Uterine electromyogram

The uterine electromyogram, recorded from the abdominal wall of a pregnant woman, also known as the electrohysterogram (EHG), has shown great potential as a cost-effective and non-invasive diagnostic tool for assessing the danger of premature birth [5, 7, 45–47]. The uterus is a highly complex non-linear dynamic system comprising billions of interconnected smooth myocyte cells which produce measurable changes in electrical potentials, reflecting the electrical activity of the uterus [48, 49]. Traditionally, the EHG was employed for non-invasive quantitative assessment of uterine contractions [45, 46, 48–50]. However, despite decades of research, the underlying physiological processes that lead to premature birth and the precise electrophysiological mechanisms driving EHG generation remain poorly understood. Nevertheless, studies have shown that the EHG appears to contain sufficient information to enable efficient non-invasive prediction of premature birth several weeks, or even months, before delivery [5, 51].

To predict premature birth using the EHG, researchers focus on two approaches. The first approach tries to differentiate between the non-labor and labor phases (which usually translates into recognition of contractions related to imminent delivery or prediction of delivery time) [5, 7, 47, 52–56]. The second, increasingly more common approach attempts to differentiate between premature EHG records obtained from pregnancies that eventually end in premature delivery (before the 37th week of gestation) and term EHG records obtained from pregnancies that eventually reach term delivery (during the 37th week of gestation or later) [51, 57, 58]. In both cases, under the traditional feature engineering approach, the features (parameters) to assess the relevant underlying physiological mechanisms are extracted from the EHG to identify the approaching labor phase, or to distinguish between premature and term records. The features may be extracted from segments of the EHG containing individual contraction burst intervals (contraction intervals), or from whole signals of individual EHG records without the need to first detect contractions and isolate, or annotate, individual contraction intervals. Moreover, there is strong evidence to suggest that segments between contraction intervals, i.e. non-contraction intervals, also known as dummy intervals, contain sufficient information to effectively separate between premature and term EHG records [57].

Various linear and non-linear signal processing techniques have been used to extract the features from the EHG, however, it is well established that the latter are better suited for assessing the underlying non-linear physiological processes [51]. The spectral and non-linear features extracted from the EHG commonly include the Peak Frequency (PF) and Median Frequency (MF) of the power spectrum [59], Peak Amplitude of the normalized power spectrum (PA) [57], and entropy measures like the widely used Sample Entropy (SE) [57, 59, 60] and approximate entropy [60–62]. Most notably, features PF, MF, and PA are easily understandable and suitable for directly and transparently assessing the shifts and intensity of the frequency content of the EHG in the frequency bands corresponding to background physiological mechanisms. On the other hand, SE is used to estimate the level of regularity, or predictability, of the signal in the selected frequency bands, thus assessing the presence or absence of underlying periodic physiological mechanisms. Moreover, a multitude of more complex and intricate entropy measures such as dispersion entropy [63, 64], bubble entropy [63, 65], fuzzy entropy [66, 67], permutation and phase entropy [67], and multivariate multiscale fuzzy entropy [68] have shown promise, yet their increased complexity may diminish their applicability and interpretability.

This proliferation of features seems to have become pronounced with the introduction of the Term-Preterm EHG Database (TPEHG DB) [59, 69] which spurred a rapid progression in the development of methods for the prediction of premature birth using various signal processing techniques and classification approaches with entire EHG records and non-linear features. The TPEHG DB includes 262 term and 38 premature EHG records, recorded early (around the 23rd week of pregnancy) or later (around the 31st week of pregnancy). However, due to the imbalance between premature and term records in this database, many researchers resorted to using over-sampling to synthesize new samples and balance the classes prior to partitioning the data into the training and test sets (Synthesis-Partition evaluation technique). For many of these methods, researchers reported spectacular classification performance results, however, a subsequent study [70] revealed that applying over-sampling before partitioning constituted a significant methodological flaw, resulting in data leakage and excessively optimistic classification performance results. Moreover, the authors re-evaluated a variety of related methods, by correctly applying over-sampling only after partitioning the data (Partition-Synthesis evaluation technique), and showed that the realistic AUC scores of those methods were closer to 0.65 [70].

The introduction of Term-Preterm EHG DataSet with Tocogram (TPEHGT DS) [57, 69] has obviated the need for over-sampling when evaluating the classification performance of methods for the prediction of premature birth, as it includes a balanced set of 13 premature and 13 term EHG records. Furthermore, premature records contain 47+47 annotated contraction and dummy intervals, while term records contain 53+53 annotated contraction and dummy intervals. Therefore, even in the case of classifying between premature and term intervals (instead of using whole signals), balancing may not be required since any balancing will have very little effect on the evaluation since the number of premature and term intervals is almost equal. However, because individual annotated intervals do not represent independent samples, and due to the small number of records in the TPEHGT DS, evaluation of methods using only this dataset may not be reliable. Additionally, it is important to note that methods utilizing annotated intervals from the TPEHGT DS may not be considered fully automated. These intervals were manually annotated by experts, first by examining the tocogram signal (which measures the mechanical uterine activity using an external tocodynamometer), and then looking for the accompanying characteristics of the EHG signals [57].

Due to the compatibility between the records of the TPEHG DB and TPEHT DS (in terms of recording protocol and recording device), several researchers have also combined the EHG records from both datasets. This was usually done to increase the number of available premature records, yet this once again introduced the challenge of dealing with an imbalanced dataset. To assess the classification performance using the records of the combined dataset, early studies [63, 71] initially employed the Synthesis-Partition evaluation technique, while recent studies [67, 72] utilizing the combined dataset have shifted towards using the recommended Partition-Synthesis evaluation technique [70].

Single-feature EHG methods.

Table 2 shows an overview of some of the most recent methods for the prediction of premature birth, or classification between premature and term records, based on a single feature extracted from the EHG. The studies in Table 2 are grouped according to the used dataset(s). The first group includes studies [57, 73] that used annotated dummy intervals of the TPEHGT DS, the second includes studies [58, 74] that used closed datasets, while only one single-feature study [75] used records of the TPEHG DB and the Partition-Synthesis evaluation technique (due to the class imbalance). Notably, none of the single-feature studies in Table 2 used a combined dataset with premature and term records of the TPEHG DB and TPEHGT DS. The most promising single-feature method [73] yielded the highest AUC of 0.85, using SE extracted from signal S2 of annotated dummy intervals of the TPEHGT DS in the frequency band B2 (2.2-3.5 Hz) [57]. Next, the method reported in [57] achieved the AUC of 0.75 using PA of signal S2 extracted from annotated dummy intervals of the TPEHGT DS in the frequency band B1 (1.0-2.2 Hz) [57]. However, these results may be less realistic since both studies [57, 73] classified between annotated intervals from only 13 premature and 13 term records of the TPEHGT DS, and neither method is fully automated because they require prior annotation of non-contraction (dummy) intervals.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. An overview of some of the most recent methods for prediction of premature birth based on the EHG based a single feature (biomarker).

The studies are grouped according to the used dataset.

https://doi.org/10.1371/journal.pone.0308797.t002

Multiple-feature EHG methods.

Table 3 shows an overview of some of the most recent methods based on multiple features extracted from the EHG. The studies are grouped according to the used datasets: studies [57, 73, 76, 77] that used only annotated intervals of the TPEHGT DS, studies [75, 78, 79] that used the records of the TPEHG DB, and studies [67, 72, 80] that used the combined dataset comprising records of both the TPEHG DB and TPEHGT DS. Note that methods which were evaluated using the Synthesis-Partition evaluation technique were excluded due to overly optimistic results according to [70].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. An overview of some of the most recent methods for prediction of premature birth based on the EHG based on multiple features.

The studies are grouped according to the used signal scope (intervals or whole signals) and dataset.

https://doi.org/10.1371/journal.pone.0308797.t003

In Table 3, the method yielding the highest AUC of 0.97 [78] used over 200 features, extracted from the Horizontal Visibility Graph (HVG), which was constructed from whole EHG signals of the TPEHG DB in the time-frequency domain. However, this transformation of the input signals into the HVG-based latent space may completely obfuscate the feature meaning, in the physiological sense, and severely diminish the much-needed insight into the underlying physiological mechanisms. Besides, their results may be overly optimistic since feature selection was performed using the whole dataset (i.e. no validation set was used and no unseen data remained for testing), which may have led to overfitting of the final feature selection to the test sets. This was also done in [76], which used annotated contraction and dummy intervals of the TPEHGT DS and yielded the same AUC of 0.97. The next best performing method [72] yielded an AUC of 0.95, using the genetic algorithm to obtain a complex feature set with 58 (out of 227) temporal, spectral, and non-linear features as well as obstetric parameters, but the authors did not report which features ended up in the optimal subset.

Notably, very few of the multi-feature methods in Table 3 used less than 10 features. In [73], the authors also evaluated the classification performance using only two meaningful features, SE of signals S2 and S3 in the band B2 (2.2-3.5 Hz), yielding an AUC of 0.89 when using annotated dummy intervals of the TPEHGT DS. Again, it is important to note that the TPEHGT DS contains only 13 premature and 13 term records, and the reported results may be less reliable, since the intervals used for training and testing may come from the same EHG records. Besides, this method is not fully automated, as it depends on using manually annotated intervals. Moreover, annotating contraction intervals or dummy (non-contraction) intervals would be extremely time-consuming and subjective in clinical practice, and therefore less reliable. Furthermore, for the method reported in [67], the best AUC of 0.89 was achieved using seven features, however, their results may be overly optimistic due to having used a single under-sampling of the term class with no repetitions.

Two of the most recent methods [79, 80] utilized complex deep learning models. However, because little to no effort was put into their explainability, these models represent complex “black-box” prediction machines which provide little to no insight into the underlying physiological mechanisms. In particular, in [80] the authors used more than 600 features, extracted from whole later EHG records of the TPEHG DB and TPEHGT DS, with the Bidirectional Long-Short Term Memory (BiLSTM) and logistic regression, yielding the AUC of 0.78. In the second study [79], the authors used 12 time-series of feature values (SE, MF, and PF), extracted from sequential overlapping segments of signals S1, S2, and S3 of EHG records of the TPEHG DB, with the Long-Short Term Memory (LSTM) or Convolutional Neural Network (CNN), yielding the AUC of only 0.58. Notably, neither of these two “black-box” approaches utilizing complex deep-learning based methods yielded higher AUC scores than those yielded by more traditional methods based on feature engineering.

The aims of this study

There is an increasing tendency towards the use overly complex methods and less understandable features which lack explainability and hardly provide any new insights into the underlying physiological mechanisms, even though the causes of premature birth are still poorly understood. Furthermore, to our knowledge, no prior studies assessed whether non-spontaneous delivery modes impact the classification performance of the methods for the prediction of premature birth based on the EHG. However, a robust approach to accurate prediction of premature birth should take into account that pregnancies may also end in other common term delivery modes like induced delivery and cesarean section.

The aims:

1. 1. To identify the most promising features, derived from the EHG signals and their power spectra in the most widely used and recently investigated frequency bands below and above 1.0 Hz, as candidate biomarkers for assessing the danger of premature birth using the sequential forward selection of features method, stabilized with the frequency-based aggregation of the selected features technique.
2. 2. To assess the classification power of the identified most promising features, using later recorded premature and term EHG records from the publicly available term-preterm EHG datasets (TPEHG DB, TPEHGT DS, and ICEHG DS), and a variety of traditional classifiers. Due to the imbalance between the number of available premature and term records, the classification performance was evaluated using the Partition-Synthesis evaluation technique.
3. 3. To develop a simple and explainable method for efficient, non-invasive, and fully automated prediction of premature birth based on the EHG, using one or at most two of the identified most promising features (biomarkers).
4. 4. To quantitatively assess the impact of common term non-spontaneous delivery modes (induced delivery, cesarean section) on the predictability of premature birth when using the developed method and the enriched representation of the term class with records from the recently introduced and publicly available ICEHG DS.

In this study, Peak Amplitude of the normalized power spectrum (PA), extracted from EHG signals in the low frequency band B0L’ (0.125-0.575 Hz), related to the known FWL frequency band, and Median Frequency (MF) of the power spectrum in the band B3 (3.5-5.0 Hz), related to the influence of the maternal heart on the uterus, were identified as the most promising features to classify between later premature and term records of the TPEHG DB, TPEHGT DS, and ICEHG DS. Furthermore, the addition of records of non-spontaneous delivery modes from the ICEHG DB to the term class did not reduce the classification performance, suggesting that these potential delivery outcomes do not significantly impact the power of the proposed method to assess the danger of premature birth. The proposed method for the prediction of premature birth is effective, simple, explainable, robust, fully automated, and appears to outperform existing single-feature and many multi-feature methods based on the EHG, as well as existing non-invasive chemical and molecular biomarkers.

EHG datasets and groups of records

For feature selection and classification, we used EHG records from the Term-Preterm EHG Database (TPEHG DB) [59, 69], Term-Preterm EHG DataSet with Tocogram (TPEHGT DS) [57, 69], and Induced-Cesarean EHG DataSet (ICEHG DS) [69, 82, 83]. The records of all three datasets are fully compatible in terms of electrode placement (E1, E2, E3, E4), signal definitions (S1, S2, and S3), sampling frequency (20 Hz), recording duration (around 30 minutes) and recording apparatus. The tocogram signal is present only in records of the TPEHGT DS, hence, it was not used in our study to ensure the consistency of the features extracted from signals available in all three datasets used in this study. Likewise, we did not use any additional clinical parameters included in the EHG records. Common clinical parameters included in the EHG records from these three datasets include maternal age, parity, number of abortions, maternal weight, placental position, and smoker status. Unfortunately, some of these were missing in several EHG records used in this study, and thus were not considered as potential features in our analysis. Besides, the results of two recent studies [79, 80] suggest that this clinical data only has a modest impact on the predictability of premature birth (in both studies, the addition of a multitude of clinical parameters improved the AUC score only by 0.04).

The TPEHG DB contains 300 EHG records, of those 19 premature and 143 term records were collected early in the pregnancy, around the 23rd week of gestation (early records), and 19 premature and 119 term records were collected later in the pregnancy, around the 31st week of gestation (later records). The TPEHGT DS contains 13 premature and 13 term records, all collected later in the pregnancy (again, around the 31st week of gestation). As such, similarly to recent related studies [67, 80], we used only later records of both dataset for consistency in terms of the time when the records were collected. Besides, the results of our previous study [81] show that features exhibit significantly different powers to separate between premature and term records when recorded early or later in the pregnancy.

The later premature records of the TPEHG DB and the TPEHGT DS were merged into group PL (19+13 = 32 Premature Later records), and the later term records of these two datasets into group TL (119+13 = 132 Term Later records). The records from groups PL and TL were initially used to identify the most promising features to distinguish between premature and term pregnancies, and to evaluate the performance of the classifiers. However, these records were collected from pregnancies with vaginal deliveries that started spontaneously. The assumption of spontaneous vaginal delivery mode at the time of collecting the EHG record for the purposes of predicting premature birth is unrealistic, since the pregnancy may end in other common delivery modes like term induced delivery or cesarean section. Therefore, for a more realistic assessment of the performance of the proposed method for automated premature birth prediction, we also utilized the recently introduced ICEHG DS. The ICEHG DS contains 126 EHG records collected from pregnancies that ended in term vaginal delivery that did not start spontaneously, necessitating labor induction (induced records), or in term delivery by emergency cesarean section, without or with prior labor induction (cesarean and induced-cesarean records, respectively). For consistency with records of both the TPEHG DB and the TPEHGT DS in terms of gestational age at the time of recording, we used only later records of the ICEHG DS. Thus, 43 later induced (group IL), eight later cesarean (group CL), and 13 later induced-cesarean records (group ICL) were subsequently used to enrich the representation of term pregnancies, i.e. that of the term class. This allowed us to evaluate the performance of binary classifiers in distinguishing between records representing premature pregnancies (group PL) and those representing term pregnancies (groups TL, IL, CL, and ICL, together comprising (119+13)+43+8+13 = 196 records), regardless of delivery mode. More importantly, it enabled us to assess whether the inclusion of the non-spontaneous delivery modes impacts the power of the proposed method to predict premature birth.

Extraction of features

In the traditional feature engineering approach, suitable features are first constructed to represent spectral and non-linear parameters of EHG signals within the selected frequency bands [51]. This is done to assess the underlying physiological mechanisms believed to be localized in different locally stationary frequency bands [57]. In this study, we employed various univariate spectral and non-linear features extracted from EHG signals within the selected frequency bands that were previously defined and used in [57, 59, 81].

Feature selection and classification

The use of the traditional feature engineering approach commonly employs a complex, high-dimensional feature space, which requires prior selection, or reduction of the number, of features to achieve acceptable classifier generalization. Even so, many methods rely on using tens or even hundredths of features, as well as “black-box” approaches based on deep learning, which may severely diminish explainability. In contrast, in this study we used at most two easily derivable and understandable features for classification of premature and term records, thereby reducing the input feature space to at most two dimensions. The most promising features were identified using the selected feature selection strategy, using multiple standard classifiers and feature sets. For feature selection and classifier performance evaluation, we initially used only later premature (group PL) and later term spontaneous records (group TL). Next, to assess the impact of common non-spontaneous term delivery modes, we repeated the feature selection and classifier performance evaluation, this time using later premature (group PL) and later term spontaneous, induced, cesarean, and induced-cesarean records (groups TL, IL, CL, and ICL). To address the premature and term class imbalance, we strictly adhered to the proposed Partition-Synthesis evaluation technique [70]. Therefore, the minority class was over-sampled in each repetition of feature selection and classification only after partitioning the data into training and test sets.

Identification of the most promising features.

The full feature set includes PF, MF, PA, and SE, extracted from 10 frequency bands for signals S2 and S3. Note however, PA was not extracted from signals S2 and S3 in bands B0 and B0b, since when extracted from these bands the feature values equal 1.0 for all EHG records. Therefore, the full feature set includes 76 (80—4) features. The resulting high-dimensional feature space is complex, and using all of the features for classification may negatively impact the classification performance due to overfitting to the training data (resulting in poor generalization), and due to possible correlation between the features (redundant information). To decrease the number of features, simplify the feature space, reduce model complexity, mitigate the potential for overfitting, and improve classification performance, researchers employ various feature selection techniques [97].

Feature selection based on the Sequential Forward Selection (SFS) method [84] is a simple wrapper feature selection search strategy. In each step, the available feature that improves the desired classification performance metric the most (selection criterion) is sequentially added to the initially empty feature subset, either until a stopping criterion is met or until all of the features have been selected. The stopping criterion is based on the minimum of the average misclassification error (MCE), i.e. the number of misclassified samples divided by the total number of samples in the training data set. However, multiple runs of the SFS yield varying feature subsets. To avoid this inherent instability in the SFS method, the selection of features is stabilized with multiple runs of the SFS [85]. The occurrences of features in the yielded subsets are aggregated across all runs of the SFS. The resulting aggregated frequency histogram of feature occurrences shows the number of times that each feature was present in all of the yielded feature subsets. The final feature subset is constructed by using only the features having the highest numbers of occurrences in the aggregated frequency histogram, whereby the number of features in the final subset is determined by the minimum of the MCE function, averaged across all runs of the SFS.

The features were initially sorted according to the p-values to separate between premature and term records of the used dataset. For each (i-th) repetition of the SFS, we used the stratified hold-out strategy to divide the dataset into the training set TR_i (80% of the dataset) and the test set TE_i (20% of the dataset). The test sets were not used for feature selection, and remained reserved for the final classification performance evaluation with N repetitions, using the identified most promising features. The average MCE of the stratified 5-fold cross-validation, using only the training set (TR_i) in each SFS repetition, was used to select the next best feature to be added to the feature subset until no more features remained to be selected. Inside the cross-validation loop for calculating the MCE in each repetition of the SFS, we used the Synthetic Minority Over-sampling technique (SMOTE) [98] to balance the premature and term classes. This was done only after splitting, or partitioning, the data into the training-validation subsets, according to the Partition-Synthesis evaluation technique [70]. The subset of selected features in each run of the SFS comprises the selected K_min features that yielded the minimum of the MCE function. Then, the aggregated frequency histogram of feature occurrences H was updated based on the subset of selected features in each SFS repetition. The number of final selected features was defined by the minimum of the average MCE function across all repetitions of the SFS.

Algorithm 1 describes the used feature selection procedure in more detail, where X is the input feature set with K features, N is the number of repetitions of the SFS, Dataset is the input set of premature and term records, function MCE returns the average cross-validation misclassification error using the specified subset of data and features, and function Index returns the position of the selected feature in the histogram. The feature selection procedure returns the aggregated frequency histogram of feature occurrences H, and the final selected feature subset Y with m features.

Algorithm 1 Pseudo code for feature selection using SFS, stabilized with the frequency-based aggregation of the selected features technique

Input: X = {x₁, x₂, …, x_K}, N = 200, and Dataset

Output: histogram H, number of final selected features m,

    and final selected feature subset Y

 H_k = M_k = 0, k ∈ [1, K] // Initialize values

 Y = Sort_p(X) // Sort X according to p-values

 for i = 1 to N do

  [TR_i, TE_i] = Partition(Dataset) // Holdout 80%/20%

  Z = ∅ // Start with empty feature subset

  for k = 1 to K do // Sequential forward selection

   

   M_k = M_k + MCE(TR_i, Z ∪ z_k)

   Z = Z ∪ z_k // Add selected feature to subset

  end for

  // Find subset yielding the lowest MCE

  K_min = arg min_k∈[1,K](MCE(TR_i, {z₁, …, z_k}))

  // Update histogram values

  for k = 1 to K_min do

   

  end for

 end for

 // Get number of final selected features

 m = arg min_k∈[1,K](M_k/N)

 // Construct final subset based on histogram peaks

 Y = ∅

 for k = 1 to m do

  

  Y = Y ∪ x_f

 end for

To reduce bias towards a particular classifier used for calculating the average MCE function, and to avoid emphasizing features extracted from less promising broad frequency bands, the whole feature selection procedure was performed multiple times, each time using one of the five selected classifiers (LDA, QDA, Naive Bayes, SVM Poly3, and SVM RBF) and one of the following four feature sets:

• Feature set 1: The four features, PF, MF, PA, and SE, extracted from signals S2 and S3 in all 10 of the selected frequency bands, except PA for S2 and S3 in bands B0 and B0b (in total 4 ⋅ 2 ⋅ 10 − 4 = 76 features);
• Feature set 2: All features from feature set 1, except those extracted from the least promising bands B0H and Bb (features extracted from these two bands yielded the least significant p-values to separate between premature and term records according to [81]) (4 ⋅ 2 ⋅ 8 − 4 = 60 features);
• Feature set 3: The four features, PF, MF, PA, and SE, extracted from signals S2 and S3, but only in bands B0L, B0L’, B3 (most promising bands according to [81]), and the broadest band B0b, except PA for S2 and S3 in band B0b (4 ⋅ 2 ⋅ 4 − 2 = 30 features);
• Feature set 4: All features from feature set 3, except those extracted from band B0b (4 ⋅ 2 ⋅ 3 = 24 features).

In total, the whole feature selection procedure was performed 20 times (for five classifiers * four feature sets), each time yielding one histogram (examples of these histograms are shown in Figs 1(a) and 2(a) in the Results section) for each distinct classifier and feature set combination. Although it is expected that each run of the feature selection procedure returns the final subset with more than two features, in this study we intentionally chose to perform the classification between premature and term record using only one, or at most two, of the most promising features. If using a single feature, or a single biomarker, for classification, the decision boundary between premature and term classes is simply represented with a single value, i.e. a threshold. In the case of using two features, the two-dimensional feature space is easily comprehensible, and the decision boundary can be very effectively visualized. Therefore, only the highest two peaks of each histogram of feature occurrences were noted (indicating the two of the most selected features, i.e. the features taking the first and the second place, for each individual classifier and feature set combination). The number of times that a feature took first place (feature A) or second place (feature B) was then counted for all combinations of classifiers and feature sets (five classifiers * four feature sets). This yielded a final histogram that shows how many times the feature took the first or second place, out of 20 runs of the feature selection procedure. These histograms are shown to summarize our results in Figs 1(b) and 2(b) in the Results section. Finally, the two features that most often took the first or second place in the final histogram were identified as the two most promising features (feature A and feature B).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. The results of feature selection using later premature (group PL) and later term spontaneous (group TL) records.

The left histogram (a) shows the counts of feature occurrences and the optimal set of features yielded by the feature selection procedure if using feature set 1 and LDA classifier. Features are ranked according to the feature occurrence count. The features from the optimal set are annotated, and the bars for the top two features (i.e. the features taking the first and second place) are shown in black. The right histogram (b) shows the total number of times that a feature took first place (above the dashed separator line) and second place (below the dashed separator line) out of 20 runs (i.e. using five classifiers and four feature sets) of the feature selection procedure.

https://doi.org/10.1371/journal.pone.0308797.g001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. The results of feature selection using later premature (group PL) and later term spontaneous, induced, cesarean, and induced-cesarean (groups TL, IL, CL, and ICL) records.

The left histogram (a) shows the counts of feature occurrences and the optimal set of features yielded by the feature selection procedure if using feature set 1 and LDA classifier. Features are ranked according to the feature occurrence count. The features from the optimal set are annotated, and the bars for the top two features (i.e. the features taking the first and second place) are shown in black. The right histogram (b) shows the total number of times that a feature took first place (above the dashed separator line) and second place (below the dashed separator line) out of 20 runs (i.e. using five classifiers and four feature sets) of the feature selection procedure.

https://doi.org/10.1371/journal.pone.0308797.g002

Visualization of average decision boundaries.

In case of using a single feature for binary classification between premature and term records, the task of all of the selected classifiers is reduced to simple thresholding of the feature value. Hence, all of the trained classifiers may be modeled with a single parameter, or threshold. Nevertheless, even in the case of a single feature, different algorithms for training the selected classifiers may produce varying thresholds, resulting in varying sensitivity-specificity trade-offs. Furthermore, the evaluation technique with multiple repetitions (N = 200), each time using different subsets of the data for training, resulted in different threshold values. Therefore, for the case of using a single feature, we calculated the average threshold across all repetitions and classifiers. This resulted in the average threshold value for the single feature, regardless of classifier. This single threshold reflects the learned decision boundary between premature and term records in a one-dimensional feature space. It can also be visualized in the two-dimensional feature space with a single straight line, perpendicular to the axis of the thresholded feature (these boundaries are shown in Figs 3(d) and 4(d) in the Results section).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Average ROC curves and decision boundaries for classification between later premature records (group PL) and later term spontaneous records (group TL).

The average ROC curves are shown for classification (a) using only feature PA, (b) using only feature MF, and (c) using two features PA+MF, for the Naive Bayes (shown in cyan), LDA (shown in magenta), and SVM Poly3 (shown in black) classifiers. The error bars for each ROC curve show the standard deviations at different false positive rates. In the scatter plot (d), the average classification decision boundaries are shown for the single-feature model (PA), averaged across all classifiers (vertical brown boundary), and for two-feature models (PA+MF) using Naive Bayes (dotted blue boundary), LDA (solid blue boundary), and SVM Poly3 (dashed blue boundary) classifiers. The histograms on the top and right sides of the scatter plot show the normalized distribution of the feature values for premature (in red) and term records (in green). The red and green dotted lines on the histograms show the fit to the normal distribution according to the feature values for premature and term records, respectively.

https://doi.org/10.1371/journal.pone.0308797.g003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Average ROC curves and decision boundaries for classification between later premature records (group PL) and all later term records (groups TL, IL, CL, ICL).

The average ROC curves are shown for classification (a) using only feature PA, (b) using only feature MF, and (c) using two features PA+MF, for the Naive Bayes (shown in cyan), LDA (shown in magenta), and SVM Poly3 (shown in black) classifiers. In the scatter plot (d), the average classification decision boundaries are shown for the single-feature model (PA), averaged across all classifiers (vertical brown boundary), and for two-feature models (PA+MF) using Naive Bayes (dotted blue boundary), LDA (solid blue boundary), and SVM Poly3 (dashed blue boundary) classifiers. The histograms on the top and right sides of the scatter plot show the normalized distribution of the feature values for premature (in red) and term records (in green). The red and green dotted lines on the histograms show the fit to the normal distribution according to the feature values for premature and term records, respectively.

https://doi.org/10.1371/journal.pone.0308797.g004

If using two features, the decision boundary curve depends on the classifier. To visualize the average decision boundary of the binary classifiers between premature and term records, we constructed a grid of feature values in the two dimensional feature space, and performed the classification using the feature values across the grid. This yielded a matrix of classifier scores, indicating the likelihood that the sample at the individual position in the grid comes from the premature class. These scores were then averaged across all repetitions, and a single contour, or isoline (at the level of the 50% likelihood that the sample belongs to the premature class), was drawn for the best performing classifiers on the scatter plot of feature values for all records of the used datasets.

The most promising selected features

Using later premature (group PL) and later term spontaneous (group TL) records of the TPEHG DB and TPEHGT DS, we first identified the most promising features based on the results of the SFS method, stabilized with the frequency-based aggregation of the selected features technique, with the specified number of repetitions (N = 200), using various classifiers and feature sets. The feature selection procedure always returned the final feature subset with more than two features, regardless of classifier and feature set. The upper half of Table 4 shows the optimal number of features returned by the feature selection procedure and the achieved average Classification Accuracy, CA, on the validation sets, when performing feature selection and classification between later premature (group PL) and later term spontaneous (group TL) records, using various classifiers and feature sets. The highest CA (83.4% ± 3.1%) was achieved with the optimal set comprising nine features returned by the feature selection procedure using the SVM Poly3 classifier and feature set 1. The average optimal number of returned features, across all classifiers, was 13.0 ± 8.0, 12.2 ± 6.6, 7.4 ± 3.2, and 6.8 ± 2.0, when using feature set 1, 2, 3, and 4, respectively. Even so, to identify the two most promising features, we always used only the two highest peaks (features taking first and second place) in the histogram of feature occurrences.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. The optimal number of features, and the achieved Classification Accuracy (CA), when performing the feature selection, and classification on the validation sets, with the specified number of repetitions (N = 200), using various classifiers and feature sets.

https://doi.org/10.1371/journal.pone.0308797.t004

The left histogram in Fig 1(a) shows the counts of feature occurrences returned by the feature selection procedure if using feature set 1 and the LDA classifier with records of the TPEHG DB and TPEHGT DS. In particular, the feature selection procedure using this feature set and LDA classifier yielded an optimal set with 11 features (according to upper part of Table 4) which are annotated. Note the two highest peaks of the histogram, i.e. the first two black bars highlighting the two features having the highest feature occurrence count. In this particular case, out of N = 200 repetitions, feature PA of signal S3 in the frequency band B0L’ (0.125-0.575 Hz), succinctly PA S3 B0L’, ended in the selected feature subset 171 times (taking first place among all of the features when using this classifier and feature set). Moreover, feature MF of signal S2 in band B3 (3.5-5.0 Hz), succinctly MF S2 B3, ended in the feature subset 30 times (taking second place). The feature selection procedure was repeated for all five classifiers and four feature sets, yielding 20 such histograms. The right histogram in Fig 1(b) shows the number of times features took first place (above the dashed separator line) and the number of times features took second place (below the dashed separator line) out of 20 runs of the feature selection procedure (using five classifiers and four feature sets). PA S3 B0L’ was always the most selected feature, regardless of classifier and feature set, hence it took first place in 20 out of 20 runs of the feature selection procedure. This is why the histogram above the dashed separator line in Fig 1(b) shows only one bar. However, the second place is shared by four features, as shown in the histogram below the dashed separator line in Fig 1(b). This means that for some classifiers and feature sets, the second most selected feature was either MF S2 B3, SE S2 B2, PA S2 B3, or SE S2 B0b. Nevertheless, MF S2 B3 took the second place among the selected features in nine out of 20 runs, which is above the counts for other features sharing the second place, as shown in Fig 1(b). Based on these results, PA S3 B0L’ and MF S2 B3 were identified as two of the most promising features for classification between later premature (group PL) and later term spontaneous (group TL) records.

The feature selection was also performed for classification between later premature (group PL) and later term spontaneous, induced, cesarean, and induced-cesarean (groups TL, IL, CL, and ICL) records of the TPEHG DB, TPEHGT DS, and ICEHG DS. For this case, the optimal number of features returned by the feature selection procedure and the achieved average CA on the validation sets are shown in the lower half of Table 4, using various classifiers and feature sets. The highest CA (85.0% ± 3.3%) was achieved with the optimal set comprising 10 features returned by the feature selection procedure using the SVM Poly3 classifier and feature set 1. The average optimal number of returned features, across all classifiers, was 12.4 ± 5.6, 11.0 ± 5.1, 6.8 ± 2.2, and 6.4 ± 1.9, when using feature set 1, 2, 3, and 4, respectively. Again, to subsequently identify the two most promising features, we always used only the features taking first and second place. Fig 2(a) shows the resulting histogram of feature occurrences returned by the feature selection procedure if using feature set 1 and LDA classifier which yielded an optimal set with 11 features (according to lower half of Table 4). Similarly to the results in Fig 1(a), PA S3 B0L’ was the most selected feature, followed by MF S2 B3. Next, Fig 2(b) shows the histogram with the number of times that features took the first place (shown in the histogram above the dashed separator line) and second place (shown in the histogram below the dashed separator line) out of 20 runs of the feature selection procedure. Again, feature PA S3 B0L’ took first place in 20 out of 20 runs of the feature selection procedure, as it was always the most selected feature, regardless of the used classifier and feature set. The second place is shared by seven features, and of those MF S2 B3 most often took second place (in seven out of 20 runs). Interestingly, regardless of the dataset used (either only using records from groups PL and TL, or using records from groups PL, TL, IL, CL, and ICL), the same two features, PA S3 B0L’ and MF S2 B3, were identified as most promising.

Classification performance

Table 5 summarizes the classification performance results (on the N = 200 leftover test sets that were not used for feature selection) for classification between later premature (group PL) and later term spontaneous (group TL) records, using one or two of the previously identified most promising features, (A) PA S3 B0L’ and (B) MF S2 B3, a variety of standard classifiers. If using only feature (A), PA, the AUC score is 0.81 ± 0.08, regardless of the used classifier. The Sensitivity (Se), Specificity (Sp), Positive Predictive Value (PPV), Negative Predictive Value (NPV), and Classification Accuracy (CA) slightly vary depending on the classifier, due to differences in the training algorithms. Using only feature (A), PA, the LDA classifier yielded the highest CA (75.1% ± 9.1%), while the Naive Bayes classifier yielded the next highest CA (74.6% ± 9.3%). If using only the second most promising feature (B), MF, the classification performance results are significantly worse. However, if using using both features (A and B), the CA and AUC are both slightly higher, compared to using a single feature. In particular, the best results were obtained using the SVM Poly3 classifier, yielding the highest CA (76.3% ± 7.8%) and AUC (0.85 ± 0.07). For two features (A and B), SVM Poly3 exhibited the highest Se (86.0% ± 18.6%) and NPV (82.6% ± 12.8%) among all classifiers, at the cost of lower Sp (66.6% ± 8.3%) and PPV (72.0% ± 6.2%), respectively, while other classifiers, particularly LDA and Naive Bayes, yielded more balanced Se and Sp (and more balanced PPV and NPV), with only slightly reduced AUC (0.84 ± 0.08). Furthermore, the average AUC and CA obtained on the testings sets (Table 5) were very close to those obtained on the training sets. In particular, using either one (A or B) or both features (A and B), the average CA and AUC using the testing sets were never lower by more than 1.8% and 0.02, respectively, compared to using the training sets, regardless of feature set and classifier, suggesting that all of these models generalize well to unseen data.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. Classification performance results for the top two features (PA and MF), and of their combination, when classifying between later premature records (group PL) and later term spontaneous records (group TL) of the TPEHG DB and TPEHGT DS.

https://doi.org/10.1371/journal.pone.0308797.t005

Similarly, Table 6 summarizes the classification performance results (on the N = 200 leftover test sets that were not used for feature selection) for classification between later premature (group PL) and later term spontaneous, induced, cesarean, and induced-cesarean (groups TL, IL, CL, and ICL) records. These results are comparable to the performance results for classification between records from groups PL and TL only (shown in Table 5). If using only the most promising feature (A), PA, the Naive Bayes and the LDA classifiers performed the best, with CA (76.5% ± 8.9%) and (76.4 ± 9.2%), respectively, and yielding the same AUC (0.81 ± 0.08). In comparison, the results for the second most promising feature (B), MF, are significantly worse. Furthermore, if using both features (A and B), SVM Poly3 exhibited the highest classification performance, yielding the highest CA (78.0% ± 7.1%) and AUC 0.86 ± 0.07, with the highest Se (88.1% ± 12.3%) and NPV (85.1% ± 11.6) at the cost of lower Sp (67.8% ± 8.1%) and PPV (73.3% ± 6.0), respectively. Similarly, the average CA and AUC using the testing sets (Table 6) were never lower by more than 1.8% and 0.01, respectively, compared to using the training sets, regardless of feature set and classifier.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 6. Classification performance results for the top two features (PA and MF), and of their combination, when classifying between later premature records (group PL) and later term spontaneous, induced, cesarean, and induced-cesarean records (groups TL, IL, CL, and ICL) of the TPEHG DB, TPEHGT DS, and ICEHG DS.

https://doi.org/10.1371/journal.pone.0308797.t006

The average decision boundaries

Fig 3 shows the average ROC curves for classification between later premature records (group PL) and later term spontaneous records (group TL), for the best three performing classifiers, i.e. Naive Bayes (shown in cyan), LDA (shown in magenta), and SVM Poly3 (shown in black) using (a) only the best performing single feature PA S3 B0L’, (b) only the second best performing feature MF S2 B3, and (c) using both features (PA + MF), as well as (d) the average decision boundaries of these classifiers, according to the results in Table 5. The associated error bars on the ROC curve plots show the standard deviations of the sensitivities (true positive rates) at different false positive rates. Notice how the ROC curves for the best performing single-feature model using only PA in Fig 3(a) and for the two-feature model using PA + MF in Fig 3(c) are very close together, indicating only a slight improvement if using both features.

The average decision boundaries in Fig 3(d) are shown along the scatter plot of feature values in the two-dimensional feature space (values for PA S3 B0L’ on the abscissa, and MF S2 B3 on the ordinate). The histograms on the top and right sides of the scatter plot show the normalized distribution of values of features PA and MF, respectively, for premature records (in red) and term records (in green). The red and green dotted lines on the histograms show the fit to the normal distribution according to the feature values for premature and term records, respectively. Note that in premature records, the value of PA is higher (top histogram), and the MF is lower (right histogram), than in term records. Furthermore, note how the term records are more clustered towards the left, and premature records more towards the right side of the scatter plot. The brown vertical decision boundary line indicates the average decision threshold using only the most promising feature PA S3 B0L’ (averaged across all classifiers), and is approximately 0.66. On average, if the value of this feature is above the threshold, the record would be classified as premature, otherwise as term. The average decision boundary for classification using only the second best performing feature, MF S2 B3, is not shown since the related classification performance using this feature was inferior, compared to using only PA. If using two features, the average decision boundary between premature and term records depends on the classifier, yet for all classifiers the decision boundaries delineate the region of premature records in the lower-right side of the visualized two-dimensional feature space. For classification using the LDA classifier, the decision boundary is a simple line, diagonally splitting the feature space into two regions. For the Naive Bayes and SVM Poly3 classifiers, their decision boundaries resemble a quadratic function and a third degree polynomial function, respectively. Note how the SVM Poly3 decision boundary extends further into the region with higher density of term records, resulting in higher Se, but lower Sp, compared to the decision boundary of the Naive Bayes classifier.

Fig 4 shows the average ROC curves for classification between later premature (group PL) and later term spontaneous, induced, cesarean, and induced-cesarean (groups TL, IL, CL, and ICL), for the best three performing classifiers, i.e. Naive Bayes (cyan), LDA (magenta), and SVM Poly3 (black) using (a) only the best performing single feature PA S3 B0L’, (b) only the second best performing feature MF S2 B3, and (c) using both features (PA + MF), as well as (d) the average decision boundaries of these classifiers, according to the results in Table 6. The ROC curves for these classifiers are quite similar to those shown in Fig 3(a)–3(c), yet for the two-feature models (PA + MF) they are slightly higher than the ROC curves for classification using only later premature and later term spontaneous records. Furthermore, the error bars, which show the associated standard deviations of sensitivities (true positive rates) at different false positive rates, are more dense than those shown in Fig 3. This is because term pregnancies are represented by more records, resulting in more diversity in the N = 200 test sets used for the final classifier evaluation.

Finally, Fig 4(d) shows the associated average decision boundaries for the same three classifiers and the same two-dimensional feature space as also shown in Fig 3(d). However, note that there are many more term records (green) since the term class is now represented with additional records from pregnancies ending in induced, cesarean, and induced-cesarean delivery modes. Furthermore, note the similarity between the decision boundaries shown in Fig 4(d), and the decision boundaries shown in Fig 3(d). In fact, for the single-feature model using only feature PA S3 B0L’, the decision threshold, averaged across all classifiers, again equals approximately 0.66, and is visualized with the solid brown vertical line in Fig 4(d).

Peak amplitude of the normalized power spectrum in the FWL band

According to the results shown in Fig 1, the most promising feature for classification between later premature (group PL) and later term spontaneous (group TL) records is PA, extracted from the horizontally oriented signal S3 in the frequency band B0L’ (0.125-0.575 Hz). These results are consistent with the results of our previous characterization study [81], which showed that feature PA S3 B0L’ exhibited the highest power, in terms of the p-values, to separate between later premature and later term records. Additionally, the results of the present study (Fig 1) show that PA S3 B0L’ is well suited for use with various traditional classifiers, since it always took the first place among the selected features, regardless of which classifier was used.

The band used for extraction of the most promising feature, PA S3 B0L’, is mainly reflective of the FWL [81], and its lower and upper bounds almost exactly correspond to the bounds of the FWL band (0.1-0.6 Hz) as defined in [49]. The findings in [81] suggest that the frequency components in the FWL band may be more important than the frequency components in the FWH band for the prediction of premature birth using later recorded EHG records, according to the p-values for features extracted from various bands related to the FWL and FWH. Again, the results of the present study are in agreement with those findings, since features extracted from the frequency band B0H’ (0.575-1.0 Hz), which was chosen specifically for its relation to the FWH, were never identified as the first, nor as the second, most promising feature for classification between later premature and later term spontaneous records. This confirms that the frequency components related to the FWL, including those below 0.34 Hz, which are commonly excluded from analysis because of overlap with frequency components due to maternal breathing [88], are important for premature birth prediction.

For comparison of our classification performance results to those of studies related to the prediction of premature birth, we used the AUC score, as it is also one of the most commonly reported performance metric for methods based on chemical and molecular biomarkers (Table 1) as well as for methods based on the EHG (Tables 2 and 3). According to the results in Table 5, if using only the most promising feature, PA S3 B0L’, for classification between later premature and term spontaneous records, the average AUC score is 0.81 ± 0.08, regardless of the used classifier. For comparison to the related methods based on the EHG, we also calculated other widely reported performance metrics CA, Se, Sp, PPV, and NPV. For the most promising feature, PA S3 B0L’, the results in Table 5 show there are slight differences between the performances of the selected classifiers in terms of CA, Se, Sp, PPV, and NPV. In particular, the best results were achieved using the LDA and Naive Bases classifiers, which yielded the average CA of 75.1% ± 9.1% and 74.6% ± 9.3%, respectively, when using later premature and later term spontaneous records. The lowest CA was achieved using the SVM Poly3 classifier, as the training algorithm settled, on average, at a less optimal threshold favoring higher Sp and NPV at the cost of lower Se and PPV, respectively.

These results suggest that the fully automated non-invasive prediction of premature birth using a single feature, PA S3 B0L’, extracted from the EHG collected around the 31st week of gestation, outperforms most recent methods for non-invasive prediction of premature birth based on chemical and molecular biomarkers shown in Table 1. In particular, all methods based on a single chemical or molecular biomarker (upper half of Table 1), yielded lower AUC scores (0.79 or less). More strikingly, using only PA S3 B0L’ (AUC of 0.81) outperforms even most methods based on multiple chemical and molecular biomarkers (lower half of Table 1). The only exception seems to be the method based on expression of multiple genes [40] that yielded only a slightly higher AUC of 0.82. As such, in comparison to chemical or molecular biomarkers, the prediction of premature birth using PA S3 B0L’ achieves better, or comparable, classification performance in terms of the AUC score, yet does not require any potentially costly sample analysis.

For comparison to related studies in the area of premature birth prediction based on the EHG, note that there is no consensus among researchers regarding a benchmark EHG dataset and signal scope (either whole signals, or annotated contraction or dummy intervals). In particular, note that none of the studies in Table 2 used a combined dataset with records of the TPEHG DB and TPEHGT DS to increase the total number of available premature samples for a more realistic assessment of classification performance. Furthermore, most studies used annotated contraction and dummy intervals instead of whole signals. Nevertheless, our results show that PA S3 B0L’ outperforms, in terms of CA and AUC, all but one of the single-feature methods in Table 2. The only exception seems to be the method reported in [73], which yielded a higher CA of 84.0% and AUC of 0.86 when using SE extracted from signal S2 in the band B2 (2.2-3.5 Hz). In that study, the best results were achieved when classifying between annotated dummy intervals in premature and term records of the TPEHGT DS. However, one has to consider that the TPEHGT DS comprises only 13 premature and 13 term EHG records, and on average each record contains about four contraction and four dummy intervals [57]. As such, the intervals from the same record are correlated, yet they may simultaneously end up in both the training and the testing set, which may have resulted in data leakage and overly optimistic results. Nevertheless, in our study we also included SE S2 B2 in two feature sets, yet it never took the first place among the selected features (Fig 1). Among the features taking the second place, it was even less frequently selected than MF S2 B3 (3.5-5.0 Hz). These results suggest that SE S2 B2 is not so promising when using a larger dataset for evaluation of the classification performance.

Comparison to the results of methods based on the EHG using multiple features in Table 3 is somewhat more tentative. Apart from the fact that most studies used different datasets for evaluation, some of the methods may suffer from potential methodological flaws related to feature selection using the whole dataset, improper use of under-sampling without repetitions, or the use of correlated intervals that supposedly represent independent samples. Nevertheless, the results in Table 5 show that the use of PA S3 B0L’ alone outperforms several methods listed in Table 3. However, it is important to note that only two of the studies [67, 80] in Table 3 used later records of the TPEHG DB and TPEHGT DS. Moreover, in [67] the authors used a test set comprising overlapping two minute intervals extracted from only four term and four premature records. However, they used only 17 out of 119 later term records of the TPEHG DB, and two later premature records from the TPEHG DB were excluded. As such, their results (CA 88.52% and AUC 0.89) may be overly optimistic. In contrast, in [80] the authors used all later records from the TPEHG DB and TPEHGT DS and analyzed whole signals, just like in our study. However, despite using a complex deep learning model with over 600 features and obstetric parameters, their method yielded an AUC score of only 0.78 (CA was not reported). Notably, another study [79] which used a deep learning approach reported an even lower AUC score of 0.58, using only records of the TPEHG DB. Interestingly, the use of a single feature, PA S3 B0L’ (AUC 0.81), outperforms both of these deep learning models.

Furthermore, our method does not perform any complex steps like prior segmentation of EHG signals to remove corrupt signals or specific non-physiological segments. In fact, we intentionally used whole records which leads to efficient and fully automated analysis. Besides, prior segmentation is normally performed in order to further improve the predictive accuracy of the model. For example, in [72] the authors relied on prior manual segmentation by two experts to remove corrupt signal segments to improve the accuracy of the model. Their results suggest that prior segmentation of signals might even improve the classification performance of the models in this study, but unfortunately at the cost of an additional complex pre-processing step. Furthermore, we are aware that the variability of EHG signals may effect the predictive performance of the model, however, the number of available EHG records is too small to perform a thorough analysis.

Nevertheless, using tens or even hundreds of features and deep learning approaches may achieve better classification performance (Table 3). Moreover, deep learning models may also obviate the need for feature engineering, and provided sufficient effort is put into their explainability, they may reveal additional relations between the underlying physiological mechanisms that may contribute to premature birth. However, obstetricians may find it favourable to use a single, easily derivable feature, powering a simple classification model requiring no complex processing steps, to quickly and effectively assess the danger of premature birth. Furthermore, in contrast to more complex models, a single feature may be more easily attributed to known underlying physiological mechanisms. As such, it provides a more effective and direct insight into the physiological changes leading to premature birth. In particular, feature PA S3 B0L’ estimates the proportion of the spectral peak amplitude [57] in the frequency band related to the FWL [81], which is believed to reflect the propagation of the electrical activity along the uterus [47, 49, 102]. Several studies have reported that the spectrum of uterine bursts shifts toward higher frequencies as labor approaches [46–48]. The trends reported in [81] suggest that in premature pregnancies, the relative power of the spectrum in the FWL band increases between the 23rd and 31st week of gestation, and this change results in significantly higher values of PA S3 B0L’ by the 31st week of gestation in premature pregnancies. The high classification power of PA S3 B0L’ (Table 5) suggests that the high relative intensity of the spectral components in the FWL frequency band is an important and effective predictor of premature birth.

Furthermore, when using a single feature for the prediction of premature birth, the decision boundary between premature and term records is simply represented by a single threshold, which may be decreased or increased to move along the ROC curve shown in Fig 3(a). The task of the classifier is then to simply threshold the value of a single feature. However, as evident from the differences in Se and Sp in Table 5, each classifier yielded a slightly different threshold due to differences in the training algorithms. To avoid bias towards any particular classifier and related training algorithm, we averaged the threshold of the single-feature models across all classifiers, resulting in a single average threshold value for assessing the danger of premature birth. For PA S3 B0L’, the average threshold, or decision boundary, is approximately 0.66, and is shown with a solid brown vertical line in Fig 3(d). If the value of PA S3 B0L’ is higher than this threshold, this indicates an increased danger of premature birth.

Median frequency in the band B3

According to the results in Fig 1, the second most promising feature is MF, extracted from the vertically oriented signal S2 in the frequency band B3 (3.5-5.0 Hz). Additional, less promising features that took the second place in the histogram in Fig 1(b) include SE S2 B2, PA S2 B3, and SE S2 B0b. Notably, all of these features were extracted from the frequency bands that include the frequency components above 1.0 Hz. Among these features, those extracted from bands B2 and B3, reflective of the influence of the maternal heart on the uterus in terms of the second and third harmonics of the maternal heart rate (MHR) seem to be particularly promising. This is in agreement with the results in [57, 73, 81], which also showed the importance of frequency components above 1.0 Hz. According to [57, 81], the MF in the band B1 (1.0-2.2 Hz) is believed to be reflective of the MHR, and similarly, B2 (2.2-3.5 Hz) and B3 (3.5-5.0 Hz) are believed to be reflective of two times the MHR or three times the MHR, respectively. Interestingly, among the features extracted from signals in the frequency bands above 1.0 Hz, MF S2 B3 yielded the lowest p-value to separate between later premature and term records [81]. This is most likely because in band B3 there is less overlap between the frequency components related to the FWH and frequency components related to the maternal heart than in bands B1 and B2. Although many features extracted from signals in the frequency bands below 1.0 Hz exhibited lower p-values [81], in our study none of them were included among the most promising features, apart from PA S3 B0L’ (Fig 1). This is most likely because those features are correlated (redundant information) to PA S3 B0L’, thus, their addition to the feature set did not improve the classification performance. Additionally, the modeled trends of MF S2 B3 in [81] suggest that in premature pregnancies MHR decreases between the 23rd and 31st week of gestation, but not in term pregnancies. Consequently, the importance of MF S2 B3 for the prediction of premature birth, according to the results in Table 1, suggests that this possible relation between premature birth and this decrease of the MHR deserves further study.

Nevertheless, if using only MF S2 B3 for classification between later premature (group PL) and later term spontaneous (group TL) records, the results in Table 5 show that this feature alone performs poorly, yielding an AUC of 0.71 or less. However, when PA S3 B0L’ is used together with MF S2 B3, the two-feature model using the SVM Poly3 classifier yielded the AUC of 0.85 ± 0.07 and CA of 76.3% ± 7.8% (lower part of Table 5). Thus, using only two features extracted from signals in non-overlapping frequency bands, each believed to be reflective of distinct underlying physiological mechanisms, resulted in the highest classification performance. The resulting two-feature model represents a slight improvement over the single-feature model that uses only PA S3 B0L’. Correspondingly, the ROC curves shown in Fig 3(a) are fairly close to those shown in Fig 3(c).

Most noteworthy, our two-feature model (PA S3 B0L’ and MF S2 B3) seems to outperform all of the chemical and molecular biomarkers in Table 1, in terms of the AUC. On the other hand, very few related studies based on the EHG used only two features (Table 3). In particular, in [73] the authors achieved the highest CA of 85.8% and AUC of 0.89 using SE in the band B2 (2.2-3.5 Hz) for signals S2 (feature 1) and S3 (feature 2), and in [57] the highest CA of 70.75% and AUC of 0.80 using PA in the band B1 (1.0-2.2 Hz) for signal S2 (feature 1) and the tocogram (feature 2). However, both studies used annotated dummy intervals from only 13 premature and 13 term records of the TPEHGT DS. As explained in the previous subsection, the evaluation of classification performance using such a small dataset and correlated intervals may be less realistic.

Additionally, in contrast to more complex models in Table 3 which rely on more than two features, the decision boundaries of our two-feature models (PA and MF) can still be very simply and effectively visualized. Fig 3(d) shows the decision boundaries, averaged across all (N = 200) evaluation runs, when using the SVM Poly3 (dashed blue curve), Naive Bayes (dotted blue curve), and LDA (solid blue line) classifiers. Note how most premature records are located in the lower right corner, “inside” the areas outlined by the average decision boundaries for these three classifiers (higher PA S3 B0L’ and lower MF S2 B3). This type of visualization makes it fairly easy to observe the differences, in terms of Se and Sp, between the individual classifiers according to the results Table 5. Although all classifiers using both features exhibit very similar CA and AUC, the LDA classifier exhibits the best balance between Se and Sp (and between PPV and NPV).

In addition, we discuss an example of a simple composite biomarker for assessing the danger of premature birth using only the weighted sum of the two feature values according to the linear decision boundary (diagonal blue line) in Fig 4(d) corresponding to a linear discriminant function: (14) where x = [PA S3 B0L’, MF S2 B3]^T, w = [5.98, -4.02]^T, and b = -12.78. Hence, if (x^Tw − b) is greater than 0, this indicates danger of premature birth, and the higher the value the higher the probability of premature delivery. This way, the predicted pregnancy outcome (premature/term) can be displayed as a mark (point) in a two dimensional diagram (ordinate—maternal resting heart rate, abscissa—peak amplitude of the normalized power spectrum in the FWL frequency band) displaying also the decision boundary line.

Implementing the predictive model into existing clinical systems would most likely start as an additional clinical investigation, probably on an elective basis at first. Later, if high correlation between the model prediction and premature birth were to be validated by a clinical trial, it can also become routine. A potential challenge would arguably be the initial acceptance by obstetricians, which would need to be convinced by the ease of use and its effectiveness, as well as garnering prior support from medical device manufacturers to support the subsequent development and take further steps towards regulatory approvals. Generally, the use of the proposed model in clinical practice would entail making a non-invasive 30-minute EHG recording around the 31rd week of gestation using a suitable recording apparatus connected to a personal computer, followed by automated digital signal pre-processing, classification, and presentation of the results. The associated direct costs are also expected to be low. The main consumables include five disposable Ag2 Cl electrodes (four placed symmetrically above and below the navel, and one reference electrode placed on the left thigh [82]). The electrodes would then be connected to a personal computer with a high-resolution multi-channel analog-to-digital converter. Due to the low computational complexity of the proposed method, no special hardware processing resources would be required.

The model would be integrated on the connected personal computer, subject to prior regulatory approval of such a diagnostic system for assessing the danger of premature birth. The obtained results using such a system could initially help obstetricians decide on further diagnostic procedures, if the pregnancy is identified at risk for premature birth. This could include more frequent ultrasounds, electronic fetal monitoring, and additional tests related to known risk factors, which would lead to more targeted management of threatened premature birth. Another approach for dealing with risk stratification is using regression analysis (i.e. to predict the pregnancy duration), or treated as a multi-class problem to stratify the EHG recording into sub-categories like extremely premature, very premature, moderate to late premature, knowing the risks for neonatal morbidity and mortality increase with the degree of prematurity [1, 51]. However, this represents a much harder problem than the proposed binary classification (premature vs term) for assessing whether the pregnancy is at risk of premature birth or not, and currently the small number of available premature EHG records would likely make the results of such risk stratification studies unreliable.

Impact of non-spontaneous delivery modes

For realistic evaluation of methods for the prediction of premature birth based on the EHG, it is important to also consider other common non-spontaneous delivery modes. At the time of recording, it is generally not known whether the pregnancy will end in delivery modes like induced labor or cesarean section. However, to our knowledge, currently existing methods for the prediction of premature birth using the EHG are based solely on classification between EHG records from pregnancies that ended in premature spontaneous or term spontaneous delivery modes, and therefore disregard the possibility of other common delivery modes.

To this end, we used the records of the ICEHG DS, alongside records of the TPEHG DB and TPEHGT DS, to additionally represent the term class with common non-spontaneous delivery modes (i.e. with groups TL, IL, CL, and ICL instead of only group TL). According to the results in Fig 2, the same two most promising features were identified, PA S3 B0L’ and MF S2 B3, as when only group TL was used to represent the term class (Fig 1). According to the results in Table 6, if using only the most promising feature (PA S3 B0L’) for classification between later premature and all later term records, the average AUC score is 0.81 ± 0.08, regardless of the used classifier. In terms of CA, the best results were achieved using Naive Bayes (CA of 76.5% ± 8.9%) and LDA classifiers (CA of 76.4% ± 9.2%). If using both of the most promising features (PA S3 B0L’ and MF S2 B3), the best results were achieved using the SVM Poly3 classifier, yielding the highest CA (78.0% ± 7.1%) and AUC (0.86 ± 0.07). Since the same two most promising features were identified, these results can be readily compared to the results in Table 5. Notably, the addition of EHG records of the ICEHG DS (groups IL, CL, and ICL) to the term class resulted in even slightly better classification performances when using PA S3 B0L’ alone, or when using both features. Furthermore, as shown in Fig 4(d), most non-spontaneous term delivery records also fall “outside” the premature regions outlined by the decision boundary lines. This demonstrates the robustness of the features PA S3 B0L’ and MF S2 B3 for the prediction of premature birth, regardless of the ultimate term delivery mode, and improves the generalizability of our findings. More importantly, these results show, for the first time, that term non-spontaneous delivery modes appear to match the term spontaneous delivery mode, and do not significantly impact the predictability of premature birth using the EHG.