Artificial intelligence for the detection of preterm labor

Electrohysterography (EHG) is a promising noninvasive measurement of uterine electrical activity, likely able to detect myometrial cell transformation in preparation for (premature) labor. Identifying useful predictors for labor from EHG signals requires automated analysis. Machine learning techniques are a viable solution to this challenge. In this project, a partnership is established between Bloomlife, manufacturing EHG sensors, and Amsterdam UMC and VU University adding clinical knowledge and understanding of signal processing through artificial intelligence. Signals measured through EHG are analyzed in a model to assess their capacity to predict premature labor based on historical cases. Standard clinical measurements are added to the model, in this way aiming for the refining of the current premature birth prediction. 

Premature birth is a significant societal problem. Yet, the pathophysiology of spontaneous premature birth remains unclear. Clinicians assess the risk of actual premature labor using ultrasound, laboratory tests and tocography (quantification of the number of contractions in time). If assessed to be at risk, hospital admission often combined with pharmacological treatment follows to prepare the baby for a premature birth and offer it the best possible start given the circumstances. However, more than 50% of those initially suspected of imminent prematurely delivery stays pregnant and delivers at term. Admission to the hospital for possible premature labor accounts for a huge burden on society, not only because of the costs of the hospital admission but also for the emotional and physical impact on the woman. In order to decrease impact of false premature labor estimations, we need to aim for new techniques to increase accuracy of the current estimation. 

To develop AI models for predicting preterm birth using EHG data, we initiated a clinical study involving women admitted to the hospital with symptoms of threatened preterm birth. Inclusion criteria were based on clinical signs such as uterine contractions, vaginal bleeding, or prelabor rupture of membranes. For each participant, a one-hour EHG recording was obtained upon admission, alongside the collection of standard clinical measurements known to be associated with preterm birth risk. The primary outcome was defined as delivery within two weeks following the EHG recording. After completing data collection, we used these inputs to train and evaluate an AI model aimed at improving the accuracy of short-term preterm birth prediction. The full clinical study was conducted over a period of approximately 3.5 years. 

The main deliverables of this project include the creation of a curated dataset of EHG recordings from a high-risk population, and the development of an AI model capable of predicting preterm birth within two weeks following the recording. This model incorporates both EHG data and standard clinical measurements to enhance predictive accuracy. In addition to these core results, several complementary studies were conducted. One such study evaluated the performance of AI models trained on EHG data from low-risk populations. The findings indicated that models developed on predominantly low-risk cohorts may not generalize effectively to real-world clinical settings, reinforcing the rationale behind the Cocoon study’s focus on high-risk hospitalized patients. 

Furthermore, two studies explored the application of survival models in preterm birth prediction. These included an external validation of the QUiPP app and the development of AI-based survival models aimed at improving generalizability across populations. Both studies utilized three distinct datasets representing at-risk patients across Europe (Netherlands, Belgium, and a broader European cohort). While the Dutch dataset was collected in-house, the others were obtained through international collaborations with clinical partners specializing in gynecology and preterm birth. The QUiPP validation study has been peer-reviewed and published in Ultrasound in Obstetrics & Gynecology, supporting its safe clinical use. The AI-based survival models demonstrated strong generalizability across populations and contributed to the advancement of AI methodologies in maternal health. This work was presented at the 23rd Internati