Enhanced accuracy in cervical dilation predictions with multi-parameter models

The risk of prediction errors in cervical dilation models is reduced by incorporating multiple clinically relevant parameters vs time alone, according to a recent study published in the American Journal of Obstetrics & Gynecology.

Intrapartum cesarean delivery (CD) is often indicated by a failure to progress in labor, which is determined by changes in cervical dilation and station over time. Future changes in dilation may be estimated using labor curves, making them relevant when providing guidelines.

Labor induction has become more common in clinical practice, making it relevant in models of labor progress. However, other relevant factors such as membrane status, effacement, and epidural anesthesia have not been considered when modeling labor curves.

To compare the accuracy of labor curves based on time alone vs those based on multiple factors, investigators conducted a longitudinal cohort study. Participants included nulliparous women with live, singleton, vertex-presenting vaginal births at 35 weeks’ gestation or later from June 1, 2017, to June 30, 2021.

A 5-minute Apgar score of 7 or higher and electronic fetal monitoring were also required for inclusion. Exclusion criteria included intensive care unit admission, shoulder dystocia, and descending dilation. Model accuracy was measured using 10-fold cross-validation.

All times were reported as relative time, measured as either negative time or forward time. The time when dilation was 10 cm was referred to as Time0 and was calculated backward. Labor curves were measured for the 20 hours prior to delivery.

Other variables measured included contractions and the cervical dilation. Contractions included the relative times of each cervical examination, membrane rupture, epidural administration, and labor induction.

The course and curve of dilation was determined using a high-order polynomial function. Gaussian processes were selected for machine learning because of their confidence intervals reflecting model uncertainty and ability to model multiple processes with shared predictors. Models with smaller prediction errors were considered more accurate.

There were 8022 births included in the final analysis, with 527 included in the external validation test. Individual trajectories of cervical dilation and fetal station over time had significant variability, highlighting the unpredictability of labor.

When basing models on a single factor using forward-time, 90% of all prediction errors were between the fifth and ninety-fifth percentiles. Prediction errors were notably wide, with expected dilation overestimated in early labor and underestimated in late labor. The mean absolute error (MAE) and root mean squared error (RMSE) were 2.122 cm and 2.504 cm, respectively.

When basing models on multiple factors, significant improvements in accuracy were observed, with a median error closer to 0. A smaller range of prediction errors was also reported, centered around 0 across time. On average, prediction errors were improved by over half when compared to single-factor forward-time models.

Improvements were also found from including multiple factors in machine learning models. The RMSE was 1.126 cm, and the MAE was 0.826 cm. Similar results were reported during external validation.

These results indicated improved prediction errors in cervical dilation models incorporating multiple factors vs those baased on time alone. Investigators concluded future models will be able to account for more explanatory factors, but clinical judgement will always be necessary since no labor curve can account for every relevant factor in every labor.

Reference

Hamilton EF, Zhoroev T, Warrick PA. New labor curves of dilation and station to improve the accuracy of predicting labor progress. American Journal of Obstetrics & Gynecology. 2024;231(1):1-18. doi:10.1016/j.ajog.2024.02.289