Definition, Uses, and Scope
A radar sleep dataset organizes radio-frequency reflections, radar-derived signals, or spatial target representations acquired in bed or sleep settings together with synchronized references and metadata for position, movement, respiration, heartbeat, sleep stage, respiratory events, or related observations. It can support contactless bed monitoring, respiratory and movement analysis, sleep-algorithm research, device evaluation, and reproducible methods. Value comes from the combination of returns, environment, hardware configuration, and credible reference, not merely a vendor-generated respiratory-rate column.
Such data are not equivalent to PSG, and not every dataset represents natural overnight sleep. Quiet wakefulness, voluntary breath holding, mechanical chests, mannequins, or brief lying experiments can validate parts of a measurement chain but cannot directly represent sleep staging or naturally occurring apnea. Users should first determine whether participants actually slept, whether synchronized PSG or another reference exists, whether the task is vital-sign estimation or clinical event recognition, and whether raw radar data are truly available.
Radar Modality, Hardware Parameters, and Measurement Meaning
Sleep studies may use continuous-wave or Doppler radar, frequency-modulated continuous-wave radar, ultra-wideband radar, or millimeter-wave systems. Continuous-wave I/Q phase can represent small displacement but generally lacks fine range separation, while modulated or ultra-wideband systems can produce range bins, range-Doppler maps, or angular information. Raw samples, spatial resolution, and the ability to separate multiple reflectors differ across modalities and must not be combined under an undifferentiated “radar signal” label.
Reproducible data need carrier or band, bandwidth, chirp slope and duration, samples per chirp, chirps and frame rate, transmit and receive channels, antenna geometry, ADC configuration, transmit power, and firmware. Mounting height, angle, bed distance, coordinate system, beam direction, and selection of human target bins are also required. Hardware configuration determines observable range, velocity, and micromotion, and later processing cannot recover critical settings that were never documented.
Scene Design, Synchronized References, and Labels
Room geometry, wall and furniture multipath, bed frame, mattress and covers, posture, occlusion, fans, bed partners, pets, and out-of-bed activity all alter reflections. A dataset should describe laboratory or home setting, one or several occupants, device line of sight, sleep position, movement periods, care interactions, and sensor repositioning. Data acquired only from quiet supine participants at a fixed distance cannot unconditionally represent side or prone positions, turns, or whole-night monitoring in real homes.
References may include PSG, effort belts, airflow, oximetry, ECG, video, or manual event logs, each supporting only the corresponding labels. Synchronization should document hardware triggers, a common clock or timestamp method, clock drift, start delay, resampling, and alignment uncertainty. If PSG and radar were not recorded simultaneously on the same night, disagreement may reflect night-to-night physiology rather than model error. Voluntary breath-hold labels must remain separate from naturally occurring sleep respiratory events.
Data Levels, File Structure, and Version Management
Radar data may be released as raw ADC samples, complex I/Q, chirp cubes, range or Doppler spectra, angular maps, point clouds, phase from selected range bins, chest-wall displacement, respiratory and cardiac waveforms, or final summary measures. Later representations are easier to use but depend more heavily on vendor algorithms and hidden settings. Publishers should identify each file's level, dimension order, complex encoding, units, scale, dropped-frame value, and the processing chain from the preceding level.
A manifest should connect participant, night, scene, radar files, reference signals, and labels, with a data dictionary, timeline, checksums, code, license, change log, and known issues. Static-clutter removal, range FFT, phase unwrapping, target tracking, filtering, downsampling, and vital-sign separation all change results and require parameters and software versions. Firmware or processing changes should create a new release rather than silently replacing files under the same names.
Signal Quality, Domain Shift, and Selection Bias
Common failures include multipath cancellation, jumping target bins, phase wrapping, gross movement overwhelming micromotion, mixtures of reflectors, dropped frames, clock offsets, and detached reference sensors. Quality labels should distinguish vacancy, out of bed, movement, occlusion, multiple people, weak signal, and technical failure instead of encoding every failure as zero respiration. A model also needs to recognize when data are uninterpretable rather than producing a smooth estimate in the absence of reliable evidence.
Small experiments often concentrate on healthy adults, one room, fixed bedding, and one chipset, while clinical samples may overrepresent particular conditions and body types. Body size, respiratory amplitude, position, covers, bed partners, device distance, and room structure all create domain shift. Distribution should be reported by participant, night, room, hardware, position, condition, and quality. A huge number of radar frames cannot substitute for enough independent people, rooms, and natural sleep nights.
Machine-learning Splits, Evaluation, and Reproducibility
Training, validation, and testing should be partitioned by participant and, when environmental generalization is tested, further isolated by room, household, or hardware batch. Adjacent windows from one night, multiple representations derived from one source cube, repeat nights from one person, and their synchronized references must not leak across sets. Clutter templates, normalization, target selection, quality filtering, and thresholds must use training information only. Tests should retain movement, absence, multiple occupants, and failures and include cross-hardware or cross-scene validation.
Evaluation should match the task. Continuous waveforms may use correlation and amplitude or rate error; event detection needs defined temporal matching, misses, false alarms, and duration error; sleep stage or bed state requires class-specific performance and confusion. Scoring only intervals from which a signal was successfully extracted overstates the system, so coverage and uninterpretable rates must also be reported. Reproducibility materials should include partition manifests, hardware settings, data and label versions, signal processing, model, random seeds, and final thresholds.
Privacy, Deployment Governance, and Clinical Limits
Radar does not capture visible images, but it can continuously infer presence in bed, movement, room occupancy, respiratory rhythms, and possible health status. “Contactless” therefore does not mean “non-personal data” or remove the need for an appropriate basis and notice. Collection and sharing should minimize raw returns and location detail according to purpose, address bed partners and cohabitants, and govern access, retention, deletion, encryption, and incidents. Wireless operation must also follow local spectrum, emission, and device-safety requirements.
Radar output can support research and assisted monitoring but does not automatically equal PSG, a technically adequate home sleep apnea test, or a clinical diagnosis. Algorithms can fail during gross movement, multiple occupancy, or weak returns, and a respiratory waveform alone cannot establish every sleep stage or event. Clinical use requires validation for the target population, environment, and hardware release, failure alerts, appropriate human review, and applicable regulatory processes. A mature dataset presents performance conditions, failure scenarios, and evidentiary limits together.