Definition, outputs, and task boundaries
Circadian phase estimation algorithms infer where an individual's endogenous circadian oscillator lies at a specified time. An output may be the clock time of a reference event such as dim-light melatonin onset or the core body temperature minimum, or it may be a phase angle, probability distribution, or uncertainty interval over a twenty-four-hour cycle. The system must name the estimated marker, date, time zone, and reference coordinate rather than returning only an "early" or "late" label.
Endogenous phase is not identical to sleep onset, wake time, sleep midpoint, or questionnaire chronotype. Sleep timing is also shaped by homeostatic sleep pressure, employment, caregiving, alarms, medication, and choice; it can become especially dissociated from the clock during shift work, travel, or free-running rhythms. Phase, period, amplitude, and the phase angle between two rhythms are separate quantities, so one phase estimate cannot establish intrinsic period or rhythm strength.
An estimate may support research, assessment of circadian rhythm sleep-wake disorders, or discussion of intervention timing, but it is not an independent diagnosis. The same wall-clock time means something different on another date, in another time zone, or across a daylight-saving transition, while similar symptoms can arise from insufficient sleep, insomnia, sleep-disordered breathing, mood conditions, or medication. Intended use must be explicit, with measured observations, inferred phase, and clinical interpretation shown as separate layers.
Reference markers and sampling conditions
The melatonin rhythm is a common human phase marker, but saliva, plasma, and urinary metabolite protocols are not interchangeable. DLMO is usually assigned when serial samples cross and remain above a prespecified absolute or individualized threshold; the selected threshold, assay detection limit, smoothing, and interpolation rule can change the result. Reports must retain specimen times, specimen type, assay platform, threshold definition, and the reason when an onset is indeterminate.
Light can suppress nocturnal melatonin, so laboratory or home DLMO collection must control ocular light exposure, sampling cadence and window, food, posture, and compliance. Exogenous melatonin, some medications, caffeine, tobacco, alcohol, oral contamination, missed samples, and out-of-range measurements can distort the profile. If the rise lies outside the sampling window or the curve is flat, multi-peaked, or noisy, the algorithm should report censoring or uncertainty rather than force a timestamp.
Core body temperature minimum, cortisol, peripheral temperature, and molecular markers can also inform phase, but they are affected by sleep, activity, meals, posture, illness, and sampled tissue. Core temperature under a constant-routine protocol better separates the endogenous component, whereas ambulatory skin temperature contains strong behavioral masking. Relationships between markers may be orderly without being fixed, and every model must be validated against the particular reference marker it claims to estimate.
Wearable inputs and the time axis
Dynamic models may drive a circadian oscillator with a person's light history, while statistical or machine-learning models may use activity, sleep, skin temperature, heart rate, or multimodal rhythms. A wrist light sensor measures illumination at the wrist, can be covered by clothing or bedding, and does not equal spectral irradiance reaching the eyes. Devices should preserve sensor placement, spectral response, saturation and dark noise, calibration information, and wear state.
Activity and sleep timing can indirectly describe regular behavior, but they can also encode a work schedule rather than the endogenous clock. When activity substitutes for light, the assumption and training population must be stated; a model trained in regular day-working adults cannot be transferred without evidence to night workers, bedbound people, people with impaired vision, children, or highly irregular schedules. Missingness, removal, charging, quiet wakefulness, and device-clock drift require distinct quality states.
The timeline must preserve offset-aware local time and UTC, named time zone, daylight-saving transitions, day boundaries, travel location, and work shift. Naive clock times create duplicated intervals when clocks move backward and gaps when they move forward. Sampling latency and drift require correction before devices are fused; after travel across longitude, phase is a state evolving by date and cannot be handled by merely shifting file timestamps.
Models, initialization, and uncertainty
Physiology-based models usually describe the phase-dependent effect of light on a circadian oscillator and advance its state through prior light history; statistical models can directly predict a reference phase from sleep midpoint, activity rhythms, or multimodal features. Both approaches depend on initialization, parameters, and history length. Initialization at a population-average phase may work in regular schedules but can create large errors in shift work, jet lag, and non-24-hour rhythms.
Phase is circular: 23:50 and 00:10 are twenty minutes apart, not twenty-three hours and forty minutes apart. Training loss, averages, uncertainty intervals, and error statistics must use circular methods or an explicit unwrapping convention. A multimodal posterior may represent two plausible phases and should not be collapsed into a physiologically implausible midpoint; outputs should retain the estimated distribution, confidence, and time of the newest supporting data.
Personalization may update a model with an earlier DLMO, long-term light-activity history, or repeat measurements, but calibration must be distinguished from genuine phase movement. The same reference assessment cannot be used both to fit a person-specific offset and to claim independent validation. Adaptive updates need bounded parameters, version records, and rollback criteria; when observations conflict strongly with the model, uncertainty should increase and review should be requested rather than silently following noise.
Data design and reference alignment
Development data should span regular schedules, delayed and advanced phases, shift work, travel, and seasons, with variation in age, sex, light environment, visual function, medications, and comorbidities. Training, tuning, and testing must be separated by participant and collection episode so adjacent days from one person do not cross partitions. If only participants with a clean DLMO profile are included, excluded participants must be described to avoid exaggerating real-world yield.
Wearable inputs and the DLMO reference must correspond to a clearly defined history window, and investigators should check whether the laboratory assessment day changed usual light, sleep, or behavior. The label algorithm should be fixed before predictions are viewed, and manual revision should be blinded to model output. Home sampling requires recorded light conditions and actual time for every tube; relying only on planned times or recall turns compliance error into apparent model error.
When sleep timing, activity, and light all enter a model, degradation under loss of each signal should be tested. Imputation must not smooth across travel, schedule changes, or daylight-saving jumps. Augmentation and simulated light histories can expand development conditions but cannot replace external testing in real populations; evaluating one model against labels generated by the same mathematical assumptions establishes internal agreement, not physiological accuracy.
Validation, comparison, and reporting
Validation should use a prespecified DLMO, core-temperature phase, or other stated target and report signed circular error, absolute error, bias, limits of agreement, and the fraction within clinically meaningful error windows. Correlation mainly measures ordering and cannot demonstrate agreement for an individual's time. Coverage, abstention rate, and calibration of uncertainty intervals are also required so apparent accuracy cannot be achieved by rejecting difficult cases.
Performance should be stratified by shift status, schedule regularity, season, age, skin tone or device optical conditions, visual function, medication, collection site, and device version, with participant-level and repeat-night errors displayed. Comparison with simple baselines such as sleep midpoint, questionnaires, or a population mean shows whether a complex model adds value. External validation should use independent sites, device lots, and laboratory assay workflows.
Longitudinal use must validate change in phase as well as each isolated estimate. Subtracting two noisy estimates compounds uncertainty, so reports must propagate error from both assessments and distinguish biological change from measurement variation. Systems used to inform intervention scheduling should also evaluate estimation latency, stale data, failure alerts, and decisions after human review, rather than presenting only retrospective mean error.
Deployment, interpretation, and safety boundaries
A deployed interface should show the target phase marker, estimate date, time zone, input coverage, latest synchronization, uncertainty interval, and major causes of missingness. Users need to know whether the result comes from biological samples, wearable proxies, or a hybrid model. Firmware, light sensors, sleep algorithms, assays, and model updates can all create longitudinal discontinuities and therefore require version tracking, bridging tests, and drift monitoring.
Light, activity, location, travel, shift, and biospecimen timestamps can reveal work and daily routines and are sensitive health and behavioral data. Collection should be minimized, with clear rules for on-device processing, upload, access, retention, deletion, and secondary training. In shared or employment settings, a low-confidence phase estimate must not drive discipline, fitness-for-duty, scheduling, or insurance decisions, and missing data must not be treated as nonadherence.
Phase estimates must not automatically generate individualized timing or doses for bright light, light avoidance, melatonin, sedatives, or stimulants. A timing error can move phase in the direction opposite to that intended, and light and medication have person-specific safety constraints. Persistent severe sleepiness, driving risk, marked mood or functional deterioration, or suspected alternative sleep or medical conditions requires appropriate professional assessment; a trustworthy algorithm presents evidence and uncertainty rather than replacing care.