Cycle-phase supervision improves transferable physiological representations from wearable data for women's health
The paper introduces CHARM, a framework that leverages menstrual cycle phase supervision alongside signal reconstruction to learn transferable physiological representations from wearable data, enabling accurate generalization to diverse reproductive states like perimenopause and menopause with significantly fewer labels.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer
The human body runs on a complex internal clock, a system of chemical signals that coordinates everything from sleep to growth. For women, one of the most significant of these rhythms is the menstrual cycle, a monthly ebb and flow driven by hormones that prepare the body for potential pregnancy. While these hormones cannot be seen, they leave a distinct trail of physical footprints. They cause subtle shifts in body temperature, alter how fast the heart beats, and change the rhythm of breathing. In recent years, small devices worn on the finger or wrist have become sensitive enough to track these nightly changes, offering a window into a woman's reproductive health without the need for blood tests or invasive procedures. However, turning these raw numbers into a clear picture of what is happening inside the body has remained a difficult challenge, especially when trying to understand different life stages like pregnancy or the transition into menopause, which often look very different from a standard monthly cycle.
Researchers at Terra Enabling Developers have developed a new way to teach computers how to read these physiological signals, creating a system that understands the underlying hormonal story rather than just memorizing patterns. They built a framework called CHARM, which stands for Cyclic Hormone-Aware Representation of the Menstrual cycle. Instead of training separate computer programs for every different reproductive state—such as pregnancy, menopause, or the use of birth control pills—the team taught a single model to recognize the fundamental rhythm of the menstrual cycle. They did this by showing the computer thousands of nights of wearable data and asking it to do two things at once. First, it had to fill in the blanks of missing data, learning how the different signals like heart rate and temperature usually fit together. Second, and most importantly, it had to learn to distinguish between the two main phases of the cycle: the follicular phase, which happens after a period, and the luteal phase, which occurs after ovulation. By forcing the computer to pay attention to these phases, the model learned to see the hormonal dynamics that drive them, creating a deep understanding of the body's state that could be applied to many different situations.
The results showed that this approach worked remarkably well. When the researchers tested the system on new users, it could predict the start of a menstrual period within three days for nearly eighty percent of cases, and it could identify the day of ovulation with similar accuracy. This was a significant improvement over older methods that relied on simple calendar math or fixed temperature thresholds, which often failed for women with irregular cycles. More importantly, the model proved to be incredibly flexible. Because it learned the core language of the menstrual cycle, it could immediately apply that knowledge to states it had never seen before. Without any additional training, the system successfully identified users who were pregnant, those going through menopause, and those using hormonal birth control. It could even detect specific symptoms like hot flashes or spotting. This ability to transfer knowledge from one situation to another is rare in artificial intelligence and suggests that the model had truly grasped the biological principles at play, rather than just memorizing specific examples.
One of the most striking findings was how much this approach saved on the need for labeled data. In the world of machine learning, teaching a computer to recognize a new condition usually requires thousands of examples where humans have already identified the answer. For conditions like menopause, where data is scarce, this is a major hurdle. The researchers found that their model could achieve high accuracy in detecting menopause using five times fewer labeled examples than traditional methods. This efficiency means that accurate health insights could be available for rare or difficult-to-study conditions much sooner. The system also proved robust; even when the temperature data was removed—a key signal for tracking hormonal changes—the model still performed well, maintaining high accuracy for pregnancy and menopause detection. This suggests that the model had learned to use the relationship between all the signals, not just one, to understand the body's state.
The study also looked at whether the computer was actually learning about hormones, or just guessing. By comparing the model's internal understanding against real hormone measurements from a separate, smaller group of participants, the researchers confirmed that the system was indeed tracking the rise and fall of key hormones like estrogen and progesterone. The model's predictions of these hormone levels were significantly sharper when it had been taught to recognize the cycle phases, proving that the extra step of phase supervision helped the computer see the biological reality more clearly. While the study relied on data from users who already owned premium wearable devices, meaning the group was not perfectly representative of the entire population, the findings offer a powerful new tool for understanding women's health. By teaching machines to listen to the body's natural rhythms, this work provides a unified foundation for tracking reproductive health across the entire lifespan, from the regularity of a monthly cycle to the profound changes of pregnancy and menopause.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.