This is a foundations post, and it is worth being direct: wrist wearables do not measure hormones. Estrogen, progesterone, LH, FSH, none of these appear directly in any commercially available wrist biosensor data stream. What the wrist provides is a set of physiological proxies that are correlated with hormonal variation across the cycle. That correlation is real, established in the literature, and useful. But the inference gap between proxy measurement and hormone state is where all the modeling work happens, and understanding what the sensors actually measure is necessary for evaluating what cycle tracking products can and cannot claim.
What the Wrist Sensors Actually Measure
A typical wrist biosensor device contains three or four physical sensing components. The most relevant for cycle inference are the thermistor, the photoplethysmograph (PPG), and the accelerometer.
The thermistor measures the temperature of the skin surface at the point of contact. This is not body core temperature and it is not the temperature used in basal body thermometry. Wrist skin temperature is substantially cooler than core temperature, typically 3 to 6 degrees Celsius lower depending on ambient conditions and vasomotor state. The relationship between wrist skin temperature and the hormonal temperature signals used in cycle tracking is indirect: progesterone's effect on the thermoregulatory set point raises core temperature, and this is detectable at the wrist as a smaller but consistent elevation in the overnight trough, when peripheral vasodilation has stabilized and ambient temperature effects are minimized.
The PPG sensor uses green or infrared LED light reflected from the skin to detect the volumetric pulse of each heartbeat at the wrist. From this signal, the device computes heart rate and, with appropriate analysis, heart rate variability (HRV), specifically the variation in the intervals between successive heartbeats. HRV is a reflection of autonomic nervous system balance, with higher HRV generally associated with greater parasympathetic tone. Estrogen and progesterone both affect autonomic regulation, and HRV patterns show cyclic variation that can be used to distinguish cycle phases in aggregate, though the signal is subtle and heavily confounded by sleep quality, stress, and fitness level.
The accelerometer captures three-axis motion. For cycle inference it serves two roles: detecting sleep onset and offset (which defines the overnight window for temperature and HRV analysis) and flagging artifact periods (high-motion episodes that contaminate temperature and PPG readings).
The Inference Gap
The inference chain from raw sensor signal to cycle phase prediction has several steps, and each step introduces a layer of uncertainty.
First, the physical sensor measurement has its own noise: thermistor precision, motion artifacts in the PPG channel, and variation from contact pressure between the device and the skin. These are hardware-level limitations that differ across device types.
Second, the mapping from physical measurement to biologically meaningful quantity requires signal processing. Converting a raw PPG waveform to an IBI sequence, applying artifact rejection, computing frequency-domain HRV features: each of these steps involves algorithms with assumptions that may not hold perfectly in every data segment.
Third, the mapping from processed physiological features to hormonal state is a statistical inference, not a direct observation. The model learns correlations between temperature and HRV patterns and cycle phase from training data, then applies those correlations to new data. The strength of this inference depends on how representative the training data is, how consistent the individual's physiology is with the population on which the model was trained, and how much individual longitudinal data the model has accumulated.
Understanding these three layers matters because they represent three different categories of limitation. Hardware noise can be addressed by improving sensors. Signal processing artifacts can be addressed through better algorithms and stricter quality gates. The fundamental inference gap between proxy physiology and hormonal state is harder to close because it is biological, not technical: wrist temperature and HRV are imperfect proxies for hormonal state, and no amount of modeling sophistication can fully compensate for that.
Where the Proxies Are Strong and Where They Are Weak
Wrist temperature is most informative during the overnight window, when it captures the thermal trough that reflects the progesterone-mediated shift across the follicular-luteal boundary. The consistency of the overnight measurement across nights makes the longitudinal signal reasonably reliable for detecting the sustained thermal elevation that characterizes the luteal phase. It is weaker for detecting ovulation in real-time (the signal lags ovulation by two to three days) and unreliable during illness, high-stress periods, or unusual thermal environments.
HRV is most informative for distinguishing the high-parasympathetic menstrual and follicular phases from the luteal phase, which tends to show a modest reduction in high-frequency HRV in some individuals. The signal is variable across people: some individuals show clear cyclic HRV patterns, others show none that are distinguishable from baseline variability. HRV is also heavily confounded by sleep quality, alcohol, and physical training load in ways that temperature is not, which means HRV contributes more noise relative to signal for some individuals than for others.
Combining temperature and HRV as joint inputs to a multi-feature model produces more robust phase estimates than either channel alone, because the failure modes of the two signals are not fully correlated. An illness event elevates both temperature and resting heart rate simultaneously, which is a distinguishable pattern from the luteal phase where temperature rises but HRV changes are modest. Using both channels also allows the model to down-weight one when the other is higher quality in a given segment.
What This Means for Claims We Make and Do Not Make
We track cycle phase, not hormone levels. Our product communicates inferred phase boundaries and confidence intervals based on wrist physiology. It does not report estrogen, progesterone, or LH values. Saying "the model estimates you are currently in the early luteal phase with moderate confidence" is accurate and reflects what the data supports. Saying "your progesterone is elevated" is not, because we have not measured progesterone and the physiological proxy we are using is several inference steps removed from it.
This also means we do not make diagnostic claims. Phase inference from wrist data is a tool for longitudinal cycle awareness. It is not a replacement for a serum hormone panel or a cycle assessment performed by a reproductive endocrinologist. The appropriate framing, and the one we use, is that our output is supplementary information that can enrich a clinical conversation, not substitute for clinical evaluation.
The wrist biosensor approach is useful precisely because it is passive and continuous. A person does not have to do anything different; they wear the device, and the data accumulates. Over multiple cycles, the pattern of physiological variation becomes a longitudinal record that reveals individual characteristics of the cycle that would be impractical to build through daily spot measurements or periodic lab draws. The inference limitations are real and we disclose them. The longitudinal value is also real, and it comes directly from the continuous nature of what the sensors actually do measure.