (SC) and Multiscale multifractal analysis (SI) and for classification data Sparse Support Vector Machine (Sparse-SVM) was utilized.

System development

This section discusses either a full-automatic or semi-automatic system that is used for feature extraction and CTG classification for system development. This second stage has 7 articles out of 50 and it is divided as follows:

• The first subclass {5/7 articles}: based on supervised classification engine.

• The second subclass {2/7 articles}: based on unsupervised classification engine.

The first subclass discusses the phase of the system. The following method is carried out in this literature: Preprocessing, Segmentation and Identification in Pasarica et al. (2017), Preprocessing, identification of parameter, clinical validation, performances indexes estimation and statistical parameters in Romano et al. (2016), preprocessing, Identifying and classify in Rotariu et al. (2014), Preprocessing and classification in Georgieva et al. (2013), Feature extraction and classification in Czabanski et al. (2015). In the second subclass, the related work focused on the proposed system in Wróbel et al. (2013) that used the weighted myriad filtering with identification and selection, and in Das, Roy & Saha (2015b) agreement between the proposed technique and visual interpretation by obstetricians was assessed using Bland-Altman analysis.

Review/survey

The third Stage includes reviews and surveys aimed at describing cardiotocography, methods and techniques used in feature extraction and cardiotocography classification. This third stage has 3 articles out of 50 and were found in Warmerdam et al. (2018); Fergus, Selvaraj & Chalmers (2018); Bhatia et al. (2017). The first in Warmerdam et al. (2018) presented Fetal HRV analysis that provided distress information on fetal, combining HRV features calculated over the entire fetal heart rate with contraction-dependent HRV features that improved the classification performance during the second stage of labor. The second in Fergus, Selvaraj & Chalmers (2018), showed a review of human cardiotocography interpretations and giving an opinion that machine learning can be used as a decision support system by obstetricians and midwives and also can provide objective results in addition to using normal practice. Finally as reported in Bhatia et al. (2017), this study presented a comparison of the cardiotocography classification system as outlined by the International Federation of Gynecology and Obstetrics (FIGO) in 2015 and the UK National Institute for Health and Care Excellence (NICE) in 2007 and 2014.

Evaluation and comparative study

This section, the fourth stage, discusses the efforts to determine the best method for feature extraction and cardiotocography classification by comparing methods based on several criteria. This fourth stage has 20 articles out of 50.

The two classification schemes were compared that depended on one of the two subclasses:

• The error rate (stage 4/ subclass 1) shown in the articles (Frigo & Giorgi, 2017; Romano et al., 2018).

• The reliability criteria group (stage 4/ subclass 2) as reported in the articles (Cömert, Kocamaz & Güngör, 2016; Magenes et al., 2016; Nagendra et al., 2017; Sahin & Subasi, 2015; Rei et al., 2015; Di Tommaso et al., 2013; Cömert & Kocamaz, 2017a; Gamboa et al., 2017; Chen et al., 2014; Arif, 2015; Garabedian et al., 2017; Pinas & Chandraharan, 2016; Ghi et al., 2016; Cömert & Kocamaz, 2017b; Yılmaz, 2016; Chinnasamy, Muthusamy & Gopal, 2013; Sundar, Chitradevi & Geetharamani, 2014; Nunes & Ayres-de Campos, 2016).

One of the contributions to this work is to prioritize the latest articles in the research literature using signal processing through content analysis from several major journals in the field of study. The number of articles obtained from the literature is shown in Fig. 3. These articles are (50) articles selected based on predetermined criteria. As depicted in Fig. 3, interest in the development of methods and systems for feature extraction and cardiotocography classification has increased accordingly over the rest.

Figure 4 shows the number of articles based on each category and publication year. A total of 20 articles explained the proposed method in the study, proposals for system development totaling of 7 articles, reviews and surveys totaling of 3 articles and evaluations and comparisons of two or more methods amounted to 20 articles.

Figure 5 shows the author’s affiliation country (the main author’s country is considered in a case study with several co-authors). Majority of the studies focused on feature extraction and cardiotocography classification are in Italy, Turkey and India. Other countries include England, the Czech Republic, Poland, Indonesia, Sweden, the Netherlands, Saudi Arabia, Portugal and Romania. The number of co-authors was 145 out of 50 articles reviewed. Co-authors come from various universities in the world with the majority coming from the fields of biomedical engineering, bioinformatics, signal processing, computer science, medicine, and pharmacy.

Datasets

In this survey, there are 3 categories of datasets discussed: publicly available, private dataset and commercial dataset. In publicly dataset, researcher can download the datasets freely, in private dataset researcher need to request the private from infirmaries, university or research Centre and for commercial dataset it can be obtained from Clinical Environment. In our survey, we got 13 sources from UCI Machine Repository, CTU-UHB Database (University Hospital), Hospital Femme Mere Enfant (HFME), CTU-UHB (Czech Technical University), Azienda Ospedaliera Universitaria Federico II, Registration for CTG is recorded at a Neoventa STAN Molndal Sweden, University Hospital of Bologna, Tertiary care University Hospital, Clinical Environment (Romano et al., 2018), Clinical Environment (Romano et al., 2016), Careggi University Hospital of Florence, Mackay Memorial Hospital, a tertiary referral center, and Maternity and Gynecological Clinic (University Hospital of Porto in Portugal).

From our survey, UCI Machine Repository is the most widely used, secondly CTU-UHB database from university hospital in Brno Czech Republic, thirdly CTU-UHB database from Czech Technical University, and lastly Hospital Femme Mere Enfant (HFME). Four of these datasets are publicly available in their website, so researchers can download the database freely. Table 2 summarizes the basic details of the cardiotocography dataset in the survey that we have conducted.

Techniques of validation

In analyzing general results, researchers evaluated independent datasets using techniques. In addition, to evaluate the classification model proposed using validation techniques (Chinnasamy, Muthusamy & Gopal, 2013). To analyses the results of generalization, researchers evaluated an independent dataset by using a technique. Basically, validation techniques are used when the aim of the study is to predict and estimate the precision of a predictive model in practice (Chinnasamy, Muthusamy & Gopal, 2013). Table 3 illustrates these validation techniques used in the included articles of our survey.

In the Table 3, there are 6 validation techniques used in this review. Firstly, Cross validation was divided into 3 techniques, there are 10-fold cross validation, k-fold cross validation and 2-fold cross validation, and note that cross validation are the most widely used in this review paper. Other validation techniques are Kappa Statistics, Paired sample t-test, confidence interval, Bland-Altman approach and Mann–Whitney test.

Performance measure

The most common way to evaluate the used of various classifiers in our survey is the Evaluation using criteria. This method can assess the effectiveness of the classification methods. Classifier performance evaluation using various measurements such as confusion matrix, RMSE, selectivity, specificity, precision, accuracy, sensitivity, f-measure, quality index, gmean, MAE, AMAE, chi-square, spearman correlation, AUC, training time, testing time, misclassification error and confidence interval. Hence in this section all the above parameters are discussed. Table 4 shows the measurement criteria implemented in the reviewed articles.

Various studies on feature extraction and cardiotocography classification are as presented in Table 4, as well as a full survey of different criteria and sub-criteria for the evaluation and benchmarking. Moreover, to evaluate the cardiotocography classification method, the criteria used are confusion matrix, RMSE, selectivity, specificity, precision, accuracy, sensitivity, f-measure, quality index, gmean, MAE, AMAE, chi-square, spearman correlation, AUC, training time, testing time, misclassification error and confidence interval. As clearly shown in Table 4, not all articles in this scope of review utilized all the above-mentioned criteria.

Motivation

From the articles that we reviewed, we found that researchers were motivated to develop and use automated methods to diagnose, feature extraction and classification of cardiotocography focused on improving early detection and rapid diagnosis, accuracy, guidelines, data sets, ONG experts, methods and techniques. This section describes the benefits obtained from the literature, mapped as benefits from the same group including quotations for each benefit for more discussion. A brief summary of the motivation for the feature extraction and cardiotocography classification is shown in Fig. 6.

Challenges

Over the past few years, there has been an increase in the field of feature extraction and cardiotocography classification, although there are still encounter difficulties and challenges in numerous important respects. Various examples of these challenges are diagnosis, parameters, guidelines, and many more. In Table 7 presents details about reference challenges.

Recommendation

This section offers several important recommendations for literature specifically recommended categories of cardiotocography classification and feature extraction as described in Fig. 7.