World Sleep Industry ENCYCLOPEDIA

Wearable Sleep Datasets

4.4.3

Definition, uses, and scope

A wearable sleep dataset organizes signals collected from devices worn on or close to the body during sleep and daily life together with device, wear, timing, quality, participant, and reference-label information. Wearables may include watches, bands, rings, adhesive patches, smart textiles, chest straps, and head-worn systems. Their sensors may measure acceleration, angular velocity, photoplethysmography or blood-volume pulse, electrocardiography, skin temperature, electrodermal activity, bioimpedance, sound, or ambient light. Such resources can support sleep-wake detection, sleep-stage estimation, circadian and schedule research, respiratory or cardiovascular signal analysis, device comparison, and algorithm reproduction.

“Wearable” describes a collection form, not a uniform measurement principle, quality grade, or clinical purpose. A table containing only activity counts, vendor-calculated heart rate, sleep stages, total sleep time, or a composite score is not equivalent to a dataset that exposes raw waveforms. Wrist heart rate is commonly derived from optical measurements and should not automatically be described as raw PPG. A dataset should state precisely whether it releases raw signals, processed signals, window-level features, device outputs, or investigator labels. These evidence layers should not be collapsed into a single notion of ground-truth sleep.

Device forms, sensors, and wear conditions

Hardware form determines tissue contact, motion artifacts, charging requirements, and tolerable wear duration. Optical paths at the wrist and finger, headband electrodes, chest electrodes, skin patches, and textile sensors do not produce interchangeable data distributions. Documentation should identify manufacturer and model, hardware revision, firmware and application versions, sensor channels, sampling rates, units, quantization, filters, automatic gain, timestamp precision, and on-device or cloud processing. When several units are used, device identifiers and their mapping to participants and nights are needed to separate unit effects from biological variation.

Placement and operation are also measurement conditions. Records should specify left or right side, dominant or nondominant wrist, ring finger, strap fit, patch position, electrode preparation, bedtime start and stop procedures, reattachment, and calibration. Device loss, off-wrist intervals, charging, bathing, skin irritation, Bluetooth interruption, and manual corrections should be logged. Temperature, perfusion, skin characteristics, hair, tattoos, sleep position, and bedding can affect particular sensors. Without this context, apparent model differences may reflect wear practice or hardware rather than sleep physiology.

Data layers, timelines, and reference labels

One resource may contain high-frequency raw waveforms, device-filtered sequences, per-second or per-epoch summaries, proprietary sleep stages, night-level summaries, and questionnaires. Each field needs a definition, unit, valid range, missing-value code, computational origin, and version. Publishers should retain the closest available representation to raw data and the code or an interpretable procedure used to derive later products. If only derived data can be released, irreversible processing should be identified. Sleep, recovery, or “stress” scores often depend on undisclosed algorithms and are not interchangeable with directly measured physical signals.

Reference labels may come from technician-scored synchronized polysomnography, a simplified EEG headband, annotated clinical events, a sleep diary, button presses, or another automated system. These sources differ in reliability and in what task they define. Alignment documentation should cover the shared time source, time zone, daylight-saving transitions, startup offsets, clock drift, resampling, segment joins, and any date shifting. Aligning only the first file row can conceal drift of minutes. Privacy-preserving date shifts must retain relative order and logic across midnight. Label epochs, sensor windows, and event start and end times require an explicit mapping.

Adherence, missing data, and quality control

Multi-night home collection is valuable because it approaches habitual conditions, but missingness is not random noise. Forgetting to wear a device, charging, discomfort, skin reaction, travel, network or synchronization failure, exhausted storage, a closed application, or removal during severe symptoms can each cause partial or whole-night gaps. Studies should report planned participants and nights, recordings actually started, successful uploads, usable wear duration, reference-label availability, and the final analyzed sample. Nonwear, quiet wakefulness, sleep, hardware failure, and algorithm rejection must be distinguishable states.

Quality control may combine signal range, saturation, discontinuities, movement, contact or off-wrist flags, agreement across sensors, and manual review. Thresholds, order of operations, and exclusion reasons should be disclosed. Zero filling, forward filling, or smoothing cannot turn absence of signal into a valid sleep observation. Removing only participants or nights that fail creates adherence and survivorship bias. Evaluation should therefore report coverage, rejection rate, missing-night frequency, and performance by quality stratum in addition to accuracy. Corrected files, realignment, and relabeling require traceable versions and change logs.

Population coverage, firmware drift, and cross-device harmonization

Representativeness depends both on recruitment and on who is willing and able to wear a device consistently. Age, disease, mobility, shift work, medication, sleeping environment, skin characteristics, peripheral circulation, and familiarity with digital devices may alter either data quality or sleep patterns. Reports should provide participant counts, nights per person, demographic and clinical composition, device distribution, and failure reasons, then examine coverage and error across relevant groups. Skin tone and socially defined race are different variables; the latter is not a precise substitute for the former.

Consumer hardware, firmware, and cloud algorithms change over time, and the computation behind a field can change while its name remains the same. Datasets should retain hardware revision, firmware, application, algorithm or export version, export date, and update batch. Measurements before and after an important update may need to be treated as separate domains. Cross-device harmonization should not compare only final sleep scores; it should examine raw or closest-to-raw signals, valid coverage, timing offsets, and reference error within shared intervals. A calibration derived for one model, placement, or population cannot be transferred without validation.

Machine-learning splits, evaluation, and reproducibility

Training, validation, and testing should be split by participant. Repeated nights from one person, members of the same household when shared context matters, overlapping windows, different sensor views of the same night, and augmented copies should remain in the same partition. Random epoch splitting produces serious leakage through individual physiology, device signatures, and adjacent temporal context. Normalization, imputation, quality thresholds, feature selection, and decision thresholds must be learned only from training data. Beyond an internal test, useful studies evaluate a different model, site, population, firmware period, or genuinely independent home setting.

Metrics must match the task. Sleep-wake and staging models need class-specific performance, confusion patterns, and agreement; night-level quantities such as total sleep time need bias, error distributions, and limits of agreement; event detectors need a stated matching tolerance and false positives per hour. Researchers should also count participants and nights for which the system emits no result, because rejecting difficult cases can create an inflated score. Reproduction materials should fix dataset and label versions, participant-level split lists, preprocessing, software environment, random seeds, final thresholds, and a locked test set. Repeatedly choosing a device or model after inspecting the test set invalidates that separation.

Privacy, licensing, and clinical boundaries

Continuous wearable records can reveal sleep schedules, activity, location patterns, cardiovascular rhythms, health status, and life events. Linkage to accounts, phones, household devices, or environmental sensors increases re-identification risk. Before collection, participants should know whether data go to a vendor cloud, who can access them, how long they are retained, whether they may support later research or model training, and how withdrawal, export, or deletion works. Releases should minimize data, remove direct identifiers, assess combined time-and-location risk, and use open, application-based, or controlled access according to sensitivity. Users must follow license, confidentiality, and no-reidentification conditions.

Consumer sleep outputs, general-wellness functions, and medical uses cannot be distinguished by appearance alone. Their intended use, claims, target population, hardware and algorithm version, validation evidence, and applicable regulatory requirements determine the boundary. An inadequately validated stage, oxygen, breathing, or heart-rate alert should not by itself diagnose or rule out disease, change treatment, or automatically replace polysomnography and professional assessment. Research reports should disclose evidentiary limits; clinical support also requires failure warnings, an understandable escalation path, and qualified review. A person with persistent sleepiness, signs of sleep-disordered breathing, or other health concerns should seek assessment from a qualified healthcare professional.