Definition, signal origin, and task boundaries
A ballistocardiogram records small changes in whole-body displacement, velocity, acceleration, or force produced by reaction to cardiac ejection and the acceleration of blood in the great vessels. Modern sleep systems place piezoelectric film, bed-leg load cells, strain gauges, pneumatic pads, or optical fibers in or around a bed, enabling continuous sensing without attaching a device to the sleeper. Feature-extraction algorithms convert the resulting mixed mechanical waveform into heartbeat locations, intervals, morphology, quality labels, and statistical features.
The raw output of a bed sensor usually contains more than BCG. It also combines respiratory loading, turns, limb movement, bed-frame vibration, and environmental disturbance. Seismocardiography recorded by a chest accelerometer, peripheral pulse waves, and electrocardiography represent different physical processes and should not be renamed interchangeably merely because each contains recurring peaks. An algorithm should state whether its sensor measures force, pressure, displacement, or acceleration, together with its axis, mounting position, and mechanical coupling path.
BCG features can support heart-rate and respiratory-rate estimation, bed occupancy and movement analysis, and input features for HRV, sleep-stage, or respiratory-abnormality models. They are not a direct recording of cardiac electrical activity. A single mechanical channel generally cannot establish QRS morphology, identify the type of ectopic beat, or prove cessation of airflow. Physiological claims must match the intended use, and screening or trend-monitoring output should not automatically become a diagnosis of arrhythmia or sleep-disordered breathing.
Sensors, coordinates, and mechanical transmission
Classical BCG nomenclature identifies systolic H, I, J, and K waves, with the J wave commonly a prominent headward deflection in a normal longitudinal record. Those definitions depend on coordinates, polarity, sensitivity, and instrument frequency response. Transmission through a mattress, deck, and multiple bed supports can invert, broaden, or resonate the waveform. Unless direction and timing relative to a defined reference have been demonstrated, a detected feature is better called a candidate mechanical heartbeat peak than an asserted J wave.
Piezoelectric films respond to dynamic pressure, bed-leg load cells measure changes in support force, pneumatic and hydraulic pads integrate loading over an area, and optical systems detect bending or strain. Their gain, bandwidth, spatial coverage, and response to static load differ. Acquisition records should identify sensor model, channel layout, sample rate, analog filtering, bed-frame and mattress construction, participant orientation, and calibration method.
Body mass, height, sleep posture, and location relative to a sensor change signal amplitude and morphology. A thick soft mattress can attenuate higher-frequency components, while springs and adjustable frames can introduce structural vibration. A multichannel array can select or fuse locally informative signals, but channels affected by the same body movement are not independent evidence. The algorithm should recognize out-of-bed, sitting on the bed edge, two-person occupancy, and pets, and should suspend person-specific output when signal ownership cannot be separated.
Preprocessing, component separation, and quality gating
Preprocessing may include DC and trend removal, antialiasing, band-pass or wavelet decomposition, mains-noise suppression, and channel normalization. Cardiac, respiratory, and movement frequency ranges overlap, so a fixed filter does not guarantee a pure cardiac component; filter phase can also move the apparent peak time. An implementation should report cutoffs, order, zero-phase status, and boundary handling, and should test amplitude and delay distortion with known inputs.
Component separation may use multiscale decomposition, principal or independent components, empirical modes, morphological templates, and multichannel spatial information. Selecting the component with greatest energy can choose respiration or a bed resonance, whereas adaptive selection in every window can make feature definitions jump over time. The selected components, parameters, and confidence should remain traceable, and systematic crossover should be tested across heart rates, breathing rates, and postures.
Turns and large movements are often much larger than the cardiac vibration, and saturation can leave prolonged recovery ringing. Quality gating should identify movement, clipping, sensor disconnection, bed impact, and low signal-to-noise intervals before heartbeat detection and should impose an appropriate recovery period. It is safer to return missing with a reason than to invent beats from periodic priors inside contaminated intervals. Valid coverage must also be reported by state so performance is not limited to selected quiet segments.
Heartbeat timing, intervals, and morphological features
Heartbeat methods include adaptive peak or envelope detection, template matching, autocorrelation, average magnitude difference, cepstral analysis, Hilbert envelopes, wavelets, and learned models. A dominant spectral peak can estimate average rate within a window without providing reliable beat times. Beat-by-beat detection is better suited to interval analysis but is more vulnerable to changing morphology. A fused estimator should define how conflicts are resolved, how confidence is formed, and when no determination is returned.
A BCG mechanical peak follows the ECG R wave, and that delay varies with ejection, vascular state, posture, and the sensor chain. Beat-level evaluation should separate a fixed or changing electromechanical delay from true missed and false detections and should examine consecutive-interval error rather than force the mechanical peak to coincide with the R wave. Timing jitter that is acceptable for window-average heart rate may be unacceptable for HRV, so the latter use needs its own validation.
Morphological features may include amplitudes, slopes, areas, durations, and interpeak intervals of candidate H-I-J-K waves, as well as energy, spectral entropy, cross-channel consistency, template distance, or learned representations. Amplitude becomes comparable across nights or beds only when both the sensor and mechanical system are calibrated. Templates should be reconfirmed after a posture change, and a change in the mattress transfer function must not be interpreted directly as a change in cardiac contractility or stroke volume.
Respiratory, movement, and sleep-related features
Low-frequency bed loading can reflect thoracoabdominal motion, allowing respiratory rate and relative amplitude to be estimated from peak intervals, envelopes, or separated components. Reference respiration should come from airflow, respiratory effort, or another validated independent channel. A respiratory waveform separated from BCG cannot serve simultaneously as the training label and validation truth. Cardiac vibration may persist during an apnea, while reduced amplitude may instead result from posture or altered mechanical coupling.
Movement features often use short-window variance, energy, zero crossings, abrupt cross-channel changes, and duration to distinguish quiet, minor movement, a turn, and bed exit. Movement is both meaningful sleep behavior and the principal artifact for cardiopulmonary features, so it should receive a separate label rather than be silently discarded. Rules are needed for event boundaries, recovery intervals, and sensor reloading to avoid splitting one turn into many false cardiac or respiratory events.
Sleep staging or respiratory-risk models may combine heart rate, interbeat variability, respiratory regularity, movement, and coupling features, but these remain indirect proxies. Models must be aligned to synchronized polysomnographic labels and account for arousals, stage transitions, and overlapping respiratory events. When all features originate from one bed signal, their correlations do not constitute multiple independent physiological observations, and good average-heart-rate performance does not replace validation of the final clinical task.
Reference annotation, validation, and generalization
Heartbeat detection should use synchronized ECG as a beat-level reference, respiratory rate should use a defined airflow or effort reference, and movement and posture may use synchronized video or a validated sensor. Devices must share a clock or have drift corrected, and annotators should be able to inspect raw reference waveforms and artifacts. If training labels come from another automated algorithm, its error and the extent of human adjudication must be disclosed.
Evaluation should jointly report beat sensitivity, positive predictive value, timing error, interval error, heart-rate bias and agreement, and the fraction of time with valid output. Low window-average error can conceal runs of missed beats, double-rate, or half-rate failures, while correlation alone does not establish interchangeability. Results should be stratified by posture, movement, bed type, single or shared occupancy, heart-rate range, and disease subgroup, with completely failed recordings included.
Data partitions must be participant-independent so adjacent windows or nights from the same person do not cross between training and test sets. External evaluation should cover different body sizes, ages, cardiovascular and sleep conditions, home environments, sensor lots, frames, and mattresses. When a model depends on resonance characteristics of particular hardware, the hardware-algorithm combination needs validation as a unit; software-only claims of device independence are insufficient.
Deployment, maintenance, and safety boundaries
A deployed system should preserve raw or auditable intermediate signals, quality flags, algorithm version, and sensor configuration so an unusual trend can be investigated. Installation should check channel loading, polarity, clock operation, and empty-bed noise, with benchmark tests after a mattress replacement, bed move, or firmware update. Longitudinal monitoring should also detect gradual changes caused by gain drift, material aging, and loose connections.
Edge processing can limit the transmission of raw bedroom data, but local caching, encryption, access, retention, and deletion policies still need definition. A shared bed contains physiological information about two people, so separation, identity binding, and consent must be clear; person-level metrics should be hidden when assignment is unreliable. Model updates should be checked on a fixed test set for regressions in heartbeat, respiration, movement, and subgroup coverage.
BCG is unobtrusive and well suited to long-term observation, but it depends strongly on the mechanical environment and commonly fails during movement. Persistent abnormalities, prominent palpitations, chest pain, syncope, or suspected apnea require an appropriate clinical evaluation rather than diagnosis from bed-derived features alone. A trustworthy algorithm treats indeterminate intervals and sensor failure as formal outputs and clearly separates the measured mechanical phenomenon, derived metrics, and clinical interpretation.