Abstract
Preterm birth is associated with significant mortality and a risk for lifelong morbidity. The complex multifactorial aetiology hampers accurate prediction and thus optimal care. A pipeline consisting of bespoke machine learning methods for data imputation, feature selection, and regression models to predict gestational age (GA) at birth was developed and evaluated from comprehensive multi-modal morphological and functional fetal MRI data from 176 control cases and 67 preterm birth cases. The GA at birth predictions were classified into term and preterm categories and their accuracy, sensitivity, and specificity were reported. An ablation study was performed to further validate the design of the pipeline. The pipeline achieves an R2 score of 0.51 and a mean absolute error of 2.22 weeks. It also achieves a 0.88 accuracy, 0.86 sensitivity, and 0.89 specificity, outperforming previous classification efforts in the literature. The predominant features selected by the pipeline include cervical length and various placental T2* values. The confluence of fast, motion-robust and multi-modal fetal MRI techniques and machine learning prediction allowed the prediction of the gestation at birth. This information is essential for any pregnancy. To the best of our knowledge, preterm birth had only been addressed as a classification problem in the literature. Therefore, this work provides a proof of concept. Future work will increase the cohort size to allow for finer stratification within the preterm birth cohort.
Introduction
Preterm birth (before 37 weeks) is the leading cause of mortality in children under five and often leads to lifelong health complications. Despite its impact, predicting when a high-risk pregnancy will actually result in delivery remains incredibly difficult. Current clinical methods rely on biomarkers like cervical length or fetal fibronectin levels, which are useful but often lack precision. While previous machine learning efforts have attempted to tackle this, they treated it as a binary classification problem (will the birth be preterm or not?). This paper addresses the gap by attempting to predict the specific Gestational Age (GA) at birth, providing clinicians with a more granular timeline for delivery planning and neonatal care.
Methods
The authors built a bespoke machine learning pipeline designed to handle the "messiness" of clinical imaging data. Key technical components include:
- Multi-modal Data Fusion: The model integrates anatomical T2-weighted MRI, functional T2* relaxometry (which measures tissue oxygenation), ultrasound metrics, and clinical history.
- Deconfounding: Since fetal organs change rapidly with age, the researchers used "internally studentized residuals" to remove the dominant effect of the gestational age at the time of the scan. This ensures the model learns signatures of preterm birth rather than just the baby's current stage of development.
- MICE Imputation: To handle missing data points across the 243 cases, they employed Multivariate Imputation by Chained Equations (MICE).
- Stacked Generalization: The core of the prediction is a "stacking" ensemble. They trained multiple base models (Random Forests, Support Vector Regression, and XGBoost) on different subsets of features. A meta-model then "learned" how to best combine the predictions of these base models.
Datasets
The study utilized a dataset of 243 singleton pregnancies aggregated from four clinical studies at King’s College London.
- 176 control cases (term births).
- 67 preterm birth cases (further stratified into late, very, and extremely preterm).
- The data included clinical records, 3T MRI scans (brain, body, and placenta), and longitudinal ultrasound measurements.
Key Insights and Results
The final meta-model (referred to as 50-KNR-RF) demonstrated strong performance, achieving an $R^2$ score of 0.51 and a Mean Absolute Error (MAE) of 2.22 weeks. When the regression results were used to categorize births as term or preterm, the model achieved 88% accuracy, 86% sensitivity, and 89% specificity.
Several key insights emerged:
- Functional Imaging is Key: Among the top 10 features selected by the model, several were derived from placental T2* values. This suggests that placental function and oxygenation levels are critical indicators of pregnancy duration.
- Regression > Classification: By training the model to predict a continuous variable (weeks), the researchers actually achieved better classification accuracy than many previous models designed specifically for classification.
- Cervical Length Remains King: Even with advanced MRI, cervical length measured via MRI remained the single most predictive feature, validating its long-standing use in traditional obstetrics.
Conclusion
The primary limitation is the sample size, particularly for "extremely preterm" cases (less than 28 weeks). Future work aims to increase the cohort size to allow for "finer stratification"—essentially training the model to distinguish between different types of preterm birth (e.g., spontaneous vs. medically induced). Additionally, the researchers hope to move toward "image-to-prediction" models that use deep learning to extract features directly from MRI scans, bypassing the need for manual measurements.
Summary updated