World Sleep Industry ENCYCLOPEDIA

Radar Sleep Staging Algorithms

4.4.1

Definition, observed phenomena, and classification boundary

A radar sleep-staging algorithm is software that uses radio-frequency reflections from a person and their changes over time to estimate wakefulness and sleep stages in successive windows. Radar can sense chest and abdominal respiratory motion, weaker heartbeat-related micromotion, turning, limb activity, and bed occupancy, then infer stages from statistical relationships between these proxies and sleep. It does not directly record electroencephalography, eye movements, or chin electromyography. Even if it outputs W, N1, N2, N3, and R, it has not observed the electrophysiologic features on which conventional staging is based.

A system may distinguish only sleep from wake, or report W/NREM/REM, W/light/deep/REM, or finer classes. Class names, epoch duration, temporal context, and merge rules must be stated with the reference mapping. Stillness is not synonymous with sleep, and regular breathing is not sufficient evidence for a particular stage. Presence detection, vital-sign estimation, sleep staging, posture recognition, fall alerts, and respiratory-event detection are distinct tasks that require separate validation. This entry addresses radar staging; it does not duplicate the preceding PPG optical chain or replace the following EEG staging mechanism.

Radar mode, installation geometry, and acquisition chain

Common architectures include continuous-wave Doppler, frequency-modulated continuous-wave, and impulse-radio ultra-wideband radar. They differ in range resolution, velocity or phase information, antenna configuration, multi-person separation, and clutter handling, so performance does not transfer merely because each device is called radar. Development records should identify the frequency band, bandwidth, transmit and receive channels, antennas and beams, frame rate, gain, raw or intermediate data format, firmware, and calibration state. The finished device must meet applicable regional radio authorization, exposure, and electromagnetic-compatibility requirements, but regulatory radio compliance does not establish valid sleep staging.

Mounting height, distance, elevation and azimuth, position relative to the bed, field of view, and obstruction alter the return. Bed frames, mattresses, bedding, canopies, furniture, walls, and room dimensions create different reflections and multipath. Instructions should define supported beds, single- or multi-person conditions, installation tolerances, clear zones, first-use calibration, recalibration after movement, and startup self-checks. If the algorithm fuses ambient light, audio, mattress pressure, or another sensor, each input, synchronization method, missing-data policy, and contribution must be disclosed. A fused system's results should not be attributed to radar alone.

Target binding, echo processing, and model pipeline

A processing chain commonly includes direct-current or static-clutter suppression, range-gate or beam selection, I/Q correction, phase processing, motion-segment identification, target tracking, and signal-quality assessment. Algorithms may extract respiratory rate and variability, breathing morphology, heartbeat-related energy, movement density, posture transitions, and occupancy sequences from torso displacement and activity, or learn directly from range-time, range-Doppler, and multi-antenna representations. Models may use past-only windows, bidirectional context, or whole-night information. They must state whether output is real-time, near-real-time, or retrospective; performance from a future-aware model cannot support an equivalent real-time alert claim.

Finding that someone is in a room and continuously binding reflections to the same sleeper are prerequisites to staging. Large movement, bed exit and return, a caregiver approaching, a partner turning, or crossing targets can cause a track switch. Every epoch should retain target-identity status, echo quality, obstruction, gross motion, missingness, and rejection. A lost target should be reacquired and the gap marked rather than carrying the previous stage forward or joining another person's data. Respiratory and cardiac components separated or reconstructed by an algorithm should likewise retain provenance and uncertainty.

Reference labels, data separation, and scenario representation

Supervised labels should preferably come from synchronized PSG scored under an identified rule version by qualified, blinded personnel, with single scoring, double-scoring agreement, or adjudication described. Radar and PSG need a traceable common clock; startup latency, drift, dropped frames, resampling, or software-filled samples can pair an echo with the wrong epoch. Class merging and invalid-segment handling should be fixed before results are examined. Validation of respiratory rate or heart rate does not establish staging accuracy, and agreement in whole-night total sleep time does not establish epoch-level stage agreement.

Training, tuning, and final testing must be separated by participant, with different nights and segments from one person kept in the same partition. Independence should also be examined by room, site, device, and acquisition batch so a model does not learn installation signatures. Data should represent the intended age, body habitus, sleep position, sleep efficiency, cardiorespiratory rhythm, medication, sleep disorder, mobility, and nighttime-care scenarios, as well as supported beds, bedding, and layouts. Results in healthy young adults sleeping alone in a laboratory do not automatically extend to children, older adults, frequent awakening, respiratory or movement disorders, inpatient rooms, or shared bedrooms.

Performance, coverage, and night-level agreement

Epoch-level evaluation should report confusion matrices, per-stage recall, precision, F1, macro averages, chance-corrected agreement, wake specificity, and participant-level distributions. It should also report the proportions of time that are localizable, identity-bound, echo-valid, rejected, and ultimately stageable. Accuracy calculated only from the cleanest segments overstates practical availability. Single-person, two-person, transient-partner, position, and setting conditions should be stratified. Model confidence needs calibration, and analysis should determine whether rejections cluster in a particular stage or participant group.

Night-level evaluation should compare total sleep time, sleep-onset latency, wake after sleep onset, sleep efficiency, stage duration or proportion, REM latency, and transitions, reporting bias and limits of agreement. High correlation does not establish individual numerical agreement, while overall accuracy can be dominated by common stages. If staging corrects a respiratory-event index or triggers a care workflow, propagation of error into the downstream output must be validated; accurate respiratory rate alone is not evidence for the complete use. Longitudinal monitoring also requires night-to-night repeatability, stability after room changes, and the ability to distinguish true physiologic change from installation drift.

Room-level failure, safety, and privacy

Bed partners, pets, caregivers, fans, airflow from climate systems, moving curtains, relocated furniture, vibrating or adjustable beds, nearby devices, and indoor multipath can create or alter returns. Heavy obstruction, position changes, movement outside the beam, weak respiratory motion, and gross movement can reduce target and vital-signal quality. Real-room stress testing should vary distance, angle, furniture, doors, windows, bedding, and interference sources, then count track loss, identity swaps, reacquisition time, false occupancy, and usable coverage per bed-night. When two people cannot be separated, the system should declare that condition unsupported or stop individual staging rather than synthesize one complete hypnogram.

Radio compliance, information security, and clinical validity are separate obligations. A device should meet applicable emission, electrical, and electromagnetic-compatibility requirements while preventing default credentials, unauthorized remote access, and update compromise. The absence of an ordinary camera image does not eliminate privacy risk: raw or processed radar can reveal occupancy, bed exits, activity, routines, and vital signs. Monitoring should be conspicuously disclosed in the sensed space, with disable and visitor controls. Collection periods, data granularity, access, retention, and secondary use should be limited, and bystander returns must not be silently added to an individual's health record.

Use boundary, version governance, and acceptance

A consumer radar stage chart may support trend observation within validated single-person and room conditions, but should not by itself diagnose or exclude a sleep disorder, direct medication changes, alter PAP, or replace safety rounds. Medical and research uses need an explicit intended purpose, PSG reference, target population, and human-review pathway. Interfaces should expose presence, identity binding, signal quality, missingness, rejection, class definitions, and unsupported conditions instead of presenting an indirect probability as a certain brain state. Concerning symptoms and emergencies still require the appropriate professional assessment and established response route.

A change to the radio front end, antenna, enclosure, mount, firmware, clutter processing, target tracker, or staging model may invalidate earlier evidence. Every result should identify hardware, firmware, model, rules, installation configuration, and application versions. Updates need bridging on frozen data, relevant subgroups, and real rooms, with rollback available. Acceptance should obtain supported scenarios, independent-test populations, per-stage performance, coverage, multi-person policy, radio compliance, security, and privacy evidence, then conduct synchronized, blinded local tests involving one sleeper, exit and return, a partner, a pet, obstruction, and network interruption. Deployment monitoring should track drift, lost and swapped targets, rejection, stage distributions, complaints, and human-review findings.