Abstract
Preterm birth prediction is important in prenatal care; however, it remains a significant challenge due to the complex physiological mechanisms involved. This study aimed to explore the feasibility of phase synchronization of multiple oscillatory components across electrohysterography (EHG) and tocodynamometry (TOCO) signals to identify preterm births using advanced machine-learning techniques. Using an open-access EHG dataset, we first assessed the degree of phase synchronization of five specified frequency ranges from 0.08 to 5.0 Hz in three individual EHG signals by constructing two distinct sets of mean phase coherence: the inclusion or exclusion of TOCO signals. We then employed two machine-learning models, XGBoost and TabNet, to classify preterm and term delivery conditions and analyze the predictive potential of these features. The models' performance was evaluated by considering varying lengths of time windows and the use of overlapping windows. Our results demonstrate the importance of lower-frequency EHG signals and synchronization patterns across the horizontal plane of the abdomen, particularly synchronization between the upper and lower regions of the uterus. Furthermore, we observed a distinctive pattern in the high-frequency band (1.0-2.2 Hz), emphasizing the important role of the lower horizontal regions with other sites in the synchronization process. Interestingly, our findings indicated that TOCO signals, while not substantially enhancing the overall prediction performance, contributed to slightly improved accuracy rates when combined with EHG signals. This study suggests the critical role of EHG signals and their intricate spatiotemporal patterns in predicting preterm birth, providing insights for the development of more accurate and efficient prediction models.