(275) Automation in Uterine Contraction Recognition: A Preliminary Step Toward Predicting Spontaneous Preterm Birth

Spontaneous preterm birth (sPTB) remains a major contributor to neonatal morbidity and mortality. Uterine contractions are a known antecedent, but their patterns and physiological significance remain poorly quantified in clinical practice. Tocometry provides a non-invasive measure of uterine activity that are widely collected but the interpretation is subjective. We aim to develop a preliminary machine learning (ML) framework that characterizes uterine contraction patterns in patients presenting with preterm contractions, as a preliminary step toward predictive modeling of sPTB.

Study Design:

We conducted a retrospective case-control study of patients with preterm contractions between 2016 and 2024. A total of 2,925 patients were screened to identify three groups: (1) those with spontaneous preterm birth (sPTB), (2) those with preterm contractions without sPTB, and (3) those without preterm contractions. Cases were included if they had intrauterine pressure recordings between 10 minutes and 2.75 hours. In total, 310 cases were included: 114 with sPTB and 196 with resolved preterm contractions. We extracted quantitative features from the tocometer time series, including the median interval between contractions, median amplitude, and proportion of waveform segments showing increasing pressure trends. These features were analyzed using logistic regression with 10-fold cross-validation to explore their discriminatory potential between sPTB and non-sPTB cases.

Results:

The model correctly identified 57 of 114 sPTB cases and 112 of 196 non-sPTB cases; however, the overall classification performance was not significant (χ², p=0.27). Representative tracings illustrated in Figure.

Conclusion:

This study demonstrates the feasibility of algorithmic analysis of uterine contraction patterns using tocometer waveforms. While not predictive in isolation, these waveform features represent a promising step toward developing more robust models for sPTB risk stratification. Future studies leveraging larger multicenter datasets and incorporating multimodal features are warranted.