World Sleep Industry ENCYCLOPEDIA

Skin Temperature Rhythm Recognition Algorithms

4.4.1

Definition, measured quantity, and task boundaries

Skin-temperature rhythm recognition algorithms estimate daily repeating patterns, phase, amplitude, stability, and fragmentation from continuous temperature measured at the wrist, finger, foot, chest, or another body-surface location. The measured quantity is the thermal state where local skin and sensor meet. It reflects internal rhythmic regulation but is also shaped by skin blood flow, environment, clothing and bedding, posture, activity, and how the device is worn.

Skin temperature is not core body temperature. Distal sites such as the hands and feet commonly warm with presleep vasodilation while core temperature tends to decline, and proximal sites on the torso have another pattern. An algorithm must identify the site and distinguish distal temperature, proximal temperature, the distal-to-proximal gradient, directly measured core temperature, and model-estimated core temperature instead of combining them under one generic “body temperature” label.

A 24-hour temperature pattern in free living contains endogenous circadian regulation together with masking by sleep, light, activity, meals, bathing, room temperature, and other behavior. Stronger attribution to an endogenous rhythm requires constant routine, forced desynchrony, or comparison with an accepted phase marker. Ordinary wearable output is better described as a daily rhythm measure and should not independently diagnose a circadian sleep-wake disorder, fever, or another disease.

Sensors, wearing conditions, and thermal coupling

A contact sensor depends on sensing-element accuracy, response time, contact area and pressure, insulation, and heat generated by the device. A loose wristband admits ambient air, an overly tight band may alter local blood flow, and removal rapidly drives the reading toward room temperature. The system should retain device model, calibration, sampling interval, site, strap state, and firmware version and should identify nonwear and reapplication.

Measurements from different sites cannot be interchanged with a fixed offset alone. Distal skin is more responsive to vasomotion and environment, proximal sites are shaped by clothing and torso heat, and left-right differences can arise from side-sleeping pressure or an exposed arm. Computing a distal-to-proximal gradient requires synchronized sensors, consistent units, and stable site definitions. When either side is missing, a population average should not fabricate an individual gradient.

A skin-temperature device may combine ambient temperature, heat flux, activity, and posture to help distinguish physiological from external change. A multisensor estimate still needs independent validation and must not be labeled a direct measurement. Factory calibration, between-device bias, long-term drift, and response lag should be checked separately with a controlled temperature reference and repeated measurements on people.

Preprocessing, missingness, and masking factors

Preprocessing commonly includes time alignment, unit checks, outlier and step detection, short-gap handling, smoothing, and wear-quality labels. Temperature already changes slowly, so aggressive smoothing can shift the timing of a presleep rise or morning decline. Interpolation can also turn removal, bathing, or charging into an artificial rhythm. Raw values, corrected values, and reasons for missingness should remain separate.

Ambient temperature, bedding microclimate, clothing, a hot bath, exercise, food, alcohol, fever, medication, menstrual phase, and peripheral vascular state can alter level or rhythm. Reduced activity and recumbency also change wrist temperature and coincide with sleep time. Available covariates should be recorded and used for stratification, adjustment, or sensitivity analysis rather than assigning every nocturnal increase to sleep or the circadian clock.

Multiday rhythm analysis needs minimum valid coverage distributed across day and night; night-only wear cannot determine a full 24-hour waveform. Missingness concentrated during daytime, exercise, or heat is not random, and simple mean imputation can distort phase and amplitude. Quality rules should prespecify a valid day, maximum gap, and cross-day stitching, and reports should state how much data was excluded.

Rhythm modeling and feature output

A fixed-period cosinor can estimate a midline level, amplitude, and acrophase, while extensions can add harmonics, participant random effects, or a variable period. It supports comparison, but actual skin-temperature profiles are often nonsinusoidal, with a broad or double peak and changes driven by sleep timing. Reports should include goodness of fit, confidence intervals, whether the period was fixed, and how phase was handled on a circular time scale.

Nonparametric methods can obtain high and low intervals and their timing, relative amplitude, interdaily stability, and intradaily fragmentation from an average daily curve or moving window without assuming a sine wave. Results still depend on sampling interval, binning, window length, and valid days. Definitions based on fixed most-active or least-active intervals in activity analysis cannot be transferred without explanation to distal temperature, whose daily direction may be reversed.

Change-point, state-space, spectral, wavelet, and machine-learning methods can track phase drift, presleep warming, or multiday patterns, but they must use circular error for the 24-hour clock. A prediction of 23:50 and a reference of 00:10 differ by 20 minutes, not almost 24 hours. Output should include uncertainty, the valid data interval, and a no-discernible-rhythm state instead of forcing a precise phase from a low-amplitude or irregular record.

Sleep association and multimodal interpretation

Near sleep onset, distal vasodilation, rising distal skin temperature, and convergence of the distal-to-proximal gradient are often observed, but environment, recumbency, and individual differences modify these relationships. Temperature can contribute a feature to a sleep-window or sleepiness model but cannot independently confirm EEG-defined sleep onset, NREM, or REM. When polysomnography is the target, synchronization and transition-specific error require direct evaluation.

Combining temperature with activity, light, heart rate, and self-reported sleep can better distinguish wear, rest, and schedule changes, but it can also create circular labels. For example, using activity to generate “sleep truth” and then including the same activity in a multimodal model overstates independent information. Training and validation should identify the source of every label and use sensor ablation to quantify the incremental contribution of temperature.

Association between rhythm phase and sleep timing does not make them interchangeable. Shift work, travel across time zones, sleep restriction, naps, and irregular schedules can temporarily misalign temperature, behavior, and the light-dark cycle. Algorithms should report temperature phase, sleep schedule, and their difference separately and validate any claim about a biological clock or phase shift against an appropriate core-temperature or melatonin phase reference.

Validation, generalization, and reporting

Sensor-level validation should measure bias, repeatability, response time, and between-device agreement against a controlled or traceable reference, while physiological validation compares the intended site or accepted phase marker. Correlation does not demonstrate agreement. Algorithmic phase also needs circular error, limits of agreement, and night-to-night repeatability, and estimated core temperature must be evaluated separately from direct core measurement.

Participant-level data partitions should represent age, sex, skin tone, body size, menopausal status, peripheral circulation, chronic conditions, medication, season, climate, clothing and bedding, and shift schedules. Randomly splitting adjacent days leaks an individual baseline. External evaluation should use other device lots, sites, and seasons and separately report nonwear, poor contact, and failure to fit a rhythm.

Reports should identify site, sensor, insulation and wearing method, sampling interval, time zone and daylight-saving handling, ambient measurements, missing-data rules, valid days, model formula, period, phase definition, and uncertainty. Personal trend comparisons also need a defined baseline window and update rule. Absolute temperature, deviation from personal baseline, and distal-to-proximal gradient should not be mixed when units or meanings are unclear.

Deployment, privacy, and safety boundaries

A real-time system must handle travel across time zones, daylight-saving changes, device clock drift, removal for charging, and windows crossing midnight so phase does not jump because of faulty time metadata. Firmware, sensor, or fitting-model updates should be checked on fixed data for discontinuities in level, phase, and valid coverage. Longitudinal trends must retain version, site, and wearing changes.

Temperature and routine data can reveal sleep times, living patterns, illness, or physiological cycles. Systems need data minimization, access control, encryption, retention, and deletion and should state which results are generated on the device. A user interface should display measurement site, data quality, change from personal baseline, and uncertainty rather than present an inferred value as precise core body temperature.

An unusual skin-temperature rhythm may result from environment, wear, a short-term schedule, or physiological change and cannot independently diagnose a circadian disorder, infection, endocrine condition, or vascular disease. Persistent fever, marked hypothermia, serious sleep difficulty, or other symptoms require appropriate clinical measurement and assessment. A trustworthy algorithm separates direct measurement, corrected skin temperature, derived rhythm, and clinical interpretation and permits an indeterminate result.