Use of sleep heart rate patterns to forecast diabetes risk in pregnancy

By Dr. Liji Thomas, MDReviewed by Lauren HardakerJun 23 2025

A new study reveals that tracking heart rate variability through simple home devices can help identify gestational diabetes weeks before standard tests, enabling earlier interventions for healthier pregnancies.

Study: Overnight maternal heart rate variability for early prediction of gestational diabetes mellitus: a retrospective cohort study. Image credit: Rabizo Anatolii/Shutterstock.com

Gestational diabetes mellitus (GDM) refers to glucose intolerance that arises in pregnancy. It must be carefully managed to prevent adverse maternal and fetal outcomes. A recent retrospective study in NPJ Women’s Health describes an artificial intelligence model based on heart rate variability (HRV) that predicts GDM risk.

Introduction

Globally, GDM is among the most typical complications of pregnancy, affecting approximately 15% of pregnancies. It increases the risk of adverse outcomes for both the mother and the baby and pushes up healthcare costs by 34% in maternal care and 49% in neonatal care.

Pregnant women are mostly screened for GDM at 24-28 weeks of pregnancy by an oral glucose tolerance test (OGTT). However, new evidence suggests fetal growth abnormalities appear earlier than this. Fetal abdominal circumference is already increased in older expectant mothers or those with excessive body weight.

Early lifestyle changes, including diet alterations and more physical activity, could prevent GDM. Interventions initiated before 14 weeks of pregnancy improve maternal and fetal outcomes.

Early diagnosis could target women needing such interventions and improve the precision of fetal growth monitoring by ultrasound, reducing the long-term risk.

GDM risk prediction is commonly based on maternal factors, including socioeconomic, family, and obstetric history of diabetes or GDM. These factors are easy to obtain but relatively inaccurate in their predictive performance. Conversely, fasting blood glucose and glycated hemoglobin (HbA1c) improve accuracy but rely on invasive sampling.

Newer noninvasive methods are needed to improve predictive accuracy. The current study sought to model baseline risk factors used in current clinical practice using machine learning algorithms and evaluate the effect of adding HRV characteristics on the model's predictive performance.

What is HRV?

HRV refers to the natural differences in the intervals between successive heartbeats, the inter-beat intervals. Wearable or conveniently attached devices measure HRV, ensuring its accessibility in pregnancy, though cost and user literacy may limit widespread adoption.

The mother’s blood volume increases by 30-40% during pregnancy, and the mean heart rate increases by 10-15% compared to the non-pregnant state. This contributes to HRV changes.

HRV also correlates with autonomic nervous function, regulating involuntary body processes like heart rate and digestion. Disruptions in the system, particularly increased sympathetic nervous activity, can lead to metabolic changes associated with the metabolic syndrome. This cluster of conditions, like high blood pressure, elevated blood sugar, excess abdominal fat, and abnormal cholesterol, often occurs together and all raise the risk of type 2 diabetes and heart disease. In pregnancy, sympathetic overactivity may prompt beta cells in the pancreas to change how they release insulin and increase the body’s resistance to it, a key driver of GDM.

There is no HRV-based predictive model for GDM in early pregnancy at present. The current study presents a machine learning model to explore the predictive value of HRV for GDM, in isolation or combined with the baseline factors for GDM risk assessment provided by the National Institutes of Health (NIH). The NIH guideline uses eight demographic and health history questions, increasing risk grades from zero to six. Only seven of the eight NIH factors were available in this study.

About the study

The study drew data from 2,748 nulliparous American women participating in the nuMoM2b database. All had standardized sleep tests conducted at home between 6 and 15 weeks of pregnancy, and GDM testing was done at 24-28 weeks.

Researchers used the NIH guidelines and three machine learning models to analyze heart rate variability (HRV) in extremely granular detail. The NIH guidelines were analyzed first to assess their performance, and then the other models were implemented.

Study findings

GDM diagnosis was more common among older and heavier mothers and was associated with higher blood pressure in early pregnancy. They were more likely to have a family history of diabetes, and to have Asian or Other ethnicities (not White, Black, Hispanic, American Indian, Native Hawaiian, or multiracial).

The NIH guideline performed poorly, with the area under the curve (AUC) being 63% for GDM prediction. A score of six reflected 10% GDM prevalence. In contrast, with a zero-risk level, the prevalence was <0.5%. This approach was associated with many false-positives as two-thirds of women in this sample were at high-risk because of a single risk factor, but with similar GDM prevalence.

The first machine learning model, using only baseline risk factors, performed better with an AUC of 69%, 6% higher than the NIH guidelines, even though both used the same risk factors. This shows the advantage of automatic feature weighting, where each factor is weighted differently depending on its importance, as judged by available data. However, such models may not be generalizable if trained on small samples.

A model that used only HRV characteristics had an AUC of 65%. In contrast, a combined model using NIH-derived baseline risk factors and HRV performed better, with an AUC of 0.73.

All three machine learning models outperformed the NIH risk assessment protocol. The combined model improved accuracy by 10% to 15% over the baseline model in younger mothers with lower body weight. Accuracy gains were lower for obese or older mothers, among whom GDM is more prevalent.

On repeated testing to ensure consistency of results, the combined model performed approximately 100% better than the HRV model in terms of precision-recall area under the curve (AUPRC), and 27% better than the baseline model. Using HRV features with clinical characteristics such as age or body weight could help identify GDM risk in different subgroups, improving the accuracy by up to 15%.

Most HRV characteristics differ in GDM vs non-GDM pregnancies, with the mean overnight heart rate being the strongest predictor of higher GDM risk. A higher mean heart rate suggests sympathetic overactivity with vagal withdrawal, a pattern characteristic of both autonomic imbalance and disrupted metabolism. Autonomic abnormalities thus appear to be very early predictors of GDM. In later pregnancy, these differences are clouded by changes caused by fetal growth and maternal hormonal alterations.

HRV metrics may also vary with gestational age, meaning the measurement timing is essential when evaluating their predictive value.

The advantage of this method is that HRV assessment requires only a discreet monitoring device, wearable or integrated into a watch, bracelet, camera, radar, or mattress. However, the latter options remain largely theoretical in this context. This discreet, continuous monitoring is possible for several nights or weeks, potentially enhancing predictive power.

Importantly, the OGTT tests used to diagnose GDM varied across participants, including 50 g non-fasting, 75 g two-hour, and 100 g three-hour fasting tests, which may affect the consistency of the diagnostic standard. The study's calibration curves also showed that the combined model reliably predicted GDM probabilities up to 15%, confirming good model calibration.

If validated, such noninvasive methods of GDM screening could prompt early pregnancy screening, as is recommended by multiple health organizations. However, issues of accessibility, affordability, and validation in diverse populations remain important challenges.

In addition to the logistic regression model, the research tested an ensemble approach combining logistic regression, support vector machines, and random forests. However, this ensemble did not significantly outperform the simpler model and added computational complexity.

Conclusions

“The findings suggest that overnight maternal HRV characteristics can be used as early predictors of GDM.” This is the first model to use HRV in GDM risk assessment. It combines it with traditional risk factors for better predictive value and uses noninvasive tools. This allows it to be done at home, increasing its utility. While promising, further work is needed to address limitations such as generalizability, device access, and diagnostic consistency across populations.

Download your PDF copy now!