Cervical Dilation Classification from Electrohysterography and Clinical Features: A Machine-Learning-Derived Digital Biomarker

Introduction: Noninvasive tracking of cervical dilation could reduce discomfort and infection risk from repeated digital examinations during labor. We present an electrohysterography (EHG)-based model framed as a digital biomarker of labor progression that leverages objective physiological signals with minimal clinical context. Methods: We analyzed 72 ten-minute single-channel EHG recordings from low-risk labor cases, yielding 648 segments of 120 s. Signals were filtered into three sub-bands. Twenty-one linear and nonlinear EHG descriptors were combined with two clinical variables, maternal age and gestational age, and two EHG-derived contraction-count features, namely counts of low (LC) and high (HC) uterine contractions, to form 25 predictors. Segments were labeled as low (1–4 cm), moderate (5–6 cm), or advanced (7–10 cm) dilation. Data were split 70/30 into training (n = 454) and independent test (n = 194) sets. Feature importance was estimated using χ2, ANOVA, and Kruskal-Wallis ranking. Thirty-three classifiers were evaluated using five-fold cross-validation within the training set, with the 10 top-ranked features. Among all models, a bagged tree ensemble achieved the highest macro-averaged F1 score and was therefore selected as the baseline classifier for this study. We then used a “Genetic Algorithm Ensemble Bagged Tree (GA-EBT)” approach, in which a binary-encoded genetic algorithm optimizes the bagged tree classifier’s feature combination using stratified five-fold cross-validation with 50 repetitions on the training set. Results: The best cross-validated model was a bagged tree ensemble. Performance plateaued at 17 predictors (median macro-F1 = 0.898) under progressive inclusion. The GA-EBT identified a four-feature subset – maternal age, gestational age, LC count, and HC count – that achieved F1, recall, precision, specificity, and accuracy of 1.000 on the independent test set for classifying cervical dilation stage (low, moderate, advanced). Conclusion: An EHG-derived digital biomarker combining a minimal set of clinical variables and EHG-derived contraction-count features enables accurate classification of cervical dilation stages from single-channel recordings. This pilot-stage classification approach showed maximal internal and independent test performance and may support real-time, noninvasive intrapartum monitoring while potentially reducing repeated digital examinations.