Exploratory Non-invasive Estimation of Cervical Dilation from Electrohysterography and Maternal Data: A Machine-Learning Proof-of-Concept Study

Abstract Cervical dilation during labor is commonly assessed through repeated vaginal examinations, which can be uncomfortable, depend on the observer, and are not suitable for continuous monitoring. This pilot study explored whether machine-learning regression could retrospectively estimate cervical dilation using non-invasive abdominal electrohysterography (EHG) recordings and maternal information. Seventy-two 10-min EHG recordings were collected during labor at two Mexican hospitals. Twenty-one EHG features were calculated from nine overlapping 120-s windows and summarized using eight temporal descriptors, generating 168 temporal EHG features. The recording-level low-contraction (LC) and high-contraction (HC) counts, along with maternal and gestational age, were subsequently appended, yielding a 172-feature vector per recording. After removing five recordings with sparse data outside the 4–8 cm dilation range, 67 records remained—50 for model development and 17 for internal testing. In the development phase, feature ranking, screening of 28 regression models, and genetic-algorithm optimization were performed. The genetic algorithm selected five EHG-derived temporal features for a squared-exponential Gaussian process regression model. During fivefold cross-validation within the development set, the optimized model achieved an RMSE of 0.97 cm, r = 0.67, and R 2 = 0.45. On the held-out internal test set, the RMSE was 1.45 cm, with r = 0.23 and R 2 = − 0.31. The model showed moderate cross-validated performance during development but limited generalization in the held-out internal test set. These findings are exploratory and do not establish performance for dystocia screening. Larger prospective studies are needed before clinical use can be considered.