Deciphering the 'due date' - predicting pregnancy duration using physiological data from wearables

Pregnant individuals are often asked about their due date. However, few births will spontaneously occur on this date. Using machine learning and physiological data from a wearable device, our study aims to forecast longer vs. shorter pregnancies relative to the clinical due date.
Deciphering the 'due date' - predicting pregnancy duration using physiological data from wearables

"When is my baby going to be born?"

When providing antenatal care to pregnant individuals and their families, I am often asked to predict when labors will start. Families may desire this information for seemingly low-stakes reasons, for example, setting up a dog sitter or the mother-in-law's flight into town.  However, for others, there can be dire consequences. Some pregnancies should not undergo labor due to obstetric or medical conditions. Many more people are living far from a hospital that can provide obstetric care due to closures of US maternity units in rural areas. This reality makes for a long, painful and/or risky trip during active labor to a hospital. Furthermore, limited (or no) maternal medical leave has serious financial ramifications for many families, leading some mothers to remain on the job until labor kicks in. We sometimes hear of births occurring on airplanes, apartment building lobbies, in Walmart or on the side of the highway. We have learned to live with this uncertainty as maternity care providers, as growing families and communities. 

In response to these questions, I sometimes (rather cheekily) blame my broken crystal ball for our professional failure in precisely predicting when an individual labor will begin and try to provide reassurance. And while signs like uterine contractions, cervical dilation or effacement (thinning) might hint that labor will one day start, in truth, many people walk around for days or weeks with some dilation prior to giving birth and many others begin laboring with no pre-labor signs at all.  

Clinically, the 'due date' is provided to each individual pregnancy, however, despite the authoritative tone of the term, this date is not a personalized estimate of an expected length of gestation.  Rather, this date is precisely 40 weeks and 0 days from the first day of the last menstrual period (if the menstrual period was known and assumes a 28-day cycle with ovulation occurring on day 14). However, due to wide variability in actual menstrual cycle length (anything from 21-36 days is considered “normal”, as well as wide variability in the day of ovulation, roughly anywhere from day 9 to day 24+ and other factors, this due date is incorrect for most women by design. As an example, even among a relatively small (n=125) study of American pregnancies that excluded preterm births or pregnancies with medical conditions, the gestational length varied by 37 days. Rather than being a personal forecast, the 40.0-week date simply reflects an average length of gestation across human populations, with the caveat that term birth may occur at any time between 37 weeks 0 days and 42 weeks (Figure 1). 

Term pregnancy implies fetal readiness for birth

"Term" pregnancy centers on the idea that the fetus, if spontaneously born at this stage, absent other obstetric complications, will transition smoothly to extrauterine life. This transition includes the ability to manage the bare necessities of newborn-hood: breathing, body temperature, blood glucose and sucking/swallowing breastmilk.  This idea is supported by the knowledge that some of the underlying physiology preparing the maternal body/tissue to give birth is triggered when hormonal signals given by the placenta and the fetus occur during maturation of key extrauterine organ systems (i.e., brain, adrenal, lungs). Considering the concept of the due date solely from the lens of 'fetal readiness' for birth, we actually observe great variation in the length of gestation. When someone reports going into labor "5 days late" or "2 weeks early" - we can reframe this idea if we consider that the timing reflects the fetus' readiness to breathe and the mothers' readiness to respond. These interwoven mechanisms usually initiate labor with a mature-enough fetus allowing the mother to give birth and lactate. At term, these births are not early or late- they are right on their own time.

Figure 1: Distribution of labor and birth in human populations and fetal maturation leading to labor onset.

Personalized due dates and managing uncertainty 

The earlier and the later birth occurs from the "term" window, the greater the risk for mothers and babies. Many of these labors/births are influenced by other pathologies (e.g., infection, diabetes, high blood pressure, fetal growth restriction). On the early side (<37 weeks) preterm labor and birth are costly in terms of life-long morbidity and mortality and health system treatment for preterm infants.  When pregnancy extends into the higher end of the range of gestation, we see more problems with larger bodied babies that can have greater difficulty being born vaginally or with tragic postdates stillbirth.  Fear held by parents and care providers alike about earlier and later gestation risks often drives decision-making, additional testing, hospitalization and other interventions in light of any biomarker or test which can prospectively predict the length of gestation.

Many studies have sought the use of ultrasound technology, uterine monitors, vaginal swabs or maternal blood based markers to predict the future onset of labor, either preterm or term. All of these studies are based on the idea that there are changes in physiology occurring in the days/weeks prior a person's perception of labor symptoms that, if detectable, might provide a useful forecast for the natural expected length of that person's pregnancy. 

For all these reasons, improving accuracy and precision of the estimated "due date" could inform the way we plan and manage uncertainty across gestation on the level of individual pregnancy care (e.g., managing work leave, travel, co-morbid conditions, avoiding labor in some high-risk pregnancies etc.).

Using wearable smart ring data to understand changes in physiology and predict longer pregnancies

The present study uses a commercially-available, non-invasive, wearable smart ring technology, which is worn on a preferred finger. The ring gathers data indicating cardiovascular and autonomic nervous system status and change in real-time. These data sync to an application and cloud storage which can be downloaded and analyzed. Informed by animal studies, we  hypothesized that physiological changes in maternal wearable data would  reflect the impending hormonal transition away from pregnancy and into parturition (labor) and would therefore be useful in identifying how much time was left in the pregnancy. 

In our study, 118 participants residing across the United States began wearing the smart ring in the early third trimester and used it throughout the remainder of their pregnancies. They also completed surveys about labor symptoms and other experiences like stress, mood and anxiety. Following birth, participants reported their birth events, timing of labor, use of labor induction, need for Cesarean birth among other outcomes postpartum. 

Of the 118, 31 gave birth prior to the due date after having labor induced, mostly due to medical or obstetric complications or had a Cesarean birth without labor. Another 39 participants began spontaneously laboring before their 40-week due date and gave birth. For the rest of the sample, 48 individuals passed their due dates, with the majority beginning laboring spontaneously (17 required labor induction after the due date).  

Using a boosted random forest, a machine learning approach, the daily averages of 30 metrics gathered by the smart ring were used to train a model. Due to the focus on the importance of the 'due date' we centered the analysis on predicting whether labor would start before the clinical due date of 40.0 weeks or if the pregnancy would naturally extend longer, passing the due date. Those participants whose pregnancy was truncated due to need for induction/delivery were excluded. Importantly, we only included data from study enrollment until 40 weeks' gestation or 4 days prior to labor onset (whichever arrived first). This step was to ensure that we were not measuring physiological changes during the labor process itself to predict the outcome. 

We found that many metrics were correlated with advancing gestation, confirming that pregnancy is not a stagnant physiological state. The smart ring metrics most useful in the models' predictions for begin pregnant after the due date were the changes in skin temperature, metabolic activity, physical activity levels and sleep patterns. Accuracy of the system was reported in terms of the area under the receiver operating curve (AUROC), which is a metric to indicate the accuracy of a classifier.  An AUROC of 0.5 indicates that the classifier is no more accurate than chance or a coin flip while an AUROC of 1.0 represents a perfect classifier.  The boosted random forest presented in this paper had an AUROC of 0.71 (Figure 2), demonstrating a moderate ability to distinguish between the groups (pregnancies likely to pass their due date from those who labored earlier in the gestation). 

Interestingly, no clinical characteristics (e.g., maternal age, number of prior births, body mass index) were useful in predicting the duration of the pregnancy, nor were the frequency of symptoms like contractions, vaginal discharge or back pain.  Finally, use of gestational age alone in the models had an AUROC of 0.58, which is little better than flipping a coin each day of pregnancy as to whether the baby is coming "early" or "late". 

Figure 2: Metrics for evaluating the model performance using smart ring data.

Future directions/implications for personalizing the due date

We are continuing to refine and expand upon this work, finding new patterns within these data that represent changes in maternal / fetal physiology. While an AUROC of 0.71 is not ready for clinical application, it is a good start to further clinical testing. It is certainly better than the current approach (which is to say - "Sorry... I wish I could tell you, let's flip a coin!").  With future iterations and refinement, perhaps one day parents will have a more reliable forecast for how long their baby needs to develop or even an alert for the potential of preterm labor.  The local and global implications of such an advancement in precision obstetric monitoring could be far reaching if it could be made reliable, affordable and scalable.  While advancing this concept challenges long-held obstetric norms and may be seen as risky, a more personalized approach to pregnancy care is long overdue

Please sign in or register for FREE

If you are a registered user on Research Communities by Springer Nature, please sign in

Subscribe to the Topic

Life Sciences > Health Sciences > Clinical Medicine > Gynecology > Obstetrics
Wearable Technology
Technology and Engineering > Biological and Physical Engineering > Biomedical Engineering and Bioengineering > Biomedical Devices and Instrumentation > Wearable Technology
Machine Learning
Mathematics and Computing > Computer Science > Artificial Intelligence > Machine Learning
Health Care
Life Sciences > Health Sciences > Health Care
  • npj Digital Medicine npj Digital Medicine

    An online open-access journal dedicated to publishing research in all aspects of digital medicine, including the clinical application and implementation of digital and mobile technologies, virtual healthcare, and novel applications of artificial intelligence and informatics.

Related Collections

With collections, you can get published faster and increase your visibility.

Clinical applications of AI in mental health care

This joint venture Collection between npj Mental Health Research and npj Digital Medicine highlights how AI can be safely, ethically, & impactfully utilized to advance our understanding of mental illnesses & improve patient care.

Publishing Model: Open Access

Deadline: Jun 22, 2024

Harnessing digital health technologies to tackle climate change and promote human health

This collection invites research on the use of digital health technologies that innovate solutions to improve sustainable health care practice and delivery.

Publishing Model: Open Access

Deadline: Apr 30, 2024