Predicting preterm births from electrohysterogram recordings via deep learning

Abstract

About one in ten babies is born preterm, i.e., before completing 37 weeks of gestation, which can result in permanent neurologic deficit and is a leading cause of child mortality. Although imminent preterm labor can be detected, predicting preterm births more than one week in advance remains elusive. Here, we develop a deep learning method to predict preterm births directly from electrohysterogram (EHG) measurements of pregnant mothers recorded at around 31 weeks of gestation. We developed a prediction model, which includes a recurrent neural network, to predict preterm births using short-time Fourier transforms of EHG recordings and clinical information from two public datasets. We predicted preterm births with an area under the receiver-operating characteristic curve (AUC) of 0.78 (95% confidence interval: 0.76-0.80). Moreover, we found that the spectral patterns of the measurements were more predictive than the temporal patterns, suggesting that preterm births can be predicted from short EHG recordings in an automated process. We show that preterm births can be predicted for pregnant mothers around their 31st week of gestation, prompting beneficial treatments to reduce the incidence of preterm births and improve their outcomes.

Fig 1. Block diagram of the three classification and regression models developed in this work.

The details of these models are provided in Methods. The clinical information model is illustrated in the upper part of the diagram, using shapes with blue outlines. This model uses clinical information, in tabular format, to predict preterm births by using logistic or linear regression, represented as a block with a blue outline and schematic illustrations below it. Preprocessing the clinical information consists of completing missing entries and normalizing the predictors, as described in Methods. The EHG model is illustrated in the lower part of the diagram, using shapes with black outlines. This model uses EHG measurements, represented by an input block with a schematic illustration below it, that are first preprocessed. This preprocessing step includes bandpass filtering (BPF) and downsampling. The preprocessed measurements are used to compute STFTs, illustrated by a block and a schematic representation, that are used as input to the RNN. This network is composed of an input layer, a BiLSTM layer, a fully connected (FC) layer, and an output layer, which are illustrated using light blue shapes with black outlines and enclosed within a dashed light blue outline. The combined model uses clinical information and EHG measurements to predict preterm births and is illustrated in the middle part of the diagram using shapes with red outlines. The dotted black outline represents the cross-validation technique employed, indicating that the operations within are applied separately for each data partition, whereas the operations outside are applied to all the data, independent of the data partition.

Fig 2. Performance of the models for predicting preterm births.

(a) ROC curves for predicting preterm births using the classification models trained with clinical information alone, EHG measurements alone, and clinical information combined with EHG measurements. (b) ROC curves for the same tasks as in (a), but using the regression models instead of the classification models. (a), (b) The performance bound is shown in both panels by a black ROC curve. The greyed area delimited by this bound indicates unattainable performance due to the uncertainty in the ground truth labels. The AUCs of the models are presented with 95% CIs.

Fig 3. Effects of information loss on the prediction of preterm births.

(a) A representative STFT of an EHG recording overlaid with the limits of the frequency bands examined. (b) AUCs obtained using the classification model trained on the various frequency bands. (c) The same STFT as in (a), but with all the columns randomly rearranged. (d) AUCs obtained using the classification model trained on STFTs with varying fractions of columns randomly rearranged. (e), The same STFT as in (a), but where ten minutes of the recording were removed. The colorbar in this panel also corresponds to panels (a) and (c). (f) AUCs obtained using the classification model trained on STFTs with varying durations. (b), (d), (f), The AUCs are presented as black dots with error bars denoting the 95% CIs.