1. A computer-implemented method for predicting likelihood of hypoxia from non-contact physiological monitoring, comprising: receiving, by a processor, a ballistocardiography signal from a non-contact sensor positioned beneath a subject support surface, wherein the ballistocardiography signal comprises cardiac component signals associated with mechanical vibrations produced by cardiac ejection and respiratory component signals associated with respiratory motion of a subject; filtering, by the processor, the ballistocardiography signal to isolate a cardiac component signal and a respiratory component signal; deriving, by the processor, a set of signal dynamics data from the cardiac component signal and the respiratory component signal, wherein the set of signal dynamics data comprises heart rate values, respiration rate values, and spectral power measurements, and wherein the set of signal dynamics data is correlated to a health status of the subject; generating, by the processor, a multidimensional time-embedded vector representation of the set of signal dynamics data using phase space reconstruction, wherein the multidimensional time-embedded vector representation projects delayed versions of the set of signal dynamics data into a multidimensional space; constructing, by the processor, a recurrence matrix from the multidimensional time-embedded vector representation, wherein the recurrence matrix identifies recurring states indicating periodicity of data values in the set of signal dynamics data; determining, by the processor, at least one dynamic feature from the recurrence matrix, wherein the at least one dynamic feature represents a parameter derived from the set of signal dynamics data that exhibits change in state over time; comparing, by the processor, the at least one dynamic feature with a historical set of signal dynamics data to establish a correlation with hypoxic conditions; and predicting, by the processor, a likelihood of the hypoxia based on the comparison, wherein a deviation of the at least one dynamic feature beyond a predefined baseline value indicates increased hypoxia risk; and generating, by the processor, an alert indicative of the likelihood of the hypoxia in response to the hypoxia risk exceeding a predefined risk threshold.

2. The method of claim 1, wherein the phase space reconstruction uses an embedding dimension between 3 and 6 and a time delay selected based on sampling characteristics of the ballistocardiography signal, and wherein each state vector in the multidimensional space comprises signal values at time t, t-12, t-24, and t-36.

3. The method of claim 1, further comprising determining that a recurrence-derived stability metric computed from the recurrence matrix falls below a predefined stability threshold indicating elevated hypoxia risk.

4. The method of claim 1, wherein the signal dynamics data further comprises: standard deviation of heart rate computed over a defined time window; coefficient of variation of heart rate; standard deviation of respiration rate computed over the defined time window; and coefficient of variation of respiration rate.

5. The method of claim 4, further comprising comparing the set of signal dynamics data against a subject-specific historical baseline derived from a prior monitoring period a subject-specific historical baseline derived from a prior monitoring period corresponding to a non-hypoxic physiological state.

6. The method of claim 5, further comprising detecting an increase in heart rate relative to the subject-specific historical baseline as an indicator of the increased hypoxia risk.

7. The method of claim 1, wherein filtering the ballistocardiography signal comprises: applying a first bandpass filter configured to pass frequencies in a range of 1 Hz to 10 Hz to isolate the cardiac component signal; and applying a second bandpass filter configured to pass frequencies in a range of 0.1 Hz to 0.5 Hz to isolate the respiratory component signal.

8. The method of claim 1, further comprising extracting respiratory waveform features from the respiratory component signal, wherein the respiratory waveform features comprise area under an curve, peak height amplitude, inspiration-to-expiration ratio, and breath-to-breath interval variability.

9. The method of claim 1, wherein the spectral power measurements comprise second harmonic power of the cardiac component signal, and wherein fluctuations in the second harmonic power over successive time windows indicate physiological instability preceding hypoxia.

10. The method of claim 1, further comprising applying a machine learning classifier to the signal dynamics data and recurrence-derived dynamic features to generate the hypoxia risk indicator, wherein the machine learning classifier comprises a support vector machine, a random forest classifier, or a neural network trained on labeled ballistocardiography data from subjects with documented hypoxic events and subjects without hypoxic events.

11. The method of claim 1, further comprising receiving oxygen saturation measurements from a secondary sensor and using the oxygen saturation measurements as a confidence metric for recalibrating the hypoxia risk indicator.

12. A system for predicting likelihood of hypoxia from non-contact physiological monitoring, the system comprising: at least one non-contact sensor positioned beneath a subject support surface and configured to acquire a ballistocardiography signal comprising mechanical vibrations produced by cardiac ejection and respiratory-induced components associated with respiratory motion of a subject; and a processor communicatively coupled to the at least one non-contact sensor, the processor being configured to: receive the ballistocardiography signal from the at least one non-contact sensor; filter the ballistocardiography signal to isolate a cardiac component signal and a respiratory component signal; derive a set of signal dynamics data from the cardiac component signal and the respiratory component signal, wherein the set of signal dynamics data comprises heart rate values, respiration rate values, and spectral power measurements, the set of signal dynamics data being correlated to a health status of the subject; generate a multidimensional time-embedded vector representation of the set of signal dynamics data using phase space reconstruction, wherein the time-embedded vector representation projects delayed versions of the set of signal dynamics data into a multidimensional space; construct a recurrence matrix from the time-embedded vector representation, wherein the recurrence matrix identifies recurring states indicating periodicity of data values in the set of signal dynamics data; determine at least one dynamic feature from the recurrence matrix, wherein the at least one dynamic feature represents a parameter derived from the set of signal dynamics data that exhibits change in state over time; compare the at least one dynamic feature with a historical set of signal dynamics data to establish a correlation with hypoxic conditions; predict a likelihood of the hypoxia based on the comparison, wherein a deviation of the at least one dynamic feature beyond a predefined baseline value indicates increased hypoxia risk; and generate an alert indicative of the likelihood of the hypoxia when the hypoxia risk exceeds a predefined risk threshold.

13. The system of claim 12, wherein the processor is configured to perform the phase space reconstruction using an embedding dimension between 3 and 6 and a time delay selected based on sampling characteristics of the ballistocardiography signal, such that each state vector in the multidimensional space comprises signal values at time t, t-12, t-24, and t-36.

14. The system of claim 12, wherein the processor is configured to compute a recurrence-derived stability metric from the recurrence matrix, and wherein the stability metric falling below a predefined stability threshold indicates elevated hypoxia risk.

15. The system of claim 12, wherein the set of signal dynamics data further comprises: standard deviation of heart rate computed over a defined time window; coefficient of variation of heart rate; standard deviation of respiration rate computed over the defined time window; and coefficient of variation of respiration rate.

16. The system of claim 15, wherein the processor is further configured to compare the signal dynamics data against a subject-specific historical baseline derived from a prior monitoring period corresponding to a non-hypoxic physiological state.

17. The system of claim 16, wherein the processor is further configured to detect an increase in heart rate relative to the subject-specific historical baseline as an indicator of elevated hypoxia risk.

18. The system of claim 12, wherein the processor is configured to filter the ballistocardiography signal by: applying a first bandpass filter configured to pass frequencies in a range of 1 Hz to 10 Hz to isolate the cardiac component signal; and applying a second bandpass filter configured to pass frequencies in a range of 0.1 Hz to 0.5 Hz to isolate the respiratory component signal.

19. The system of claim 12, wherein the processor is further configured to extract respiratory waveform features from the respiratory component signal, the respiratory waveform features comprising area under the curve, peak height amplitude, inspiration-to-expiration ratio, and breath-to-breath interval variability.

20. The system of claim 12, wherein the processor is further configured to apply a machine learning classifier to the signal dynamics data and recurrence-derived dynamic features to generate the hypoxia risk indicator, wherein the machine learning classifier comprises a support vector machine, a random forest classifier, or a neural network trained on labeled ballistocardiography data from subjects with documented hypoxic events and subjects without hypoxic events.

Description:SYSTEMS AND METHODS FOR PREDICTING HYPOXIA USING NON-CONTACT BALLISTOCARDIOGRAPHY

CROSS-REFERENCE TO RELATED APPLICATIONS
This application is a patent of addition to Indian Patent Application Number 202541049786, filed on 21 May 2025, the entire contents of which are incorporated herein by reference.
FIELD OF INVENTION
The present disclosure relates to non-contact physiological monitoring systems utilizing ballistocardiography signal processing, and more particularly to a system and method for predicting hypoxia through recurrence analysis and time-embedded feature extraction from ballistocardiography-derived cardiac and respiratory waveforms.

BACKGROUND
Continuous monitoring of blood oxygen saturation levels in clinical environments presents a technical challenge rooted in signal acquisition constraints. Hypoxia, characterized by oxygen saturation levels falling below 90% for sustained periods, can cause irreversible damage to brain tissue and vital organs if not detected and addressed promptly. Current monitoring approaches rely on contact-based sensors, such as pulse oximeters employing photoplethysmography (PPG), which require physical attachment to a subject’s finger, earlobe, or other peripheral site. These contact-based sensors generate SpO2 measurements only after hypoxia has already manifested, providing reactive rather than predictive information. The contact requirement introduces additional technical limitations: sensor displacement during subject movement, signal degradation from poor peripheral perfusion, and subject discomfort that leads to sensor removal during extended monitoring periods, resulting in gaps in the physiological data stream.
Existing approaches to respiratory and cardiac monitoring include ballistocardiography (BCG) systems that capture mechanical vibrations produced by cardiac ejection and respiratory motion through non-contact sensors positioned beneath a subject’s mattress or within bedding materials. These BCG-based systems extract heart rate and respiration rate from filtered frequency bands of the raw vibration signal, typically isolating cardiac components in the 1-10 Hz range and respiratory components in the 0.1-0.5 Hz range. Some systems further process these signals to derive heart rate variability metrics or detect apnea events based on respiratory signal absence. However, these existing BCG processing pipelines treat the extracted vital signs as independent scalar time series, applying threshold-based alerting when instantaneous values exceed predefined limits.
SUMMARY
This summary is provided to introduce a selection of concepts in a simplified form that are further described below in the detailed description. This summary is not intended to identify key features or essential features of the claimed subject matter, nor is it intended to be used as an aid in determining the scope of the claimed subject matter.
According to an aspect of the present disclosure, a computer-implemented approach for predicting the likelihood of hypoxia using non-contact physiological monitoring is disclosed. In operation, a processor receives a ballistocardiography (BCG) signal acquired from a non-contact sensor positioned beneath a subject support surface, where the BCG signal captures mechanical vibrations generated by cardiac ejection as well as respiratory-induced components arising from respiratory motion of the subject. The processor processes the BCG signal to separate a cardiac component signal from a respiratory component signal and derives a set of signal dynamics data from these components. The signal dynamics data include physiological parameters such as heart rate values, respiration rate values, and spectral power measurements, which collectively correlate with a health status of the subject.
The processor further generates a multidimensional time-embedded vector representation of the signal dynamics data using phase space reconstruction, in which delayed versions of the signal dynamics data are projected into a multidimensional space to capture temporal evolution. Based on this representation, a recurrence matrix is constructed to identify recurring physiological states that indicate periodicity within the signal dynamics data. From the recurrence matrix, the processor determines one or more dynamic features that represent parameters derived from the signal dynamics data and that vary over time. These dynamic features are compared against historical signal dynamics data to establish correlations with hypoxic conditions. Using this comparison, the processor predicts a likelihood of hypoxia, where deviations of the dynamic features beyond predefined baseline values indicate an increased risk of hypoxia. When the predicted hypoxia risk exceeds a predefined risk threshold, the system generates an alert indicative of the likelihood of hypoxia.
According to other aspects of the present disclosure, the method may include one or more of the following features. The embedding dimension may be set to 4 and the time delay may be set to 12 samples, such that each state in the phase space representation comprises values at time t, t-12, t-24, and t-36. The signal dynamics data may further comprise standard deviation of heart rate, coefficient of variation of heart rate, standard deviation of respiration rate, and coefficient of variation of respiration rate, wherein elevated values of these variability metrics during a monitoring window indicate increased hypoxia risk. The method may further include comparing the signal dynamics data against a subject-specific historical baseline derived from a prior monitoring period when oxygen saturation levels were within a normal range above 90%. The method may further include extracting respiratory waveform features comprising area under the curve, peak height amplitude, inspiration-to-expiration ratio, and breath-to-breath interval variability from the respiratory component. The method may further include performing pattern recognition on the BCG signal to distinguish between respiratory events including normal breathing, coughing, gasping, and snoring based on waveform morphology templates derived from training data. The method may further include receiving oxygen saturation measurements from a secondary sensor and using the oxygen saturation measurements as a confidence metric for recalibrating the hypoxia risk indicator. The spectral power measurements may include fast Fourier transform (FFT) coefficients and signal-to-noise ratios of dual peaks in the cardiac component. The method may further include categorizing extracted features into high-weighted features, medium-weighted features, and low-weighted features, wherein heart rate variability metrics and second harmonic power fluctuations are categorized as high-weighted features.
According to another aspect of the present disclosure, a system for predicting hypoxia onset from non-contact physiological monitoring is provided. The system includes a non-contact sensor configured to be positioned beneath a subject support surface and to capture a ballistocardiography (BCG) signal comprising mechanical vibrations produced by cardiac ejection and respiratory motion of a subject. The system includes a processor coupled to the non-contact sensor. The system includes a memory storing instructions that, when executed by the processor, cause the processor to perform operations. The operations include filtering the BCG signal into a cardiac component isolated within a frequency band of 1-10 Hz and a respiratory component isolated within a frequency band of 0.1-0.5 Hz. The operations include deriving signal dynamics data from the cardiac component and the respiratory component, wherein the signal dynamics data comprises heart rate, respiration rate, and spectral power measurements including second harmonic power of the cardiac component. The operations include generating a multidimensional time-embedded vector representation of the signal dynamics data using phase space reconstruction with a defined embedding dimension and time delay. The operations include constructing a recurrence matrix from the time-embedded vector representation, wherein the recurrence matrix identifies recurring states in the cardiopulmonary system. The operation includes determining at least one dynamic feature from the recurrence matrix, wherein the at least one dynamic feature represents a parameter derived from the set of signal dynamics data that exhibits change in state over time. The operation includes comparing, by the processor, the at least one dynamic feature with a historical set of signal dynamics data to establish a correlation with hypoxic conditions. The operation includes predicting a likelihood of the hypoxia based on the comparison, wherein a deviation of the at least one dynamic feature beyond a predefined baseline value indicates increased hypoxia risk and generating an alert indicative of the likelihood of the hypoxia when the hypoxia risk exceeds a predefined risk threshold.
According to other aspects of the present disclosure, the system may include one or more of the following features. The non-contact sensor may comprise a piezoelectric sensor array embedded within a mattress substrate. The system may further include an alert interface configured to transmit the predictive alert to a clinical monitoring station. The memory may further store subject-specific historical baseline data comprising signal dynamics data from a prior monitoring period, and the operations may further include comparing current signal dynamics data against the subject-specific historical baseline data to detect deviations indicative of cardiopulmonary deterioration. The operations may further include applying a machine learning classifier trained on labeled BCG data from subjects with documented hypoxic events and subjects without hypoxic events, wherein the machine learning classifier receives the signal dynamics data and the recurrence-derived dynamic features and outputs the hypoxia risk indicator. The machine learning classifier may comprise a support vector machine, a random forest classifier, or a neural network. The system may further include a secondary sensor configured to measure oxygen saturation, wherein the operations further include using oxygen saturation measurements from the secondary sensor to adjust confidence weighting of the hypoxia risk indicator. In an embodiment the secondary sensor may be contactless. In yet another embodiment, the secondary sensor may not be contactless.
The foregoing general description of the illustrative embodiments and the following detailed description thereof are merely exemplary aspects of the teachings of this disclosure and are not restrictive.
DRAWINGS
Embodiments of the present disclosure will now be described, by way of example only, with reference to the following diagrams wherein:
FIG. 1 illustrates a system for predicting likelihood of hypoxia, in accordance with an embodiment of the present disclosure;
FIG. 2A illustrates a flow chart depicting steps of a method for predicting likelihood of hypoxia, in accordance with an embodiment of the present disclosure;
FIGs. 3A to 3E illustrate graphical representations of the vitals of five subjects in non-hypoxic state, in accordance with an embodiment of the present disclosure;
FIGs. 3F to 3J illustrate graphical representations of the vitals of five subjects in hypoxic state, in accordance with an embodiment of the present disclosure;
FIGs. 4A to 4E illustrate graphical representations of second harmonics analysis for a set of BCG waveform data corresponding to five subjects in non-hypoxic state, in accordance with an embodiment of the present disclosure;
FIGs. 4F to 4J illustrate graphical representations of second harmonics analysis for a set of BCG waveform data corresponding to five subjects in hypoxic state, in accordance with an embodiment of the present disclosure;
FIGs. 5A to 5E illustrate recurrence plots plotted for a set of signal dynamics data corresponding to five subjects in non-hypoxic state, in accordance with an embodiment of the present disclosure; and
FIGs. 5F to 5J illustrate recurrence plots plotted for a set of signal dynamics data corresponding to five subjects in hypoxic state, in accordance with an embodiment of the present disclosure.

DETAILED DESCRIPTION
The following description sets forth exemplary aspects of the present disclosure. It should be recognized, however, that such description is not intended as a limitation on the scope of the present disclosure. Rather, the description also encompasses combinations and modifications to those exemplary aspects described herein.
The signal processing architectures in current systems do not capture the complex interdependencies between cardiac and respiratory dynamics that precede hypoxic deterioration, nor do they model the temporal evolution of signal morphology that distinguishes compensatory physiological responses from stable baseline states.
The technical shortcomings of current BCG-based monitoring systems stem from their reliance on one-dimensional time series analysis that discards phase-space information embedded in the raw waveforms. When a subject’s cardiopulmonary system transitions toward a hypoxic state, the body initiates compensatory mechanisms including increased respiratory effort, elevated heart rate, and altered breathing patterns that manifest as changes in signal power distribution, waveform morphology, and inter-signal timing relationships. These multi-dimensional dynamics cannot be adequately characterized by scalar metrics such as instantaneous heart rate or respiration rate alone. Furthermore, existing systems lack mechanisms for comparing current signal dynamics against subject-specific historical baselines in a manner that accounts for the recurrence structure of physiological states over time. The absence of time-embedded phase-space reconstruction and recurrence-based feature extraction in current BCG processing pipelines prevents these systems from identifying the characteristic instability patterns that precede hypoxic events, limiting their utility to reactive detection rather than predictive alerting.
The system and method described herein comprise hardware components and a software pipeline configured to predict hypoxia risk using non-contact ballistocardiography (BCG) sensing. The hardware components may include a sheet positioned beneath a mattress, where the sheet contains sensing elements including non-contact ballistocardiography (BCG) sensors to detect micro-movements of a subject resting on the mattress. The hardware components may further include a processor coupled to the sheet, a memory storing instructions and data, and a communication module configured to transmit data to remote devices or servers. The processor may be a local processor disposed of within a bedside unit, a remote server processor, or a combination thereof operating in a distributed computing arrangement. The memory may store raw BCG waveform data, derived features, template libraries, trained machine learning model parameters, and threshold values for alert generation.
The software pipeline may include a signal acquisition module, a filtering module, a feature extraction module, a machine learning inference module, and an alert generation module. The signal acquisition module may receive time-series BCG waveform data from the sheet at a defined sampling rate and may buffer the received data with associated timestamps. The filtering module may process the buffered BCG waveform data to separate cardiac component signals from respiratory component signals and to suppress noise and movement artifacts. The feature extraction module may derive signal dynamics data including heart rate data, respiration rate data, and spectral power data from the filtered BCG waveform data. The feature extraction module may further generate multidimensional time-embedded vectors and recurrence matrices from the signal dynamics data. The machine learning inference module may receive the extracted features and generate a physiological instability indicator representing a likelihood of hypoxia. The alert generation module may compare the physiological instability indicator against a stored threshold and may generate an alert signal when the threshold is exceeded. In an example, throughout the present disclosure the term "hypoxia" refers to a state of medical emergency where tissues are deprived of adequate oxygen supply over a certain period of time. The term "ballistocardiogram (BCG) waveform" refers to signal representing vibrations caused by motions of the body produced by various activities of the cardiovascular and respiratory system such as blood pumping via heart, inhalation-exhalation of air by lungs and so on. Further, as used herein, a set of signal dynamics data refers to physiological parameters derived from ballistocardiogram (BCG) waveform data, including heart rate data, respiration rate data, spectral power data, and statistical or temporal representations thereof. At least one dynamic feature refers to one or more metrics derived from the set of signal dynamics data, including statistical measures, recurrence-based measures, spectral measures, or trend-based measures that characterize physiological variability or instability. A physiological instability indicator (also referred to herein as a hypoxia risk indicator) refers to an output generated based on the at least one dynamic feature and represents a likelihood or risk of hypoxia for a subject over a current or future time window.
The system and method address technical difficulties associated with extracting clinically relevant information from composite BCG waveforms captured via non-contact sensing. A BCG waveform captured from a subject resting on a mattress contains superimposed signals from multiple physiological sources including cardiac mechanical activity, respiratory chest and abdominal motion, and voluntary or involuntary body movements. Isolating the cardiac component signals and the respiratory component signals from the composite BCG waveform presents a signal processing challenge because the frequency bands of these components may overlap and because movement artifacts may corrupt portions of the waveform. The system addresses signal isolation through bandpass filtering configured to attenuate frequencies outside target ranges for cardiac and respiratory activity. The system addresses movement artifact suppression through artifact detection logic that identifies segments of the BCG waveform exhibiting amplitude excursions or spectral characteristics inconsistent with physiological signals, and through gating logic that excludes or down-weights such segments during feature extraction. The non-contact sensing constraint introduces additional technical difficulty because the BCG sheet does not directly contact the subject's skin, resulting in signal attenuation and coupling variability depending on mattress properties, subject weight, and subject position.
The technical nature of the system and method combines digital signal processing, pattern recognition, and machine learning inference to transform raw BCG waveform data into a physiological instability indicator. Digital signal processing operations may include bandpass filtering, windowing, Fast Fourier transform (FFT) computation, power spectral density estimation, and peak detection. Pattern recognition operations may include comparison of extracted waveform features against stored templates representing hypoxic and non-hypoxic respiratory patterns and may include recurrence analysis to identify recurring states within multidimensional representations of the signal dynamics data. Machine learning inference operations may include applying a trained classification model to the extracted features and recurrence-derived metrics to generate the physiological instability indicator. The trained classification model may be a gradient boosting model, a random forest model, a support vector machine, a neural network, or another supervised learning model trained on labeled BCG waveform samples paired with ground truth SpO2 measurements.
The present system relies on non-contact sensing via BCG sensing to assess hypoxia risk, whereas traditional approaches use contact sensors such as pulse oximeters to determine hypoxia. A pulse oximeter measures blood oxygen saturation (SpO2) by transmitting light through tissue and detecting absorption changes caused by oxygenated and deoxygenated hemoglobin. Pulse oximetry provides a direct measurement of SpO2 but requires continuous contact with the subject, typically via a finger clip or adhesive sensor. Subjects may remove or dislodge contact sensors due to discomfort, limiting the continuity of monitoring. The present system captures BCG waveform data without requiring any sensor to contact the subject's body, enabling continuous monitoring even when the subject moves or repositions during rest. The present system does not directly measure SpO2 but instead analyzes patterns within the BCG waveform that correlate with physiological changes preceding hypoxic events, thereby providing an early warning capability that contact-based SpO2 monitoring does not provide.
Referring to FIG. 1, Referring now to FIG. 1, a network 100 implementation of a system 100 for predicting hypoxia risk using non-contact ballistocardiography is disclosed. The system 102 comprises a processor 108, a memory 112, a communication module (not shown in fig.), and interface components 110 including a user interface and an admin interface. In an embodiment, the network 100 includes a system 102, one or more user devices 104-N (for example but not limited to one or more user devices 104-1, 104-2…104-N) associated with one or more users. Although the present disclosure is explained considering that the system 102 is implemented on a server, it may be understood that the system 102 may be implemented in a variety of computing systems, such as a laptop computer, a desktop computer, a notebook, a workstation, a virtual environment, a mainframe computer, a server, a network server, a cloud-based computing environment. In one implementation, the system 102 may be implemented in a cloud-based computing environment in which the system is configured to execute or interact with remotely located applications. Further, system 102 and user device 104 may communicate through the network 106. Examples of the user devices 104 may include, but are not limited to, a portable computer, a personal digital assistant, a handheld device, and a workstation. In an embodiment, sensing element 114 may be coupled to the system 102. The system 102 may be coupled to the sending element 114 via the network 106. In an embodiment, the sensing element 114 may comprise BCG sensing elements.
In one implementation, the network 106 may be a wireless network, a wired network, or a combination thereof. The network 106 can be implemented as one of the different types of networks, such as intranet, local area network (LAN), wide area network (WAN), the internet, and the like. The network 106 may either be a dedicated network or a shared network. The shared network represents an association of the different types of networks that use a variety of protocols, for example, Hypertext Transfer Protocol (HTTP), Transmission Control Protocol/Internet Protocol (TCP/IP), Wireless Application Protocol (WAP), and the like, to communicate with one another. Further the network 106 may include a variety of network devices, including routers, bridges, servers, computing devices, storage devices, and the like.
In one embodiment, the system 102 may include at least one processor 108, an input/output (I/O) interface 110, a memory 112, and one or more modules explained later in the description. The at least one processor 108 may be implemented as one or more microprocessors, microcomputers, microcontrollers, digital signal processors, Central Processing Units (CPUs), state machines, logic circuitries, and/or any devices that manipulate signals based on operational instructions. Among other capabilities, the at least one processor 108 is configured to fetch and execute computer-readable instructions stored in the memory 112.
The I/O interface 110 may include a variety of software and hardware interfaces, for example, a web interface, a graphical user interface, and the like. The I/O interface 110 may allow the system 102 to interact with the user directly or through the client devices 104. Further, the I/O interface 110 may enable the system 102 to communicate with other computing devices, such as web servers and external data servers (not shown). The I/O interface 110 can facilitate multiple communications within a wide variety of networks and protocol types, including wired networks, for example, LAN, cable, etc., and wireless networks, such as WLAN, cellular, or satellite. The I/O interface 110 may include one or more ports for connecting a number of devices to one another or to another server.
The memory 112 may include any computer-readable medium or computer program product known in the art including, for example, volatile memory, such as static random-access memory (SRAM) and dynamic random-access memory (DRAM), and/or non-volatile memory, such as read only memory (ROM), erasable programmable ROM, flash memories, hard disks, Solid State Disks (SSD), optical disks, and magnetic tapes. The memory 112 may include routines, programs, objects, components, data structures, etc., which perform particular tasks or implement particular abstract data types. The memory 112 may include programs or coded instructions that supplement applications and functions of the system 102. In one embodiment, the memory 112, amongst other things, serves as a repository for storing data processed, received, and generated by one or more of the programs or the coded instructions.
In an embodiment, a sheet may be positioned beneath a mattress surface and may comprise BCG sensing elements 114 configured to detect mechanical vibrations transmitted through the mattress. In an embodiment, the processor 108 may be disposed within a bedside unit coupled to the sheet, or the processor may be disposed within a remote server, or processing operations may be distributed between a local processor and a remote server processor. The memory may be local memory coupled to the local processor, remote memory coupled to the remote server processor, or a combination thereof. The communication module may couple the local processor to the remote server processor and may couple the processor to the user interface and the admin interface. The user interface may present alert notifications and physiological trend data to clinical staff. The admin interface may provide configuration controls for threshold values, alert routing parameters, and system operational settings.
In an embodiment, a bed frame supports a mattress, and the sheet may be positioned between the bed frame and the mattress or between the mattress and a mattress topper. The sheet may be positioned such that the sheet underlies a torso region of a subject resting on the mattress. Cardiac mechanical activity of the subject produces micro-movements of the subject’s body resulting from blood ejection from the heart and subsequent vascular recoil. Respiratory activity of the subject produces micro-movements of the subject’s chest and abdomen resulting from lung expansion and contraction. The micro-movements propagate through the mattress material and may get transmitted to the sheet. The mattress material acts as a mechanical transmission medium that attenuates high-frequency vibrations while transmitting low-frequency vibrations associated with cardiac and respiratory activity. The transmission between the subject and the sheet depends on mattress stiffness, mattress thickness, subject weight, and subject position on the mattress.
The sheet may comprise one or more sensing elements configured to convert mechanical displacement or pressure variations into electrical signals. In an embodiment, the sensing elements may be a combination of one or more: of piezoelectric sensors, piezoresistive sensors, capacitive sensors, or fiber optic sensors. In some embodiments, the sheet comprises an array of sensing elements distributed across a sensing area to capture spatial variations in mechanical activity. In an embodiment, the sheet comprises a single sensing element or a linear arrangement of sensing elements. The sensing elements may generate analog electrical signals proportional to detected mechanical displacement or pressure. The analog electrical signals may be conditioned by amplification circuitry and anti-aliasing filter circuitry disposed of within the sheet or within a signal conditioning module coupled to the sheet. The output signal format may be a time-series waveform representing amplitude variations over time, where each sample of the time-series waveform corresponds to a measured amplitude value at a defined time instant.
In an embodiment, an acquisition module of the system 102 may receive the time-series waveform from the sheet and may perform analog-to-digital conversion, if the output received from the sensing elements of the sheet is analog in nature. In an embodiment, the acquisition module may sample the time-series waveform at a sampling rate selected to capture frequency content associated with cardiac and respiratory activity. In an embodiment, the sampling rate may be in a range from 50 samples per second to 1000 samples per second. In an example, the acquisition module may buffer received samples in a circular buffer or a linear buffer disposed in local memory. The acquisition module may associate each buffered sample or each buffered segment of samples with a timestamp indicating the time of acquisition. The timestamp may be derived from a local clock synchronized to a network time source or may be derived from an unsynchronized local clock. The acquisition module may segment the buffered samples into frames of defined duration for subsequent processing.
With continued reference to FIG. 1, the memory 112 may store raw BCG waveform data received from the acquisition module for a defined retention period. The memory may store derived features extracted from the raw BCG waveform data, including heart rate values, respiration rate values, spectral power values, and recurrence-derived metrics. The memory may store the derived features in association with timestamps to enable trend analysis over time. Trend analysis may compare current derived feature values against historical derived feature values for the same subject to identify changes in physiological state. The memory may store template libraries containing reference waveform patterns representing hypoxic and non-hypoxic respiratory states. The memory may store trained machine learning model parameters including weights, biases, and configuration values. The memory may store threshold values used by the alert generation module to determine when to generate alert signals. In some embodiments, the memory comprises non-volatile storage for persistent retention of historical data and model parameters, and the memory comprises volatile storage for buffering of current waveform data and intermediate processing results.
Referring to FIG. 2, a method for predicting hypoxia risk using non-contact ballistocardiography begins with receiving a set of BCG waveform data from the non-contact sensor. The non-contact sensor may be positioned beneath a mattress such that the non-contact sensor does not contact the skin or clothing of a subject resting on the mattress. The non-contact sensor may detect mechanical vibrations transmitted through the mattress material as the subject's body undergoes micro-movements associated with physiological activity. The set of BCG waveform data may comprise a time-series of amplitude values sampled at a defined sampling rate, where each amplitude value represents a magnitude of detected mechanical displacement or pressure at a corresponding time instant. The method may further include receiving the set of BCG waveform data continuously during a monitoring session or may receive the set of BCG waveform data in discrete segments corresponding to defined time windows.
Ballistocardiography captures mechanical signals produced by the body in response to cardiovascular and respiratory activity. When the heart contracts and ejects blood into the aorta, the momentum transfer between the blood mass and the vascular system produces a mechanical recoil force that propagates through the body. The mechanical recoil from cardiac ejection causes micro-movements of the torso that can be detected by a sensor positioned beneath the subject. The BCG waveform includes a cardiac component that reflects the timing and force characteristics of each heartbeat. The cardiac component of the BCG waveform exhibits a repeating pattern corresponding to the cardiac cycle, with waveform features that correlate with ventricular ejection, aortic valve closure, and diastolic filling phases. The amplitude and morphology of the cardiac component may vary depending on cardiac output, vascular compliance, and the mechanical coupling between the subject and the sensor.
Respiration-induced body motion also contributes to the BCG waveform. As the lungs expand during inhalation, the chest wall and abdomen move outward, and as the lungs contract during exhalation, the chest wall and abdomen move inward. The respiratory motion produces low-frequency oscillations in the BCG waveform that are superimposed on the cardiac component. The respiratory component of the BCG waveform reflects the breathing rate, breathing depth, and breathing pattern of the subject. The respiratory component may exhibit variations in amplitude corresponding to tidal volume changes and may exhibit variations in cycle duration corresponding to respiratory rate changes. In addition to the cardiac component and the respiratory component, the BCG waveform may include movement artifacts resulting from voluntary or involuntary body movements such as limb repositioning, coughing, or postural adjustments. Movement artifacts may produce transient amplitude excursions or sustained signal disturbances that differ in frequency content and morphology from the cardiac and respiratory components. In an embodiment, the method may further include processing the received set of BCG waveform data to separate the cardiac component, the respiratory component, and the movement artifacts for subsequent analysis.
The method may include filtering the set of BCG waveform data to obtain a filtered BCG waveform comprising cardiac component signals and respiratory component signals. Filtering may be performed using bandpass filter configurations selected to isolate frequency content associated with each physiological component. A first bandpass filter may be configured to pass frequencies associated with cardiac activity, where the passband may span a range from approximately 0.5 Hz to approximately 10 Hz to capture the fundamental frequency and harmonic content of the cardiac component. In an example, a second bandpass filter may be configured to pass frequencies associated with respiratory activity, where the passband may span a range from approximately 0.1 Hz to approximately 0.5 Hz to capture the fundamental frequency of normal breathing rates. In some embodiments, the system filters the BCG signal to extract a respiratory waveform in a 1 Hz to 3 Hz frequency range for primary feature extraction, where the 1 Hz to 3 Hz range captures respiratory effort signals and respiratory-related modulations that may be more informative for hypoxia prediction than the fundamental respiratory rate alone. The bandpass filters may be implemented as finite impulse response (FIR) filters, infinite impulse response (IIR) filters, or combinations thereof. FIR filters may provide linear phase response characteristics that preserve waveform morphology. IIR filters may provide steeper roll-off characteristics with lower computational cost. The filter coefficients may be selected based on the sampling rate of the BCG waveform data and the desired passband and stopband characteristics.
In an embodiment, the filtered BCG waveform may comprise a cardiac component signal output from the first bandpass filter and a respiratory component signal output from the second bandpass filter. In some embodiments, the filtered BCG waveform further comprises a respiratory effort signal extracted from the 1 Hz to 3 Hz frequency range, where the respiratory effort signal captures higher-frequency modulations associated with breathing depth, breathing pattern irregularities, and respiratory distress indicators. The cardiac component signal, the respiratory component signal, and the respiratory effort signal may be stored in memory as separate time-series for subsequent feature extraction operations.
Filtering is performed to achieve component isolation, noise reduction, artifact suppression, and stable feature extraction. Component isolation separates the cardiac component signals from the respiratory component signals, enabling independent analysis of heart-related and breathing-related physiological activity. Without component isolation, features extracted from the composite BCG waveform would reflect a mixture of cardiac and respiratory contributions, reducing the specificity of extracted features for detecting respiratory-related changes associated with hypoxia risk. Noise reduction attenuates frequency content outside the physiological frequency bands of interest, including high-frequency electrical noise, sensor noise, and environmental vibration noise. Artifact suppression reduces the influence of transient disturbances caused by subject movement, coughing, or external mechanical impacts on the extracted features. Bandpass filtering attenuates artifact energy that falls outside the passband frequencies, although large-amplitude artifacts may still produce residual disturbances within the passband that require additional artifact detection and gating logic. Stable feature extraction depends on consistent isolation of the cardiac and respiratory components across varying signal conditions, subject positions, and mattress coupling characteristics. Filtering with fixed passband characteristics provides a standardized signal conditioning stage that reduces variability in the input to subsequent feature extraction operations, enabling more consistent and reproducible extraction of heart rate data, respiration rate data, spectral power data, and waveform morphology features.
In an embodiment, a set of signal dynamics data is derived from the set of BCG waveform data. In this regard, after receiving BCG waveform data from the at least one sensor, the set of BCG waveform data (for example, the data pertaining to the heart rate and respiration rate) are analyzed to derive the set of signal dynamics data. In this regard, the term "signal dynamics data" as used herein refers to data that changes over a time period. It may be appreciated that the set of signal dynamic data is related to both the cardiac component and the respiratory components of the BCG waveforms. In this regard, it may be appreciated that the set of BCG waveform data is collected over a time period at successive time intervals i.e., in a time series form. As the set of signal dynamics data is derived from the set of BCG waveform data, the set of signal dynamics data is also in time series form.
In an embodiment, the set of signal dynamic data pertains to at least one of: heart rate variability, respiration depth, inspiration/expiration rate, inspiration-expiration ratio, spectral shifts in power level of the BCG waveform data and morphology of the BCG waveform data. In this regard, heart rate variability (HRV) is a quantitative measure of the variation in time intervals between successive heartbeats and is calculated by analyzing the time differences between consecutive heartbeats from the waveforms related to cardiac activities of the BCG waveform data. Similarly, respiration depth, inspiration/expiration rate, inspiration-expiration ratio can be determined from the waveforms related to respiratory activities of the BCG waveform data. It may be appreciated that the spectral shifts in power level of the BCG waveform data refers to changes in the frequency distribution and amplitude of the BCG waveform data over time and can provide insights into cardiovascular and respiratory health. Similarly, morphology of the BCG waveform data refers to information on shape, amplitude, peaks and valleys of BCG waveforms. Notably, the set of signal dynamics data is derived using a suitable technique. For example, the set of signal dynamics data is derived using at least one of: standard deviation of normal-to-normal intervals (SDNN), root mean square of successive differences (RMSSD), frequency-domain metrics, Hilbert transform, empirical mode decomposition (EMD), wavelet transform (WT), using approximate entropy (ApEn) and sample entropy (SampEn). The technical advantage of deriving the set of signal dynamics data from the set of BCG waveform data is that the set of signal dynamics data provides rich, high-resolution insights into heart and lung function for early detection of likelihood of hypoxia.
In an example, the method includes analyzing cardiac harmonic power components of a BCG-derived cardiac signal using time frequency analysis. In BCG signals, the fundamental and first harmonic frequency components include contributions from respiratory motion and body movement, and therefore may not represent cardiac activity in isolation. Accordingly, the method evaluates higher-order harmonics, including at least a second harmonic and optionally a third harmonic, which contain a higher proportion of cardiac-specific mechanical energy. The estimated heart rate is converted into a fundamental cardiac frequency defined as HR/60, and harmonic frequencies are determined as integer multiples of the fundamental frequency, including a second harmonic at two times the fundamental frequency and a third harmonic at three times the fundamental frequency. A Short-Time Fourier Transform (STFT) is applied to the BCG signal to obtain a time-frequency representation. For each harmonic, the method defines a narrow frequency band centered on the corresponding harmonic frequency, wherein the bandwidth is selected based on the estimated heart rate and the frequency resolution of the STFT to capture the harmonic spectral peak while excluding adjacent frequency components. The power within each harmonic frequency band is computed over successive time windows. Variations in harmonic power over time, including changes in second and/or third harmonic power, are quantified using statistical measures such as standard deviation, coefficient of variation, or range over a defined observation period. Variability in cardiac harmonic power is indicative of physiological instability and may correlate with impending or ongoing hypoxia events.
In an embodiment, the method includes comparing current signal dynamics data values against baseline signal dynamics data values established during a stable monitoring period for the same subject. The method may include detecting an increase in heart rate by at least 15 beats per minute (BPM) before or during a hypoxia event compared to a stable baseline heart rate, where the stable baseline heart rate may be computed as a mean or median heart rate value over a preceding reference period during which the subject exhibited stable physiological state. The method may include detecting an increase in respiration rate by at least 5 respirations per minute before or during a hypoxia event compared to a stable baseline respiration rate, where the stable baseline respiration rate may be computed as a mean or median respiration rate value over a preceding reference period. Detection of heart rate increases exceeding 15 BPM or respiration rate increases exceeding 5 respirations per minute relative to baseline may contribute to the determination of the physiological instability indicator and may trigger or weight subsequent pattern recognition and machine learning inference operations.
In an example, the method may include comparing extracted features from the filtered BCG waveform data against templates of known hypoxic and non-hypoxic patterns to identify respiratory states associated with elevated hypoxia risk. A template library may store reference waveform patterns derived from labeled training data, where each template represents characteristic signal features observed during a defined respiratory state or physiological condition. The template library may include templates representing non-hypoxic respiratory patterns observed during stable breathing in subjects with normal SpO2 levels, and templates representing hypoxic respiratory patterns observed during periods preceding or concurrent with documented hypoxia events where SpO2 dropped below 90%. Template creation may involve extracting feature vectors from labeled BCG waveform segments in the training data, where each labeled segment corresponds to a known respiratory state confirmed by concurrent SpO2 measurement or clinical observation. The extracted feature vectors may include respiration rate values, respiration rate variability metrics, respiratory waveform amplitude values, area under the respiratory waveform curve, peak height values, breath-to-breath interval variability, spectral power distribution across respiratory frequency bands, and recurrence-derived metrics. Template creation may aggregate feature vectors from multiple labeled segments representing the same respiratory state to generate a representative template, where aggregation may involve computing mean values, median values, or centroid positions in the feature space. The template library may store each template as a feature vector with associated metadata indicating the respiratory state label, the number of training samples contributing to the template, and confidence bounds or variance estimates for each feature dimension.
In an example, the method may include computing a similarity metric between a current feature vector extracted from the filtered BCG waveform data and each template stored in the template library to determine which respiratory state the current feature vector most closely resembles. Similarity metrics may include distance measures and correlation measures. Distance measures may include Euclidean distance, Manhattan distance, Mahalanobis distance, or cosine distance computed between the current feature vector and each template feature vector. Euclidean distance may quantify the straight-line separation between two points in the feature space. Mahalanobis distance may account for correlations between feature dimensions and differences in feature variance by incorporating a covariance matrix derived from the training data. Correlation measures may include Pearson correlation coefficient or Spearman rank correlation coefficient computed between the current feature vector and each template feature vector, where higher correlation values indicate greater similarity. The method may include generating a decision output based on the computed similarity metrics, where the decision output may indicate the respiratory state label of the template exhibiting the highest similarity to the current feature vector, or the decision output may indicate a probability distribution over respiratory state labels based on relative similarity values. The decision output may contribute to the determination of the physiological instability indicator by indicating whether the current respiratory state resembles a hypoxic pattern or a non-hypoxic pattern.
In an embodiment, the system may differentiate between different respiratory events including snoring, coughing, gasping, and choking through template matching and pattern recognition trained on labeled data. The template library may include templates representing snoring patterns characterized by periodic low-frequency vibration signatures superimposed on the respiratory waveform. The template library may include templates representing coughing patterns characterized by transient high-amplitude jerky interference patterns that interrupt the normal respiratory rhythm. The template library may include templates representing gasping patterns characterized by dual-peak waveforms within individual breath cycles, where the dual peaks reflect irregular inspiratory effort associated with respiratory distress. The template library may include templates representing choking patterns characterized by absent or severely attenuated respiratory waveform amplitude combined with elevated movement artifact activity. Pattern recognition trained on labeled data may enable the system to classify current BCG waveform segments into respiratory event categories, where classification into gasping or choking categories may elevate the physiological instability indicator due to the association of these respiratory events with hypoxia risk.
In an embodiment, the method may include generating multidimensional time-embedded vectors from the signal dynamics data to enable analysis of complex physiological system behavior that cannot be adequately represented by one-dimensional time series alone. Time series data derived from the filtered BCG waveform, including respiration rate values, heart rate values, spectral power values, and waveform amplitude values, may be transformed into multidimensional representations using phase space reconstruction techniques. Phase space representation with time embedding adds complexity to time series analysis by projecting delayed versions of the same signal into multidimensional space. The multidimensional representation captures temporal dependencies and dynamic patterns within the signal that would not be apparent from examination of instantaneous signal values alone. The method may include applying time embedding to respiratory component signals, cardiac component signals, respiratory effort signals extracted from the 1 Hz to 3 Hz frequency range, or derived metrics such as respiration rate time series and heart rate time series. The resulting multidimensional time-embedded vectors may serve as input to recurrence analysis operations and machine learning inference operations for determining the physiological instability indicator.
The phase space representation may use an embedding dimension and a delay parameter to construct the multidimensional representation from a one-dimensional time series. In some embodiments, the phase space representation uses an embedding dimension of 4 and a delay of 12 to create a 4-dimensional space representation of the signal. For a time series denoted as a(t), where t represents a time index, the 4-dimensional embedded state at time t may be represented as the vector [a(t), a(t-12), a(t-24), a(t-36)], where each element of the vector corresponds to the signal value at a different time offset. The delay value of 12 specifies the number of sample intervals between successive elements of the embedded vector, such that the embedded vector captures signal values spanning a temporal window of 36 sample intervals. The embedding dimension of 4 specifies the number of delayed versions of the signal that are concatenated to form each embedded vector. The selection of embedding dimension and delay parameters may depend on the sampling rate of the signal, the characteristic time scales of the physiological dynamics being analyzed, and the complexity of the underlying system. In some embodiments, the embedding dimension may be selected based on false nearest neighbor analysis or other embedding dimension estimation techniques. In some embodiments, the delay parameter may be selected based on autocorrelation analysis or mutual information analysis to identify a delay value that captures meaningful temporal structure without excessive redundancy between embedded vector elements.
The multidimensional time-embedded vectors enable path-based state comparison where states of the physiological system at different times are compared by evaluating not just current signal values but also delayed versions to assess the path the system followed to arrive at the current state or the path the system is likely to take from the current state. For the states of the system at time t1 and time t2 to be considered similar under the time-embedded representation, the signal values a(t1) and a(t2) must be comparable, and the delayed signal values must also be comparable such that a(t1-12) is similar to a(t2-12), a(t1-24) is similar to a(t2-24), and a(t1-36) is similar to a(t2-36). Path-based state comparison enables identification of recurring dynamical patterns that reflect the trajectory of the physiological system through state space rather than merely the instantaneous position of the system. The method may include generating multidimensional time-embedded vectors from multiple feature sources to capture different aspects of physiological state. Primary feature sources may include BCG respiratory waveform analysis, where the respiratory component signal or the respiratory effort signal extracted from the 1 Hz to 3 Hz frequency range may be embedded into multidimensional vectors to capture respiratory dynamics. Secondary feature sources may include heart rate variability (HRV) metrics derived from the cardiac component signal, where inter-beat interval time series or instantaneous heart rate time series may be embedded into multidimensional vectors to capture cardiac dynamics. Ancillary feature sources may include movement artifact signals and activity pattern signals derived from the BCG waveform, where movement-related signal components may be embedded to capture subject activity state and postural stability.
In an embodiment, a recurrence matrix is generated from the multidimensional time-embedded vectors to quantify the recurrence structure of the physiological system dynamics. The recurrence matrix may be constructed by computing pairwise distances between all pairs of time-embedded vectors within a defined analysis window. For a sequence of N time-embedded vectors denoted as v(1), v(2), ..., v(N), the recurrence matrix R may be an N×N matrix where each element R(i,j) represents a measure of similarity or distance between the time-embedded vector v(i) and the time-embedded vector v(j). The pairwise distance computation may use Euclidean distance, where the distance d(i,j) between v(i) and v(j) is computed as the square root of the sum of squared differences across all dimensions of the embedded vectors. In some embodiments, the pairwise distance computation may use other distance metrics including Manhattan distance, maximum norm distance, or weighted distance metrics that account for different scaling or variance across feature dimensions. The recurrence matrix may be converted from a distance matrix to a binary recurrence matrix by applying a threshold operation, where R(i,j) is set to 1 if the distance d(i,j) is less than or equal to a threshold value ε, and R(i,j) is set to 0 if the distance d(i,j) exceeds the threshold value ε. The threshold value ε may be selected based on a percentage of the maximum distance observed in the distance matrix, based on a fixed distance value determined from training data, or based on adaptive thresholding that maintains a target recurrence rate. The binary recurrence matrix encodes which pairs of time points exhibit similar physiological states under the time-embedded representation.
In an embodiment, the recurrence matrix may be visualized as a recurrence plot, which provides a graphical representation of the times at which the physiological system's state recurs. The recurrence plot may display the recurrence matrix as a two-dimensional image where each pixel at position (i,j) corresponds to the matrix element R(i,j). In an embodiment, a recurrence analysis is applied on the multidimensional time-embedded vector for identifying recurring states and determining at least one dynamic feature from amongst the set of signal dynamics data. In this regard, when recurrence analysis is applied on the multidimensional time embedded vector to identify repetitive patterns of data values represented in vector form, over time. The recurrence analysis can be expressed in matrix wherein
Ri,j={1,if ‖Xi-Xj ‖<€ 0,otherwise
wherein € is a threshold for closeness which defines how close two points in the phase space must be to be considered recurrent. The threshold of closeness is determined empirically or based on statistical analysis of the set of signal dynamics data, ensuring that it captures meaningful recurrences without being overly sensitive to noise determined based on volume of the. The recurrence analysis can also be expressed in recurrence plot, comprising scattered points representing changes in data values over time. The term "recurring states" refers to instances where a data value (expressed in multidimensional vector form) in the multidimensional phase space is similar to a previously observed data value, indicating periodicity in the set of signal dynamics data correlated to a health status of the subject under observation. The term "dynamic feature" refers to measurable characteristics or parameter derived from the set of signal dynamics data that exhibit significant changes in their state over time, as identified through recurrence analysis. These changes can be correlated to physiological conditions, such as fluctuations in blood oxygen saturation (SpO2), to predict hypoxia. For example, the set of signal dynamics data may include HRV and the inspiration-expiration ratio. Using a suitable representation technique, a multidimensional time-embedded vector is constructed for HRV and inspiration-expiration ratio data. When the constructed multidimensional time-embedded vector is analyzed, HRV data shows significant changes in its state, as indicated by increased instability in the recurrence plot, while the inspiration-expiration ratio remains relatively stable. This suggests that HRV is the at least one dynamic feature which can be correlated with the likelihood of hypoxia. Moreover, unlike static features (which remain relatively constant, such as mean J-peak amplitude, average I-J interval and so on), dynamic features evolve and aid in real-time monitoring and early detection of conditions like hypoxia. In other words, static features, which remain relatively constant, are considered to be baseline characteristics, whereas dynamic features change over time, enabling real-time monitoring and early detection of conditions like hypoxia.
In an embodiment, at least one dynamic feature is determined by comparing a set of signal dynamics data with a historical set of signal dynamics data derived from a historical set of ballistocardiogram (BCG) waveform data and corresponding historical blood oxygen saturation (SpO₂) data, to correlate the set of signal dynamics data with blood oxygen saturation (SpO₂) levels. In this regard, the historical set of signal dynamics data refers to a dataset derived from previously recorded BCG waveforms and their corresponding SpO₂ measurements. Notably, the historical set of BCG waveform data, the corresponding historical SpO₂ data, and the historical set of signal dynamics data may be stored and retrieved from a data repository. The historical set of signal dynamics data serves as a reference or baseline for comparison to identify how changes in present signal dynamics data relate to changes in SpO₂ levels, thereby enabling detection of physiological conditions such as hypoxia.
In this regard, by comparing the present set of signal dynamics data with the historical set of signal dynamics data, at least one dynamic feature may be identified, and a correlation between the present signal dynamics data and blood oxygen saturation (SpO₂) levels may be established. The comparison may be performed by analyzing patterns, trends, or anomalies in the present set of signal dynamics data and matching such patterns with similar patterns observed in the historical set of signal dynamics data under known SpO₂ conditions. To ensure accurate comparison, noise in the present set of signal dynamics data may be mitigated using suitable preprocessing techniques, including smoothing or filtering. Accordingly, a dynamic feature is identified when a parameter derived from the present signal dynamics data exhibits a significant change that is similar to a change observed in the historical set of signal dynamics data under a specific SpO₂ condition. Such identification enables accurate detection of dynamic features, such as heart rate variability (HRV), for predicting likelihood of hypoxia. For instance, HRV has been observed to exhibit increased instability during hypoxic events, as reflected by changes in recurrence plot characteristics, thereby enabling early detection of hypoxia.
In an embodiment, at least one dynamic feature is indicative of likelihood of hypoxia when a deviation in value of one or more parameters from amongst the set of signal dynamics data, as compared to a predefined baseline value, is observed. The predefined baseline value refers to a nominal or standardized value range associated with normal physiological function for a given parameter. For example, a predefined baseline value of HRV measured using SDNN may range from 50 ms to 150 ms. As another example, a predefined baseline value for an inspiration-expiration ratio may range from 1:1.5 to 1:2. If the inspiration-expiration ratio for a subject deviate beyond this range, such as 2:1 or 1:5, such deviation may indicate abnormal respiratory activity, which may be associated with increased likelihood of hypoxia. The technical effect achieved is accurate assessment and analysis of derived signal dynamics data to determine likelihood of hypoxia.
In an embodiment, the at least one dynamic feature indicative of likelihood of hypoxia comprises an increase in standard deviation values, coefficient of variation values, and/or a max-min difference calculated for the set of BCG waveform data, as compared to predefined nominal values. An increase in standard deviation and/or coefficient of variation calculated for the set of BCG waveform data may indicate increased heart rate and respiration rate, which are physiological responses associated with elevated oxygen demand and potential hypoxia. For example, during hypoxia or in a pre-hypoxic state, heart rate may increase by at least 15 beats per minute and respiration rate may increase by at least 5 respirations per minute. Use of standard deviation and coefficient of variation enables mitigation of noise and transient artifacts present in heart rate and respiration rate measurements, thereby providing a more reliable analysis of the set of BCG waveform data. Further, a max-min difference calculated for the set of BCG waveform data provides insight into the magnitude of variation in physiological parameters over time. Such variations may be correlated with SpO₂ changes to predict likelihood of hypoxia.
In an embodiment, changes in heart rate, changes in respiration rate, irregular cardiac rhythms, or irregular breathing patterns may indicate respiratory distress and may further suggest likelihood of hypoxia. To quantify such changes for identification of the at least one dynamic feature, suitable statistical techniques may be applied to the set of BCG waveform data, including calculation of standard deviation and coefficient of variation for heart rate and respiration rate. In this manner, a relationship between the set of BCG waveform data and likelihood of hypoxia may be established. For example, during hypoxic conditions, standard deviation and coefficient of variation of heart rate and/or respiration rate tend to increase. Accordingly, observation of an upward trend in these values may indicate a possible onset of hypoxia. Additionally, changes in one or more of orientation, propagation, or trend of BCG waveform characteristics in relation to physiological events may also be correlated with likelihood of hypoxia. Considering such changes aids in minimizing noise and errors arising from artifacts, such as sensor drift, while improving personalized monitoring by adapting models to individual waveform variations.
In an embodiment, likelihood of hypoxia is predicted based on the at least one identified dynamic feature. When the at least one dynamic feature is identified using recurrence analysis applied to the set of signal dynamics data, such dynamic feature may be associated with hypoxic conditions. For example, increased HRV values or increased inspiration-expiration ratios may indicate increased physiological load on the cardiopulmonary system, which is related to increased oxygen demand and may suggest insufficient oxygen availability in the subject. Accordingly, likelihood of hypoxia may be predicted prior to an actual hypoxic event, thereby enabling timely intervention and improved healthcare outcomes.
In this regard, from the BCG waveform data, which comprises BCG signals, heart rate data, and respiration rate data, heart rate and respiration rate values may be extracted and expressed as a multidimensional time-embedded vector. Trends in heart rate and respiration rate values may be identified from the multidimensional representation, and at least one dynamic feature may be derived from such trends. The dynamic feature, including standard deviation values, coefficient of variation values, and/or max-min differences calculated for the set of BCG waveform data, may be used to distinguish hypoxic states of a subject from non-hypoxic states, such as by identifying hypoxic regions and non-hypoxic regions within a recurrence plot. Further, by tuning a filter window length, trends may be analyzed over time using a baseline window and a comparison window. Use of multiple filter windows enables accurate prediction of hypoxia onset in advance. For example, heart rate and respiration rate trends may be analyzed using a 30-minute baseline window and a 15-minute comparison window, where median changes in heart rate and respiration rate are tracked at regular intervals, and both absolute and percentage changes are computed to assess physiological fluctuations over time.
Furthermore, likelihood of hypoxia may also be predicted based on analysis of morphology and spectral power shifts in the BCG signal. In this regard, recurrence analysis may be applied to spectral components of the BCG signal, and recurrence plots may be generated. For example, a recurrence plot may be generated for a second harmonic power component of the BCG signal associated with cardiac activity. The BCG signal may be filtered using a signal processing module to extract the cardiac component, after which spectral power of the second harmonic frequency may be computed as a continuous time series. Recurrence plots may then be generated using phase space representations and time embedding of the time series. Changes in recurrence plot characteristics may correspond to changes in the subject’s physiological state, such as transition from a non-hypoxic state to a hypoxic state. A recurrence plot represents repeated patterns in a time-embedded multidimensional representation of the signal dynamics data, thereby enabling detection of subtle physiological changes indicative of hypoxia.
In an embodiment, the method may include receiving the at least one dynamic feature and a recurrence matrix and generating a physiological instability indicator using a machine learning model trained to distinguish between physiological states associated with elevated hypoxia risk and physiological states associated with normal respiratory function. The machine learning model may receive as input a feature vector comprising dynamic features derived from the recurrence matrix, including recurrence rate, determinism, laminarity, diagonal line metrics, entropy measures, and relative fractional area of stable regions. The machine learning model may further receive as input the recurrence matrix itself or a compressed representation thereof suitable for processing by convolutional neural network architectures. Additionally, the machine learning model may receive signal dynamics data comprising heart rate data, respiration rate data, spectral power data, and derived metrics including standard deviation and coefficient of variation computed over defined time windows. The machine learning model may generate as output the physiological instability indicator, which represents likelihood of hypoxia and may comprise a continuous probability value, a discrete classification label, or a risk score mapped to defined risk tiers.
Training of the machine learning model may be performed using supervised learning with labeled BCG waveform samples corresponding to hypoxic and non-hypoxic states. Each labeled BCG waveform sample may be paired with a corresponding SpO₂ measurement obtained from a reference pulse oximetry device, wherein the SpO₂ measurement serves as ground truth. A BCG waveform sample may be labeled as hypoxic when the corresponding SpO₂ measurement falls below 90% for a defined duration, and may be labeled as non-hypoxic when the corresponding SpO₂ measurement remains at or above 90%. Temporal alignment may be applied such that the BCG waveform sample corresponds to a time window preceding the SpO₂ measurement by a defined prediction horizon, thereby enabling training of models configured to predict future hypoxic events based on current BCG patterns. The training dataset may be derived from retrospective clinical data collected from subjects monitored using both BCG sensing and pulse oximetry, including subjects with documented hypoxic events.
In an embodiment, window labeling logic may segment continuous BCG waveform recordings into analysis windows of defined duration, wherein each analysis window is assigned a hypoxia label based on SpO₂ measurements occurring within or following the analysis window. An analysis window may be labeled as hypoxic if any SpO₂ measurement within a defined future time horizon falls below a threshold value, or if an average SpO₂ measurement within the future time horizon falls below the threshold value. Different labeling criteria may be applied for models trained to predict hypoxia at different prediction horizons. The labeled dataset may be divided into training, validation, and test subsets to enable model development and performance evaluation. In some embodiments, the dataset may be partitioned by subject identifier to prevent data leakage, such that data from a given subject appears in only one of the subsets.
The machine learning model may comprise a Support Vector Machine (SVM) classifier, a Random Forest classifier, a convolutional neural network (CNN), a deep neural network (DNN), a Long Short-Term Memory (LSTM) network, or an ensemble combining multiple model types. A Support Vector Machine classifier may be trained to detect abnormal patterns indicating impending hypoxia by finding a hyperplane in the feature space that separates feature vectors associated with hypoxic states from feature vectors associated with non-hypoxic states. The SVM classifier may use a radial basis function kernel or a polynomial kernel to handle non-linear separability in the feature space. A Random Forest classifier may be trained to detect abnormal patterns indicating impending hypoxia by constructing an ensemble of decision trees, where each decision tree is trained on a bootstrap sample of the training data and makes predictions based on a random subset of input features. The Random Forest classifier may aggregate predictions from individual decision trees using majority voting for classification or averaging for probability estimation. A CNN architecture may process the recurrence matrix as a two-dimensional image input, where convolutional layers extract spatial features from the recurrence plot structure and fully connected layers map the extracted features to the physiological instability indicator. A DNN architecture may process a feature vector comprising dynamic features and signal dynamics data through multiple fully connected layers with non-linear activation functions. In some embodiments, the CNN or DNN architecture may use initial weights set based on severity classification, where a preliminary classification stage determines an initial severity index for the current physiological state and the initial severity index influences weight initialization or weight scaling applied to subsequent network layers. The CNN or DNN architecture may be trained to predict likelihood of hypoxia in the next 30 to 60 minutes based on current BCG waveform patterns and derived features. An LSTM architecture may process sequential BCG waveform data or sequential feature vectors to capture temporal dependencies across successive time windows. The LSTM architecture may be configured to predict SpO₂ waveforms 5 to 30 minutes in advance, where the predicted SpO₂ waveform may be analyzed to identify anticipated hypoxic events based on predicted SpO₂ values falling below 90%. The LSTM architecture may employ a structure configured for sequential physiological signal prediction, where LSTM cells maintain hidden state representations that encode information from preceding time steps and gate mechanisms control the flow of information through the network.
In an embodiment, the machine learning model may comprise an ensemble machine learning approach with multiple classification stages. A first classification stage may determine an initial severity index based on current BCG waveform features and derived metrics, where the initial severity index categorizes the current physiological state into severity tiers ranging from stable to severely compromised. A second classification stage may determine a compensatory type classification based on patterns of physiological responses observed in the BCG waveform, where the compensatory type classification categorizes the current state as cardiac compensatory, pulmonary compensatory, cardio-pulmonary compensatory, or deviant compensatory. Cardiac compensatory classification may indicate that the subject exhibits elevated heart rate and cardiac output changes as a primary response to physiological stress. Pulmonary compensatory classification may indicate that the subject exhibits elevated respiration rate and respiratory effort changes as a primary response to physiological stress. Cardio-pulmonary compensatory classification may indicate that the subject exhibits combined cardiac and pulmonary responses. Deviant compensatory classification may indicate that the subject exhibits atypical or irregular compensatory patterns that do not conform to standard cardiac or pulmonary response profiles. The initial severity index and the compensatory type classification may be used for initial weight setting in subsequent neural network layers, where the weights applied to different feature inputs may be adjusted based on the determined severity and compensatory type to improve prediction accuracy for subjects exhibiting different physiological response patterns. In some embodiments, the machine learning model may incorporate reinforcement learning to modify predictions based on actual SpO₂ values obtained from concurrent pulse oximetry monitoring. The reinforcement learning approach may compare predicted physiological instability indicator values against actual SpO₂ measurements and may adjust model parameters to reduce prediction errors over time. In some embodiments, the reinforcement learning approach may further incorporate intervention information indicating clinical actions taken in response to alerts, where the model may learn to account for the effects of interventions such as oxygen therapy, medication administration, or postural adjustments on subsequent physiological state trajectories.
In an embodiment, the method may include generating an alert signal when the physiological instability indicator exceeds a stored threshold. The stored threshold may be a configurable value maintained in memory and accessible to an alert generation module. The stored threshold may be set by clinical staff via an admin interface based on clinical protocols, subject population characteristics, or institutional preferences for alert sensitivity and specificity. In some embodiments, the stored threshold may be a single value that defines a boundary between alert and non-alert states, where the alert generation module generates the alert signal when the physiological instability indicator equals or exceeds the stored threshold and does not generate the alert signal when the physiological instability indicator falls below the stored threshold. In some embodiments, multiple stored thresholds may define multiple alert tiers corresponding to different levels of hypoxia risk, where exceeding a first threshold may trigger a low-priority alert, exceeding a second threshold may trigger a medium-priority alert, and exceeding a third threshold may trigger a high-priority alert. The alert generation module may compare the physiological instability indicator against the stored threshold each time a machine learning inference module generates an updated physiological instability indicator value, thereby enabling continuous monitoring with alert generation responsive to changes in the physiological instability indicator over time.
The alert signal may be routed to user devices, admin devices, dashboards, nurse stations, and hospital systems based on configured alert routing parameters. A communication module may transmit the alert signal to a user interface device associated with clinical staff responsible for the monitored subject, where the user interface device may be a mobile device, a tablet, a workstation terminal, or a dedicated alert notification device. The communication module may transmit the alert signal to a dashboard application that displays physiological monitoring data and alert status for multiple subjects in a clinical unit, thereby enabling centralized monitoring by nursing staff or clinical supervisors. The communication module may transmit the alert signal to a nurse station terminal or a central monitoring station where clinical staff can view and acknowledge alerts. The communication module may transmit the alert signal to hospital information systems including electronic health record systems, clinical decision support systems, or hospital-wide alert management systems via interface modules configured to communicate using standard healthcare data exchange protocols. Escalation logic may govern routing and prioritization of alert signals based on alert tier, time elapsed since alert generation, and acknowledgment status. In some embodiments, a low-priority alert may be routed initially to the user interface device of an assigned clinical staff member, and if the alert is not acknowledged within a first time interval, the escalation logic may route the alert to additional clinical staff members or to a supervisor. In some embodiments, a high-priority alert may be routed simultaneously to multiple recipients including the assigned clinical staff member, a nurse station, and a supervisor, and may trigger audible or visual alarm indicators at the nurse station. The escalation logic may track acknowledgment status for each generated alert and may continue escalation routing until an acknowledgment is received or until a maximum escalation level is reached. Alert routing parameters and escalation logic may be configurable via the admin interface to accommodate different clinical workflows, staffing patterns, and institutional alert management policies.
The method may include calculating an area under the curve (AUC) of a respiratory waveform as a feature correlating with tidal volume, where larger breaths produce a larger area under the curve. The AUC calculation may be performed on a respiratory component signal or a respiratory effort signal extracted from filtered BCG waveform data. For each breath cycle identified within the respiratory waveform, the method may include computing the AUC by integrating the waveform amplitude over the duration of the breath cycle. The integration may be performed using numerical integration techniques including trapezoidal integration or Simpson's rule applied to discrete sample values within breath cycle boundaries. Breath cycle boundaries may be determined by identifying successive zero crossings, successive local minima, or successive local maxima within the respiratory waveform. The computed AUC value for each breath cycle may reflect the volume of air displaced during the breath, where breath cycles with larger tidal volumes produce respiratory waveforms with greater amplitude excursions sustained over longer durations, resulting in larger AUC values. The method may include tracking AUC values across successive breath cycles to identify trends in breathing depth over time. A decrease in AUC values over successive breath cycles may indicate shallow breathing or reduced respiratory effort, and an increase in AUC values over successive breath cycles may indicate compensatory deep breathing in response to physiological stress. Changes in AUC values over time may contribute to determination of the physiological instability indicator by reflecting changes in respiratory effort and tidal volume that precede hypoxic events.
The method may include analyzing peak height and amplitude variation within the respiratory waveform to detect periodic breathing patterns and to monitor changes in diaphragmatic amplitude. Peak height may be measured as an amplitude difference between a local maximum and a preceding or following local minimum within each breath cycle of the respiratory waveform. The method may include computing peak height values for successive breath cycles and analyzing variation in peak height values over a defined observation window. Variation in peak height may be quantified using standard deviation, coefficient of variation, or range of peak height values computed across multiple breath cycles. Elevated variation in peak height may indicate periodic breathing patterns characterized by alternating periods of deep and shallow breathing. Periodic breathing patterns may manifest as cyclic fluctuations in peak height values, where sequences of high peak height values alternate with sequences of low peak height values. The method may include detecting periodic breathing patterns by analyzing periodicity of peak height variation using autocorrelation analysis or spectral analysis applied to a time series of peak height values. The method may include monitoring increased diaphragmatic amplitude in response to hypoxia, where the body attempts to take deeper breaths as a compensatory mechanism to increase oxygen intake. Increased diaphragmatic amplitude may manifest as elevated peak height values in the respiratory waveform compared to baseline peak height values established during stable breathing periods. The method may compare current peak height values against baseline peak height values to detect compensatory increases in respiratory effort that may precede or accompany hypoxic events.
In an embodiment, the method may include applying Fast Fourier Transform (FFT) analysis to the respiratory waveform to assess signal-to-noise levels and to detect dual peaks indicating gasping or abnormal breathing patterns. FFT computation may transform windowed segments of the respiratory waveform from a time domain to a frequency domain representation, producing a frequency spectrum that reveals the distribution of signal power across different frequencies. The method may include analyzing the frequency spectrum to identify a dominant frequency component corresponding to a fundamental respiratory rate and to identify harmonic components at integer multiples of the fundamental frequency. The method may include computing a signal-to-noise ratio by comparing power at the dominant respiratory frequency and its harmonics against power at frequencies outside the respiratory frequency bands. A high signal-to-noise ratio may indicate clean, regular breathing with minimal interference from noise or artifacts, while a low signal-to-noise ratio may indicate degraded breathing quality, irregular breathing patterns, or the presence of interfering signals from coughing, movement, or other sources. The method may include detecting dual peaks within the frequency spectrum or within individual breath cycles of the respiratory waveform that indicate gasping or abnormal breathing patterns. Dual peaks may occur when a subject gasps during a breath cycle, producing an irregular inspiratory pattern with two distinct inhalation efforts within a single breath cycle. FFT analysis may reveal dual peaks as additional frequency components or as spectral broadening around the fundamental respiratory frequency. Time-domain analysis of individual breath cycles may reveal dual peaks as two local maxima within a single breath cycle where normal breathing would produce a single local maximum. Detection of dual peaks may elevate the physiological instability indicator due to the association of gasping patterns with respiratory distress and hypoxia risk.
In an embodiment, the method may include performing periodicity analysis to detect periodic breathing characterized by alternating periods of deep and shallow breathing or brief pauses indicating respiratory instability. Periodic breathing may manifest as cyclic variations in respiratory waveform amplitude, breath-to-breath interval duration, or both. The method may include analyzing the respiratory waveform to identify patterns where sequences of deep breaths with high amplitude alternate with sequences of shallow breaths with low amplitude. The method may include analyzing breath-to-breath interval durations to identify patterns where sequences of rapid breaths with short intervals alternate with sequences of slow breaths with long intervals or with brief pauses where breathing temporarily ceases. Brief pauses in breathing may be detected as intervals between successive breath cycles that exceed a threshold duration, where the threshold duration may be set based on a baseline breath-to-breath interval for the subject. The method may include quantifying periodicity by computing autocorrelation of a respiratory waveform amplitude envelope or a breath-to-breath interval time series, where periodic breathing produces autocorrelation peaks at lags corresponding to a period of cyclic variation. The method may further include quantifying periodicity by computing spectral analysis of the respiratory waveform amplitude envelope, where periodic breathing produces spectral peaks at frequencies corresponding to a cyclic variation rate. Detection of periodic breathing patterns may indicate respiratory instability associated with conditions including sleep apnea, Cheyne-Stokes respiration, or impending respiratory failure, and may contribute to elevation of the physiological instability indicator.
In an embodiment, the method may include assessing signal-to-noise ratio of the respiratory waveform as a measure of breathing quality. Signal-to-noise ratio assessment may be performed in the frequency domain by comparing power at respiratory frequency bands against power at non-respiratory frequency bands, or may be performed in the time domain by comparing amplitude of respiratory waveform features against amplitude of baseline noise or residual signal variations. A high signal-to-noise ratio may indicate that the respiratory waveform exhibits clear, well-defined breath cycles with minimal contamination from noise sources including sensor noise, environmental vibration, or movement artifacts. A low signal-to-noise ratio may indicate that the respiratory waveform is degraded by noise or artifacts, reducing the reliability of features extracted from the waveform, or may indicate abnormal breathing patterns where the respiratory signal itself exhibits irregular or attenuated characteristics. In some embodiments, the method may include using signal-to-noise ratio as a quality metric to gate feature extraction, where features extracted from waveform segments with signal-to-noise ratio below a quality threshold may be excluded from subsequent analysis or may be assigned reduced weight in determination of the physiological instability indicator. In some embodiments, the method may include tracking signal-to-noise ratio over time to detect degradation in breathing quality that may precede hypoxic events, where a declining trend in signal-to-noise ratio may indicate progressive respiratory compromise.
The method may include applying capnography-inspired waveform analysis to a BCG-derived respiratory waveform, where the shape of the BCG respiratory waveform changes depending on breathing depth, coughing, or other respiratory events similar to how capnography uses pressure differential and air pressure curves to assess respiratory effort. Capnography generates air pressure curves from nasal prong sensors that reflect respiratory effort, breathing pattern, and respiratory distress indicators based on the shape of the pressure waveform over each breath cycle. The BCG respiratory waveform may exhibit analogous shape variations that reflect breathing characteristics, where the waveform shape changes based on breathing depth, coughing during a breath cycle, or presence of respiratory distress. The method may include analyzing waveform shape features including symmetry of inhalation and exhalation phases, sharpness or smoothness of waveform transitions, and presence of inflection points or secondary peaks within breath cycles. Asymmetric waveforms where an inhalation phase differs in duration or shape from an exhalation phase may indicate respiratory effort imbalance or airway obstruction. Sharp waveform transitions may indicate forceful breathing, whereas smooth waveform transitions may indicate relaxed breathing. Secondary peaks or inflection points within breath cycles may indicate coughing, gasping, or other respiratory events that interrupt normal breathing patterns. The method may include extracting waveform shape features using morphological analysis techniques including template matching, curve fitting, or parameterized waveform models that characterize the shape of each breath cycle. The extracted waveform shape features may contribute to determination of the physiological instability indicator by reflecting respiratory effort characteristics and respiratory event patterns associated with hypoxia risk.
In an embodiment, the method may include computing breathing synchronicity and periodicity metrics including breath-to-breath interval variability, similar to heart rate variability, to indicate autonomic nervous system dysfunction or respiratory instability. Breath-to-breath interval variability may be computed as variation in time intervals between successive breath cycles over a defined observation window. The method may include computing breath-to-breath interval variability using time-domain metrics including standard deviation of breath-to-breath intervals, root mean square of successive differences in breath-to-breath intervals, and a percentage of successive breath-to-breath interval differences exceeding a threshold value. The method may include computing breath-to-breath interval variability using frequency-domain metrics by applying spectral analysis to a breath-to-breath interval time series and computing power in defined frequency bands. Elevated breath-to-breath interval variability may indicate irregular breathing patterns associated with respiratory instability or autonomic nervous system dysfunction affecting respiratory control. Reduced breath-to-breath interval variability may indicate rigid, inflexible respiratory patterns that may also be associated with physiological compromise. The method may include analyzing synchronicity between cardiac and respiratory components by comparing timing relationships between heartbeats and breath cycles. In some embodiments, the method may include computing respiratory sinus arrhythmia metrics that quantify modulation of heart rate by respiratory phase, where reduced respiratory sinus arrhythmia may indicate autonomic dysfunction. Breathing synchronicity and periodicity metrics may contribute to determination of the physiological instability indicator by reflecting regularity and autonomic control of respiratory function.
The method may include establishing a valid baseline for each subject over long-term monitoring to enable detection of deviations from a subject's normal physiological state and to account for individual variations and adaptations. Baseline establishment may involve computing reference values for respiratory features including respiration rate, AUC, peak height, breath-to-breath interval variability, and waveform shape metrics during periods when the subject exhibits a stable physiological state without respiratory distress. Baseline values may be computed as mean, median, or percentile values of respiratory features over a defined baseline establishment period. The baseline establishment period may span multiple hours or multiple days of monitoring to capture a subject's typical respiratory patterns across different activity states, sleep stages, and postural positions. The method may include establishing separate baseline values for different postures, where baseline values computed when the subject is supine may differ from baseline values computed when the subject is in lateral or prone positions. Posture-specific baseline establishment may account for influence of posture on respiratory mechanics and BCG signal coupling characteristics. The method may include updating baseline values over time using adaptive baseline tracking that incorporates recent stable measurements while maintaining robustness against transient disturbances. Long-term monitoring may enable detection of physiological adaptations that result in changes to baseline parameters, where subjects with chronic respiratory conditions may exhibit higher resting heart rate or altered respiratory patterns compared to healthy subjects. The method may include comparing current respiratory feature values against established baseline values to detect deviations that may indicate physiological deterioration or impending hypoxia, where deviations exceeding defined thresholds relative to baseline may contribute to elevation of the physiological instability indicator.
In an embodiment, the method may include detecting coughing events by analyzing movement artifacts and body jerks within the BCG signal that produce characteristic interference patterns distinguishable from normal respiratory and cardiac activity. Coughing produces a forceful expulsion of air accompanied by rapid contraction of respiratory and abdominal muscles, generating a transient high-amplitude mechanical disturbance that propagates through the body and couples to a BCG sheet. The mechanical disturbance from a cough event may manifest in the BCG signal as a jerky interference pattern characterized by a sharp amplitude excursion followed by oscillatory decay, where the amplitude excursion exceeds an amplitude range of normal respiratory waveform features by a factor of two or more. The method may include identifying cough events by detecting transient amplitude excursions in the BCG signal that exceed a cough detection threshold computed as a multiple of a baseline signal amplitude or as a multiple of a standard deviation of signal amplitude over a preceding reference window. The method may include further characterizing cough events by analyzing a temporal profile of the amplitude excursion, where cough events exhibit a rapid rise time on the order of tens to hundreds of milliseconds followed by a decay period on the order of hundreds of milliseconds. The method may include distinguishing cough events from other transient disturbances including limb movements or postural adjustments by analyzing frequency content of the transient disturbance, where cough events produce frequency content concentrated in a range associated with respiratory muscle contraction and thoracic vibration. Detection of coughing events may indicate respiratory irritation, airway obstruction, or respiratory infection, and repeated coughing events within a defined observation window may contribute to elevation of the physiological instability indicator due to the association of persistent coughing with respiratory distress and increased hypoxia risk.
For example, the method may include using movement artifacts and body jerks detected in the BCG signal as indicators of pain, coughing, or respiratory distress. Movement artifacts may be identified by detecting signal segments where BCG waveform amplitude, frequency content, or morphology deviates from expected characteristics of cardiac and respiratory components. Body jerks associated with pain responses may produce transient amplitude excursions similar to cough events but may exhibit different temporal profiles and may be accompanied by sustained postural changes reflected in baseline shifts of the BCG signal. The method may include analyzing patterns of movement artifacts over time to distinguish between isolated movement events and sustained movement activity indicative of restlessness or discomfort. Subjects experiencing pain or respiratory distress may exhibit increased frequency of movement artifacts as they shift position in an attempt to find a more comfortable posture or to relieve breathing difficulty. The method may include computing a movement artifact frequency metric representing a number of detected movement artifact events per unit time, where elevated movement artifact frequency may indicate subject distress. The method may include computing a movement artifact intensity metric representing average amplitude or energy of detected movement artifact events, where elevated movement artifact intensity may indicate forceful movements associated with severe discomfort or respiratory struggle. The movement artifact frequency metric and the movement artifact intensity metric may be incorporated as features in determination of the physiological instability indicator, where elevated values of these metrics may contribute to increased hypoxia risk assessment.
In an embodiment, the method may include identifying gasping events through detection of dual-peak waveforms within individual breath cycles of a respiratory waveform. Normal breathing produces a respiratory waveform with a single peak corresponding to maximum chest expansion during inhalation followed by a single trough corresponding to minimum chest position during exhalation. Gasping produces an irregular inspiratory pattern where a subject makes two or more distinct inhalation efforts within a single breath cycle, resulting in a dual-peak waveform where two local maxima appear within a time interval that would normally contain a single inhalation peak. The method may include detecting dual-peak waveforms by analyzing the respiratory waveform to identify breath cycles containing multiple local maxima separated by an intermediate local minimum that does not reach a baseline level associated with full exhalation. The method may include applying peak detection to the respiratory waveform and may identify dual-peak events when two successive local maxima occur within a time interval shorter than an expected breath cycle duration and when an intermediate local minimum between the two peaks remains above a threshold amplitude relative to a preceding exhalation trough. Detection of dual-peak waveforms may indicate gasping associated with respiratory distress, airway obstruction, or oxygen deprivation, and may contribute to elevation of the physiological instability indicator due to the association of gasping patterns with impending or ongoing hypoxia.
In an embodiment, the method may include differentiating snoring from other respiratory events through analysis of vibration signatures within the BCG signal. Snoring produces periodic low-frequency vibrations resulting from oscillation of soft tissues in an upper airway during inspiration, and these vibrations couple to a BCG sheet as a characteristic interference pattern superimposed on a respiratory waveform. A vibration signature of snoring may manifest as a modulation of respiratory waveform amplitude during an inspiratory phase, where the modulation exhibits a frequency in a range from approximately tens of hertz to a few hundred hertz corresponding to oscillation frequency of upper airway tissues. The method may include detecting snoring by applying high-pass filtering or bandpass filtering to the BCG signal to isolate a frequency range associated with snoring vibrations and by analyzing the filtered signal for periodic amplitude patterns that correlate with an inspiratory phase of a respiratory cycle. The method may include computing a snoring intensity metric representing amplitude or power of a detected vibration signature, and may compute a snoring frequency metric representing the number of breath cycles exhibiting snoring vibration signatures within a defined observation window. The method may include distinguishing snoring from coughing by analyzing temporal characteristics of detected events, where snoring produces sustained vibration patterns that persist throughout inspiratory phases of multiple successive breath cycles, while coughing produces isolated transient disturbances that interrupt normal respiratory rhythm. Snoring detection may indicate upper airway obstruction that can contribute to intermittent hypoxia, and persistent snoring patterns may contribute to determination of the physiological instability indicator in subjects at risk for obstructive sleep apnea or other conditions associated with airway compromise.
For example, the method may include detecting indicators of pneumonia by analyzing interference patterns in respiratory waveform curves that reflect crackles or fluid-related sounds in the lungs. Pneumonia causes accumulation of fluid, mucus, or inflammatory exudate in lung tissue and airways, and the presence of fluid produces characteristic acoustic phenomena during breathing that can be detected through mechanical coupling to the BCG signal. Crackles are discontinuous, non-musical sounds produced when collapsed or fluid-filled airways open during inspiration, and crackles may manifest as brief, high-frequency transient disturbances superimposed on a respiratory waveform. The method may include detecting crackles by analyzing the respiratory waveform for transient high-frequency components that occur during an inspiratory phase and that exhibit characteristics distinguishable from normal respiratory waveform features. The method may include applying high-pass filtering to the respiratory waveform to isolate frequency content above a normal respiratory frequency range and may detect crackle events as transient amplitude excursions in the filtered signal that exceed a crackle detection threshold. Presence of fluid in the lungs may also affect smoothness and regularity of a respiratory waveform, where fluid-related interference may produce irregular waveform morphology with reduced signal-to-noise ratio compared to respiratory waveforms from subjects with clear lungs. The method may include computing a waveform irregularity metric representing deviation of the respiratory waveform from a smooth template waveform, where elevated waveform irregularity may indicate fluid accumulation or other pulmonary pathology. Detection of crackle-like interference patterns or elevated waveform irregularity may indicate pneumonia or other conditions affecting lung function, and may contribute to elevation of the physiological instability indicator due to the association of pneumonia with increased hypoxia risk.
In an embodiment, the method may include performing trend analysis via time-embedded vectors to track hypoxia risk over time and to determine whether a subject's physiological state is improving or deteriorating. Trend-based risk tracking may involve computing the physiological instability indicator at successive time intervals throughout a monitoring session and analyzing trajectories of physiological instability indicator values over time. The method may include storing a time series of physiological instability indicator values in memory, where each stored value is associated with a timestamp indicating when the value was computed. The method may include analyzing the time series of physiological instability indicator values to identify trends including increasing trends where successive values exhibit a rising pattern, decreasing trends where successive values exhibit a falling pattern, and stable trends where successive values remain within a defined variance range. An increasing trend in the physiological instability indicator may indicate that the subject's hypoxia risk is worsening over time, suggesting that current clinical interventions may be insufficient or that the subject's underlying condition is progressing. A decreasing trend in the physiological instability indicator may indicate that the subject's hypoxia risk is improving over time, suggesting that clinical interventions such as medication administration, oxygen therapy, or postural adjustments are producing a beneficial effect. Trend analysis may enable medication titration by providing clinical staff with information regarding whether a subject's physiological state is responding to treatment, where an improving trend may support continuation of current medication dosing and a worsening trend may prompt adjustment of medication dosing or initiation of additional interventions.
In an embodiment, the method may include applying trend detection algorithms to a time series of physiological instability indicator values to quantify direction and magnitude of trends. Trend detection algorithms may include linear regression analysis where a regression line is fitted to the time series of physiological instability indicator values and a slope of the regression line indicates trend direction and rate of change. A positive slope may indicate a worsening trend with increasing hypoxia risk, and a negative slope may indicate an improving trend with decreasing hypoxia risk. Magnitude of the slope may indicate the rate at which hypoxia risk is changing, where larger magnitude slopes indicate more rapid changes in physiological state. The method may include applying moving average analysis to smooth short-term fluctuations in physiological instability indicator values and to reveal underlying trends that may be obscured by measurement noise or transient physiological variations. The method may include computing a short-term moving average over a window of recent values and a long-term moving average over a window spanning a longer historical period, where crossover events between short-term and long-term moving averages may indicate trend reversals. The method may include generating trend-based alerts when trend analysis indicates sustained worsening of hypoxia risk, where a sustained worsening trend may be defined as an increasing trend that persists for a defined minimum duration or that exceeds a defined cumulative change threshold. The trend-based alerts may be distinct from threshold-based alerts generated when the physiological instability indicator exceeds a stored threshold, thereby enabling clinical staff to receive early warning of deteriorating trends before the physiological instability indicator reaches alert threshold levels.
In an embodiment, the method may include handling posture changes and sleep stage proxies using movement and activity patterns detected within the BCG signal. Posture changes may produce characteristic movement artifact signatures in the BCG signal as a subject shifts from one position to another, and the method may include detecting posture change events by identifying sustained movement artifact activity followed by changes in baseline signal characteristics. The method may include maintaining posture-specific baseline values and may select an appropriate baseline for comparison based on a detected current posture. Sleep stage proxies may be derived from movement and activity patterns, where periods of reduced movement activity may indicate deeper sleep stages and periods of increased movement activity may indicate lighter sleep stages or wakefulness. The method may include adjusting feature extraction parameters or interpretation thresholds based on inferred sleep stage to account for normal physiological variations associated with different sleep stages. The method may include detecting respiratory muscle adaptation over long-term monitoring by analyzing changes in respiration rate curve morphology. Respiratory muscle strengthening or stiffening may cause the respiration rate curve to become less smooth with increased skew, where the skew reflects asymmetry between inspiratory and expiratory phases of breathing. The method may include computing skewness metrics for the respiratory waveform over successive analysis windows and may track changes in skewness over long-term monitoring periods spanning multiple days or weeks. An increasing trend in respiratory waveform skewness may indicate respiratory muscle adaptation associated with chronic respiratory effort or underlying pulmonary pathology. The method may include monitoring changes in area under the curve for the respiration waveform to detect overall volumetric increases in tidal volume over long-term monitoring. Tidal volume area-of-curve tracking may involve computing area under the respiratory waveform curve for successive breath cycles and analyzing trends in computed area values over extended monitoring periods. An increasing trend in tidal volume area may indicate compensatory increases in lung capacity or breathing depth in response to chronic hypoxic stress or respiratory compromise.
In an embodiment, the method may include handling missing data segments that occur when the BCG signal is unavailable or unreliable due to subject absence from the bed, sensor malfunction, or severe artifact contamination. Missing data segment detection may involve monitoring signal quality metrics including signal amplitude, signal-to-noise ratio, and artifact frequency, and identifying time intervals where the signal quality metrics fall below defined thresholds indicating that the BCG signal does not contain reliable physiological information. The method may include flagging missing data segments and may exclude features extracted from missing data segments from trend analysis and physiological instability indicator computation. The method may include implementing reinitialization logic to resume normal monitoring operations after missing data segments end. Reinitialization logic may involve re-establishing baseline values when duration of a missing data segment exceeds a defined threshold, where extended absence from monitoring may result in physiological state changes that render previously established baselines invalid. Reinitialization logic may involve applying a stabilization period after the subject returns to the bed, where features extracted during the stabilization period may be used to update baseline values before resuming alert generation. The method may include interpolating across short missing data segments to maintain continuity of trend analysis, where interpolation may use linear interpolation, spline interpolation, or model-based prediction to estimate feature values during the missing data segment based on feature values before and after the segment. The method may further include adjusting confidence levels associated with the physiological instability indicator based on recency and completeness of available data, where physiological instability indicator values computed from data with recent missing segments may be assigned reduced confidence compared to values computed from continuous uninterrupted data streams.
Referring to Fig. 2, at step 202, the method includes receiving a ballistocardiography signal from a non-contact sensor positioned beneath a subject support surface. Further, at step 204, the ballistocardiography signal to isolate a cardiac component signal and a respiratory component signal is filtered. At step 206, a set of signal dynamics data from the cardiac component signal and the respiratory component signal is derived. Further the method includes, generating a multidimensional time-embedded vector representation of the set of signal dynamics data using phase space reconstruction, at step 208. At step 210, a recurrence matrix from the multidimensional time-embedded vector representation is constructed. Further, at step 212, at least one dynamic feature from the recurrence matrix is determined. Furthermore, at step 214, the at least one dynamic feature with a historical set of signal dynamics data is compared, to establish a correlation with hypoxic conditions. Further, at step 216 a likelihood of hypoxia based on the comparison is predicted. Finally at step 218, an alert indicative of the likelihood of the hypoxia is generated.
FIGs. 3A to 3E illustrate graphical representations of vitals of five subjects in non-hypoxic state, in accordance with an embodiment of the present disclosure. As shown, the vitals recorded are heart rate, respiration rate and blood oxygen saturation (SpO2). In this regard, the vitals (i.e., heart rate, respiration rate and blood oxygen saturation (SpO2)) are recorded consistently over time. Notably, the values for the vitals are shown in Y-axis and the time of recording are shown in X-axis. As shown in FIGs. 3A to 3E, the graphs for the heart rate, and the respiration rate show no sudden increase of values. Similarly, the graphs for blood oxygen saturation (SpO2) demonstrate no sudden decrease of values. Thus, the concerned subjects are in a normal i.e., non-hypoxic state.
FIGs. 3F to 3J illustrate graphical representations of vitals of five subjects in hypoxic state, in accordance with an embodiment of the present disclosure. With respect to FIGs. 3F to 3J, the vitals recorded are heart rate, respiration rate and blood oxygen saturation (SpO2). The vitals are recorded consistently over time. Notably, the values for the vitals are shown in Y-axis and the time of recording are shown in X-axis. As shown, the graphs for the heart rate, and the respiration rate show sudden increase of values at certain time t1. Similarly, the graphs for blood oxygen saturation (SpO2) demonstrated sudden decrease of values correspondingly, at time t1, when the heart rates and the respiration rates show sudden increase. Thus, it can be asserted that when blood oxygen saturation (SpO2) decreases the heart rate and respiration rate increases. Such response indicates that the subjects are under deprivation of required oxygen and may go into hypoxic state in near future.
FIGs. 4A to 4E illustrates graphical representations of second harmonics analysis for a set of BCG waveform data corresponding to five subjects in non-hypoxic state, in accordance with an embodiment of the present disclosure. As shown, the fluctuations in the second harmonic power of the set of BCG waveform data namely, heart rate, for five subjects. In this regard, the X-axis represents time, and Y-axis represents power value of second harmonics of heart rate. Notably, the graphical representations show no sudden fluctuation in power value with change in time. Therefore, it can be concluded that the five subjects are in non-hypoxic state.
FIGs. 4F to 4J illustrate graphical representations of second harmonics analysis for a set of BCG waveform data corresponding to five subjects in hypoxic state, in accordance with an embodiment of the present disclosure. As shown, the graphical representations show sudden fluctuation in power value with change in time. Therefore, it can be concluded that the five subjects are in hypoxic state.
Referring to FIGs. 5A to 5E, illustrated are recurrence plots plotted for a set of signal dynamics data corresponding to five subjects in non-hypoxic state, in accordance with an embodiment of the present disclosure. In this regard, the recurrence plots are used to identify parameters from amongst the set signal dynamics data to quantify results from recurrence analysis applied on multidimensional time-embedded vector (constructed from the set signal dynamics data using at least one representation technique). Notably, from the recurrence plots provides insight on time steps at which the subject’s previous state recurs. Notably, the recurrence plots shown in FIGs. 5A to 5E are plotted for change in spectral shift in power level of the BCG waveform (i.e., the second harmonics analysis of heart rate plotted in FIGs. 4A to 4E). It may be appreciated that the recurrence plots can be plotted from any parameter from amongst the set of signal dynamics data. As shown, grey regions in the recurrence plots indicate stable region and black regions in the recurrence plots indicate unstable regions indicating hypoxia onset. In FIGs. 5A to 5E, the grey regions are widely observed in the recurrence plots. So, it can be concluded that the subjects are in non-hypoxic state.
FIGs. 5F to 5J illustrate recurrence plots plotted for a set of signal dynamics data corresponding to five subjects in hypoxic state, in accordance with an embodiment of the present disclosure. Notably, the recurrence plots shown in FIGs. 5F to 5J are plotted for change in spectral shift in power level of the BCG waveform (i.e., the second harmonics analysis of heart rate plotted in FIGs. 4F to 4J). Notably, the hypoxia can be clearly distinguished by looking at a relative fractional area of the grey region with respect to the black region. It may be appreciated that when the grey region is less than 0.75 of total plot area, then it can be concluded that the concerned subjects are in hypoxia state. As shown, the black regions are widely observed in the recurrence plots. So, it can be concluded that the subjects are in hypoxic state.
Modifications to embodiments of the invention described in the foregoing are possible without departing from the scope of the invention as defined by the accompanying claims. Expressions such as “including”, “comprising”, “incorporating”, “consisting of”, “have”, “is” used to describe and claim the present invention are intended to be construed in a non-exclusive manner, namely allowing for items, components or elements not explicitly described also to be present. Reference to the singular is also to be construed to relate to the plural. Numerals included within parentheses in the accompanying claims are intended to assist understanding of the claims and should not be construed in any way to limit subject matter claimed by these claims.

EXPERIMENTAL PART
An experiment on BCG waveform data collected for five subjects was conducted. The collected BCG waveform data was filtered, and cardiac component (waveforms related to cardiac activity) and respiratory component (waveform related to respiratory activities) were extracted. Such extracted data primarily pertains to heart rate and respiratory rate respectively.
The standard deviation, coefficient of variation of heart rate (HR), and max – min HR value during hypoxia and non-hypoxia state was computed and noted in tabulation 1 given below.
Tabulation 1
Subject ID standard deviation: non-hypoxia region standard deviation: hypoxia region coefficient of variation: non-hypoxia region coefficient of variation: hypoxia region max – min HR: non-hypoxia region max – min HR: hypoxia region
DZ_001 3.14 17.29 0.04 0.16 12.19 51.13
DZ_002 1.58 7.95 0.02 0.10 8.72 22.14
DZ_003 2.06 7.76 0.03 0.09 7.95 23.67
DZ_004 1.40 6.95 0.01 0.07 6.61 16.89
DZ_005 1.37 10.10 0.03 0.13 6.03 34.88
From the tabulation 1, it was observed that during hypoxia state, the standard deviation, the coefficient of variation, and max-min values of heart rate was increased. Thus, it was concluded that increases in the standard deviation, the coefficient of variation, and max-min values of heart rate was indicative on likelihood of hypoxia.
Similarly, the standard deviation, coefficient of variation of respiration rate (RR) and max – min RR value during hypoxia and non-hypoxia state was computed and noted in tabulation 2 given below.
Tabulation 2
Subject ID Standard Deviation: non-hypoxia region Standard Deviation: hypoxia region coefficient of variation: non-hypoxia region coefficient of variation: hypoxia region max – min RR: non-hypoxia region max – min RR: hypoxia region
DZ_001 1.72 1.58 0.08 0.05 5.88 6.90
DZ_002 0.91 3.88 0.03 0.13 3.13 12.03
DZ_003 0.87 1.87 0.06 0.09 3.53 7.41
DZ_004 0.78 2.51 0.03 0.08 3.90 10.43
DZ_005 0.95 2.17 0.05 0.09 4.33 8.36
From the tabulation 2, it was observed that during hypoxia state, the standard deviation, the coefficient of variation, and max-min values of respiration rate were generally increased. Thus, it was concluded that increases in the standard deviation, the coefficient of variation, and max-min values of respiration rate were indicative of likelihood of hypoxia.
The method steps described herein may be reordered where technically compatible, and the specific sequence of steps presented in the description represents one embodiment that does not limit the scope of the disclosure to that particular sequence. Steps that do not have data dependencies on preceding steps may be performed in parallel or in alternative sequences without departing from the technical principles described herein. Steps that produce intermediate results used by subsequent steps may be performed in sequences that preserve the data dependencies while allowing flexibility in the ordering of independent operations. The scope of the disclosure covers equivalents to the described embodiments, where equivalents include alternative implementations that achieve substantially the same function using substantially the same means to produce substantially the same result. Equivalents may include alternative filtering techniques that achieve component isolation between cardiac and respiratory signals, alternative feature extraction techniques that derive signal dynamics data from filtered BCG waveforms, alternative time embedding techniques that generate multidimensional representations from time series data, alternative recurrence analysis techniques that quantify dynamic features from embedded representations, and alternative classification techniques that generate physiological instability indicators from extracted features. The specific numerical values, threshold values, frequency ranges, time intervals, and other parameters presented in the description represent exemplary values that may be adjusted based on sensor characteristics, subject population characteristics, clinical requirements, and deployment constraints without departing from the scope of the disclosure.
A number of implementations have been described. Nevertheless, it will be understood that various modifications may be made without departing from the spirit and scope of the disclosure. Accordingly, other implementations are within the scope of the following claims.
, Claims:CLAIMS
1. A computer-implemented method for predicting likelihood of hypoxia from non-contact physiological monitoring, comprising:
receiving, by a processor, a ballistocardiography signal from a non-contact sensor positioned beneath a subject support surface, wherein the ballistocardiography signal comprises cardiac component signals associated with mechanical vibrations produced by cardiac ejection and respiratory component signals associated with respiratory motion of a subject;
filtering, by the processor, the ballistocardiography signal to isolate a cardiac component signal and a respiratory component signal;
deriving, by the processor, a set of signal dynamics data from the cardiac component signal and the respiratory component signal, wherein the set of signal dynamics data comprises heart rate values, respiration rate values, and spectral power measurements, and wherein the set of signal dynamics data is correlated to a health status of the subject;
generating, by the processor, a multidimensional time-embedded vector representation of the set of signal dynamics data using phase space reconstruction, wherein the multidimensional time-embedded vector representation projects delayed versions of the set of signal dynamics data into a multidimensional space;
constructing, by the processor, a recurrence matrix from the multidimensional time-embedded vector representation, wherein the recurrence matrix identifies recurring states indicating periodicity of data values in the set of signal dynamics data;
determining, by the processor, at least one dynamic feature from the recurrence matrix, wherein the at least one dynamic feature represents a parameter derived from the set of signal dynamics data that exhibits change in state over time;
comparing, by the processor, the at least one dynamic feature with a historical set of signal dynamics data to establish a correlation with hypoxic conditions; and
predicting, by the processor, a likelihood of the hypoxia based on the comparison, wherein a deviation of the at least one dynamic feature beyond a predefined baseline value indicates increased hypoxia risk; and
generating, by the processor, an alert indicative of the likelihood of the hypoxia in response to the hypoxia risk exceeding a predefined risk threshold.
2. The method of claim 1, wherein the phase space reconstruction uses an embedding dimension between 3 and 6 and a time delay selected based on sampling characteristics of the ballistocardiography signal, and wherein each state vector in the multidimensional space comprises signal values at time t, t-12, t-24, and t-36.

3. The method of claim 1, further comprising determining that a recurrence-derived stability metric computed from the recurrence matrix falls below a predefined stability threshold indicating elevated hypoxia risk.

4. The method of claim 1, wherein the signal dynamics data further comprises:
standard deviation of heart rate computed over a defined time window;
coefficient of variation of heart rate;
standard deviation of respiration rate computed over the defined time window; and
coefficient of variation of respiration rate.

5. The method of claim 4, further comprising comparing the set of signal dynamics data against a subject-specific historical baseline derived from a prior monitoring period a subject-specific historical baseline derived from a prior monitoring period corresponding to a non-hypoxic physiological state.

6. The method of claim 5, further comprising detecting an increase in heart rate relative to the subject-specific historical baseline as an indicator of the increased hypoxia risk.

7. The method of claim 1, wherein filtering the ballistocardiography signal comprises:
applying a first bandpass filter configured to pass frequencies in a range of 1 Hz to 10 Hz to isolate the cardiac component signal; and
applying a second bandpass filter configured to pass frequencies in a range of 0.1 Hz to 0.5 Hz to isolate the respiratory component signal.

8. The method of claim 1, further comprising extracting respiratory waveform features from the respiratory component signal, wherein the respiratory waveform features comprise area under an curve, peak height amplitude, inspiration-to-expiration ratio, and breath-to-breath interval variability.

9. The method of claim 1, wherein the spectral power measurements comprise second harmonic power of the cardiac component signal, and wherein fluctuations in the second harmonic power over successive time windows indicate physiological instability preceding hypoxia.

10. The method of claim 1, further comprising applying a machine learning classifier to the signal dynamics data and recurrence-derived dynamic features to generate the hypoxia risk indicator, wherein the machine learning classifier comprises a support vector machine, a random forest classifier, or a neural network trained on labeled ballistocardiography data from subjects with documented hypoxic events and subjects without hypoxic events.

11. The method of claim 1, further comprising receiving oxygen saturation measurements from a secondary sensor and using the oxygen saturation measurements as a confidence metric for recalibrating the hypoxia risk indicator.

12. A system for predicting likelihood of hypoxia from non-contact physiological monitoring, the system comprising:
at least one non-contact sensor positioned beneath a subject support surface and configured to acquire a ballistocardiography signal comprising mechanical vibrations produced by cardiac ejection and respiratory-induced components associated with respiratory motion of a subject; and
a processor communicatively coupled to the at least one non-contact sensor, the processor being configured to:
receive the ballistocardiography signal from the at least one non-contact sensor;
filter the ballistocardiography signal to isolate a cardiac component signal and a respiratory component signal;
derive a set of signal dynamics data from the cardiac component signal and the respiratory component signal, wherein the set of signal dynamics data comprises heart rate values, respiration rate values, and spectral power measurements, the set of signal dynamics data being correlated to a health status of the subject;
generate a multidimensional time-embedded vector representation of the set of signal dynamics data using phase space reconstruction, wherein the time-embedded vector representation projects delayed versions of the set of signal dynamics data into a multidimensional space;
construct a recurrence matrix from the time-embedded vector representation, wherein the recurrence matrix identifies recurring states indicating periodicity of data values in the set of signal dynamics data;
determine at least one dynamic feature from the recurrence matrix, wherein the at least one dynamic feature represents a parameter derived from the set of signal dynamics data that exhibits change in state over time;
compare the at least one dynamic feature with a historical set of signal dynamics data to establish a correlation with hypoxic conditions;
predict a likelihood of the hypoxia based on the comparison, wherein a deviation of the at least one dynamic feature beyond a predefined baseline value indicates increased hypoxia risk; and
generate an alert indicative of the likelihood of the hypoxia when the hypoxia risk exceeds a predefined risk threshold.

13. The system of claim 12, wherein the processor is configured to perform the phase space reconstruction using an embedding dimension between 3 and 6 and a time delay selected based on sampling characteristics of the ballistocardiography signal, such that each state vector in the multidimensional space comprises signal values at time t, t-12, t-24, and t-36.

14. The system of claim 12, wherein the processor is configured to compute a recurrence-derived stability metric from the recurrence matrix, and wherein the stability metric falling below a predefined stability threshold indicates elevated hypoxia risk.

15. The system of claim 12, wherein the set of signal dynamics data further comprises:

standard deviation of heart rate computed over a defined time window;
coefficient of variation of heart rate;
standard deviation of respiration rate computed over the defined time window; and
coefficient of variation of respiration rate.

16. The system of claim 15, wherein the processor is further configured to compare the signal dynamics data against a subject-specific historical baseline derived from a prior monitoring period corresponding to a non-hypoxic physiological state.

17. The system of claim 16, wherein the processor is further configured to detect an increase in heart rate relative to the subject-specific historical baseline as an indicator of elevated hypoxia risk.

18. The system of claim 12, wherein the processor is configured to filter the ballistocardiography signal by:
applying a first bandpass filter configured to pass frequencies in a range of 1 Hz to 10 Hz to isolate the cardiac component signal; and
applying a second bandpass filter configured to pass frequencies in a range of 0.1 Hz to 0.5 Hz to isolate the respiratory component signal.

19. The system of claim 12, wherein the processor is further configured to extract respiratory waveform features from the respiratory component signal, the respiratory waveform features comprising area under the curve, peak height amplitude, inspiration-to-expiration ratio, and breath-to-breath interval variability.

20. The system of claim 12, wherein the processor is further configured to apply a machine learning classifier to the signal dynamics data and recurrence-derived dynamic features to generate the hypoxia risk indicator, wherein the machine learning classifier comprises a support vector machine, a random forest classifier, or a neural network trained on labeled ballistocardiography data from subjects with documented hypoxic events and subjects without hypoxic events.