Predicting gestational age at birth in the context of preterm birth from multi-modal fetal MRI

Introduction

Preterm birth is defined as a live birth before 37 completed weeks of gestation [1]. It is estimated that every year 13.4 million babies are born prematurely, corresponding to a global preterm rate of around 9.9% [2]. Prematurity is the leading cause of mortality among children under 5 years accounting for 17.7% of the 5.3 million yearly deaths in this age group [3]. Complications associated with preterm birth are also the leading cause of neonatal mortality, accounting for 36% of these deaths [3]. The chances of survival of preterm babies are directly related to their gestational age (GA) at birth, with survival chances increasing from less than 18% for babies born at 22 weeks to over 95% for babies born at 29 weeks or later [4–6]. Despite advances in perinatal and neonatal care [4–9] survival critically depends on every additional week in-utero.

A continuous rise in survival rate has not translated into a decrease of the short- or long-term morbidity associated with preterm birth [8–10]. Short-term outcomes of premature birth include infections, bronchopulmonary dysplasia, retinopathy, necrotising enterocolitis, and brain disorders [11]. Long-term consequences include an increased risk of neuropsychiatric disorders such as psychosis, neurodevelopmental disabilities such as cerebral palsy and neuromotor dysfunction, adverse sensory outcomes such as hearing and visual impairment, as well as disabilities encompassing learning, cognition, and behaviour [10, 12, 13]. Similar to mortality rates, the incidence and severity of short- and long-term consequences of preterm birth are inversely related to GA at birth [11, 14, 15]. GA at birth is also correlated to social aspects later in life such as income and education level [15].

Reducing the incidence of preterm birth and the impact of its consequences would not only alleviate the burden on individual patients and their families, but also on entire healthcare systems, since the lifetime cost of preterm births in the USA (in 2016) was estimated to be $25.2 billion [16]. Unsurprisingly, a review of the literature on the economic consequences of preterm birth found a prevailing inverse relation between economic costs and GA at birth, regardless of methodology, date, or country of publication [17].

Preterm birth is classified into three subcategories: extremely preterm (less than 28 weeks), very preterm (28 to 32 weeks), and late preterm (32 to 37 weeks) [1], with further categorisation by clinical presentation: medically induced (or iatrogenic) and spontaneous [18]. While maternal and fetal indicators for iatrogenic preterm birth are well characterised and include conditions such as pre-eclampisa and fetal growth restriction (associated with 30.1% of cases) [19, 20], the aetiologies underlying spontaneous preterm birth are complex, varied, and poorly understood [21]. Causes include - but are not restricted to-infection or inflammation, vascular disease (leading to uterine ischaemia), uterine overdistention, and cervical injury. The latter can be a consequence of LLETZ procedures, cervical cone biopsies for abnormal smear tests, and injuries resulting from emergency C-sections in previous pregnancies [19, 22]. However, definitive causes are registered for only 50% [23, 24] of cases. As such, spontaneous preterm birth should more broadly be considered a syndrome resulting from multiple intricate causes [19, 25].

Despite this complexity, several risk factors have been identified [19, 21, 26] (see Table 1) and are useful, both to provide insights and to help identify at-risk women. The wide variety of factors thereby matches the aetiological complexity of preterm birth. Even within the same clinical subtype, some factors can have opposite effects. For example, low maternal body mass index (BMI) is a risk factor for fetal growth restriction but protective against preeclampsia, whereas these roles are reversed for maternal obesity [27].

Currently there are three leading indicators used in clinical practice to identify women at high risk. The strongest predictor is a history of previous preterm birth or cervical surgery or injury (32% chance of recurrent preterm birth) [19, 28]. The other two biomarkers are mid-trimester cervical length below 25mm [28, 29], measured via vaginal ultrasonography; and the presence of more than 50ng/mL fetal fibronectin, a glycoprotein usually absent in cervicovaginal fluid from 18 weeks of gestation and an indicator of choriodecidual disruption. The absence of any of these factors suggests the likelihood of delivering within the following 7 days is only around 1% [19, 28]. These factors have been combined within clinical practice to improve their predictive capabilities [30, 31]. In other analyses, the combination of these predictors also reduced the average cost of high-risk pregnancies. [32, 33].

Another modality with good potential to investigate preterm birth is Magnetic Resonance Imaging (MRI). Fetal MRI is used as a complementary modality to the commonly used ultrasound screening due to its higher resolution, operator independence and suitability for use on women with a higher BMI. It is also non-invasive with no evidence indicating any risk to the fetus or mother [34–36]. Another key advantage of MRI is that it offers multiple complementary contrasts that can support comprehensive functional evaluation of fetal and maternal tissues [37]. Available contrasts include T2-weighted anatomical imaging, T2* relaxometry (which provides an indirect measure of oxygenation [38]), diffusion MRI (which can quantify alterations in tissue microstructure [39–42]), flow measurements and T2 relaxometry. Past studies have largely focused on individual organs such as investigating changes to lung [43], thymus volumes [44], or assessing placental microstructure by measuring T2* and ADC values [45]. One study measured umbilical vein T2 values as a potential marker of intrauterine growth restriction [46]).

While predictive machine learning (ML) models have enjoyed an ever-increasing popularity, preterm birth has only been addressed as a classification problem. Models based on electronic health records, uterine electromyography and transvaginal ultrasound [47] reported accuracies of approximately 0.77 [48–50], while studies based on electrohysterography reported values above 0.94 [51–53], with the latter, however, only including records of women with recorded contractile activity [53, 54].

Machine learning applied to structural MR measurements has been successful at predicting GA at the time of the scan during pregnancy. For example, Convolutional Neural Networks trained on fetal brain MRI have been able to outperform current clinical methods to estimate GA at the time of scan [55, 56]. A different study managed to obtain a mean absolute error of 6.1 days by developing bespoke features from 3D ultrasound and using a regression forest for prediction [57].

For this work, a stacking approach was chosen to predict GA at the time of birth. Stacking is an ensembling technique that consists of combining the predictions of individual base models by training a meta-model [58]. Stacking was introduced by Wolpert in 1992 to improve the predictions and generalisability of individual classification models [59]. In 1996 Breiman [60] showed that stacking was also suitable for regression problems, while in 1999 Ting and Witten [61] generalised the technique further by stacking three different types of base models and exploring different meta-models than the ones used in previous work. Ensemble methods such as stacking have the statistical advantage of reducing the risk of overfitting to the training data by taking into account the predictions of all the base models, as well as the representational advantage of expanding the space of available models by combining the base models into meta-models [62].

In recent years, stacking has been successful at various tasks such as genomic prediction [63], protein interactions prediction [64], or prostate cancer detection [65]. These works take advantage of more recent ML learning models, e.g., Yi et. al. [64] use Support Vector Machines and XGBoost models as part of their base models, while Wang et. al. explore using a Random Forest as their meta-model [65].

The present study combines a uniquely rich MR data acquisition including both anatomical and functional scans of multiple fetal organs, and multimodal MRI of the placenta, with a ML pipeline based on stacking. To the best of our knowledge, this is the first work to leverage the advantages of stacking methods together with a comprehensive multi-modal data set to predict GA at birth.

Methods

This section contains a detailed outline of the development of the ML pipeline introduced in this work. The pipeline was designed to address the challenges presented by the data. These include: a large number of derived features relative to the number of training examples, data imbalance, and missing data. These problems were addressed through feature selection, balanced training, and feature imputation. Throughout the development of the pipeline, different design options were investigated including changing data threshold for imputation, and models for feature selection and regression. The end product is a meta-model where predictions are stacked to obtain a final predicted GA at birth. The last subsection describes an ablation study, which investigates the impact of each component. Fig 1 illustrates the workflow of the project. The reader is invited to refer to it repeatedly to complement the description that follows.

• Download figure
• Open in new tab

Fig 1. Schematic representation of the project pipeline.

The boxes in a darker shade denote steps of the pipeline where different design options were explored. The top part of the figure shows the flow of the pipeline with fixed designed options, while the bottom part explicitly indicates the different design choices available for each step.

Summary of Derived Features

In addition to imaging derived features, demographics, obstetric and medical history of the patients (including previous pregnancies, miscarriages and preterm births) were recorded as well as any relevant information from the current pregnancy such as diagnosis of pre-eclampsia, gestational diabetes, fetal growth restriction or any other fetal or maternal pathology. Finally, the outcomes of the pregnancy were obtained, including gestational age at birth, birth weight-centile and any occurrence of major complications. Collectively the full set of features used by our models was summarised as follows (see S1 Table for more details):

1. Clinical variables: e.g. number of previous preterm deliveries, and maternal body mass index.

2. Structural MRI metrics: describing sizes of structures e.g. volumes of different brain regions or bi-parietal diameter of the foetal head; these were extracted from both the anatomical and functional scans

3. Functional MRI metrics: statistics derived from T2* distributions of the placenta, brain and lungs.

4. Ultrasound metrics: from both anomaly and growth ultrasounds - manually extracted by a trained sonographer e.g. the fetal head circumference, femur length.

Results

Data Exploration

Table 2 shows key demographics, clinical information, and outcomes, divided into preterm and control cohorts. Specifically, the data set consisted of 176 control cases and 67 preterm cases. The distribution of the data according to the four temporal categories was 72.4% term, 14.8% late preterm, 7% very preterm, and 5.8% extremely preterm. Fig 2 shows the distribution of the five continuous features and outcomes included in Table 2, namely GA at scan, maternal BMI at scan, maternal age, GA at birth, and birth weight centile. The pairwise relationship between these is also plotted. For a statistical summary of the data set see S2 Table.

• Download figure
• Open in new tab

Fig 2. Data exploration.

Distribution and pairwise relationship between GA at scan, maternal BMI at scan, maternal age, GA at birth, and birth weight centile in the data set.

View this table:

• View inline
• View popup
• Download powerpoint

Table 2. Key demographics, clinical information, and outcomes of participants.

Meta-model

The best performing meta-model achieved an R² score of 0.45 and a mean absolute eror (MAE) of 2.55 weeks on the validation set. This performance was achieved using the following settings in the pipeline. First, features with more than 50% missing values were discarded. Then, features with 50% or less missing values were imputed using the MICE algorithm with a KNR as its regression model and a RF was used for feature selection. In order of importance, the selected features were:

1. Cervical length measured from the sagittal plane of MR scan.

2. Mean whole placental T2* value measured from the MR scan.

3. End-diastolic flow measured from the growth ultrasound.

4. T2* brain to placenta ratio measured from the MR scan.

5. Bi-parietal diameter measured from the anomaly ultrasound.

6. Placental T2* kurtosis value measured from the MR scan.

7. Fetal head circumference measured from the anomaly ultrasound.

8. Brain T2* kurtosis value measured from the MR scan.

9. Estimated fetal weight at growth ultrasound.

10. Whole brain T2* volume value measured from the MR scan.

Fig 3 a) shows the mean decrease in impurity corresponding to each of these features. This is the metric used by Random Forests to assign importances to each feature [84].

• Download figure
• Open in new tab

Fig 3. Feature importances and predictions of the meta-model.

a) Feature importance score of the 10 most important features used to train the base models in the pipeline. b) Predictions made by the meta-model 50-KNR-RF on the test set, colourised according to their true preterm temporal category.

After training RF, SVR, and XGBoost models with every non-empty subset of these 10 features, the 14 models with the highest R² score on the validation set were used as input for a RF meta-model. Specifically, the 14 base models that provided the features used by the meta-model were all SVRs, trained with different subsets of the selected features. In what follows, this meta-model will be referred to by abbreviating its components, i.e. 50-KNR-RF.

The metrics used for evaluation on the test set were the R² score and the mean absolute error (MAE) measured in weeks. The cases were labeled as term (≥ 37 weeks) or preterm (< 37 weeks), according to the GA predicted by the meta-model, and accuracy, sensitivity, and specificity were also reported. On the test set 50-KNR-RF achieved an R² score of 0.51 and a MAE of 2.22 weeks. After labeling each subject as term or preterm according to their predicted GA at birth, the meta-model achieved a 0.88 accuracy, 0.86 sensitivity, and 0.89 specificity. The predictions made by 50-KNR-RF on the test set are depicted in Fig. 3 b). It is worth noting that the model only misclassified one of the preterm cases and two term ones.

Discussion

Comprehensive multi-modal fetal data and ML models combine synergistically to predict GA at delivery. The developed pipeline acknowledges and addresses key challenges such as imbalances and missing features in the data set, both of which are common when investigating preterm birth.

As mentioned in the Introduction, preterm birth had so far only been addressed as a classification problem. The pipeline constructed in this work attempts to make more precise predictions by developing a regression model to predict GA at birth. In turn, these predictions can be classified as term or preterm to compare the performance of 50-KNR-RF to other classification models in the literature. Table 4 shows a comparison between 50-KNR-RF and the best performing models obtained by other recent studies. Out of the 5 models, 50-KNR-RF has the highest accuracy and specificity. The only model that achieves a higher sensitivity than 50-KNR-RF is one of the models developed by by Esty et al. [48]. However, the 0.93 sensitivity achieved by that model comes with the trade-off of having an accuracy and specificity lower than 0.73. In contrast, 50-KNR-RF achieved a score greater than 0.85 in every metric.

To the best of our knowledge, the only other study on predicting GA at delivery using ML is the one by Heinsalu et al. [85], where they investigated models using a simpler version of the pipeline displayed in this work. Their best performing model achieves an R² score of 0.66 and a MAE of 1.60, but their implementation suffers from data leakage at the imputation stage which makes these results unreliable. Nevertheless, the framework they established is valuable and served as the basis of the present work.

This study provides a proof of concept, but the clinical implementation of a reliable model that could predict the timing of delivery would have important benefits. These include ensuring women are transferred to appropriate neonatal care facilities. A timely transfer helps reduce neonatal mortality and decrease costs [37]. Another crucial example is the targeting of therapies to mitigate the effects of prematurity. Especifically, cortiscosteroids administration can help reduce intra-ventricular haemorraghe and promote lung maturity. The timing of this therapy is highly important, since it works best when administered within a week before delivery, and repeated doses increase the risk of adverse effects such as reduction in birthweight [86].

The results of the ablation study in Table 3 are helpful to understand the importance of every element of the pipeline. In experiments 1) and 2) it can be seen that even though XGBoost has a built-in mechanism to be trained using data sets with missing values, the results achieved with minimal preprocessing steps are among the worst in the study. In experiments 3) and 4) RF, SVR, and XGBoost models were trained using the data set obtained after imputation, without a feature selection selection step, and using upsampling in the case of experiment 4). On average, results in these experiments were better than in experiments 1) and 2) but upsampling did not translate in a significant difference between experiments 3) and 4). Contrastingly, the impact of upsampling in the whole pipeline can be appreciated by comparing the meta-model obtained in experiment 5) and 50-KNR-RF. The only difference between these two experiments is the use of upsampling during training. The sensitivity score of 0.43 obtained by the meta-model in experiment 5) evidences its inadequacy to address data imbalance, while the 0.86 score achieved by 50-KNR-RF shows that upsampling is effective in tackling this problem. Lastly, it can also be observed that 50-KNR-RF generalises well, i.e. it is not overfitting. 50-KNR-RF achieves an R² score of 0.4530 and a MAE of 2.5489 weeks on the validation set and an R² score of 0.5143 and a MAE of 2.2202 weeks on the test set. This is line with the literature, that suggests that ensemble models tend to generalise well [59]. The reason why the SVR model in experiment 6) has lower scores than its counterparts in experiments 3) and 4) in spite of being the best model prior to the ensembling of 50-KNR-RF is because it is not generalising well. It is the best performing model on the validation set, where it achieves an R² score of 0.4878 and a MAE of 2.6418 weeks, but its performance on the test set is significantly worse. Combining the predictions of the 14 best models via 50-KNR-RF helps to solve this problem.

The features selected as the most important are in line with the literature. The importance of cervical length as a predictor in clinical practice is reflected by its consistent use in the models. Placental features obtained from MRI scans were other prominent features, which is in line with the current understanding of the mechanisms leading to iatrogenic preterm birth [45, 87]. The most common clinical indicator, number of previous preterm births, was not a predominant feature. This could be explained by the decision of approaching the problem as a regression instead of a classification one.

While this data set could be considered large given the comprehensive data acquisition, including a fetal MR scan in a cohort of women requiring a high level of medical care, its size is an important limitation for ML methods. The few examples of extremely preterm subjects available during training help explain the poor performance on the test set for this category. Taking into account the performance of 50-KNR-RF in the validation and test sets, the meta-model seems to generalise well. However, the size of the data set makes it hard to predict if the performance of the meta-model would generalise to new data.

Another limitation is the lack of information on the clinical presentation of preterm birth for every patient in the data set. Iatrogenic and spontaneous preterm births have different aetiologies and training separate models for each case could not only yield better predictions, but also help improve the understanding of each clinical presentation by differentiating their most predictive features. Future work will focus on such subgroups and on extracting relevant phenotypes associated with the different types of preterm birth.

Data obtained on a 1.5T and a 0.55T scanner were available for this study. However, these were not included as there is not a straightforward way to extrapolate the signals acquired by scanners with different magnetic field strengths [88]. Future experiments that include these types of data could test the adequacy of the elements of the pipeline, such as the method of internally studentised residuals, to make accurate predictions regardless of field strength.

There are other directions future research can take to expand or improve the methodology presented in this work. The implementation of the models is ready to benefit from larger or more complete data sets. Adding features known for their predictive power, such as quantitative fibronectin measurements, could improve the results. The performance of the meta-model demonstrates that structural and functional information obtained from MRI can be used to predict GA at delivery. An interesting direction is to make predictions directly from the images making use of deep learning techniques, bypassing the problem of missing data and the need of time-consuming measurements made by experienced clinicians. These techniques have been explored to classify preterm and term patients by automatic measurements of cervical length from transvaginal ultrasound [49] and to estimate GA at scan from fetal brain MRI [55, 56].

MRI remains an expensive modality, however, with an increasing use of fetal MRI, the pipeline presented in this study helps to address a question essential for any pregnancy, and can find an application regardless of the indication of the scan. One of the fundamental contributions of this work is that it shows that fetal MR data acquired as part of diagnostic care or research can be used to obtain useful predictions on the GA at delivery, which in turn can inform the care provided to all pregnancies.

Author Contributions

Conceptualization: Diego Fajardo-Rojas, Emma Robinson, Jana Hutter.

Data Curation: Megan Hall, Daniel Cromb, Lisa Story.

Formal Analysis: Diego Fajardo-Rojas.

Investigation: Diego Fajardo-Rojas.

Methodology: Diego Fajardo-Rojas, Emma Robinson, Jana Hutter.

Software: Diego Fajardo-Rojas.

Supervision: Mary A. Rutherford, Lisa Story, Emma Robinson, Jana Hutter.

Validation: Diego Fajardo-Rojas.

Visualization: Diego Fajardo-Rojas.

Writing – Original Draft Preparation: Diego Fajardo-Rojas.

Writing – Review & Editing: Emma Robinson, Jana Hutter.

Supporting information

S1 Fig. Details of First Preprocessing Steps. Number of subjects kept after each initial preprocessing step.

• Download figure
• Open in new tab

S1 Algorithm. Pseudocode of the MICE algorithm.