Identifying high-risk pre-term pregnancies using the fetal heart rate and machine learning

Introduction: Fetal heart rate (FHR) monitoring is one of the commonest and most affordable tests performed during pregnancy worldwide. It is critical for evaluating the health status of the baby, providing real-time insights into the physiology of the fetus. While the relationship between patterns in these signals and adverse pregnancy outcomes is well-established, human identification of these complex patterns remains sub-optimal, with experts often failing to recognise babies at high-risk of outcomes such as asphyxia, growth restriction and stillbirth. These outcomes are especially relevant in low- and middle-income countries where an estimated 98% of perinatal deaths occur. Pre-term birth complications are also the leading cause of death in children <5 years of age, 75% of which can be prevented. While advances have been made in developing low-cost digital solutions for antenatal fetal monitoring, there is still substantial progress to be made in developing tools for the identification of high-risk, adverse outcome pre-term pregnancies using these FHR systems. In this study, we have developed the first machine learning algorithm for the identification of high-risk preterm pregnancies with associated adverse outcomes using fetal heart rate monitoring. Methods: We sourced antepartum fetal heart rate traces from high-risk, pre-term pregnancies that were assigned at least one of ten adverse conditions. These were matched with normal pregnancies delivered at term. Using an automated, clinically-validated algorithm, seven distinct fetal heart rate patterns were extracted from each trace, subsequently filtered for outliers and normalized. The data were split into 80% for model development and 20% for validation. Six machine learning algorithms were trained using k-fold cross-validation to identify each trace as either normal or high-risk preterm. The best-performing algorithm was further evaluated using the validation dataset based on metrics including the AUC, sensitivity, and specificity at three distinct classification thresholds. Additional assessments included decision curve analysis and gestational age-specific and outcome-specific performance evaluations. Results: We analysed antepartum fetal heart rate recordings from 4,867 high-risk, pre-term pregnancies with adverse outcomes and 4,014 normal pregnancies. Feature extraction and preprocessing revealed significant differences between the groups (p<0.001). The random forest classifier was the most effective model, achieving an AUC of 0.88 (95% CI 0.87-0.88). When evaluating specific adverse outcomes, the median AUC was 0.85 (IQR 0.81-0.89) and the model consistently exceeded an AUC of 0.80 across all gestational ages. The model's robustness was confirmed on the validation dataset with an AUC of 0.88 (95% CI 0.86-0.90) and a Brier score of 0.14. Decision curve analysis showed the model surpassed both the treat-none and treat-all strategies over most probability thresholds (0.11-1.0). Performance metrics when using the Youden index were as follows: sensitivity 76.2% (95% CI 72.6-80.5%), specificity 87.5% (95% CI 83.3-91.0), F1 score 81.7 (95% CI 79.6-83.9), and Cohen's kappa 62.8 (95% CI 59.6-66.4), indicating high discriminative ability between pregnancy outcomes. Conclusions: Our study successfully demonstrated machine learning algorithms are capable of identifying high-risk preterm pregnancies with associated adverse outcomes through fetal heart rate monitoring. These findings demonstrate the potential of machine learning in enhancing the accuracy and effectiveness of antenatal fetal monitoring, particularly for high-risk cases where timely intervention is crucial. This algorithm could substantially improve pregnancy outcome prediction and consequently, maternal and neonatal care, especially in low- to middle-income countries where the burden of adverse outcomes is high.