Objective. To quantify agreement between continuously measured distal skin temperature and laboratory dim-light melatonin onset (DLMO) as estimates of circadian phase in free-living adults, and to identify conditions under which agreement degrades.
Design. Prospective method-comparison study with paired measurements. Each participant completed laboratory DLMO assessment while wearing the study device continuously under ordinary living conditions.
Participants. Sixty-eight adults (n = 68).
Measurements. Continuous distal skin temperature and triaxial accelerometry at 1 Hz, processed by a per-participant limit-cycle oscillator model to yield a circadian phase estimate. Reference standard: salivary DLMO under dim-light laboratory conditions.
Main outcome. Absolute difference, in minutes, between model-derived phase estimate and laboratory DLMO on paired sessions.
Results. Median absolute error was 19 minutes. Interquartile range 11–34 minutes; 95% CI for the median 15–24 minutes; Bland-Altman limits of agreement −58 to +50 minutes. Agreement degraded in participants reporting irregular ambient temperature exposure; median error 38 minutes in that subgroup (n = 21) against 15 minutes in the remainder (n = 47).
Conclusions. In free-living adults, continuous distal skin temperature supports circadian phase estimation with a median error small relative to the precision required for light and sleep scheduling. Irregular ambient thermal exposure is the principal identified moderator and should be treated as a boundary condition of the method rather than as residual noise.
Registration. Not registered.
Circadian phase determines when alertness bottoms out, when sleep will consolidate, and which direction a given light exposure will shift the internal clock. For workers on rotating or night schedules, phase and schedule are routinely misaligned, and the interventions that correct the misalignment — timed light, timed sleep, timed caffeine — are only effective if applied relative to the individual's actual phase rather than to clock time.
Dim-light melatonin onset is the accepted reference standard for circadian phase. It requires repeated saliva or plasma sampling under controlled dim-light conditions over several hours, typically in a laboratory. The protocol is accurate and impractical: it cannot be repeated often enough to track a phase that moves across a rotation, and it cannot be deployed at the scale of a workforce.
Distal skin temperature offers a candidate proxy. Distal vasodilation rises ahead of sleep onset and the distal temperature minimum tracks the core body temperature minimum, which holds a stable phase relationship to DLMO. Unlike photoplethysmography, the signal is low-power enough to sample continuously for days, which is what a moving phase estimate requires.
The open question is not whether the relationship exists under laboratory conditions, where it is established, but whether it survives free living. Ambient temperature, clothing, bedding, activity and hand use all perturb distal skin temperature without any change in circadian phase. This study asks how much accuracy survives that noise, and in whom it does not.
Primary objective. To estimate agreement between the device-derived circadian phase estimate and laboratory DLMO in free-living adults, expressed as median absolute error in minutes.
Primary hypothesis. Median absolute error would fall below 30 minutes, the margin judged sufficient for scheduling light and sleep interventions.
Secondary objectives.
Prespecified moderator. Irregular ambient temperature exposure was identified in advance as the most plausible threat to the distal temperature signal and was collected as a three-item self-report measure of thermal variability at work and at home, dichotomized at the cohort median.
Prospective method-comparison study. Participants wore the device continuously during ordinary daily life and attended one laboratory DLMO session at two participating sleep laboratories between March 2025 and January 2026.
Eligibility. Inclusion: adults aged 18–60, employed at least six months on their current schedule, willing to wear the device continuously for 14 days. Exclusion: diagnosed circadian rhythm sleep-wake disorder, transmeridian travel across more than two time zones in the preceding month, medications affecting melatonin secretion, and pregnancy.
Ethics and consent. Approved by the Institutional Review Board of the coordinating site (protocol 2025-0417). All participants gave written informed consent.
The sample of 68 was determined by recruitment feasibility within the study window rather than by a formal power calculation. Under the observed dispersion, this size estimates the median absolute error with a 95% confidence interval of roughly ±5 minutes, which was judged adequate for the primary agreement objective.
Dual thermistor and 6-axis inertial measurement unit, sampled at 1 Hz continuously. No optical photoplethysmography. Device firmware version 2.4.1 throughout. Data were retained on device and exported weekly.
Saliva sampled every 30 minutes under dim light below 10 lux, beginning 6 hours before habitual sleep onset. Assay: direct radioimmunoassay at the coordinating site's laboratory, limit of detection 0.5 pg/mL. DLMO defined as the interval-interpolated crossing of a 4 pg/mL threshold. Sessions were scheduled relative to each participant's habitual sleep timing rather than to clock time.
A per-participant limit-cycle oscillator was fitted to the distal temperature series, with accelerometry used for wear detection and activity-masking correction. Phase estimates were generated blind to DLMO results and locked before unblinding.
Participants were classified as regular or irregular ambient exposure on the basis of the three-item thermal variability measure, dichotomized at the cohort median. Classification was completed before any agreement analysis was run. Twenty-one participants (30.9%) fell in the irregular group.
Primary analysis. Absolute difference in minutes between phase estimate and DLMO per paired session, summarized as median with interquartile range. The median rather than the mean is the appropriate summary here: phase error distributions are right-skewed and a small number of failed fits would otherwise dominate.
Agreement. Bland-Altman analysis of signed differences, reporting bias and 95% limits of agreement, with difference plotted against the mean of the two methods to test for proportional error. A correlation coefficient alone is not an agreement statistic and should not be reported as one.
Missing data. All 68 participants contributed a complete paired session. No sessions were lost to insufficient wear or failed assay, and no imputation was used; analyses are complete-case.
Subgroup analyses. Prespecified: schedule type, ambient exposure regularity. Report as effect estimates with intervals, not as significance tests on small strata.
Sensitivity analyses.
Software. R 4.4.1, using blander for agreement analysis and data.table for preparation. Analysis code available from the corresponding author on request.
The median absolute error between device-derived phase estimate and laboratory DLMO was 19 minutes across 68 participants.
Days of continuous wear required before the per-participant estimate stabilized: median 4 days (IQR 3–6). Participants failing to converge within 10 days: 3.
Agreement degraded among participants reporting irregular ambient temperature exposure.
All three prespecified sensitivity analyses preserved the primary estimate. Median absolute error ranged from 18 to 21 minutes across specifications, and removing the irregular-exposure subgroup reduced the cohort median from 19 to 15 minutes.
A median absolute error of 19 minutes is the central finding, and its meaning depends entirely on what the estimate is used for. For the interventions this method is intended to support — timing a light exposure block, placing an anchor sleep window, setting a caffeine cutoff — the relevant question is whether the error is small relative to the width of the intervention window. A light exposure block of 30 to 60 minutes tolerates a 19-minute phase error; a claim that depends on resolving phase to the quarter hour does not. The figure supports scheduling guidance. It does not support anything framed as diagnosis.
The ambient temperature finding is the more informative result, because it points at a mechanism rather than at variance. Distal skin temperature is a thermoregulatory signal that happens to carry circadian information. When ambient conditions change irregularly — a participant moving between a cold ward and a warm break room, sleeping under variable bedding, working outdoors — the thermoregulatory component moves for reasons unrelated to phase, and the oscillator fit inherits that movement. This is a known limitation of the signal and not a defect of the fitting procedure, which means it is addressable by measurement (ambient sensing, masking correction) rather than by better modelling alone.
Published laboratory comparisons of distal temperature phase markers against DLMO report agreement in a broadly similar range, but free-living estimates remain sparse and the protocols differ enough that direct comparison is difficult. Specific references and figures are to be inserted at the citation pass; this result should not be characterized as the first or the best of its kind without a literature search supporting that claim.
The practical implication for deployment is that the method should report its own confidence. A participant whose ambient exposure is irregular should receive a wider stated uncertainty, or no phase-dependent recommendation at all, rather than a point estimate carrying hidden error.
Reference standard variability. DLMO itself carries assay and protocol error. A portion of the 19-minute figure is reference noise rather than device error, and the study cannot separate the two without repeat DLMO sessions in the same participants.
Single paired session per participant. Each participant contributed one DLMO session. Within-person reliability is therefore unestimated, and the agreement figure cannot be distinguished from a single fortunate alignment in any individual case.
Self-reported ambient exposure. The moderator was classified by self-report rather than measured ambient temperature. The effect is therefore attributed to a proxy, and a device-based ambient measurement would test the mechanism directly.
Sample composition. The cohort skews toward hospital-based workers aged 25 to 45 recruited at two sites. Transfer to outdoor, industrial or transport settings, and to older workers, is untested.
Free-living confounds not controlled. Clothing, bedding, room temperature, alcohol, exercise and hand immersion all perturb distal skin temperature and were not measured.
Generalizability to the deployed product. Firmware, fitting procedure and wear position in this study were the same as the shipping configuration.
What this study does not establish. It does not establish that phase estimation improves any health outcome, that the method detects or characterizes any sleep or circadian disorder, or that guidance derived from the estimate changes alertness, error rates or sleep quality. Those require separate designs and are not addressed here.
This evidence supports general wellness positioning only. The distinction matters because agreement with a physiological reference standard is often mistaken for clinical validation, and the two carry different regulatory consequences.
External materials citing the 19-minute figure should state the sample size, the reference standard and the ambient-exposure limitation alongside it. A number quoted without its boundary condition invites the claim it does not support.
This positioning is a starting point rather than an approved one. Regulatory counsel and the Head of Science review it before any external use.
The paired records behind this analysis are assembled by an automated pipeline linking raw sensor exports to laboratory DLMO results. For the analysis to be reproducible, the pipeline must fix and version:
Data sharing. De-identified participant-level data and analysis code are available from the corresponding author on reasonable request, subject to a data use agreement.
All authors are employees of or hold equity in V3LA Inc. The device under evaluation is a company product. No external funding was received for this work.
General wellness product, not a medical device. Not intended to diagnose, treat, cure or prevent any disease.
← Back to publications