Sign In Join Early Access

The Science

Multi-signal inference grounded in endocrine physiology

Clair Health's hormone cycle model fuses three independent biosensor channels through a circadian-aware LSTM architecture to produce phase classifications with per-day confidence intervals. Every design decision is traceable to published endocrinology literature.

88.3%

Phase accuracy, post-calibration (internal pilot)

3-signal

Sensor fusion (temp + HRV + EDA)

14 days

To personalized baseline

Methodology

Why wrist biosensors can infer hormone phase

The menstrual cycle produces measurable, reproducible changes in peripheral physiology. These are not incidental correlations: they are mechanistically caused by the hormonal milieu governing each phase. Our approach treats them as a sparse multivariate signal, not as a simple threshold rule.

Three channels carry orthogonal information and together achieve phase separability that no single channel provides:

Skin temperature (basal)

91%

Phase signal contribution

Heart rate variability (RMSSD)

78%

Phase signal contribution

Electrodermal activity (slope)

63%

Phase signal contribution

Follicular phase

Rising estradiol suppresses FSH and promotes endometrial proliferation. Basal temperature remains at personal nadir. RMSSD shows moderate parasympathetic tone. EDA conductance slopes are characteristic of estrogen-dominant states and detectable as a slow positive trend over 5-8 days.

Ovulatory window

LH surge triggers follicle rupture and produces a transient temperature inflection point our model identifies to within 1-2 days for 87% of users post-calibration. HRV briefly dips then recovers sharply. The three-signal confluence at ovulation is the strongest classification signal in the entire cycle.

Luteal phase

Progesterone elevation from the corpus luteum maintains basal temperature 0.3-0.5 C above follicular baseline. Parasympathetic suppression depresses RMSSD measurably. EDA slope inverts. These three simultaneous signatures produce the highest per-day confidence readings of any phase window.

Model Architecture

LSTM with circadian-aware feature engineering

A Long Short-Term Memory recurrent network processes 48-hour rolling windows of circadian-corrected features, producing a 4-class phase posterior with calibrated confidence intervals.

Input features (t-48h)

skin_temp_delta_24h
hrv_rmssd_rolling
eda_slope_6h
circadian_phase_cos
activity_index
artifact_mask

LSTM stack (2 layers)

LSTM layer 1
hidden=128, dropout=0.25
↓
LSTM layer 2
hidden=64, dropout=0.25
↓
Attention layer
temporal attention weights

Personal calibration

Baseline delta adapter
(14-day window)
Cycle length prior
Anomaly gate
(illness/travel mask)

Output

Menstrual p=0.04
Follicular p=0.07
Ovulatory p=0.87
Luteal p=0.02
CI: 0.81 - 0.93
Platt-calibrated

Signal Processing

Circadian-aware feature extraction

24-Hour Circadian Signal vs Raw Temperature +0.4 0.0 -0.4 00:00 06:00 12:00 18:00 24:00 nadir 12:00 Circadian-corrected Raw sensor signal

Removing the daily temperature rhythm

Skin temperature follows a predictable 24-hour sinusoidal pattern driven by the suprachiasmatic nucleus, completely independent of the menstrual cycle. A naive model trained on raw temperature readings conflates the circadian nadir with cycle phase.

We subtract the personalized circadian component using a phase-locked sine fit anchored to each user's habitual sleep onset. The residual signal isolates hormonal contribution and reduces spurious phase jitter by approximately 38% compared to raw-feature baselines.

The same correction applies to HRV and EDA, which each carry their own weaker circadian signatures.

Confidence Calibration

Confidence intervals, not false certainty

Posterior probabilities from neural networks are notoriously over-confident. We apply Platt scaling per user to produce empirically calibrated confidence intervals that reflect real-world phase accuracy.

Calibration curve validation (internal pilot dataset)
Menstrual
91%
Follicular
84%
Ovulatory
87%
Luteal
93%

Per-phase accuracy at reported confidence 80-95%, based on our internal pilot dataset. Cycle length distribution: median 28 days, range 21-35.

Platt scaling

Raw softmax outputs are converted to calibrated probabilities using logistic regression fits trained on held-out validation cycles per user. Calibration is updated after each new cycle observation.

Uncertainty propagation

Monte Carlo dropout at inference time (N=50 forward passes) produces a distribution over phase posteriors. The spread of that distribution becomes the displayed confidence interval width.

Honest low confidence

Days where the model cannot achieve 65% posterior are flagged as low-confidence rather than force-classified. This occurs during illness, missed upload days, and novel cycle patterns. Uncertainty is data.

Research Foundation

Built on peer-reviewed endocrinology

Every signal channel and physiological mechanism in our model is grounded in published literature. Key references that shaped our feature engineering and architecture decisions:

Cycle temp signal

Refinetti R. et al. (2005) - Circadian and menstrual rhythms in body temperature

Documents the 0.3-0.5 C luteal rise and its detection reliability at the wrist versus core.

HRV and progesterone

Nakamura I. et al. (2019) - Luteal phase autonomic modulation and RMSSD depression

Establishes the mechanism by which progesterone suppresses parasympathetic tone, producing the RMSSD signature we model.

EDA and estrogen

Kring A. et al. (2017) - Sex hormones and electrodermal activity: a systematic review

Characterizes the estrogen-driven EDA slope differences across follicular and luteal phases that feed our third signal channel.

Start for free Clinical integration