World Sleep Industry ENCYCLOPEDIA

Actigraphy Accelerometer Modules

6.1.1

Definition and position in a system

An actigraphy accelerometer module is a hardware and low-level firmware unit that converts body movement into a digital sequence of acceleration measurements. It may be used in wrist actigraphs, sleep research devices, bedside monitors, or other wearable systems. The sensing element is commonly a three-axis microelectromechanical accelerometer, while the module may also integrate power management, a clock, temperature sensing, memory, a microcontroller, and wired or wireless interfaces. Its direct output consists of motion and gravity components along each axis, or on-chip products such as activity magnitude, orientation, and interrupt events. Sleep, wake, and derived sleep parameters are later inferences made by an algorithm using time windows, placement, and sometimes other data; they are not physiological states directly measured by the accelerometer.

Actigraphy exploits the general tendency for movement to decrease during sleep, enabling relatively low-burden observation across multiple days. It differs from polysomnography, which uses electroencephalography, eye movements, muscle activity, and other signals to score sleep stages, and it differs from a step counter that records only selected movement events. A sensor advertised as low power with a hardware “sleep/wake” interrupt is usually describing an electronic power state, not a validated classification of human sleep and wakefulness. Whether a module can support wellness trends, research measurement, or clinical assistance depends on the intended use and end-to-end evidence for the finished system, not on the component name.

Sensor performance and mechanical integration

Selection requires joint consideration of full-scale range, sensitivity, resolution, noise density, zero-g bias, temperature drift, cross-axis sensitivity, linearity, output data rate, and effective bandwidth. A wider range can prevent clipping during vigorous movement but may reduce resolution for subtle movement. A higher sampling rate or lower-noise mode can increase power use and data volume. Output data rate is not identical to usable signal bandwidth; filters, decimation, and internal low-power sampling can cause aliasing or miss brief motion if not characterized. Nominal bit depth, acceleration per least significant bit, and theoretical resolution cannot substitute for measurements of assembled-system noise, mounting stress, and dynamic response.

Orientation, stiffness, and location within the finished device affect the signal. Printed-circuit-board flex, solder stress, enclosure resonance, strap tightness, left versus right wrist, dominant hand, skin contact, and rotation can all alter output. The design should define a coordinate system and mounting tolerances, record anatomical location and orientation, and be verified through drop, vibration, temperature, humidity, sweat, cleaning, and extended-wear conditions. Even when an algorithm uses vector magnitude, angle relative to gravity, or multiaxis features, clipping, bias, and rotation handling still require testing; an assertion of orientation independence is not a reason to omit mechanical verification. Skin-contact materials, edges, heat, and strap design are system safety and adherence issues rather than properties of the sensing die.

Acquisition chain, time integrity, and power

A reproducible acquisition chain records the units, axis order, sign convention, range, sampling rate, filters, activity thresholds, firmware version, and buffer mode for raw or processed data. Static response to gravity, bias, and axis consistency should be checked, with a documented factory, field, or autocalibration procedure. An on-chip first-in-first-out buffer can let the host processor sleep and read data in batches, but overflow, readout latency, packet reordering, and missed interrupts must be detectable. If a module exports only proprietary activity counts rather than raw samples, documentation should explain how counts are formed, where they saturate, and whether they remain comparable after a model change. An unauditable count should not be assumed equivalent to an identically named measure from another platform.

Multi-night recording is particularly dependent on time integrity. The system must address real-time-clock drift, failed phone synchronization, time-zone and daylight-saving changes, restart events, low battery, wireless interruption, and files that cross midnight, while preserving original timestamps and a correction log. A power budget should cover the sensor, processor, storage, indicators, and radio and make the trade-off between endurance and measurement quality explicit. If advertised endurance is obtained by reducing the actual sampling rate, the algorithm needs renewed validation. Missing, clipped, or duplicate samples, nonwear, charging, and deliberate removal require distinct quality flags and must not be silently imputed as sleep. Battery and storage warnings should arrive early enough for an operator to act before data are lost.

Converting movement into sleep measures

Traditional actigraphy often compresses continuous samples into activity for fixed epochs and applies a threshold, a weighted combination of neighbouring epochs, or a classifier to label sleep and wake. Newer methods may use raw multiaxis acceleration, changes in arm angle, individual timing patterns, light, or heart-rate information. Regardless of sophistication, the system should disclose its inputs, preprocessing, epoch length, sleep-window method, sensitivity setting, post-processing rules, and version. Bedtime and rise-time logs, event markers, or light measurements can help distinguish time in bed from the opportunity for sleep, but manual edits require an audit trail and should be retained separately from the fully automatic result.

Outputs may include estimated sleep onset, total sleep time, wake after sleep onset, sleep efficiency, rest-activity rhythm, and naps, but every report needs the operational definitions. Quiet reading in bed, motionless wakefulness during insomnia, paralysis, or limited mobility can be labelled sleep. Turns during sleep, periodic limb movements, childhood movement, or unstable wear can be labelled wake. Algorithms often have high sensitivity for sleep epochs but much lower specificity for wake epochs, so a high overall accuracy can largely reflect the prevalence of sleep in the validation sample. Movement alone does not directly identify electroencephalographic sleep stages, cortical arousals, dreams, apnea, or hypopnea, and nocturnal movement alone cannot diagnose periodic limb movement disorder.

Validation design and boundaries of use

The object of validation is the complete configuration: sensor model, range and sampling, mechanical location, firmware, signal processing, algorithm version, wear instructions, and intended population. Replacing the sensor, lowering the sample rate, changing the strap, applying an adult model to children, or moving from the nondominant wrist to a mattress can change performance and requires bridging evidence or revalidation. Technical verification examines time, units, noise, data loss, and repeatability. Analytical validation compares outputs with an appropriate reference. Clinical validation establishes that a measure safely supports a specified use in a specified population. A component data sheet or high correlation with a different device cannot complete the latter two tasks.

Sleep-wake algorithms are commonly evaluated against simultaneous polysomnography. Epoch-level reports should include sensitivity for sleep, specificity for wake, and a confusion matrix, while nightly outcomes should report bias and limits of agreement for total sleep time, sleep-onset latency, wake after sleep onset, and sleep efficiency. Correlation alone can hide a fixed bias, and pooling every epoch can hide poor performance for particular participants. Validation should represent the intended ages, body sizes, movement abilities, and wear conditions, as well as insomnia, shift work, daytime sleep, and relevant disorders, and should disclose exclusions, failures, and nonwear. Actigraphy can assist assessment of multi-day sleep patterns and some circadian rhythm questions, but it does not replace indicated clinical evaluation or polysomnography.

Quality, regulation, and data governance

Development should translate the intended use into testable requirements and establish risk analysis, design review, supplier controls, traceable hardware and firmware, production tests, change control, and complaint and adverse-event processes. Wellness use and claims of medical diagnosis or treatment may follow different regulatory paths. Once algorithms or connectivity are added, compliance normally concerns the finished device and its software lifecycle rather than the isolated accelerometer. Calibration references, fixtures, and acceptance limits should be traceable. Lot changes and substitute parts require assessment of bias, noise, power, clock behaviour, and downstream algorithm results.

Continuous movement can disclose sleep routines, work shifts, occupancy, and health-related patterns, making it sensitive information. The system should minimize collection, state the purpose, control access, protect data in transit and at rest, set retention periods, support appropriate deletion and export, and maintain audit logs. Wireless pairing, firmware updates, third-party components, and cloud interfaces belong in threat modelling, vulnerability response, and support-life planning; necessary security controls should not be disabled merely to extend battery life. Research and health-care deployment should also address consent, secondary use, international transfers, and disposition after withdrawal. Removing a name does not necessarily prevent re-identification from a long, precisely timestamped behavioural record.

Selection, deployment, and interpretation

Procurement should work backward from the intended use rather than begin with the largest bit depth or longest advertised battery life. General sleep-routine trends may prioritize comfort, endurance, and access to raw data, whereas clinical support or research may prioritize a validated configuration, accurate time, quality flags, a locked algorithm, and auditable export. Candidate modules can be compared with static multi-orientation tests, controlled motion, temperature and battery boundaries, repeat placement, and synchronized reference recordings. An acceptance protocol should prespecify limits for missingness, clock error, clipping, unit consistency, endurance, recovery, and lot-to-lot variation, and raw files should be retained for review.

Deployment includes instruction on wear, charging, event recording, skin discomfort, and fault reporting. Investigators and users should avoid compulsive review of a sleep score that can itself increase anxiety. Reports should identify the device, location, number of valid nights, nonwear rules, algorithm and version, data quality, and uncertainty. An unusual single night should first prompt review of wear and context, not an immediate disease interpretation. Lifecycle maintenance includes assessing firmware and algorithm changes, calibration sampling, cybersecurity updates, equivalent spare parts, and migration when a component is discontinued. The value of a high-quality module is a stable, traceable movement signal; clinical meaning still comes from a transparent algorithm, evidence in the applicable population, and qualified interpretation rather than the sensor's marketing label.