Deep Learning-Based Inference of Uterine Contractions from Maternal ECG

Background Preterm birth and stillbirth have multifaceted causes, many of which remain unknown. Factors such as maternal age, sexually transmitted infections, and genetic abnormalities are notably implicated. Globally, there has been minimal improvement in preterm birth rates, leading to substantial economic losses and underscoring the need for innovative approaches.

Objective This study aims to analyze and detect uterine contractions by applying deep learning techniques to physiological data derived from maternal electrocardiograms (ECGs). Leveraging ECGs, which are easily obtainable from wearable devices, allows for effective and convenient monitoring of uterine contractions outside clinical settings. Since uterine contractions are a critical indicator for assessing the risk of preterm birth, our algorithm has the potential to facilitate early identification of high-risk individuals, thereby contributing to timely interventions and improved maternal-fetal outcomes.

Methods Participants meeting all inclusion criteria and none of the exclusion criteria were recruited from patients admitted to or attending outpatient services at Saintpaulia Misao Ladies Hospital in Gifu, Japan, between December 6, 2023, and July 31, 2024. Our deep learning model was developed using maternal ECG data, uterine contraction waveforms obtained from cardiotocograms (CTGs). The collected data were divided into training and evaluation datasets. The AI-generated uterine contraction waveforms from the developed model were compared with ground truth labels obtained from the fetal monitoring devices. This study was reviewed and approved by the Ethics Committee of the Graduate School of Medicine at Gifu University and was conducted as a collaborative research effort with nonat Inc. (the lead research institution) and Saintpaulia Misao Ladies Hospital (joint research institution).

Results Finally, 73 participants took part in this study, and 57 datasets were used for algorithm development and evaluation. Multiple measurements were taken on the same subjects on different days only with their consent. Our analysis demonstrated a strong correlation between the uterine contraction waveforms estimated by the developed model and the ground truth waveforms from CTGs, achieving an average correlation across all evaluation data (Pearson correlation coefficient = 0.53).

Conclusion This is the world’s first trial to detect uterine contractions from maternal ECGs using deep learning algorithms. In this study, we successfully developed a deep learning algorithm capable of accurately inferring uterine contraction waveforms from maternal ECGs. Given that ECGs can be easily obtained using wearable devices, this approach may provide healthcare professionals with objective and precise information on uterine contractions—a crucial indicator of preterm labor—even in resource-limited settings outside hospital environments.