The standard method for detecting ovulation through body temperature relies on a single daily measurement, taken orally or vaginally, before getting out of bed. The thermal shift caused by progesterone after ovulation is real and measurable at roughly 0.2 to 0.5 degrees Celsius above follicular baseline. One reading per day, however, leaves a lot of room for error. A night of disrupted sleep, an early alarm, or a mild viral infection can shift a basal body temperature reading by enough to obscure or fabricate a phase transition. Continuous wrist temperature measurement changes the data density: a wrist sensor sampling at one-minute intervals generates hundreds of readings across a single overnight period, which allows the signal to be averaged and short-term noise to be reduced in ways that daily spot measurements cannot support.
The Physiology of the Ovulatory Temperature Shift
The LH (luteinizing hormone) surge triggers follicular rupture and ovulation approximately 24 to 36 hours after its onset. The temperature rise that follows is not a direct consequence of LH itself. It comes from progesterone, secreted by the corpus luteum that forms at the follicle site after ovulation. Progesterone raises the body's thermoregulatory set point through action on the hypothalamus. The thermal shift typically appears two to three days after ovulation and persists for the duration of the luteal phase, then drops at menstruation when the corpus luteum degrades and progesterone falls.
At the wrist, this manifests as a sustained elevation of the overnight temperature trough: the lowest point in the daily circadian arc. The shift is smaller at the wrist than at core sites because wrist skin temperature tracks peripheral blood flow, which is modulated by ambient temperature, clothing, and vasomotor activity. In our internal pilot dataset, the median overnight wrist temperature trough during the luteal phase runs approximately 0.15 to 0.25 degrees Celsius above the same participant's mid-follicular baseline, though individual ranges vary considerably from person to person.
Some participants show a pre-ovulatory thermal dip in the two to three days preceding the LH peak, attributed to an estrogen-driven increase in peripheral vasodilation. We observe this feature in roughly half of the cycles with good overnight coverage in our current dataset. It is a useful supporting signal, but not reliable enough to use as a primary trigger for ovulation detection.
How the Model Detects the Luteal Shift
Isolating the luteal-phase temperature signal requires separating it from at least two overlapping sources of variation: the circadian rhythm (which drives a daily temperature arc of 0.4 to 0.6 degrees at the wrist) and short-term noise from environmental temperature, body movement, and skin perfusion changes. Individual baseline drift across the cycle also needs to be corrected; some people show a gradual rise through the follicular phase that has nothing to do with ovulation.
Our LSTM model processes overnight temperature segments after alignment on sleep onset, extracted from the continuous wrist stream. Using overnight windows rather than full-day traces significantly reduces contamination from ambient temperature and daytime activity. Within each overnight segment, we further standardize to remove between-night differences attributable to room temperature variation.
The model outputs phase probabilities across four main cycle stages. For ovulation inference specifically, we track the transition probability from follicular to luteal across consecutive nights. A crossing event is flagged when the luteal probability stays above a per-individual learned threshold across at least three consecutive overnight segments. This prevents a single aberrant night from generating a false positive ovulation call.
Coverage Requirements and What Happens Without Them
Making a reliable ovulation inference requires sustained overnight data without major gaps in the peri-ovulatory window. Our current coverage threshold is a minimum of five consecutive nights with overnight data density above 70 percent of the expected sampling rate. If a participant removes the device for more than two consecutive nights during what the model estimates to be the peri-ovulatory window, we report that ovulation cannot be confirmed for that cycle rather than issuing an uncertain estimate.
This matters because the peri-ovulatory window in a typical cycle spans roughly seven to ten days. Missing two to three nights in that window means losing a meaningful fraction of the temperature record precisely when the signal is most informative. We surface the coverage gap to the user so they understand why no inference was generated, rather than presenting a low-confidence prediction that looks authoritative but rests on weak data.
Known Failure Modes
Illness is the most common source of false positives. A fever from a viral infection causes an overnight temperature elevation that closely resembles the luteal thermal shift in shape and duration. We cross-check resting heart rate alongside temperature as a secondary signal: a simultaneous elevation in resting heart rate, more than approximately seven beats per minute above the individual's recent baseline, is consistent with a systemic illness response rather than a progesterone effect. This heuristic catches the majority of fever events in our dataset, but the model does not always resolve ambiguous cases correctly when the heart rate elevation is modest.
Shift work and highly irregular sleep schedules break the overnight segmentation logic because sleep onset timing varies by several hours day to day. For these participants, the temperature records during their non-standard sleep periods still contain useful signal, but the alignment step becomes much noisier, which widens the uncertainty estimates throughout the cycle prediction.
Short luteal phases, those lasting fewer than nine days, present a different structural problem. The temperature elevation exists for less time before the next menstrual cycle begins, and in some cases the model detects the shift only retrospectively, after menses onset, rather than during the luteal phase itself.
What the Limits Mean for How We Report Results
Because wrist temperature is an indirect proxy for the progesterone-driven thermal shift, we do not report ovulation timing as a precise date. The output is a probability window, typically three to five days wide, with an explicit uncertainty indicator. When confidence is low, the display reflects that directly rather than presenting a point estimate with hidden uncertainty beneath it.
We are not positioning this as a replacement for LH test strips or serum hormone measurement. Those methods answer a different question: they detect the LH surge directly, before ovulation occurs, and can be used for same-day timing decisions. What continuous wrist monitoring contributes is a retrospective confirmation signal, and, across multiple cycles, a longitudinal record of ovulation timing patterns that is not practical to build through urine testing alone. The comparison is not which tool is better; they are addressing different time horizons.
Individual Calibration Across Cycles
The 0.2 to 0.5 degree shift range that appears in basal thermometry literature describes a population distribution, not a reliable threshold for any specific person. Some individuals consistently show a shift of only 0.1 degree; others show more than 0.4 degrees every cycle. A universal threshold applied across people introduces systematic false negatives for low-shift individuals and false positives for high-variability individuals.
Our model learns an individual's follicular and luteal temperature baselines from earlier cycles and applies person-specific thresholds rather than population constants. This calibration requires data. Ovulation inferences from the first cycle should be treated as exploratory, reflecting the model's prior distribution rather than a calibrated individual estimate. By the third cycle, the model has accumulated enough within-person data to apply thresholds that reflect that person's actual thermal signature around the luteal transition.
We communicate this clearly during onboarding and in the in-app cycle summary view. First-cycle output is framed as the model learning a baseline, not as a finished ovulation report. Users who understand why the first cycle is less certain tend to stay engaged through the calibration period rather than dismissing the product because the first prediction was imprecise.
The core value proposition of continuous wrist temperature for ovulation tracking is not precision at the level of individual hours. It is density and passivity. A single measurement taken while half-asleep carries more noise than 400 measurements taken across the same night. Over multiple cycles, that density translates into a richer picture of an individual's thermal signature around ovulation than any diary-based method can produce. The accuracy limits are real, and they are built into how the product communicates results rather than papered over with confidence language the data does not support.