World Sleep Industry ENCYCLOPEDIA

Millimeter-Wave Radar Sleep Monitoring Algorithms

4.4.1

Definition and classification boundary

Millimeter-wave radar sleep monitoring algorithms process range, velocity, angle, amplitude, and phase sequences from bedside or overhead radar. They use small thoracoabdominal motion, gross body movement, and changes in target position to estimate bed occupancy, respiration, heartbeat, turns, sleep versus wake, and optional respiratory events or sleep stages. Their practical attraction is continuous, low-burden acquisition without a sensor attached to the skin or wrist. Radar nevertheless observes changes in electromagnetic reflections produced by the person and the room; it does not directly record EEG, EOG, EMG, airflow, or oxygen saturation.

The term millimeter wave usually describes carrier frequency, whereas FMCW, continuous wave, multiple-input multiple-output, and pulsed ultra-wideband describe waveform or antenna arrangements. These concepts can overlap. This entry covers the general sensing and algorithm stack built on millimeter-wave hardware; the neighboring UWB entry focuses on recovering respiration and heartbeat from UWB returns. A radar module, generic presence detector, specialized radar sleep-staging model, and complete diagnostic system are also not synonyms. Classification should follow the actual input, output, and intended use rather than the marketing label.

From radar returns to usable signals

An FMCW system normally mixes transmitted and received signals to form a beat frequency related to target range. Range transformation, antenna calibration, static-clutter suppression, and target detection turn raw ADC samples into representations such as range–Doppler or range–angle maps. The algorithm selects and continually tracks the range cell associated with the sleeper's chest and abdomen, then extracts sub-wavelength displacement from the phase of the complex return. Phase unwrapping, baseline correction, band-pass filtering, time–frequency analysis, autocorrelation, and harmonic suppression can separate respiration from the weaker heartbeat component.

Large turns temporarily corrupt the micromotion phase, so motion gating should precede target reacquisition and a stable recovery window. Multi-antenna beamforming may help separate a bed partner from background reflectors, but it does not guarantee reliable multi-person separation in every room. Each segment should retain target range, signal-to-noise ratio, motion state, track identity, and valid coverage. Reporting an unknown state during occlusion, saturation, or uncertain identity is more defensible than filling an entire night by smoothing or carrying the last value forward.

Separating tasks and reporting results

Evidence for one monitoring task cannot be borrowed by another. Bed entry and exit depend mainly on target presence and gross motion. Respiratory rate comes from periodic thoracoabdominal displacement. Heart rate relies on much smaller mechanical motion that is more vulnerable to respiratory harmonics. Sleep–wake classification generally combines movement and respiratory variability, while sleep staging infers latent states from these surrogate features. Stable respiratory-rate measurement therefore does not establish accurate heart rate, arousal, REM, or N3 classification, and it does not identify the type or clinical cause of an apnea.

Outputs should pair every value or event with a timestamp, confidence, valid duration, and reason for rejection, and should distinguish “not present” from “not measurable.” Total sleep time, sleep-onset latency, sleep efficiency, and wake after sleep onset require an explicit sleep–wake rule. A respiratory event index divided by total recording time should not be presented as though it were a polysomnography-derived AHI using actual sleep time. User-facing interfaces should expose coverage and algorithm version instead of reducing an uncertain night to a single deceptively precise score.

Evidence and validation design

Studies with synchronized polysomnography show that radar can estimate sleep–wake state, respiratory vital signs, or respiratory events for particular devices, installations, and populations, but performance varies materially by task and cohort. One sleep–wake study, for example, reported different results in healthy young adults and patients referred for sleep examination. A 60 GHz FMCW event study also found appreciable false positives and limited event sensitivity. More recent small clinical datasets show promise for event-level models and sleep-duration estimation, but they do not justify automatic transfer to different radars, rooms, or unrepresented populations.

Validation should use synchronized, manually scored reference signals and report metrics separately for the intended task. Continuous respiratory and heart rates require error distributions, Bland–Altman agreement, and valid coverage. Occupancy, movement, and event detection require sensitivity, specificity, precision, recall, false alarms per night, and latency. Staging requires per-stage F1, confusion matrices, and kappa, while night-level sleep measures require bias and limits of agreement. Cross-validation on a development dataset is not a substitute for participant-separated external testing over multiple nights in real homes.

Bedroom deployment and failure boundaries

Installation instructions should specify height, distance, tilt, field of view, and acceptable obstruction relative to the bed, with an initial quality check. Multipath from walls and metal furniture, moving fans or curtains, pets, bed partners, heavy or shifting bedding, posture changes, nocturnal exits, and rearranged rooms can all change the return distribution. A weak return is not “no breathing,” and a target outside the coverage area is not necessarily “out of bed.” Sensor failure, low quality, and a genuine physiological event need separate states in the data model.

Real acceptance testing should cover supine, lateral, and prone postures; different body sizes and respiratory patterns; and single-person, two-person, and intruder conditions. Useful operational measures include reacquisition time after a turn, proportion of the night rejected, identity swaps, and the longest continuous outage. Firmware, radio parameters, mounts, signal processing, and model updates can all alter performance, so versions must be traceable and reversible. If output controls lighting, care alarms, or treatment workflows, tests must also cover network loss, power loss, false alarms, and manual override.

Clinical use, privacy, and cybersecurity

Consumer trend tracking, institutional care prompts, and medical screening or diagnosis carry different levels of risk. An algorithm that claims to detect sleep apnea, aid diagnosis, or drive treatment needs a defined intended use, risk management, and evidence in the target population under the applicable device and software framework. Home sleep apnea results should not become diagnoses or treatment decisions solely through automated scoring. People with significant cardiorespiratory disease, neuromuscular weakness, suspected hypoventilation, chronic opioid use, a history of stroke, or severe insomnia require the appropriate professional assessment pathway.

Radar does not capture an ordinary visible image, yet it can continuously infer bedroom occupancy, routines, movement, and respiratory health; it is not inherently anonymous. Products should explain the sensing boundary to the sleeper, bed partner, visitors, and care recipients, provide a pause or local disable control, and minimize retention of raw returns and derived health data. Device identification, controlled configuration, data protection, restricted interfaces, signed updates, and security-state logging belong throughout the lifecycle. Cloud accounts, family sharing, and institutional fleet views need role-based access, audit trails, and deletion controls.

Procurement, acceptance, and future directions

Procurement and research comparisons should define the task, population, and setting before reviewing frequency and bandwidth, antenna channels, field of view, maximum target count, edge or cloud processing, raw-data access, missing-data rules, update policy, and radio compliance. A displayed average accuracy is not enough for acceptance. Buyers should request room-, participant-, and night-level results, failure examples, exclusion rates, train–test separation, and confidence intervals, then reproduce performance with synchronized references in the intended installation environment.

Likely development directions include more robust multi-target tracking, physics-informed learning, adaptation across rooms, joint respiratory and cardiac modeling, calibrated uncertainty, and fusion with oximetry, audio, mattress pressure, or wearable signals. Fusion can reduce the blind spots of one sensor, but it also adds time-synchronization, identity-matching, privacy, and maintenance burdens. A mature system is not one that forces every curve to remain continuous; it is one that knows when it is reliable, when it must degrade gracefully, and when a person or standard test must take over.