Definition, targets, and output levels
Ballistocardiographic heart-rate and respiration inversion algorithms estimate cardiac and respiratory activity from composite vibration recorded by mattress, bed-leg, chair, or other mechanical sensors. Inversion here means working backward from an observation produced jointly by the body and supporting structure to latent physiological rhythms; it is not direct reading of an electrocardiogram or airflow. The input commonly mixes cardiac-ejection reaction forces, thoracoabdominal breathing motion, gross movement, and the response of the bed.
Outputs fall into at least three levels: average heart or respiratory rate over a window, beat-by-beat or breath-by-breath times and intervals, and a reconstructed continuous cardiac or respiratory waveform. A window-average rate can be obtained without accurate individual-event timing and therefore does not establish suitability for HRV. Unless independently validated, the phase and amplitude of a reconstructed waveform should not be interpreted directly as tidal volume, airflow, or contractile force.
The algorithm should also output timestamps, valid windows, signal quality, confidence, and reasons for missingness. BCG can support long-term trending and unobtrusive monitoring, but it is not synonymous with ECG, respiratory inductance plethysmography, nasal pressure, or an oronasal thermal airflow channel. Arrhythmia classification, apnea typing, and clinical alarms are higher-level tasks requiring separate sensor evidence and validation.
Observation model and identifiability
A bed signal can be viewed as an additive or nonlinear mixture of cardiac vibration, respiratory loading, movement, environmental noise, and the sensor-mattress transfer function. Heartbeats are generally faster than breathing, but the bands do not separate perfectly. Respiration modulates cardiac amplitude, baseline, and interval, while structural resonance can create harmonics. A fixed frequency band alone cannot guarantee that every extracted component comes from the intended physiological source.
Sensor type determines what is observed. Piezoelectric film emphasizes dynamic pressure, load cells measure support force, pneumatic or hydraulic pads integrate load over an area, and accelerometers reflect local or bed acceleration. Body mass, posture, mattress thickness, frame stiffness, and position relative to the sensor change gain, polarity, and morphology. Model parameters therefore should not move directly from one bed design to another without validation.
Multichannel systems can exploit spatial differences for best-channel selection, weighted fusion, or separation of bed partners, but identifiability has limits. When two rhythms converge, two people turn together, or structural vibration is shared, the sources may not be uniquely recoverable. The system should distinguish empty bed, one person, multiple occupants, pets, and sitting on the bed edge; suspending person-specific rates is safer than forced demixing when ownership is uncertain.
Preprocessing, windows, and movement gating
Common preprocessing includes antialiasing, detrending, band-pass filtering, wavelet or empirical-mode decomposition, channel normalization, and mains-noise suppression. Every filter can alter peak timing, envelope, and amplitude, and a real-time causal filter has a different delay from an offline zero-phase filter. An implementation should fix the sample rate, cutoffs, order, boundary rules, and internal resampling, and should test distortion with synthetic and measured signals.
Window duration controls temporal resolution and robustness. A longer window stabilizes a spectral, autocorrelation, or cepstral peak but smooths rapid change; a shorter window is more responsive but contains fewer cardiac or respiratory cycles and is more vulnerable to transient noise. Sliding step, overlap, the interval represented by an output timestamp, and change-point handling should be explicit. Strongly overlapping windows must not be treated as statistically independent examples.
Turns, limb movement, bed impacts, and sensor saturation are usually much larger than cardiopulmonary components. Movement gating should identify contaminated intervals, ringing recovery, disconnection, and low signal-to-noise before inversion and retain the reason for rejection. Periodic priors should not manufacture smooth vital signs through long contamination. Performance reporting must include valid coverage and the longest continuous gap, not only error among successful windows.
Heart-rate and beat-interval inversion
Average heart rate can be estimated from a spectral peak in a cardiac band, a short-time spectrum, autocorrelation, average magnitude difference, or a cepstral period. These methods may tolerate morphology variation but can select a harmonic, subharmonic, respiratory harmonic, or bed resonance. Search ranges, harmonic logic, and constraints from the previous window should be confidence-aware so a physiological prior does not conceal a real rapid change or algorithm failure.
Beat-level inversion commonly locates mechanical beats with peak or envelope detection, template matching, wavelets, clustering, dynamic programming, or learned models. The BCG peak follows the ECG R wave, and the electromechanical delay varies. Evaluation should specify matching rules and quantify missed and false beats, timing error, and consecutive-interval error. Agreement in one-minute average heart rate can conceal alternating misses, double counting, or timing jitter that makes the result unsuitable for HRV.
Algorithms may fuse channels or period estimators and reject outliers using physiological continuity, but the component errors can be correlated before fusion. Candidate estimates, final selection, and uncertainty should remain traceable. Atrial fibrillation, frequent ectopy, paced rhythms, tachycardia, and weak mechanical signals require separate evaluation. Without it, a BCG rate may be useful for trends but cannot demonstrate that rhythm is normal.
Respiratory-rate and waveform inversion
Respiratory inversion may use low-frequency loading, modulation of cardiac amplitude, an envelope, principal components, channel differences, a spectral peak, or breath-by-breath peaks and troughs. Window-frequency methods yield average rate, peak-trough methods yield cycle and phase, and continuous-waveform reconstruction additionally requires amplitude and morphology checks. Each output level needs separate validation; small rate error does not prove that the waveform reliably measures the depth of individual breaths.
The respiratory reference should match the intended use and may be nasal pressure or thermal airflow, thoracoabdominal effort, carbon dioxide, or manually reviewed polysomnography. Airflow, respiratory effort, and BCG bed loading represent different physiological quantities and can disagree during obstruction, paradoxical breathing, mouth breathing, or hypopnea. Validation should prespecify the reference channel, breath onset and offset annotation, synchronization, and handling of missing reference data.
An absent visible respiratory oscillation may indicate apnea, but it may also result from side posture, distance from the sensor, mattress attenuation, postmovement recovery, or selection of the wrong component. Respiratory-rate inversion cannot type obstructive or central events from low amplitude alone. Apnea detection additionally requires airflow, effort, duration, and an appropriate sleep denominator. “No reliable respiratory estimate” and “respiratory event detected” must remain distinct states.
Joint cardiorespiratory estimation and validation design
Heartbeat and breathing are not independent noise sources. Respiratory sinus arrhythmia, mechanical amplitude modulation, and cardiopulmonary coupling can aid joint estimation but can also create circular information. A joint model should identify which features originate from the same raw channel rather than present mutually derived signals as independent evidence. When joint optimization improves average performance, it should also show that one output or its coverage was not sacrificed for the other.
Heart-rate reference should use synchronized ECG R waves or adjudicated beat intervals, while respiratory reference should be an independent channel aligned with the target definition. All devices need a common clock or drift correction. Participant-level data partitions and external testing should cover intended ages, body sizes, conditions, postures, beds, device lots, and single- and two-person settings, including movement and poor-quality periods rather than excluding them after inspection.
Reports should separately present bias, absolute error, limits of agreement, response to changes, event-level sensitivity and positive predictive value, and valid-time fraction for heart and respiratory rates. Correlation is not a substitute for agreement, and window-average error is not a substitute for beat- or breath-level performance. Results should be stratified by rate range, posture, movement, sleep stage, and disease group, including nights with no output or persistent harmonic errors.
Deployment, monitoring, and clinical boundaries
Before deployment, installation direction, loading of every channel, empty-bed noise, clock behavior, and bed configuration should be checked, with a new baseline after a mattress or frame change. On-device processing can reduce bandwidth and bedroom privacy exposure, but sufficient quality flags and auditable excerpts should remain. Model, filter, or firmware updates need regression tests on fixed data for numerical continuity, latency, coverage, and subgroup performance.
Real-time alerting requires defined output delay, tolerance for continuous missingness, duplicate-alert suppression, and sensor-failure messaging. Brief gaps during a turn may be expected, but prolonged gaps must not silently display the previous value; interpolated values must also be distinguishable from measurements. Shared-bed use requires appropriate consent and identity handling, and a partner’s cardiopulmonary rhythm must not be assigned to the target user when separation is unreliable.
BCG inversion can complement long-term trend observation, but it should not independently rule out arrhythmia, respiratory depression, or sleep apnea. Chest pain, syncope, prominent palpitations, persistent abnormal breathing, or clinical deterioration require appropriate medical assessment rather than waiting for confirmation from a bed algorithm. A trustworthy system clearly separates sensor observations, derived rates, quality, and intended-use boundaries and treats indeterminate output as a valid result.