Identifying menstrual metrics as personal health markers: Age trends and individual footprints in temperature across 5674 cycles.

The WAVES algorithm computes (i) daily physiological data through menstrual cycles, here, basal body temperature, in addition to (ii) the estimated ovulation day, here, based on the identification of the date of peak cervical mucus. The algorithm inputs these variables, identifies each different menstrual cycle and its phases, and leverages the data to generate the metrics described in table S1 and illustrated in Fig. 1 . Overall, the following 32 metrics are captured for each cycle, grouped into four sets: Inputs, represented in blue, are the daily temperature measures and cervical mucus code. Within the WAVES algorithm, four sets of metrics are extracted per menstrual cycle (the main metrics within each set are illustrated in the box on the right and are described in detail in the text). Metrics (means and SD for each participant) are returned as outputs, which are represented in orange on the left bottom panel. BBT, basal body temperature. 1) The menstrual cycle temperature levels include five metrics: the mean temperature level of the whole cycle overall and of the follicular and luteal phases separately based on the raw temperature trend. The maximum and minimum values of the whole cycle are calculated as the average of the highest five and lower five temperature values, respectively. 2) The temporal dimensions of the cycle contain seven metrics (moments of the menstrual cycle): the timing of the nadir, corresponding to the number of days after last menses until the minimum of the 3-day smoothed temperature curve is reached; the timing of the maximum, corresponding to the number of days after last ovulation until the maximum of the 3-day smoothed temperature curve is reached (acrophase). In addition to be presented in number of days into the menstrual cycle phases, timing of the minimum and maximum are also presented in angles of a 360° menstrual cycle. In this framework, the first day of menses marks 0°, the estimated ovulation day (based on the cervical mucus marker) marks 180°, and the first day of the next menses marks 360°. Follicular days of the cycle are then recorded from 0° to 180° and luteal days from 180° to 360°. This normalization strategy may enable visualization of some trends that could be masked by the high variability in the number of days in the follicular phase and in cycle duration. Other metrics in this category are the period of the cycle calculated as the number of days from the first day of menses of one cycle to the first day of menses of the next cycle and the duration of the follicular and luteal phases, calculated as the number of days from the first day of menses to estimated day of ovulation (mucus peak) and from ovulation to the first day of the next menses. 3) The variations in the temperature signal within each cycle are captured in 14 metrics: the average slope of the raw curve for the whole cycle and skewness and kirtosis present in the signal. Metrics for the amplitude (peak to trough) of the smoothed curve, smoothness, and SD metrics of the raw curve for the whole cycle and separately within the follicular and luteal phases are calculated. Metrics are also included in this set for how steep are the slopes of temperature decrease from menses to the day of nadir, the increase from the day of nadir to the acrophase, and the decrease from the acrophase to the next menses. 4) The temperature curve shape (six metrics), characterized by how well cosine, square, zigzag, Gaussian, sawtooth, and half-sine curves model the raw menstrual cycle temperature wave, which can capture whether a certain condition alters the overall shape of the wave. The dataset used for this analysis contained 4654 cycles from 671 individuals between 18 and 35 years and 1020 cycles from 110 individuals between 35 and 42 years old. As illustrated in Fig. 2 , there are differences in the menstrual rhythm of temperature between the two age groups. Of the 32 metrics, 27 were found to be significantly different between age groups (table S2). First, the older group presented significantly higher metrics of temperature levels, with higher mean, maximum five values, lowest five values, and the average of both the follicular and luteal phases compared with the younger group. Second, temporal metrics also differed between the two groups, with the older group having a shorter cycle period due to the shortening of the follicular, but not of the luteal phase relative to the younger group. Consequently, the timing of the nadir and the maximum (acrophase) of the curve occurred earlier in the cycle in the older age group. Third, multiple metrics capturing the variations across the cycle differed between the groups. Amplitude was lower in the older than the younger group across the entire cycle and in both cycle phases independently. Fourth, for all of the models of curves tested, the fit qualities were better for the older group. The average temperature across days of the menstrual cycle for the younger age group (18 to 35 years old) is represented as a blue curve and for the older age group (35 to 42 years old) as an orange curve. The SD is represented by a shadow of the same color around each line. Linear mixed models were run to assess whether detectable changes occurred in the menstrual cycle temperature metrics across multiple cycles, as individuals aged. As shown in Tables 1 to 4 , the absolute level of 16 of the 32 menstrual metrics varied significantly with aging within individuals. For example, every year, the duration of the follicular phase shortened (−0.118 day/year), the lowest temperature values (minimum of 5) increased (+0.004°C/year), and the amplitude of the menstrual cycle rhythm decreased (−0.002°C/year) within an individual. While the days of the nadir and acrophase within a cycle were earlier with aging, the cycle angle at which these events occurred did not significantly change with aging. Also, for all of the models of curves tested, the fit qualities were better with aging. Individual linear mixed models were run for each metric predicting age, using an α of 0.05 divided by 32 as the adjusted threshold of significance to account for multiple tests. Individual linear mixed models were run for each metric predicting age using an α of 0.05 divided by 32 as the adjusted threshold of significance to account for multiple tests. Individual linear mixed models were run for each metric predicting age using an α of 0.05 divided by 32 as the adjusted threshold of significance to account for multiple tests. Individual linear mixed models were run for each metric predicting age using an α of 0.05 divided by 32 as the adjusted threshold of significance to account for multiple tests. Linear mixed models were run to assess whether menstrual cycle temperature metrics become increasingly irregular (i.e., more variable) from cycle to cycle, as an individual ages. Tables 1 to 4 show that multiple metrics have a significant change in the SD, reflecting greater variability, with aging. Most changes in variability in the metrics have a positive coefficient, indicating that variability increases with aging. For example, the cycle period became less regular, associated with the luteal phase duration becoming less regular, with aging. Also, eight metrics capturing variations in the cycle, including of the amplitude, average slope, smoothness, and SD of both the luteal and follicular phases showed less regularity with aging. On the other hand, the fitness quality, or coefficient of determination ( r 2 ) of the square wave, zigzag, cosine, and half-sine curves became less variable with aging. The intraclass correlation coefficients (ICCs) were calculated to determine the presence of a personal footprint in the menstrual metrics’ absolute levels and in their regularity, with a higher ICC reflecting greater variability between individuals ( Tables 1 to 4 ). Regarding the absolute level of metrics, the highest ICCs (>0.7) were found in the metrics capturing the level of the menstrual cycle temperature wave such as the mean values of the follicular and luteal phases and the lowest and highest five points of the curve. Moderate to high ICCs (>0.5) were observed in the duration of the cycle and its follicular and luteal phases, followed by the day of the nadir temperature (ICC = 0.46). The other metrics had ICCs ~0.3 and below, including the indicators of variations within the temperature signal of the menstrual cycle (amplitude, SD, and smoothness) and the indicators of goodness of fit for different curves. Regarding the cycle-to-cycle regularity in the metrics, the ICCs were almost all moderate to high. The highest ICC was found in the cycle period (0.76), and moderate to high ICCs (>0.5) were found in all of the metrics capturing temperature levels, wave shape [ R 2 (coefficient of determination) of cosinor, square, zigzag, etc.], temporal metrics (moments of the cycle), and in almost all of the metrics of the variability set.

Here, we presented WAVES, a signal-agnostic algorithm that extracts 32 metrics across four dimensions of menstrual cycle patterns present in time series of physiological data. WAVES was used on a large dataset of daily basal body temperature time series across 5674 menstrual cycles from individuals ranging from 18 to 42 years old. Age was found to be significantly associated with differences in several menstrual metrics, including in their average level and cycle-to-cycle variability. Last, high ICCs, suggesting an individual-specific footprint, were found in several menstrual cycle metrics, highlighting the importance of considering menstrual signals as individualized health markers. The WAVES algorithm was specifically tailored to characterize menstrual cycle patterns in physiological signals. It inputs the menses and ovulation dates, a time series of a signal influenced by the menstrual cycle (e.g., body temperature), and outputs 32 metrics across 4 dimensions: level, temporal structure, variations, and wave shape. The algorithm is available so that other investigators can have access to a free, easy-to-use, and transparent tool to generate menstrual cycle metrics, which can then be tested as potential biomarkers. The metrics have been identified and developed with the aim of capturing and reproducing what has already been identified as relevant aspects to quantify in menstrual cycle rhythms, including the oscillatory shape of the temperature curve, its amplitude, and the temporal dimension of the curve such as the duration of each phase and the acrophase ( 21 , 28 ). In addition, other metrics were added to test other aspects of the menstrual cycle rhythm that may be influenced by changes in an individual’s physiology. We have also provided two ways to calculate the acrophase and timing of the nadir of the curve, one as the day number into the cycle and the other one as the degree angle into the cycle. The variability in cycle duration means that day 10 of the menstrual cycle might be in the mid-follicular phase of certain cycles with a long follicular phase but in the early luteal phase of cycles with a short luteal phase. Determining the degree angle into the cycle addresses this variability and, in some cases, may more accurately represent position in a cycle versus a simple day number assignment. There are two possible ways to use these metrics to identify biomarkers. First, at the population level, metrics can be compared between individuals with or without a particular characteristic, disease, or treatment. Second, a personalized approach can be applied to track metrics across multiple cycles within an individual to capture changes in metrics that coincide with the development of a new condition or changes over time, such as with aging and/or the approach of menopause. Given the numerous pathways connecting menstrual cycle regulation loops to other physiological health systems, many disorders or treatments may affect menstrual cycle metrics, which could in turn be used as biomarkers. Nevertheless, it may be strategic to start examining menstrual cycle metrics in conditions that are known to disrupt the reproductive hormones and organs and that would benefit from easily accessible diagnostic tools, such as polycystic ovary syndrome (PCOS), endometriosis, or ovarian cancer. With the democratization of consumer devices tracking physiological signals, there is an expansion of datasets available that are suitable to serve this purpose and allow researchers to investigate women’s health biomarkers with tools such as WAVES to facilitate future discoveries. WAVES-generated metrics were used here to characterize aging effects on the menstrual cycle patterns in basal body temperature. The WAVES algorithm was applied to a rich dataset collected during a study with an accelerated longitudinal design ( 33 ). In both analyses, comparing menstrual cycle temperature metrics between younger and older individuals and within-individual trajectories over time, we found aging-related effects on the menstrual cycle temperature rhythm metrics. The group of individuals between 35 and 42 years old presented a higher average temperature level across their menstrual cycles than did the group of individuals aged 18 and 35 years, with the age-related difference being greater for the lower temperature values (+0.06°C) than for the higher values of the curve (+0.04°C). Consistent with this finding, the amplitude of the menstrual cycle temperature rhythm was lower in the older compared with the younger group. While we focused on comparing two age groups, younger and older, who were all premenopausal, we acknowledge that age likely exerts nonlinear effects closer to menopause [occurring at a median age of 51.5 years ( 16 )], which should be investigated in future studies using spline or breakpoint modeling approaches. The approach to menopause is associated with changes in temperature regulation, reflected most obviously in the emergence of hot flashes and night sweats in the majority of women ( 48 ). Our results support and extend those of a prior study that we conducted in a different sample of women that used wearable rings to track skin temperature across the menstrual cycle using just one menstrual cycle of data for most participants and modeling it with a cosinor ( 28 , 41 ). In that study, we also found a higher mean of the fitted curve or mesor of temperature in a group of individuals aged 42 to 55 years compared with younger individuals (18 to 35 years) ( 41 ). For the within-individual longitudinal models here, that included, on average, 7.3 menstrual cycles per individual, we found a subtle but significant increase (+0.004°C/year) in the lowest five temperature values of the cycle over time. There were also trends toward an increase in the highest five points and temperature average, although these trends did not reach significance, suggesting that an overall upward shift in the temperature curve may gradually occur over many years, clearly detectable with the between-group analyses, and consisting of an increase of the basal body temperature before ovulation, combined with a smaller temperature rise in the luteal phase. Another effect of aging that we found was a shortening of the follicular phase, defined as the interval from the first day of menses to the day of peak cervical mucus ( 34 ). This finding supports previous studies that used other ovulation detection proxies such as at-home LH kits or thresholds of temperature increases ( 21 , 22 ) and showed that there is a shortening of the follicular phase over the years, resulting in shorter menstrual cycles overall ( 21 ). In association with these temporal structure changes, the acrophase and nadir days were shifted in time, occurring earlier into the cycle in the older compared with the younger group. However, when looking at the within individual analyses, while the cycle days of nadir and acrophase occurred earlier with aging, the cycle angle at which the nadir and acrophase occurred remained unchanged. This result suggests that the normalized metrics for the timing of when the nadir and acrophase occur relative to the 360° menstrual cycle are less affected by short-term aging effects. Future work could follow up these results to examine timing of the nadir and acrophase days and cycle angles relative to reproductive hormone levels in relation to aging within an individual. A somewhat unexpected result was that the temperature waves were better modeled (shown by a higher r 2 ) by all the periodic curves tested (cosine, square, zigzag, sawtooth, and Gaussian) in the older group compared to the younger group. Although we had anticipated the possibility of the signal resembling more one type of curve in the younger reproductive years and evolving toward another shape in the later reproductive years, that was not supported. Possibly, the shortening of the follicular phase in the older age group may have contributed to the better fit of at least some of the models applied here. As part of our within-subject analyses, we examined whether menstrual cycle temperature metrics became increasingly irregular (i.e., more variable) from cycle to cycle, as an individual ages. Within an individual, the temperature amplitude and overall level and the amplitude, SD, and smoothness specifically from the luteal phase presented more variability with aging. Also, the cycle period became more variable with aging, associated with greater variability in luteal phase duration. This greater variability in various metrics could have multiple causes, which are beyond the scope of what we could measure in the current dataset and should be investigated in future. For example, there may be a rise in conditions with aging, such as obesity and/or stress, which are associated with irregular menstrual cycles ( 49 ). Here, applied to the aging use-case, the WAVES algorithm proved capable of detecting well-established associations with age, such as follicular phase and overall cycle shortening, in addition to detecting previously unknown associations. Last, an important aspect in menstrual health is the interindividual variability. Studying the menstrual cycle is already complex, and it can be tempting to stop at the group level and generalize results to all menstruating individuals. However, a wide variety of experiences exist around menstrual cycles, with women having very long to very short cycles; very regular to very irregular; and associated or not with patterns of pain, symptoms, or sleep disruptions ( 21 , 46 , 50 ). Here, we took advantage of the within-individual longitudinal analyses to quantify how much of the variability in each menstrual metric was related to the individual, in other words, to detect the existence of a personal footprint in the menstrual metrics of temperature. Some characteristics of the menstrual cycle and the amount of variability in these characteristics have a large proportion of their variability explained by the individual, including the metrics related to the temperature level, both mean temperature level across the entire cycle and specifically for each menstrual phase and the minimum and maximum five values. In addition, the metrics related to the temporal structure of the cycle, such as the duration of the full cycle and its follicular and luteal phases, were strongly influenced by the individual. These findings suggest that, across cycles from the same individual, each person has their own temperature baseline measurement around which the menstrual fluctuations are organized and are highly stable. One obvious bias to the individual’s footprint in the temperature level is the fact that each individual used the same thermometer in the same location over time, which differed from other participants. However, it is likely that physiological reasons also contribute to the individual trait in menstrual cycle body temperature rhythms given that the partial inheritability of basal body temperature has previously been shown in twin studies ( 51 ). Also, a large study conducted on 243,506 temperature measurements in 35,488 patients at an outpatient clinics concluded that “individuals’ baseline temperatures showed meaningful variation that was not due solely to measurement error or environmental factors.” ( 52 ). Our findings indicate that the temperature level attained in specific portions of the menstrual cycle temperature rhythm is a personal trait. However, the variations within cycle (amplitude, smoothness, SD, kurtosis, and skewness) and the shape of the cycle wave (cosine, square, zigzag, etc.) were not found to reflect a personal footprint, suggesting that these aspects of the curve are mostly affected by short-term fluctuations or contextual factors rather than stable personal characteristics. Our results also show that the similarity in menstrual cycle temperature rhythms from cycle to cycle is a personal trait, with some individuals presenting a high amount of variability from cycle to cycle and others having very regular and similar profiles of menstrual cycle temperature rhythms, for all of the four dimensions: the level, temporal structure, menstrual temperature variations, and wave shape. The results concerning the temporal structure are in line with prior studies showing that the menstrual cycle duration is partially controlled by genetic factors ( 53 ) and also by environmental factors ( 51 ), which are likely to vary only moderately for an individual over multiple cycles. To our knowledge, there are no prior reports of a personal footprint in the other dimensions of the menstrual cycle temperature rhythm. Overall, these findings show that certain metrics of the menstrual cycle patterns in temperature may be particularly relevant to consider at the individual level across multiple cycles, first establishing the individual’s baseline and regularity levels, to allow the assessment of changes over time in relation to aging or health conditions. This study benefits from a remarkably large and rich dataset, which contains not only the relatively common self-reports of menses dates but also the daily basal body temperature and cervical mucus measurements, in a large sample of individuals across months to years. This dataset allowed us to apply a mixed between and within longitudinal analytical approach to the data and to bring new insights and perspectives about the relationship between menstrual health and aging during the reproductive span. However, this study also presents certain limitations. The study population, because of the original purpose of the study around fertility, does not contain menstrual cycles from individuals younger than 18 or older than 42 years and also excluded individuals with fertility problems. While it makes it even more interesting that several metrics show significant aging variations across this restricted timeframe and population, studies integrating larger age ranges and broader populations including those with menstrual cycle problems may allow further understanding of how menstrual cycle health evolves from menarche to menopause and how menstrual cycle characteristics differ for those with menstrual associated disorders such as PCOS. In addition, the present study did not assess whether the menstrual metrics evolved after stopping taking hormonal contraceptives over time or after postpartum and lactation periods ended. Future works could assess this, aiming to identify how long these specific situations continue to influence menstrual cycle health after they end. In addition, while the 35 years old cutoff traditionally used in research and clinical was used here, future works on menstrual health and aging would benefit from collecting and using the final menstrual period (menopause) as a reference point, in line with a personalized approach to women’s health. Another limitation in the study is related to the data collection, as participants self-measured their body temperature, with each participant using their own thermometer and chosen location of measurement. As mentioned in the discussion of the ICC result, this element limits the interpretability of the interindividual comparisons. Last, the dataset did not contain broader sociodemographic, anthropometric, or biological data, which would have been valuable both for investigating associations with menstrual cycle metrics and for adjusting as potential confounding factors. Future studies are called for to pursue this effort of investigating the relationships between menstrual cycle health and overall health. The present findings underline the potential of the menstrual cycle to inform health, which could lead to the use of menstrual metrics as indicators and markers of health conditions. Regarding aging, further work could help refine whether the present findings reflect the aging of the temperature regulatory system in an individual in general or of the reproductive system specifically. For example, if specific menstrual cycle metrics were found to be associated with the decline in the pool of ovarian follicles, then monitoring temperature variations would represent a less invasive way to monitor this natural process that is key to fertility planning. In conclusion, the WAVES algorithm is available for researchers to extract metrics from menstrual cycle patterns in biological signals and test whether they are significantly associated with specific health characteristics, diseases, or treatments. Applied here to a large dataset of basal body temperature measured daily across thousands of menstrual cycles, we found that aging is associated with measurable changes in the menstrual cycle of temperature across the reproductive stage. In addition, we found individual-specific footprints in multiple characteristics of the menstrual cycle of temperature and in their cycle-to-cycle regularity. This work suggests that the WAVES algorithm can be used for advancing digital biomarker discovery and highlights the relevance of a personalized approach in the development of next-generation tools for women’s health.

The menstrual cycle has previously been described as “a vital sign from menarche through menopause, an underused but powerful tool for understanding gynecological and general health” ( 1 ). This infradian (periodicity of greater than 24 hours) biological rhythm is tightly interconnected to an individual’s health. On the one hand, menstrual cycles, through the rhythmic expression of hormones which have ubiquitous receptors ( 2 ), have the potential to influence physiology, including metabolic and immune systems ( 2 , 3 ). On the other hand, nutritional, emotional, or physical stressors have the capacity to disrupt menstrual cycles’ rhythms and reproductive functioning ( 4 – 6 ). Across the reproductive life stage, a woman living in the United States would have, on average, 450 menstrual cycles ( 7 ), out of which ~3.2 cycles result in pregnancy ( 8 ). Yet, most of the focus on menstrual health—including research, medical training, customers apps, and patents—are centered solely on the reproductive aspect and fail to leverage these 99% nonconceptive menstrual cycles as health indicators. In the field of biological rhythms research, rhythm metrics are regularly used to quantify how a personal trait, health condition, and treatment influence the rhythm. An example of an ultradian rhythm (periodicity less than 24 hours) routinely monitored for diagnostics is the cardiac rhythm. The interval between the points of its electric signal are measured, and a long interval between points Q and T is used as a diagnostic tool, indicating a specific disease, or can be the consequence of a specific treatment ( 9 ). Another example of biological rhythm applications comes from circadian rhythm research. In this field, the period, amplitude, acrophase, and mesor of physiological signals such as hormones ( 10 ), genetic expression ( 11 ), and temperature ( 12 ) are typically used to compare the circadian rhythm characteristics between groups of individuals of diverse ages ( 10 ), work shifts ( 12 ), body composition ( 13 ), clinical conditions, and other factors. Consequently, given the identification of reliable associations between circadian rhythm metrics and health conditions, certain metrics have been proposed for use as diagnosis markers of metabolic disorders ( 13 ). The rhythm metrics of the menstrual cycle also have the potential to provide information about an individual’s health; however, major gaps remain, including (i) the identification of the most relevant metrics, (ii) the discovery of which of these metrics are associated with health conditions, and (iii) the understanding of the extent to which variations in menstrual signals reflect individual footprints therefore requiring a personalized approach. The menstrual cycle is a hypothalamic-pituitary-ovarian regulatory loop that orchestrates the physiological adaptations that are needed to prepare for a potential pregnancy. The female oocytes, or eggs, originate from a pool of follicles that is produced and reaches its maximum around 20 weeks of gestational life ( 14 ). As no further follicular production occurs in life, the pool decreases across the reproductive life span, which begins with the first menstrual period, or menarche, occurring around 12.25 years old ( 15 ), and continues until the onset of menopause, which occurs at a median age of 51.5 years old ( 16 ). The decrease in follicles is not linear; by 30 years old, about 12% of the follicular reserve remains, which drops to about 3% by age 40 ( 17 ). In parallel, the myometrial and endometrial function of the uterus declines with aging, which can affect fertility and pregnancy outcome ( 18 ). There is a slow but steady decline in fertility in women between 30 and 35 years, after which there is an accelerated decline ( 19 ). In light of these and other physiological changes linked to aging, age 35 years old is typically used as a cutoff in women’s health research and pregnancy care ( 20 ). The menstrual cycle is composed of two main menstrual phases: the follicular phase and the luteal phase. During the follicular phase, the release of gonadotropin-releasing hormone from the hypothalamus triggers the anterior pituitary to release follicle-stimulating hormone, which in turn stimulates the development of follicles until typically one becomes dominant. In parallel, estrogen levels rise, promoting endometrial proliferation and thickening of the uterus lining. The anterior pituitary releases luteinizing hormone (LH), which triggers the release of the egg from the dominant follicle (ovulation), marking the onset of the luteal phase of the cycle. The follicle envelope degrades into the corpus luteum, releasing high amounts of progesterone that prepares the endometrium for a potential implantation. In the absence of fertilization, progesterone and estrogen levels decline, leading to the breakdown of the endometrial lining and onset of menstruation, occurring, on average, 12.4 days [95% confidence interval (CI): 7 to 17] after ovulation ( 21 ). The median duration of the cycle overall is 29.3 days ( 21 ), which varies significantly both in duration and regularity between individuals ( 21 , 22 ). Most of the variability in duration is in the follicular phase, which has an average duration of 16.9 days (95% CI: 10 to 30) ( 21 , 22 ). In addition to coordinating the activity of reproductive organs, the menstrual cycle influences physiology more broadly. Heart rate, heart rate variability, and body temperature are examples of biological signals that follow menstrual patterns ( 23 – 25 ). Among the physiological signals that reflect menstrual patterns, temperature is one of the most extensively described ( 26 ) and has a long history of use in natural family planning ( 27 ). Specifically, basal body temperature (measured in the morning after waking, before any activity) oscillates across the menstrual cycle ( 28 ), with a rise in temperature evident after plasma progesterone levels start to increase and reaching a maximum during the mid-luteal phase before declining around the time of menses ( 26 ). This temperature oscillation is an indicator of ovulatory cycles, since it is detected in 98% of cycles in which ovulation was detected with the gold standard direct method of transvaginal ultrasonography ( 29 ). If fertilization occurs, then body temperature increases to an even greater level than in nonconceptive cycles and remains high throughout the pregnancy ( 30 ). Menstrual-related variations in temperature have been historically leveraged as a marker of ovulation ( 26 ). The ovulation date can be retrospectively identified on the basis of menstrual cycle temperature oscillations ( 28 , 31 ), and the changes in minimum, maximum, and mean of temperature can be leveraged for early pregnancy detection ( 30 ). Daily measurements of basal body temperature, taken upon awakening in a consistent location (oral, vaginal, and anal) ( 22 ) ( 27 ) or strategically extracted from high-frequency temperature monitoring from wearables ( 32 ), reflect menstrual variations. The current analysis takes advantage of a remarkably rich dataset of daily basal body temperature measurements, collected in the 1990’s, which contains 5674 menstrual cycles of data from 753 individuals across seven centers in Europe. Data were collected from the present cohort following an accelerated longitudinal design, as participants’ age at study start spanned between 18 and 40 years old, and they were each followed during several menstrual cycles. In other words, although participants were not followed during their entire reproductive stage, the compilation of relatively short time series from multiple individuals results in a dataset containing longitudinal data covering most of the reproductive life span. This design also reduces cohort effects, where individuals share a similar experience or are affected by the same external factor, potentially biasing the analysis ( 33 ). Multiple types of data were collected, including daily temperature measures, daily vaginal mucus secretions, age, reproductive history, and sexual activity of participants ( 34 ). From this dataset, multiple articles have advanced our understanding about the fertile menstrual window ( 35 ) and the evolution of fertility with aging ( 36 ). The dataset was also recently used in an analysis supporting the existence of an internal menstrual clock ( 37 ). Here, we advance the use of this dataset to develop an algorithm that can track how menstrual variations in temperature relate to physiology beyond reproductive health. One aspect of life that influences physiology broadly and is therefore likely to be reflected in metrics of the menstrual cycle is aging. Aging encompasses the progressive degeneration of physiological functions, in parallel with the decrease in reproductive capacity and survival rate ( 38 ). However, the aging of the organs and their functions does not necessarily mirror chronological aging. In the interest of expanding not only the life span but also the health span, the identification of indicators of physiological aging has gained interest ( 39 ). Prior work has started discerning associations between aging and menstrual cycle characteristics. For instance, others have shown there is a shortening of the follicular phase over the years, resulting in shorter menstrual cycles overall ( 21 ). The menstrual cycle rhythms of resting heart rate and heart rate variability have also been shown to have a smaller amplitude with aging ( 40 ). The metrics that are typically used in circadian science (period, amplitude, acrophase, and mesor) have been derived from menstrual variations in skin temperature data collected via wearables ( 28 ), and an association was found with aging ( 41 ). Specifically, the mesor of the menstrual cycle temperature rhythm, corresponding to the mean level of the temperature wave modeled by a cosinor, was higher in a group of midlife individuals (42 to 51 years old, between late reproductive stage and early menopausal transition) than in a group of younger participants (18 to 35 years old, peak reproductive stage) ( 41 ). These examples show that multiple dimensions of menstrual cycles of temperature evolve with aging, including the overall level of the rhythm, its temporal organization and duration, and the temperature variation itself. Other metrics—including best fitting waveform, slope of the temperature trend, and smoothness of the rhythm—may also change across the reproductive life span; however, they have not yet been examined. It is also possible that aging influences the amount of variability in menstrual cycle metrics from one cycle to another, making them less regular and more variable, not only in regard to their period but also to their amplitude, acrophase, and mean level. Notably, while the term “period” is often used to refer to the bleeding segment of the menstrual cycle, we encourage the use of “menses” or “menstruation,” and to reserve period as reference to the menstrual cycle length or duration (the number of days between the first day of menses of one cycle and the first day of menses of the next cycle), since the period of a rhythm refers to the time to complete a cycle. Although the menstrual cycle is typically described as 28 days long with ovulation on day 14, research based on large datasets indicate that this is more the exception than the rule, as only 12.4% of individuals present 28-day cycles, with ovulation day spread across 10 days of those 28-day cycles ( 22 ). Researchers therefore point toward the relevance of the specific menstrual cycle duration and ovulation days for an individual’s own cycle when aiming to conceive ( 22 ). In addition to this variability of cycle duration between individuals, there is also variability between individuals in the regularity of menstrual cycles ( 42 ) that may not only have repercussions for fertility but has also been shown to be predictive of future cardiovascular disorders even after adjusting for potential confounders ( 43 ). This interindividual variability also manifests in menstrual-related symptoms including pain ( 44 ), mood variations ( 45 ), or sleep disturbances ( 46 ). There is also individual variability in the distribution of symptoms across the cycle, for example, with some individuals reporting sleep problems around ovulation, others around menses, and others not presenting specific menstrual patterns ( 46 ). All these elements underline the importance of considering interindividual differences when considering the menstrual cycle in research and care and using a personalized approach to identify previously unknown biomarkers related to the menstrual cycle. The identification of metrics capturing multiple dimensions of a given physiological signal that presents menstrual patterns can provide a valuable tool to identify previously unknown biomarkers such as in basal body temperature time series. The derived metrics could be used to detect subtle changes associated with various conditions, treatments, behaviors, or broader biological processes including, but not limited to, aging. Furthermore, understanding the degree to which these metrics are reflecting individual traits could offer critical insight into the need for personalized approaches in menstrual health care. Several prior studies have characterized the rhythmic structure of menstrual temperature variations and their relation to ovulation timing. Classical fertility awareness methods, such as those described by Billings ( 47 ), retrospectively identify ovulation based on a sustained postovulatory temperature rise of ~0.3° to 0.5°C, reflecting the thermogenic effect of luteal-phase progesterone. More recent modeling efforts have applied harmonic or cosinor analyses to continuous temperature data collected from wearable sensors to quantify the amplitude, phase, and mesor of menstrual temperature rhythms ( 28 ). These models typically fit a sinusoidal waveform to temperature measurements across the cycle, enabling the extraction of rhythm characteristics that can be compared across individuals or conditions. Complementary work by Ecochard and colleagues ( 37 ) demonstrated evidence for an internal circamonthly timing system regulating the ovarian cycle, providing further support for the application of rhythmic modeling approaches to menstrual physiology. In the present work, we built upon this foundation by developing the WAVES (women’s health assessment through variability in endocrine-related signals) algorithm, which integrates both classical and rhythmic metrics within a unified framework and computes multidimensional features that capture temperature level, temporal structure, intracycle variability, and waveform shape, thereby extending existing methods to enable large-scale, reproducible menstrual cycle biomarker discovery. We used the previously introduced dataset ( 34 ) to (i) develop an extensive set of metrics grouped in four sets to capture the multiple dimensions across level, variations, temporal aspect, and overall shape of the menstrual temperature signal that may be influenced by health changes; (ii) assess how these metrics differ in those younger versus older than 35 years; and (iii) use within-individual menstrual cycle data to determine individual specificity of these metrics with aging. We hypothesized that certain characteristics of the menstrual cycle would evolve with aging, with a shortening of the follicular phase and overall cycle period, an increase in menstrual cycle variability, and a global increase of the mean of the menstrual cycle temperature. We also hypothesized that features of the temperature curve and the metrics that capture them represent individual-specific characteristics. In other words, we propose that each individual will show a consistent absolute level and degree of variability in temperature metrics from one cycle to the next. If confirmed, then this would highlight the importance of a personalized approach to menstrual health. These two hypotheses are not mutually exclusive: An individual footprint may exist (for example, a woman who consistently has long and regular menstrual cycles), yet age-related changes can still emerge over time (for example, her cycles gradually becoming shorter and less regular).

The database used in the present study is the “Fertili” dataset previously described by Colombo and Masarotto ( 34 ). Of the original 6175 menstrual cycles from 857 participants in the overall dataset, cycles that were conceptive or lacked sufficient daily temperature and cervical mucus records were excluded, leaving 5674 nonconceptive cycles from 753 participants for analysis. The average age at study entry was 29.5 ± 3.9 years, and each participant contributed a median of four cycles. Cycle distribution was highly right-skewed (interquartile range = 2 to 9; mean ± SD = 7.2 ± 7.9 cycles; range = 1 to 63). On average, participants had 1 ± 1 pregnancies before study entry, and 30% reported a past history of using hormonal contraceptives. The original objective of this dataset was to advance knowledge about the fertile window and natural family planning accuracy to prevent or target pregnancy ( 34 ). Participants were recruited across seven European centers (Milan, Verona, Lugano, Düsseldorf, Paris, London, and Brussels). The detailed description of the population per center is available elsewhere ( 34 ). Inclusion criteria were as follows: the use of natural family planning methods, between 18 and 40 years old, in a stable relationship, having had at least one menses after a pregnancy or lactation, and not taking medication affecting hormones or fertility and not having (nor the partner) disorders causing subfertility or infertility. Couples with habits of mixing protected and unprotected sex or not meeting the inclusion criteria were excluded from the study. Participants gave consent, and the study followed the tenets of the Declaration of Helsinki and was approved by the Ethical Committees of Fondazione Lanza (Padua, Italy) ( 34 ). During the study, participants were instructed to perform basal body temperature measurements in a consistent location (anal, vaginal, or oral) using a mercury thermometer with a precision of 0.05°C before engaging in any activity upon awakening in the morning ( 34 ). In addition, participants scored daily their cervical mucus or sensation into categories proposed by Billings and colleagues ( 47 ) as the Billings Ovulation Method in 1972 [1 = dry, rough, and itchy feeling or nothing felt; nothing seen, 2 = damp feeling; nothing seen, no mucus, 3 = damp feeling; mucus is thick, creamy, whitish, yellowish, not stretchy/elastic, sticky, 4 = wet, slippery, smooth feeling; mucus is transparent, similar to raw egg white, stretchy/elastic, liquid, watery, reddish (with some blood)]. From this report, the day of mucus peak was identified as a proxy for ovulation detection, defined retrospectively as the last day with best quality mucus (clear, stretchy, and lubricative) by sensation or appearance, for each cycle as previously described in work using this dataset ( 34 ) and as described in ( 47 ). This method has been validated against LH surge and ultrasound-confirmed ovulation, occurring within 1 day of true ovulation in 75% of cycles ( 29 ). While participant self-observation introduces some subjectivity, the mucus peak provides a reliable, noninvasive marker of ovulation that is well suited for retrospective phase delineation in large naturalistic datasets. Last, participants logged daily if they had unprotected intercourse and their menses dates. Individuals reported this information through charts and were asked to do these reports daily until confirmation of a pregnancy onset. Prior works on this dataset implemented a filtering protocol to exclude cycles with excessive missing data based on whether sufficient temperature and cervical mucus observations were available to reliably identify an ovulation reference day ( 34 ). The same filtering was applied here, resulting in a dataset containing an average of 22.8 ± 7.1 temperature data points per cycle of 29.0 ± 4.4 days equal to 79% of daily measures. Daily basal body temperature data were smoothed using a 3-day moving average, a standard approach in menstrual rhythm modeling ( 28 ) that minimizes single-day fluctuations while preserving the underlying physiological pattern. Missing values within the smoothing window were handled using the mean of available values, and windows containing only missing values were left as missing. Outlier filtering was applied before metric extraction: All entries coded as missing-value flag and temperature values < 33.0°C were set to missing values. Approximately 4% of all temperature readings were affected (1.19% flagged as missing and 2.8% below 33°C). Only cycles with complete and reliable temperature and mucus information were retained, yielding 5674 nonconceptive, high-quality cycles from 753 participants after filtering. All subsequent metrics were computed on cycles meeting predefined completeness thresholds, ensuring quality of derived rhythmic features. The WAVES algorithm stands for women’s health assessment through variability in endocrine-related signals. It extracts comprehensive, structured, physiologically meaningful metrics from daily time series data, in this case, basal body temperature, for each menstrual cycle. The first day of self-reported menses marks the onset of a cycle, which runs until the last day before the first day of the next menses. Ovulation day was estimated in the current study as the peak day of cervical mucus quality ( 47 ). The luteal phase was defined as the interval beginning the day after the mucus peak and ending on the day before the start of the next menses. Four sets of metrics are extracted to capture the different dimensions of the menstrual cycle temperature rhythm. First, metrics capture the dimension of the temperature level, with the absolute overall temperature values within a cycle and within each cycle phase. Second, metrics capturing the temporal dimension of the menstrual rhythm are extracted, including the duration of the cycle and phases and the acrophase and day of the nadir, which are respectively the temporal position of the maximum and the minimum of the temperature curve. Third, a set of metrics capturing the variations and slopes of each cycle was used to quantify the sharpness and regularity of transitions across the cycle. Fourth, a last set of metrics was generated to determine the shape of the curve, including cosine, square, zigzag, sawtooth, half-sine, and Gaussian shapes. The goodness of fit, or r 2 , ranges between 0 and 1 and takes a higher value when the model tested explains an important proportion of the variability in the data. For each metric, with the goal of capturing the most physiologically meaningful value, we evaluated the trade-off between using the raw temperature data to have the highest possible fidelity and applying smoothing to reduce the influence of outliers and short-term noise. Therefore, some metrics were extracted from the raw and others from a smoothed version of the data from a 3-day rolling average. A full list of metrics and calculation details is provided in table S1. Figure S1 shows an example of an individual’s basal body temperature across multiple menstrual cycle and metrics extracted for each cycle. The algorithm was developed in Python and is available in open access. Participants were first grouped by age older or younger than 35 years, a typical cutoff in women’s health and reproduction studies ( 20 ). To compare the value of each metric between groups (18 to 35 years old versus 35 to 42 years old), Kruskal-Wallis tests were performed as variables did not present a normal distribution, and Bonferroni correction was applied for multiple tests using α = 0.05/32 as the adjusted threshold of significance to account for multiple tests. To assess whether metrics evolved with age within individuals, mixed linear models were performed using age as predictor and the value of each metric as outcome, with subject ID nested with collection site and teacher in charge of the data collection (“CODICE” variable), integrated as a random factor. For this analysis, age was mean-centered for each individual, and Bonferroni correction for multiple tests was applied. A low P value indicates that the metric tested does evolve with age, and the coefficient informs in how many units and in what direction. In addition, to assess whether metrics are more regular from cycle to cycle at a younger than an older reproductive age, the SD of the metrics were calculated via a rolling window of six consecutive cycles over the sequence of available measurements within each individual. When less than four of the six consecutive cycles were available, the window was not retained. Linear mixed-effects models were fitted, with each feature’s 6-month SD as outcome, using age as predictor, with subject ID as a random factor. Age was mean-centered for each 6-month window, and Bonferroni correction for multiple tests were applied. The ICC was extracted from the linear mixed models and used to determine whether there is more diversity between individuals than within individuals. The ICC can be read as a percentage of the variance in the metric that is attributable to between-individual differences. A high ICC therefore indicates greater variability between than within individuals, suggesting the presence of a stable personal signature, or “footprint,” in the menstrual metrics level and/or in their variability ( 54 ). The statistical analysis and WAVES algorithm were written in Python v 3.10 using the packages statsmodels, sklearn, scipy, and pandas.