Concept, target, and classification boundary
“Based on melatonin” is the catalog shorthand here for rhythm prediction models based on melatonin. These models use the endogenous human melatonin rhythm, or validated related proxy signals, as a key state variable to estimate an individual’s circadian phase, phase change, and biological-night window. They fall under WSIC 4.4.4, Sleep Digital Twins and Personalized Prediction Models. Typical outputs include dim-light melatonin onset, rhythm peak or offset, the phase angle relative to sleep timing, an uncertainty interval for current phase, and a probabilistic forecast of future phase under a specified light–schedule scenario.
This entry is not synonymous with melatonin medication, a dietary supplement, a one-time laboratory test, or every sleep-rhythm model. A measured melatonin concentration, a phase fitted from sparse samples, a phase inferred indirectly from wearables, and a modeled change after an intervention occupy different evidentiary levels and must be distinguished in interfaces and reports. Circadian phase is also not the same as sleep quality, the cause of insomnia, or daytime function. Even an accurate phase estimate does not automatically determine an individual dose, prove that a treatment works, or replace clinical assessment of a circadian rhythm sleep–wake disorder.
Biological rationale, markers, and input data
Melatonin is regulated by the circadian system and acutely suppressed by light, so its rise under controlled dim conditions is an important reference for endogenous human circadian phase. Models may use serial or interval plasma or salivary melatonin, or a principal urinary metabolite. Concentration ranges, temporal smoothing, and assay properties differ by matrix, so thresholds cannot be transferred directly. DLMO is a common phase label, but the result depends on fixed versus baseline-relative thresholds, sampling interval, interpolation, quantification limits, and handling of anomalous points. Other markers such as peak and offset require different temporal coverage.
A hormone series alone may not predict future state. Inputs may also include ocularly relevant light exposure, sleep and wake, activity, posture, wrist or skin temperature, timezone, shift work, naps, meals, caffeine, medication, exogenous melatonin, illness, and device state. Light is a time series, not merely average daily illuminance; sensor location, spectrum, occlusion, and indoor–outdoor differences affect its ability to represent circadian stimulus. Activity and temperature can improve observability but remain proxies. A model should disclose when it has been calibrated by an actual biomarker and when phase is inferred only from behavior and sensors.
Sampling protocol, data engineering, and quality control
Phase-label quality depends on sampling design. The protocol should set a window around habitual sleep that is long enough to capture the expected rise and record planned and actual sample time, light, food or oral contamination, volume, storage temperature, freeze–thaw history, assay batch, and values below quantification. Home collection needs clear instructions and evidence of compliance with timing and light. Laboratory and home results should not be treated as equivalent without bridging evidence. If sampling begins too late, is too sparse, or fails to capture the rise, an indeterminate result or a wide uncertainty interval may be more valid than a forced precise clock time.
Data engineering should harmonize local time, Coordinated Universal Time, timezone, and daylight-saving rules while preserving the true sequence across midnight. Device streams need clock-drift and missing-data correction, and light, activity, sleep logs, and hormone samples should share one timeline. Lineage is needed for participant duplicates, assay-batch shifts, unit conversion, threshold-algorithm versions, and outlier rules. Exogenous melatonin can make the measured curve no longer represent endogenous secretion alone; medications that affect production or metabolism, acute light exposure, schedule change, and illness also need recording and protocol-based handling. Missingness should distinguish an uncollected sample, inadequate volume, assay failure, and sensor non-wear.
Model families, personalization, and scenario forecasts
Models include threshold or curve-fitting approaches, mechanistic circadian oscillators, state-space or Bayesian filters, machine learning, and hybrids. Threshold methods derive phase from an observed curve and are transparent but depend on a complete rising segment and threshold choice. Mechanistic models use light input and oscillator state to simulate phase advances or delays and suit scenario analysis, but are sensitive to parameters and light measurement. Data-driven models can combine activity, temperature, and sleep history, yet may learn population schedule correlations rather than endogenous timing. A hybrid can use sparse melatonin samples to calibrate a mechanistic state and continuous sensing to update it, but should identify which parameters are individualized and which remain population priors.
Personalization is not merely retrospective fitting to the same person. A baseline period may estimate usual phase, amplitude, light sensitivity, or behavioral regularity, followed by validation in a later independent period. Scenario forecasts should state assumptions about light timing and intensity, sleep opportunity, travel direction, timezone, and adherence, and present several plausible trajectories rather than one certain answer. A model that also simulates exogenous melatonin adds uncertainty from absorption, formulation, dose, administration time, and individual metabolism. Such a research model should not become automatic dosing advice without evidence and regulatory support appropriate to that use.
Validation, metrics, and generalizability
Validation should center on a melatonin phase reference matched to the intended use and separate training, tuning, and testing by participant. Reports may use circular-time error, absolute phase error, bias, uncertainty-interval coverage, unavailable-output rate, and error in phase change. They should also examine longitudinal tracking within individuals, day-to-day changes, and classification on either side of relevant clinical or operational thresholds. High population correlation does not mean that individual errors are small, and ordinary linear statistics can mislead when clock time wraps across midnight. A model that forecasts future phase should be validated by forecast horizon; fitting current DLMO does not validate prediction of later change.
External validation should cover age, sex, chronotype, season, latitude, shift and travel scenarios, regular and irregular schedules, sleep disorders, medication use, and measurement devices. Biomarker-calibrated and uncalibrated modes should be reported separately, with degradation and rejection tested when light, activity, temperature, or diary data are missing. A model trained mainly in healthy adults with regular schedules cannot be generalized from average performance alone to children, older adults, pregnancy, night workers, or people with psychiatric or neurological conditions. Assay laboratory, reagent batch, sensor position, and firmware changes also need examination for nonphysiological phase drift.
Applications, interaction, and safety limits
Models can support circadian research, personal biological-time profiles, comparison of jet-lag and shift-work scenarios, sleep-window prompts, research on timed light plans, and rhythm-state updates in a digital twin. User-facing output should display the source of the estimate, latest calibration, error range, data gaps, and scope, and express critical times in both local and biological time. If advice assumes that a user receives or avoids light on schedule, the system should record what occurred rather than assume perfect adherence. Travel, daylight-saving transitions, and plans spanning midnight require explicit error-resistant presentation.
An overconfident phase prompt could lead a user to schedule bright light, sleep, driving, or work at the wrong time, or be mistaken for a hormone-treatment instruction. Diagnostic use, treatment decisions, automatic control of a medical device, or high-risk scheduling require stronger clinical evaluation, human-factors validation, professional oversight, and an applicable regulatory determination. The model should reject or request new measurements when data are insufficient, the assay curve is anomalous, exogenous melatonin was recently used, or inputs fall outside the training domain. Marketing should not call an indirect inference a measured hormone value or promise precise hormone regulation, an insomnia cure, or guaranteed elimination of jet lag.
Privacy governance, updates, and future direction
Melatonin concentration, inferred phase, schedules, light, location, and shift information can jointly reveal health and daily routines. Collection should be minimized and purpose-limited, identity mappings separated, access controlled, and retention and deletion defined. Cross-study or institutional sharing should identify authorization for samples and derived models, re-identification, and commercial use. Precise sleep–work schedules in a report may create occupational or personal risk even after hormone values have been converted into phase. Retraining data require lineage for consent, source, measurement protocol, and version.
Model updates should distinguish a real phase change from an algorithmic change, using fixed regression sets and longitudinal bridging to prevent an unexplained jump for the same individual after upgrade. Future work includes active sampling with fewer specimens, fusion of continuous sensing with sparse biomarkers, interpretable individual light response, online uncertainty calibration, privacy-preserving multi-center learning, and incorporation of environmental and behavioral adherence into a digital twin. Reduced sampling burden does not remove the need for independent validation against measured phase in new devices, populations, and settings. Model complexity cannot turn a proxy into biological ground truth.