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

Technical Summary: Predicting Gestational Age at Birth from Multi-modal Fetal MRI

Problem Statement
Preterm birth (delivery before 37 weeks) is a leading cause of neonatal mortality and long-term morbidity. While survival rates have improved, the incidence of complications remains high, and the etiology of spontaneous preterm birth is complex and poorly understood. Current clinical prediction relies heavily on history (e.g., prior preterm birth) and specific biomarkers (cervical length, fetal fibronectin), but these often lack sensitivity. Furthermore, existing machine learning (ML) approaches for preterm birth have primarily treated the problem as a binary classification task using electronic health records or ultrasound, rather than predicting the precise gestational age (GA) at birth. There is a need to leverage comprehensive multi-modal fetal MRI data to predict GA at birth, which could inform antenatal planning and therapeutic timing.

Methodology
The authors developed a bespoke machine learning pipeline to predict GA at birth using data from 426 singleton pregnancies (333 control/term, 93 preterm). The dataset included clinical records, manually extracted ultrasound metrics, and multi-modal fetal MRI data (anatomical T2-weighted and functional T2* relaxometry).

The pipeline addressed three primary challenges: high dimensionality relative to sample size, class imbalance, and missing data. The workflow included:

1. Data Preprocessing:
   • Deconfounding: To prevent data leakage and isolate subtle signatures from the dominant effect of gestational age at scan, GA was linearly regressed from all features using internally studentised residuals.
   • Imputation: Missing data (up to 50% threshold) was handled using Multivariate Imputation by Chained Equations (MICE), comparing Weighted K-Nearest Neighbours (KNR) and Random Forests (RF) as regression models.
   • Normalization: Features were normalized (mean 0, std 1).
2. Feature Selection: To mitigate overfitting, 10 features were selected using either Linear Regression or Random Forests.
3. Model Training (Stacking):
   • Base Models: A diverse ensemble of Random Forests (RF), Support Vector Regression (SVR), and XGBoost was trained on every non-empty subset of the 10 selected features.
   • Meta-Model: The predictions of the top-performing base models (selected via validation set $R^2$) were combined using a meta-model (either Linear Regression or RF).
   • Class Imbalance: The pipeline explored upsampling preterm cases in the training set to address the skewed distribution.
4. Validation: Performance was evaluated using stratified 10-fold cross-validation. An ablation study was conducted on the best-performing fold to isolate the contribution of each pipeline component (deconfounding, imputation, upsampling, feature selection, and stacking).

Key Results
The best-performing configuration (denoted 50-KNR-RF) utilized a 50% missing data threshold, KNR for imputation, RF for feature selection, and an RF meta-model.

• Regression Performance: The pipeline achieved a mean $R^2$ score of 0.13 and a Mean Absolute Error (MAE) of 2.74 weeks across folds.
• Classification Performance: When predictions were thresholded at 37 weeks to classify term vs. preterm, the model achieved 0.77 accuracy, 0.59 sensitivity, and 0.82 specificity.
• Feature Importance: The most consistently selected features included cervical length, mean placental T2, placental T2 lacunarity, and bi-parietal diameter.
• Ablation Findings:
  • Upsampling significantly improved sensitivity (from <0.6 to >0.66) but increased MAE in individual models; however, the final stacked meta-model achieved a balance of high accuracy and sensitivity without the trade-off seen in single models.
  • Deconfounding did not improve performance for XGBoost in this specific context compared to raw data, likely due to XGBoost's inherent robustness.
  • The full pipeline (stacking + upsampling + feature selection) outperformed all ablated versions.
• Subgroup Analysis: Performance was consistent for nulliparous and multiparous women but degraded in women with a history of preterm birth (MAE 3.18 weeks, Accuracy 0.66), likely due to the small sample size (n=32) in this subgroup.

Significance and Claims
The paper positions this work as a proof of concept demonstrating that multi-modal fetal MRI, combined with stacking-based machine learning, can predict gestational age at birth.

• Novelty: To the authors' knowledge, this is the first study to address preterm birth prediction as a regression problem using comprehensive multi-modal MRI data and a stacking ensemble approach. Previous literature largely treated this as a classification problem using different data modalities.
• Clinical Utility: The authors state that the current performance (MAE ~2.7 weeks) is insufficient for guiding narrow timing interventions (e.g., corticosteroid administration, which requires a <1-week window) but is suitable for broader antenatal planning, such as ensuring transfer to appropriate neonatal care facilities.
• Limitations: The authors acknowledge that the model explains only a small proportion of the variance ($R^2$ 0.13), likely due to the small cohort size, the heterogeneity of preterm birth etiologies (spontaneous vs. iatrogenic), and the lack of longitudinal feature data. They note that the model struggles with class imbalance and generalization to specific high-risk subgroups.
• Future Work: The authors propose increasing the cohort size for finer stratification, separating models for iatrogenic and spontaneous preterm birth, and potentially moving toward deep learning approaches that predict directly from images to bypass manual feature extraction and missing data issues.

The paper concludes that while the model is not yet a standalone diagnostic tool, it validates the potential of fetal MRI data to provide useful predictions for pregnancy management.