World Sleep Industry ENCYCLOPEDIA

Actigraphy Detection Algorithms

4.4.1

Definition, task levels, and intended-use boundary

Actigraphy detection algorithms use acceleration and orientation changes measured at the wrist, ankle, waist, or another site to identify movement, its intensity, and activity–rest patterns. In sleep applications, a continuous signal is commonly transformed into epoch-level activity features before estimating sleep windows, sleep–wake states, and summaries such as total sleep time. The sensor detects mechanical motion, not sleep as defined by brain electrical activity.

Movement detection, activity counts, sleep–wake classification, and sleep-summary calculation are distinct tasks. A turn can produce large movement without a cortical awakening, while a person lying awake without moving can be labeled asleep. An algorithm should state its output level, temporal resolution, sensing location, applicable ages, and setting; a downstream sleep label must not be presented as a directly measured sleep stage.

Actigraphy is well suited to low-burden recording across many days and nights and can support assessment of schedules, insomnia, circadian patterns, and insufficient sleep. It can also estimate sleep time for some home testing workflows. It does not replace polysomnography when that test is indicated, and limb motion alone cannot diagnose periodic limb movement disorder, epilepsy, or REM sleep behavior disorder. Consumer sleep scores and clinical actigraphy are not automatically interchangeable.

Sensing, activity representation, and classification models

A triaxial accelerometer records specific force on each axis. Sample rate, range, resolution, analog or digital filtering, and sensor orientation all alter the waveform. Gravity can support posture estimation, while dynamic components describe movement; common features include vector magnitude, axis variation, angle change, zero crossings, variance, and spectral energy. Saturation, clipping, or packet loss during acquisition cannot be fully repaired downstream.

Early devices often compressed raw acceleration into dimensionless activity counts, but counts from different manufacturers, models, filters, and epoch lengths are not a universal unit. Traditional methods such as Cole–Kripke and Sadeh weight counts from the current and neighboring epochs and apply a threshold. Raw-signal approaches may use arm angle, heuristics, state models, or machine learning. Sharing a formula name does not guarantee identical count inputs or implementation.

Temporal context suppresses isolated impacts and represents persistence around sleep onset or awakening, but it creates edge effects and latency. Postprocessing often relabels short segments according to sustained-sleep or sustained-wake rules. A threshold favoring sleep will generally increase sleep sensitivity at the expense of wake specificity. Model complexity cannot remove the fundamental information gap between movement and sleep physiology.

Preprocessing, wear detection, and sleep windows

Preprocessing should verify timestamps, time zone, daylight-saving changes, sample intervals, units, axis orientation, and device-clock drift before applying documented filtering, gravity calibration, and epoch generation. Resampling and epoch aggregation change the amplitude and duration of brief movement, so thresholds cannot be transferred directly across epoch lengths. Raw, cleaned, excluded, and inferred data should remain traceable.

Device removal, loose wear, charging, leaving a sensor on a bedside table, and genuine prolonged stillness can produce similar signals. Nonwear detection may combine sustained low variability with temperature, contact or capacitance sensing, and diaries, but each method has errors. A system should distinguish nonwear, missing signal, valid immobility, and unknown state rather than automatically filling an absent interval as sleep.

A sleep window may be derived from an event marker, sleep diary, light, angle change, activity pattern, or multiple signals. Restricting classification to a predefined time-in-bed window is a different task from finding the main sleep episode in a full 24-hour record. Shift work, naps, fragmented sleep, bed rest, and intervals crossing midnight require multiple candidate windows and preserved evidence for human review.

Validation design and performance metrics

Epoch-by-epoch sleep–wake validation usually uses synchronized manually scored polysomnography as the reference and prespecifies how PSG epochs map to algorithm epochs. A clock offset of tens of seconds can misalign brief awakenings, so studies need common triggers, synchronization records, or auditable retrospective alignment. A diary provides time-in-bed and behavioral context but is not an equivalent truth for sleep physiology.

Evaluation should report sleep sensitivity, wake specificity, balanced accuracy, F1 or another metric appropriate for class imbalance, and the confusion matrix. Overall accuracy is often dominated by the large number of sleep epochs and can hide poor detection of wake during the night. Threshold selection belongs in an independent validation set, and participant-level distributions should accompany pooled-epoch results.

Total sleep time, sleep-onset latency, wake after sleep onset, sleep efficiency, and sleep onset and offset need bias, limits of agreement, and error distributions. Correlation does not establish agreement, and a near-zero average error may cancel errors in opposite directions. One-night PSG assesses epoch performance, while multiday home recording assesses adherence and rhythm outputs; neither evidence type substitutes for the other.

Error sources, population differences, and generalization

Quiet wakefulness is the characteristic source of false sleep in actigraphy and may be prominent with insomnia, pain, bed rest, or reading. Turns during sleep, tremor, caregiver handling, and vibration transferred from a bed partner can create false wake. Wrist activity is not the same as whole-body activity, and wrist restraint, assistive devices, paralysis, or abnormal movement disrupt usual mappings. Quality control must permit an indeterminate state instead of forcing every epoch into two classes.

Age, development, body size, handedness, wear side, mobility, neurologic disease, medication, and sleep disorders change movement distributions. Children, older adults, inpatients, shift workers, and healthy adults should not be assumed to share one threshold. Validation samples should cover the intended population, including low sleep efficiency, daytime sleep, and natural behavior at home, and should report performance for important subgroups.

Data must be partitioned by participant, with adjacent nights from one person kept in the same partition; otherwise personal movement signatures leak into testing. External evaluation should also vary sites, device lots, and operating workflows and test transfer across models or firmware. A model trained on one proprietary activity-count definition cannot be applied to another count stream or to raw acceleration without validation.

Output interpretation, transparent reporting, and version governance

Algorithm output should identify valid wear time, missing fraction, the source of each sleep window, threshold or model version, and confidence. Sleep onset, offset, and total sleep time can support longitudinal pattern assessment, while sleep stages or microarousals inferred from acceleration alone need additional evidence. Interfaces should visually and semantically separate measured movement, algorithmic inference, and aggregate scores.

A reproducible report includes device and placement, sample rate and range, calibration and filters, the activity-count definition, epoch length, nonwear rules, sleep-window method, classifier, postprocessing, missing-data handling, and software version. Excluded nights, the proportion manually edited, and complete failures should also be disclosed. Reporting only an accuracy value or brand name is insufficient for reproduction.

Firmware, filtering, count conversion, or cloud-model updates can shift a longitudinal baseline. Deployment therefore needs version locking and regression tests, followed by a bridging assessment when an update occurs and visible marking of dates that are not directly comparable. Input distributions, failure rates, and subgroup bias should be monitored, but online data must not silently rewrite clinical results.

Clinical use, privacy, and safety

In clinical workflows, actigraphy should be interpreted with history, sleep diaries, symptoms, and appropriate testing. A multiday trend may reveal schedule irregularity or divergence from subjective reports, but it cannot establish the cause by itself. If the output will change diagnosis, treatment, or a safety decision, the system needs validation and regulatory status appropriate to that intended use.

Continuous activity records can disclose sleep timing, work routines, presence at home, and illness behavior, making them sensitive health and lifestyle data. Collection should follow data-minimization principles and explain local or cloud processing, access, retention, export, and deletion. Data used for retraining or sharing require appropriate authorization and protection against reidentification from temporal patterns.

Algorithm failure should not produce a falsely complete night or deterministic advice. The system should degrade explicitly during low battery, detachment, clock errors, inadequate signal, or use outside the target population and preserve a path for human review. Persistent insomnia, marked daytime sleepiness, unusual nocturnal behavior, injury risk, or respiratory symptoms warrant professional assessment and cannot be excluded by “normal movement.”