Physiological characteristic monitoring data processing method, device and computer readable storage medium

Abstract

The application discloses a physiological characteristic monitoring data processing method and device and a computer readable storage medium, and the method comprises the following steps: acquiring radar in-phase and quadrature signals, positioning a target thoracic cavity distance unit, extracting phase information, and generating a micro-motion signal; performing signal separation on the micro-motion signal based on a preset frequency interval, extracting independent respiratory signals and heartbeat signals; extracting a heartbeat cycle sequence of the heartbeat signal to generate a heart rhythm state feature; calculating the timing correlation of the respiratory signal and the heartbeat cycle sequence to generate a respiratory-heartbeat correlation feature; inputting the heart rhythm state feature and the respiratory-heartbeat correlation feature into a preset classification network to perform feature fusion and inference operation, and outputting a sinus rhythm type. The application realizes high-fidelity decoupling of a non-contact heart-lung signal, can accurately identify respiratory and non-respiratory sinus arrhythmia, effectively avoids false health warnings, and breaks through the cognitive barriers of professional medical indicators.

Description

Technical Field

[0001] This invention relates to the field of vital sign monitoring technology, and in particular to a method, device and computer-readable storage medium for processing physiological characteristic monitoring data. Background Technology

[0002] Sinus rhythm is the standard physiological manifestation of a normal heartbeat. When the automaticity of the sinoatrial node changes, it can cause sinus arrhythmia. In clinical medicine, sinus arrhythmia is mainly divided into respiratory and non-respiratory types: respiratory sinus arrhythmia is usually a marker of cardiovascular health and is a normal physiological compensation that requires no intervention; while non-respiratory sinus arrhythmia is mostly caused by pathological factors (such as sinoatrial node dysfunction, autonomic nervous system disorders, or organic heart diseases such as coronary heart disease and hypertension), requiring timely warning and identification of the cause to avoid missing the optimal treatment window for underlying diseases. Therefore, accurate classification and monitoring of sinus arrhythmia has extremely important clinical and daily health management value.

[0003] However, in the process of large-scale implementation of existing technologies in daily home monitoring scenarios, the following problems still exist: First, traditional sinus rhythm diagnosis relies heavily on contact-based devices such as electrocardiograms (ECGs) to assess cardiac electrical conduction. These devices are not only complex and expensive, but also require users to wear electrodes or restraints for extended periods, resulting in significant discomfort and making them incompatible with the long-term, continuous, and imperceptible sleep and health monitoring scenarios in the home environment.

[0004] Secondly, although frequency modulated continuous wave (FMCW) millimeter-wave radar has been introduced into the monitoring field due to its advantages such as strong penetration and immunity to light, most existing solutions only extract and analyze heart rate variability (HRV) indicators, lacking the ability to deeply identify and classify sinus arrhythmias. More critically, there is currently no universally accepted optimal range for HRV indicators in medicine, and its values ​​are easily affected by factors such as user age, gender, hormones, and lifestyle. This model of directly outputting complex clinical indicators that heavily rely on professional physician interpretation to ordinary users results in a severe lack of practicality and guidance for existing solutions in home settings.

[0005] Finally, most existing solutions use a single signal processing algorithm to make a rough judgment on arrhythmia, resulting in extremely low accuracy in type differentiation, which can easily trigger false alarms and cause unnecessary medical anxiety for users.

[0006] Therefore, there is an urgent need in this field for a novel method for processing physiological characteristic monitoring data to achieve highly comfortable, accurate, and intelligent classification and early warning of sinus rhythm. Summary of the Invention

[0007] This application provides a physiological characteristic monitoring data processing method, aiming to achieve highly comfortable, high-precision, and intelligent classification and early warning of sinus rhythm state.

[0008] To achieve the above objectives, embodiments of this application provide a method for processing physiological characteristic monitoring data, including: Acquire radar in-phase orthogonal signals, locate the target's thoracic cavity range cell based on the radar in-phase orthogonal signals and extract phase information to generate micro-motion signals; Based on a preset frequency range, the micro-motion signal is separated to extract the respiratory and heartbeat signals; Extract the heartbeat cycle sequence from the heartbeat signal, and generate heart rhythm state features based on the heartbeat cycle sequence; Calculate the temporal correlation between respiratory signals and heart rate cycle sequences to generate respiratory-heart rate correlation features; The heart rhythm characteristics and breathing-heartbeat correlation characteristics are input into a preset classification network to perform feature fusion and inference operations, and output the sinus rhythm type.

[0009] In one embodiment, acquiring radar in-phase orthogonal signals, locating the target thoracic cavity range cell based on the radar in-phase orthogonal signals and extracting phase information, and generating micro-motion signals, includes: Extract the environmental background reference signal within a preset time window, and perform static clutter filtering on the radar in-phase quadrature signal based on the environmental background reference signal to generate a denoised signal; A distance fast Fourier transform is performed on the denoised signal to generate a distance spectrum matrix. Based on the constant false alarm rate detection algorithm, the distance spectrum matrix is ​​traversed, and the distance cell corresponding to the maximum signal energy is locked as the target chest cavity distance cell. The initial phase information of the target thoracic cavity distance unit is extracted, and the phase entanglement in the initial phase information is eliminated based on the temporal phase difference rule to generate a micro-motion signal.

[0010] In one embodiment, static clutter filtering is performed on the radar in-phase quadrature signal based on the environmental background reference signal to generate a denoised signal, including: extracting the mean of the in-phase component and the mean of the quadrature component of the environmental background reference signal within a set number of frames; subtracting the mean of the in-phase component from the in-phase component of the radar in-phase quadrature signal, and subtracting the mean of the quadrature component from the quadrature component to generate a denoised signal.

[0011] In one embodiment, the distance spectrum matrix is ​​traversed based on a constant false alarm rate (CFAR) detection algorithm, and the distance cell corresponding to the maximum signal energy value is locked as the target chest cavity distance cell. This includes: for the cell to be detected in the distance spectrum matrix, a preset number of reference cells located on both sides of the cell to be detected are selected to calculate the average power; a detection threshold is set based on a preset false alarm probability and average power; and the distance cell with power greater than the detection threshold and the maximum energy is extracted as the target chest cavity distance cell.

[0012] In one embodiment, the micro-motion signal is generated by eliminating phase entanglement in the initial phase information based on the time-series phase difference rule. This includes: calculating the difference between the phase of the current frame and the phase of the previous frame; subtracting a preset period compensation value from the phase of the current frame in response to the difference being greater than a preset positive threshold; and adding the preset period compensation value to the phase of the current frame in response to the difference being less than a preset negative threshold, thereby generating the micro-motion signal.

[0013] In one embodiment, signal separation is performed on the micro-motion signal based on a preset frequency range to extract the respiratory signal and heartbeat signal, including: The micro-motion signal is subjected to variable mode empirical decomposition to generate multiple eigenmode functions with independent center frequencies; Based on preset heart rate frequency ranges and preset respiratory rate frequency ranges, multiple intrinsic mode functions with independent center frequencies are filtered to extract target heart rate mode functions and target respiratory rate mode functions; Signal reconstruction is performed on the target heart rate mode function and the target respiratory rate mode function respectively to generate heartbeat signals and respiratory signals.

[0014] In one embodiment, a variable mode empirical decomposition process is performed on the micro-motion signal to generate multiple eigenmode functions with independent center frequencies. This includes: configuring the number of decomposition parameters to three, configuring the penalty factor to two thousand, and constructing a variable mode empirical decomposition objective function; performing iterative decomposition on the micro-motion signal based on the variable mode empirical decomposition objective function to generate multiple eigenmode functions with a number equal to the number of decomposition parameters.

[0015] In one embodiment, based on a preset heart rate frequency range and a preset respiratory rate frequency range, multiple intrinsic mode functions with independent center frequencies are screened to extract target heart rate mode functions and target respiratory rate mode functions, including: extracting intrinsic mode functions with center frequencies in the range of 0.6 Hz to 2 Hz as target heart rate mode functions; and extracting intrinsic mode functions with center frequencies in the range of 0.1 Hz to 0.6 Hz as target respiratory rate mode functions.

[0016] In one embodiment, the heartbeat cycle sequence of the heartbeat signal is extracted, and heart rhythm state features are generated based on the heartbeat cycle sequence, including: Peak detection is performed on the heartbeat signal based on the statistical characteristics of a preset sliding window, and the peak timestamp sequence is extracted. Calculate the time difference between adjacent peak timestamps to generate a heartbeat cycle sequence; Calculate time-domain statistical features and / or frequency-domain power features based on the heartbeat cycle sequence, and generate heart rhythm state features based on the time-domain statistical features and / or frequency-domain power features.

[0017] In one embodiment, peak detection is performed on the heartbeat signal based on the statistical characteristics of a preset sliding window, and the peak timestamp sequence is extracted, including: Calculate the mean and standard deviation of the heartbeat signal within a preset sliding window; construct a dynamic detection threshold based on the mean and standard deviation of the waveform; extract the timestamps corresponding to the peaks whose amplitudes are greater than the dynamic detection threshold; and generate a peak timestamp sequence.

[0018] In one embodiment, the time-domain statistical features include at least one of the following: the percentage of adjacent period differences exceeding the limit, the standard deviation of the periodic sequence, and the root mean square of adjacent period differences. Frequency domain power characteristics include at least one of low-frequency power, high-frequency power, and the ratio of low-frequency to high-frequency power. Based on the heartbeat cycle sequence, time-domain statistical features and / or frequency-domain power features are calculated, and based on the time-domain statistical features and / or frequency-domain power features, cardiac rhythm state features are generated, including: The time-domain statistical features and frequency-domain power features are calculated based on the heartbeat cycle sequence, and the extracted time-domain statistical features and frequency-domain power features are concatenated by vector dimension to generate heart rhythm state features.

[0019] In one embodiment, calculating time-domain statistical features based on the heartbeat cycle sequence includes: The percentage of adjacent heartbeat cycle differences whose absolute value exceeds a preset time difference threshold is calculated as the percentage of adjacent cycle differences exceeding the limit. Calculate the standard deviation of the heartbeat cycle sequence and use it as the standard deviation of the cycle sequence. Calculate the root mean square of the difference between adjacent cycles in the heartbeat cycle sequence, and use it as the root mean square of the difference between adjacent cycles.

[0020] In one embodiment, calculating the time-frequency power characteristics based on the heartbeat cycle sequence includes: Calculate the integral value of the power spectral density of the heartbeat cycle sequence in the preset low-frequency range, and use it as the low-frequency power. Calculate the integral value of the power spectral density in the preset high-frequency range, and use it as the power in the high-frequency band; Calculate the ratio of low-frequency power to high-frequency power, and use this as the low-frequency to high-frequency power ratio.

[0021] In one embodiment, the temporal correlation between respiratory signals and heart rate cycle sequences is calculated to generate respiratory-heart rate correlation features, including: Perform time-length alignment processing on the heartbeat cycle sequence to align the data length of the heartbeat cycle sequence with that of the respiratory signal in the time dimension, and use the aligned heartbeat cycle sequence as the target cycle sequence. Standardization processing is performed on the respiratory signal and the target periodic sequence respectively to generate standard respiratory signal and standard periodic sequence that conform to the preset distribution state; Calculate the cross-correlation function between the standard respiratory signal and the standard periodic sequence, and extract the peak value of the cross-correlation coefficient; The target duration that is greater than the preset correlation threshold in the result of the statistical cross-correlation function is calculated as the proportion of the target duration in the total monitoring duration, and the cross-correlation compliance duration percentage is generated. By combining the peak value of the cross-correlation number with the proportion of cross-correlation duration meeting the target, a respiratory-heartbeat correlation feature is generated.

[0022] In one embodiment, a time-length alignment process is performed on the heartbeat cycle sequence to align the data length of the heartbeat cycle sequence with that of the respiratory signal in the time dimension. The aligned heartbeat cycle sequence is then used as the target cycle sequence. This process includes: using linear interpolation to fill in the missing data between adjacent data points of the heartbeat cycle sequence to generate a target cycle sequence with the same data length as the respiratory signal.

[0023] In one embodiment, standardization processing is performed on the respiratory signal and the target periodic sequence respectively to generate a standard respiratory signal and a standard periodic sequence that conform to a preset distribution state. This includes: calculating the sequence mean and the sequence standard deviation of the respiratory signal and the target periodic sequence respectively; subtracting the sequence mean from the corresponding signal sequence and dividing by the sequence standard deviation to convert them into standard normal distribution signals with a mean of zero and a standard deviation of one, which are used as the standard respiratory signal and the standard periodic sequence.

[0024] In one embodiment, extracting the peak value of the cross-correlation coefficient includes: normalizing the result of the cross-correlation function, taking the absolute value of the normalized cross-correlation sequence, and extracting the maximum value among the absolute values ​​as the peak value of the cross-correlation coefficient.

[0025] In one embodiment, cardiac rhythm features and respiratory-heartbeat correlation features are input into a preset classification network to perform feature fusion and inference operations, and output sinus rhythm type, including: The heart rhythm state features are combined with the breathing-heartbeat correlation features into a static feature vector. Feature embedding processing is performed on the static feature vector to generate an embedded static vector. Obtain the temporal feature sequence corresponding to the static feature vector, perform position encoding and dimension mapping on the temporal feature sequence, and generate a temporal feature encoding vector; The embedded static vector and the temporal feature encoding vector are fused together and input into the shared encoder in the classification network to perform self-attention feature extraction, generating a deep feature representation. Multi-task inference operations are performed based on deep feature representation to output the sinus rhythm type.

[0026] In one embodiment, the classification network includes a first classification head and a second classification head; Multi-task inference operations are performed based on deep feature representation to output sinus rhythm type, including: Input the deep feature representation into the first classification head and output the first state probability vector; In response to the first state probability vector being represented as normal sinus rhythm, normal sinus rhythm is output as the sinus rhythm type; In response to the first state probability vector representing sinus arrhythmia, the second classifier is activated, the deep feature representation is input into the second classifier, and the second state probability vector is output. Based on the second state probability vector, the target abnormality type is determined in respiratory sinus arrhythmia and non-respiratory sinus arrhythmia, and the target abnormality type is output as the sinus rhythm type.

[0027] In one embodiment, feature embedding processing is performed on a static feature vector to generate an embedded static vector, including: performing standardization processing on the static feature vector and then inputting it into a fully connected embedding layer to output an embedded static vector of uniform dimension.

[0028] In one embodiment, position encoding and dimension mapping are performed on the temporal feature sequence to generate a temporal feature encoding vector, including: using sine-cosine position encoding to mark the temporal position of the temporal feature sequence containing respiratory signals and target periodic sequences, and after standardization, converting it into a temporal feature encoding vector of uniform dimension through a fully connected layer.

[0029] In one embodiment, the embedded static vector and the temporal feature encoding vector are fused, and the fused feature vector is input into the shared encoder in the classification network to perform self-attention feature extraction and generate a deep feature representation. This includes: concatenating the embedded static vector and the temporal feature encoding vector by dimension to generate a fused feature vector; inputting the fused feature vector into the shared encoder containing a multi-head self-attention layer and a feedforward network layer, and extracting features through attention weight calculation and nonlinear fitting to generate a deep feature representation.

[0030] To achieve the above objectives, this application also proposes a physiological feature monitoring data processing device, including a memory, a processor, and a physiological feature monitoring data processing program stored in the memory and executable on the processor. When the processor executes the physiological feature monitoring data processing program, it implements the physiological feature monitoring data processing method as described above.

[0031] To achieve the above objectives, embodiments of this application also propose a computer-readable storage medium storing a physiological characteristic monitoring data processing program. When the physiological characteristic monitoring data processing program is executed by a processor, it implements the physiological characteristic monitoring data processing method as described above.

[0032] The physiological characteristic monitoring data processing in this application has the following beneficial effects: 1. Achieved high-fidelity and robust contactless extraction of low-level physiological signals; This application acquires radar in-phase orthogonal signals and integrates dynamic environmental background cancellation, target space locking based on constant false alarm rate (CA-CFAR), and temporal phase differential compensation mechanism. It not only completely eliminates the physical constraints of traditional contact devices such as electrocardiograms (ECG) on users, but also adaptively eliminates static clutter and phase entanglement distortion in complex and dynamic sleep environments, thereby accurately restoring the real chest cavity micro-displacement signal.

[0033] 2. Achieved cardiopulmonary decoupling of mixed micro-motion signals, solving the problem of modal aliasing; This application employs the Variable Mode Decomposition (VMD) algorithm, combined with a preset penalty factor and physiological frequency bands (0.1-0.6Hz respiratory band, 0.6-2Hz heartbeat band) to perform signal separation. Compared to traditional Empirical Mode Decomposition (EMD), it not only achieves non-recursive adaptive decoupling of respiratory, heartbeat, and noise components in the frequency domain, completely solving the mode aliasing problem, but also losslessly removes out-of-band clutter while preserving the physical meaning of the waveform.

[0034] 3. A bridge for mapping heart rate variability characteristics to connect with mature medical evaluation systems was constructed; This application employs a dynamic sliding window mechanism to robustly extract heartbeat peaks and maps the extracted continuous heartbeat cycle sequence to the standard medical "RR interval sequence." Based on this, the system further calculates time-domain statistical features such as pNN50, SDNN, and RMSSD, as well as frequency-domain power features such as LF, HF, and LF / HF. This design enables seamless integration of radar wave-based physical displacement data into a mature clinical heart rate variability diagnostic system, endowing non-contact radar signals with profound medical characterization value.

[0035] 4. A quantitative identification method for respiratory / non-respiratory arrhythmias based on the cardiopulmonary coupling mechanism was realized; Traditional electrocardiogram (ECG) monitoring often struggles to distinguish between harmless physiological arrhythmias and harmful pathological arrhythmias. This application, based on the medical principle of "heart rate fluctuating periodically with respiration" in respiratory sinus arrhythmia, extracts the degree of cardiopulmonary synchrony as a quantifiable mathematical feature by aligning the data lengths of respiratory signals and heartbeat cycle sequences, and calculating the normalized cross-correlation coefficient and the percentage of time the target was achieved. This provides a crucial analytical dimension for accurately blocking normal physiological warnings and targeting the identification of pathological arrhythmias.

[0036] 5. Based on the multi-task Transformer architecture, a hierarchical and accurate diagnosis and false alarm prevention and early warning system were implemented to reduce costs and increase efficiency; The classification network in this application adopts an advanced architecture of "static feature embedding + temporal feature encoding + shared encoder + dual-task output head". On the one hand, it perfectly realizes the self-attention deep fusion of multimodal heterogeneous data (discrete static statistics and continuous dynamic waveforms); on the other hand, its condition-triggered dual-classification head design of "first-level judgment of abnormality, second-level judgment of etiology" can "short-circuit" the calculation in time when a normal heart rhythm is diagnosed, which greatly saves the computing power on the end side. At the same time, the system only triggers pathological warnings for the most probable "non-respiratory sinus arrhythmia", completely shielding physiological alarms caused by normal respiratory compensation, perfectly eliminating the user's "electronic medical anxiety", and forming a perfect health monitoring closed loop.

[0037] 6. It enables the intuitive output of complex medical data, breaking down the cognitive barriers of professional medical indicators; This application achieves accurate qualitative differentiation between respiratory and non-respiratory sinus arrhythmias through deep fusion and classification network inference of multidimensional features (static heart rate features and dynamic respiratory-heartbeat correlation features). Compared to traditional heart rate variability (HRV) monitoring devices, which can only output obscure low-level statistical parameters, resulting in the inherent limitation that requires physicians with professional medical theoretical knowledge to interpret them, this solution can directly output highly intuitive final heart rhythm type results at the device. This design allows ordinary users, without any professional medical knowledge, to clearly and accurately understand their own cardiac health status, greatly improving the universality and interactive experience of home health monitoring devices. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0039] Figure 1 This is a module structure diagram of an embodiment of the physiological characteristic monitoring data processing device of the present invention; Figure 2 This is a flowchart illustrating an embodiment of the physiological characteristic monitoring data processing method of the present invention; Figure 3 This is a schematic diagram of the overall architecture of the classification network analysis module in the physiological characteristic monitoring data processing method of the present invention.

[0040] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0041] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0042] It should be noted that when ordinal numbers such as "first" and "second" are mentioned in the embodiments of this application, they are only used to distinguish different objects and do not indicate a specific order or degree of importance, unless the context clearly specifies otherwise. Furthermore, the "connection" or "coupling" described in the embodiments of this application includes not only direct physical connections but also indirect connections or electrical / communication connections via an intermediate medium.

[0043] like Figure 1 As shown, Figure 1 This is a schematic diagram of the structure of the physiological characteristic monitoring data processing device 1 of the hardware operating environment involved in the embodiment of the present invention.

[0044] The physiological characteristic monitoring data processing device 1 (hereinafter referred to as "the device") in this application embodiment may, in physical form, take the form of, but is not limited to, a server (including cloud servers, server clusters, edge computing nodes), a high-performance workstation, a personal computer (PC), a mobile terminal, an IoT gateway, or a dedicated embedded processing device. The device is configured to execute the physiological characteristic monitoring data processing method provided in this application embodiment.

[0045] like Figure 1 As shown, the device may include a memory 11, a processor 12, a communication interface 13, and a system bus 14.

[0046] The memory 11 is used to store computer programs (or instructions) and data required for the operation of the device.

[0047] The memory 11 includes at least one type of readable storage medium. The readable storage medium includes non-volatile memory (NVM), such as solid-state drive (SSD), hard disk drive (HDD), flash memory, optical disk, or other magnetic / optical storage media; the readable storage medium may also include volatile memory, such as random access memory (RAM) or cache.

[0048] More importantly, the memory 11 stores the operating system, the database, and the physiological characteristic monitoring data processing program 10 involved in this application.

[0049] Processor 12 is the core of the device's operation and control center.

[0050] Specifically, processor 12 may be one or more central processing units (CPUs), microprocessors (MCUs), digital signal processors (DSPs), or field-programmable gate arrays (FPGAs). In embodiments involving artificial intelligence, big data processing, or image rendering, processor 12 may also include an artificial intelligence acceleration chip (such as an NPU, TPU) or a graphics processing unit (GPU) for performing parallel vector or tensor operations.

[0051] The processor 12 uses the system bus 14 to read the physiological characteristic monitoring data processing program 10 in the memory 11, and implements each step of the physiological characteristic monitoring data processing method provided in this application embodiment by parsing and executing the program instructions.

[0052] Communication interface 13 (or network interface) is used to enable communication and interaction between the device and other electronic devices (such as clients, third-party servers, and sensor nodes).

[0053] Specifically, the communication interface 13 may optionally include a wired interface (such as an Ethernet interface, fiber optic interface, or USB interface) or a wireless interface (such as a Wi-Fi module, cellular mobile communication module, Bluetooth module, or NFC module). This interface supports various standard communication protocols, including but not limited to TCP / IP, HTTP / HTTPS, UDP, MQTT, and RPC.

[0054] System bus 14 can be a Peripheral Component Interconnect Standard (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This bus is used to transfer instruction and data streams between processor 12, memory 11 and communication interface 13.

[0055] Optionally, the device 1 may also include a user interface (not shown) for human-computer interaction. The user interface may include a display unit (such as an LCD screen, OLED screen, or touch screen) and an input unit (such as a keyboard, mouse, or microphone).

[0056] Those skilled in the art will understand that Figure 1 The structure shown does not constitute a physical limitation on the physiological characteristic monitoring data processing device 1. Depending on the specific application scenario, the device may include fewer or more components than shown, or combine certain components, or use different component arrangements.

[0057] exist Figure 1 In the operating environment shown, processor 12 calls the physiological characteristic monitoring data processing program 10 stored in memory 11 and is configured to perform the following operations: Acquire radar in-phase orthogonal signals, locate the target's thoracic cavity range cell based on the radar in-phase orthogonal signals and extract phase information to generate micro-motion signals; Based on a preset frequency range, the micro-motion signal is separated to extract the respiratory and heartbeat signals; Extract the heartbeat cycle sequence from the heartbeat signal, and generate heart rhythm state features based on the heartbeat cycle sequence; Calculate the temporal correlation between respiratory signals and heart rate cycle sequences to generate respiratory-heart rate correlation features; The heart rhythm characteristics and breathing-heartbeat correlation characteristics are input into a preset classification network to perform feature fusion and inference operations, and output the sinus rhythm type.

[0058] Furthermore, the processor 12 may also be configured to perform refined steps of the physiological feature monitoring data processing method in any of the following embodiments.

[0059] Based on the hardware architecture of the aforementioned physiological characteristic monitoring data processing device, an embodiment of the physiological characteristic monitoring data processing method of the present invention is proposed. The physiological characteristic monitoring data processing method of the present invention aims to achieve highly comfortable, high-precision, and intelligent classification and early warning of sinus rhythm status.

[0060] Reference Figure 2 And Figure 3, Figure 2 Figure 3 is a schematic diagram of the overall architecture of the classification network analysis module in this invention, which is an embodiment of the physiological characteristic monitoring data processing method of this invention.

[0061] As shown in Figure 2, the processing of this physiological characteristic monitoring data includes the following steps: S10. Acquire radar in-phase orthogonal signals, locate the target's thoracic cavity range unit based on the radar in-phase orthogonal signals, extract phase information, and generate micro-motion signals.

[0062] Specifically, the processor receives the raw observation data collected by the radar, performs spatial mapping and phase demodulation processing, and generates continuous and seamless physical displacement data, thereby achieving non-contact, high-fidelity extraction of underlying vital sign fluctuations.

[0063] In one optional embodiment, acquiring radar in-phase orthogonal signals, locating the target's thoracic cavity range cell based on the radar in-phase orthogonal signals, extracting phase information, and generating micro-motion signals can be achieved through the following steps S11-S13: S11. Extract the environmental background reference signal within the preset time window, and perform static clutter filtering on the radar in-phase quadrature signal based on the environmental background reference signal to generate a denoised signal.

[0064] Specifically, step S11 is the background interference suppression process in the radar signal preprocessing stage. In this step, the processor receives the environmental radar signal collected during a specific time period as a reference input, performs clutter filtering, and generates a clean denoised signal that filters out interference from reflections of static objects, thereby achieving adaptive calibration and removal of environmental background noise.

[0065] In some embodiments, step S11 can be implemented by the following steps S111-S112: S111. Extract the mean values ​​of the in-phase and quadrature components of the environmental background reference signal within a set number of frames.

[0066] Specifically, before the user goes to bed, the processor collects a segment of environmental information with N frames (e.g., lasting 1 minute; if the frame rate is set, N is 60 × 20) as an environmental background reference signal. In this process, the background signal of the nth frame can be represented as... The processor performs mean filtering on the aforementioned environmental background reference signal through its internal calculation module, calculating the average amplitude of the background noise using the following formula: .

[0067] in, , , represent the mean values ​​of the in-phase component I and the quadrature component Q of the environmental signal within N frames, respectively.

[0068] By utilizing this mean filtering mechanism, the system accurately extracts the static components that remain constant in space, thereby constructing a reliable background noise model.

[0069] Furthermore, the system monitors the status through mechanisms such as point cloud tracking. If it detects that the target has left the bed or that there has been a significant change in the environment (such as the movement of tables and chairs), the system will re-trigger the aforementioned background signal acquisition and averaging mechanism. This dynamic update mechanism adapts to the dynamic changes in the environment and ensures the long-term stability of the clutter suppression effect.

[0070] S112. Subtract the mean value of the in-phase component from the in-phase component of the radar quadrature signal, and subtract the mean value of the quadrature component from the quadrature component to generate a denoised signal.

[0071] In actual monitoring, when a real-time radar in-phase orthogonal signal containing the vital signs of the target is received, the processor performs signal purification through the mean cancellation mechanism.

[0072] Specifically, after the target is detected entering the bed, the complex form of the radar in-phase orthogonal signal acquired by the processor at time t is: In this stage, the processor uses a mean cancellation mechanism to suppress clutter by subtracting the pre-calculated mean of the in-phase component from the in-phase component of the real-time acquired signal. and subtract the mean of the orthogonal components from the orthogonal components. The specific calculation formula is as follows: .

[0073] in, This represents the generated denoised signal. It is the imaginary unit.

[0074] Through the algebraic subtraction operation in the complex field described above, the processor directly cancels out the reflection interference caused by static objects such as bed frames and wardrobes from the underlying physical signal, thereby generating a clean, denoised signal. .

[0075] Based on this, to prevent macroscopic human movements from masking subtle physiological signals, the processor is configured to perform body movement filtering: if the radar detects body movements such as turning over or raising an arm, and the system determines that the set criteria of at least three range cell jumps and a signal amplitude fluctuation of at least 50% are met, subsequent analysis is paused. The system will wait for the signal to stabilize for 3 seconds, i.e., after meeting the recovery criteria of stable range cells and amplitude fluctuation of less than or equal to 10%, before restarting signal acquisition and analysis. Through this conditionally triggered acquisition blocking mechanism, the system effectively eliminates nonlinear noise introduced by vigorous movement, further ensuring the purity of the signal sent to the subsequent cardiopulmonary separation stage.

[0076] S12. Perform a distance fast Fourier transform on the denoised signal to generate a distance spectrum matrix. Based on the constant false alarm rate detection algorithm, traverse the distance spectrum matrix and lock the distance cell corresponding to the maximum signal energy as the target chest cavity distance cell.

[0077] Specifically, step S12 is the process of locating and extracting the target spatial coordinates. In this step, the processor receives the time-domain denoised signal, performs distance-dimensional spectral transformation and adaptive threshold detection, thereby generating the distance coordinates of the location of the chest cavity, thus achieving accurate locking of effective signal sources containing respiratory and heartbeat characteristics.

[0078] In some embodiments, step S12 can be implemented by the following steps S121 to S123: S121. For the cell to be detected in the distance spectrum matrix, select a preset number of reference cells located on both sides of the cell to be detected and calculate the average power.

[0079] Specifically, the processor first performs a 128-point Fast Fourier Transform (FFT) to denoise the time-domain signal of each received channel. Convert to distance spectrum The formula for its calculation is: .

[0080] This outputs a two-dimensional distance-time matrix, where each row of the matrix corresponds to the time-domain signal of a distance unit.

[0081] Subsequently, the processor uses the Cell-Averaging CFAR (CA-CFAR) algorithm for target selection. In this process, for the target cell (CUT) in the range spectrum matrix, the processor avoids guard cells and selects N reference cells on each side (e.g., N=16), and then uses the formula: ; Calculate the average power of the above reference unit. Based on this local mean calculation theory, the system dynamically assesses the background clutter energy level around the unit under test, laying a data foundation for subsequent adaptive noise stripping.

[0082] S122. Set the detection threshold based on the preset false alarm probability and average power.

[0083] Based on this, the processor uses the acquired average power to set the detection threshold T. The specific calculation is represented by the formula... Among them, the proportionality coefficient ,and This represents the preset false alarm probability configured in the system (e.g., configured as...). ).

[0084] By introducing a preset false alarm probability for exponential and scaling operations, this dynamic threshold setting mechanism effectively overcomes the limitations of a globally fixed threshold under environmental fluctuations, thereby defining an accurate mathematical boundary between minute thoracic physiological micro-motion signals and random environmental noise.

[0085] S123. Extract the distance unit with the highest energy and power greater than the detection threshold as the target thoracic cavity distance unit.

[0086] In response to a dynamically set detection threshold T, the processor performs a traversal filtering of the distance spectrum matrix, retaining those that meet the conditions. The distance unit is determined by physical constraints. Under the condition that the target is in a lying position, the only area of ​​the human body that experiences continuous and periodic mechanical displacement due to breathing and heartbeat is the chest cavity. Therefore, the micro-motion signal energy generated in this area is the most significant. Based on this physiological and physical prior mechanism, the processor further extracts the specific distance unit (bin) with the highest corresponding energy from all retained units that exceed the threshold. In this way, the system automatically eliminates other minor clutter reflection sources in space (such as fluctuations in the reflection cross-sectional area caused by clothing obstruction or changes in sleeping posture during sleep), achieving absolute locking of the true physical location of the human chest cavity.

[0087] S13. Extract the initial phase information of the target thoracic cavity distance unit, eliminate phase entanglement in the initial phase information based on the time-series phase difference rule, and generate micro-motion signal.

[0088] Specifically, step S13 is the demodulation and continuous reconstruction process of physiological micro-motion features. In this step, the processor receives the initial phase information with jump characteristics extracted from the target thoracic cavity distance unit, performs phase difference and periodic compensation actions in time sequence, thereby generating continuous phase features that reflect the real physical displacement, thus achieving high-fidelity extraction of physiological micro-motion signals containing respiratory and heart rhythm.

[0089] In some embodiments, step S13 can be implemented by the following steps: S131. Calculate the difference between the phase of the current frame and the phase of the previous frame.

[0090] Step S131: Calculate the difference between the phase of the current frame and the phase of the previous frame. Specifically, the processor extracts the initial phase information of the located target range unit (bin). Due to the physical characteristics of radar signal transmission and observation line-of-sight, the initial phase often exhibits phase jumps, and this phase entanglement can cause severe distortion of signals reflecting physiological characteristics. To handle this entangled phase, the processor calculates the phase difference between adjacent frames using differential calculation. The specific mathematical formula is as follows: Based on this time-difference mechanism, the system accurately quantifies the instantaneous phase change amplitude of the signal between adjacent sampling points, providing basic judgment parameters for subsequent identification and compensation of distortion jumps.

[0091] S132. In response to the difference being greater than a preset positive threshold, subtract a preset period compensation value from the current frame phase.

[0092] After acquiring the aforementioned phase difference value, the processor dynamically monitors its numerical boundaries. If the difference exceeds a preset positive threshold (e.g., a positive folding critical point), the system determines that the signal has experienced positive entanglement distortion. At this point, the processor is configured to perform a reverse compensation action, subtracting a preset periodic compensation value (e.g., ...) from the current frame phase. (Radians). By utilizing this algebraic subtraction mechanism, the system mathematically smooths out the positive abrupt changes caused by the periodicity of the radar wavelength, allowing the waveform trajectory to return to the true physical displacement trend.

[0093] S133. In response to the difference being less than a preset negative threshold, add a preset period compensation value to the phase of the current frame to generate a micro-motion signal.

[0094] Correspondingly, in response to a difference less than a preset negative threshold (e.g., a negative folding critical point), the system determines that the signal has experienced negative entanglement distortion. The processor then performs a positive compensation operation, adding a preset periodic compensation value (e.g., ...) to the current frame phase. radian).

[0095] Through this bidirectional condition-triggered phase differential compensation mechanism, the system effectively eliminates all... The fuzzy transition ultimately yields a smooth and continuous phase signal. The reconstructed continuous phase directly maps the mechanical displacement of the thoracic cavity caused by human respiration and heartbeat, thus generating the micro-motion signals used for subsequent cardiopulmonary separation and depth feature extraction.

[0096] S20. Perform signal separation on the micro-motion signal based on the preset frequency range, and extract the respiratory signal and heartbeat signal.

[0097] Specifically, step S20 is the decoupling process of physiological mixed signals. In this step, the processor receives continuous micro-motion signals containing superimposed information of respiration and heartbeat, performs frequency domain or time-frequency domain separation and extraction, thereby generating independent respiration and heartbeat signals, thus achieving precise separation of micro-motion features of different physiological frequency bands of the target.

[0098] In some embodiments, step S20 can be implemented by the following steps S21 to S23: S21. Perform variable mode empirical decomposition on the micro-motion signal to generate multiple eigenmode functions with independent center frequencies.

[0099] Specifically, step S21 is a frequency-domain adaptive decoupling process for complex physiological micro-motion signals. In this step, the processor receives micro-motion signals containing multiple physiological rhythms and random interference superimposed on each other, and performs non-recursive variable mode empirical decomposition to generate multiple eigenmode functions that are independent of each other in the frequency domain and have sparse characteristics. This achieves high-fidelity separation of the intertwined physical displacement features in the mathematical dimension.

[0100] In some embodiments, step S21 can be implemented by the following steps: S211. Configure the number of decomposition parameters to three, configure the penalty factor to two thousand, and construct the objective function of variable modal empirical decomposition.

[0101] Specifically, the processor targets the target thoracic cavity micromotion signal after clutter suppression and phase demodulation. Configure the core operating parameters of the Variational Modal Empirical Decomposition (VMD) algorithm.

[0102] In this process, the processor sets the number of decompositions, k, to three. This parameter setting mechanism precisely covers the three independent physical components: respiration, heartbeat, and noise. Simultaneously, the processor applies a penalty factor... By setting it to 2,000, this numerical mechanism system achieves the best mathematical balance between signal fidelity and frequency separation accuracy.

[0103] Furthermore, the processor is configured to synchronously set the convergence tolerance to To control internal iteration efficiency and avoid wasted computing power and convergence risks caused by too many or too few iterations, the processor constructs a variable-modal empirical decomposition objective function based on the above parameters: ; The corresponding constraints are set as follows: ; in, This represents the k-th intrinsic mode function (IMF) component. The center frequency of the k-th component is... For time derivative, For convolution operations, This is a penalty factor.

[0104] Preferably, the core parameters are set as follows: ① The number of decompositions k=3, which just covers the three components of breathing, heartbeat, and noise; ② A penalty factor of 2000 can balance signal fidelity and separation accuracy; ③ The convergence tolerance is set to control the iteration efficiency and avoid too many or too few iterations; ④ Frequency screening range: heart rate 0.6Hz-2Hz, respiratory rate 0.1Hz-0.6Hz.

[0105] S212. Based on the variable mode empirical decomposition objective function, perform iterative decomposition on the micro-motion signal to generate multiple intrinsic mode functions equal to the number of decomposition parameters.

[0106] After constructing the objective function, the processor uses continuous micro-signals. With full input, the objective function is iteratively solved under constrained variational conditions. Leveraging the non-recursive analytical mechanism of variable mode empirical decomposition, the system adaptively decouples highly overlapping physiological mixed signals in the frequency domain, ultimately generating three sparse intrinsic mode functions (IMFs). Each generated IMF converges to its specific center frequency. Furthermore, it possesses a definite finite bandwidth, thereby enabling the high-fidelity decomposition of complex physical displacement signals into multiple independent sub-signals in a mathematical dimension.

[0107] Based on this, the system further executes step S22 to screen specific physiological frequency bands: S22, based on the preset heart rate frequency range and the preset respiratory rate frequency range, screens multiple intrinsic mode functions with independent center frequencies to extract target heart rate mode functions and target respiratory rate mode functions.

[0108] Specifically, step S22 is the targeted identification and extraction process for a specific physiological frequency band. In this step, the processor receives multiple intrinsic mode functions with independent center frequencies generated by variable mode empirical decomposition, performs frequency matching and screening based on preset physiological frequency band boundaries, thereby generating target mode functions corresponding to the actual heartbeat and respiratory rhythm, thus achieving accurate locking of effective vital sign components and automatic removal of out-of-band clutter.

[0109] In some embodiments, step S22 can be implemented by the following steps S221 to S222: S221. Extract the intrinsic mode function with a center frequency in the range of 0.6 Hz to 2 Hz as the target heart rate mode function.

[0110] Specifically, the processor iterates through the multiple intrinsic mode functions (EMFs) generated by the above decomposition, extracting their respective center frequencies. Based on the prior physiological properties of normal human heart rhythm, the processor performs a strict boundary filtering process, selecting specific EMFs whose center frequencies fall entirely within the closed interval of 0.6Hz to 2Hz (0.6 Hz to 2 Hz). Utilizing this high-frequency matching mechanism, the system accurately identifies the weak but high-frequency thoracic displacement component caused by the mechanical pulsation of the heart, thereby generating the target heart rate mode function.

[0111] S222. Extract the intrinsic mode functions with center frequencies in the range of 0.1 Hz to 0.6 Hz as the target respiratory rate mode function.

[0112] Similarly, based on the prior physiological frequency band setting of normal human breathing rhythm, the processor performs low-frequency filtering to select another intrinsic mode function whose center frequency falls within the range of 0.1Hz to 0.6Hz (0.1 Hz to 0.6 Hz).

[0113] Based on this frequency domain isolation mechanism, the system effectively isolates the macroscopic fluctuation displacement component caused by lung inflation, identifying it as the target respiratory rate mode function. Following this, after the two-stage extraction of heartbeat and respiration, the remaining intrinsic mode function, which does not belong to any target physiological frequency band, is automatically identified as a meaningless noise component and directly removed, thus cutting off the interference path of environmental noise for the subsequent classification network.

[0114] S23. Perform signal reconstruction on the target heart rate mode function and the target respiratory rate mode function respectively to generate heartbeat signals and respiratory signals.

[0115] After screening and noise removal of the target mode functions, the processor directly extracts the time-domain waveform data of the target heart rate mode function and the target respiratory rate mode function to perform output reconstruction. Since the intrinsic mode functions obtained by the VMD algorithm are themselves independent continuous waveforms in the time domain with specific physical meaning, the system uses this isomorphic mapping mechanism to directly calibrate and output them as the final target respiratory signal. and heartbeat signals In this way, the system can achieve high-precision, distortion-free extraction and separation of weak and originally deeply intertwined cardiopulmonary physiological fluctuation signals without the constraint of any contact sensors.

[0116] S30. Extract the heartbeat cycle sequence of the heartbeat signal and generate heart rhythm state features based on the heartbeat cycle sequence.

[0117] Specifically, step S30 is the quantification and multidimensional mapping process of cardiac rhythm dynamic features. In this step, the processor receives the pure heartbeat signal after reconstruction by modal empirical decomposition, extracts its periodic interval data in the time domain, and performs statistical and frequency domain transformation operations in the feature space to generate cardiac rhythm state features that reflect the law of heart rate variability. This realizes the conversion of continuous one-dimensional physiological signals into multidimensional medical feature vectors that can be directly read by the classification network.

[0118] In some embodiments, step S30 can be implemented by the following steps S31 to S33: S31. Perform peak detection on the heartbeat signal based on the statistical characteristics of a preset sliding window, and extract the peak timestamp sequence.

[0119] Specifically, step S31 is the time-domain adaptive anchoring process of the cardiac mechanical pulsation point. In this step, the processor receives the aforementioned heartbeat signal and performs peak screening based on statistical indicators within a local time window, thereby generating a peak timestamp sequence representing the physical moment of each effective heartbeat. This enables the robust locking of the true physiological contraction peak even in radar signals with slight baseline drift or residual noise.

[0120] In some embodiments, step S31 can be implemented by the following steps S311 to S312: S311. Calculate the mean and standard deviation of the heartbeat signal within a preset sliding window.

[0121] Specifically, the processor slides a pre-defined time window (e.g., a local time window with a fixed number of frames or seconds) along the time axis to capture local heartbeat signal segments. Within this specific sliding window, the processor, through its internal statistical module, performs arithmetic mean and variance calculations on the discrete signal points within the segment, obtaining the waveform mean and standard deviation representing that local interval. Utilizing this local characteristic statistical mechanism, the system dynamically quantifies the local energy baseline and amplitude fluctuation severity of the current signal segment, thereby completely eliminating the easily invalidated global fixed threshold and providing a mathematical reference that is updated in real time for the generation of subsequent dynamic detection benchmarks.

[0122] S312. Construct a dynamic detection threshold based on the waveform mean and waveform standard deviation, extract the timestamps corresponding to the peaks whose amplitudes are greater than the dynamic detection threshold, and generate a peak timestamp sequence.

[0123] After acquiring the aforementioned local statistical parameters, the processor uses the waveform mean as the baseline anchor point and combines the waveform standard deviation with pre-configured sensitivity weighting coefficients to perform algebraic weighted calculations, thereby dynamically constructing a dynamic detection threshold specific to the current sliding window. Subsequently, the processor performs extreme value comparison within the current time window, filtering out local maxima (i.e., peaks) where the signal amplitude is strictly greater than the dynamic detection threshold, and accurately extracting the timestamp record point corresponding to the peak. Relying on this dynamic threshold anti-interference mechanism that adaptively adjusts with waveform fluctuations, the system effectively overcomes the defects of missed or false detections easily caused by traditional globally fixed thresholds, accurately eliminates spurious peak interference caused by residual spur noise, captures the physical peak coordinates representing each real heart contraction, and then arranges all verified time nodes sequentially to generate a peak timestamp sequence for constructing physiological characteristic cycles.

[0124] S32. Calculate the time difference between adjacent peak timestamps to generate a heartbeat cycle sequence.

[0125] Specifically, step S32 is the discretization and quantization process of the instantaneous heart rate cycle. In this step, the processor receives the extracted continuous peak timestamp sequence and performs adjacent difference subtraction in the time domain to generate a heartbeat cycle sequence that reflects the duration of each continuous heartbeat interval. This converts the absolute coordinate points in the time dimension into low-level relative interval data that can be used for in-depth analysis of heart rate variability.

[0126] In one embodiment, the processor accurately records the exact occurrence time of each effective peak parsed from the heartbeat signal based on a pre-set overall detection window (e.g., set to twenty seconds to ensure that the window contains about twenty complete heartbeat cycles) and a local sliding window (e.g., set to five seconds to contain about five heartbeat cycles while taking into account the real-time responsiveness of the system and the stability of feature extraction), thus extracting a series of consecutive peak timestamps. , where n is the total number of detected peaks.

[0127] After obtaining the aforementioned timestamp coordinates, the processor performs a timing difference operation on adjacent peak timestamps on the timeline, calculating the time difference between the occurrence of the next peak and the occurrence of the previous peak. The underlying mechanism of this operation can be understood through the formula... To provide accurate mathematical expressions, among which Characterizing the first The time interval between consecutive heartbeats.

[0128] It should be further noted that in standard medical electrocardiogram (ECG) analysis, the heartbeat interval is usually defined by the time difference between adjacent ventricular depolarization waves (R waves) (i.e., the RR interval sequence). In this embodiment, the processor maps the heartbeat cycle sequence acquired based on radar micro-motion signals to the traditional medical RR interval sequence in terms of physiological characteristics. Relying on the electromechanical coupling mechanism of strict temporal synchronization between the electrophysiological activity inside the heart and the mechanical pumping displacement of the external thoracic cavity, the system accurately reconstructs the firing cycle of cardiac electrical signals by capturing the time difference of physical and mechanical pulsations on the human body surface. By establishing this cross-modal physical and physiological equivalence, the system can directly and seamlessly integrate the generated heartbeat cycle sequence into the mature heart rate variability (HRV) medical evaluation system in a completely contactless monitoring scenario, thus laying a solid medical foundation for the subsequent extraction of highly clinically valuable cardiac rhythm characteristics.

[0129] S33. Calculate time-domain statistical features and / or frequency-domain power features based on the heartbeat cycle sequence, and generate heart rhythm state features based on the time-domain statistical features and / or frequency-domain power features.

[0130] Specifically, step S33 is the multidimensional feature engineering extraction process for the heartbeat interval data. In this step, the processor receives the aforementioned heartbeat cycle sequence (equivalent to the RR interval sequence in traditional medicine), performs statistical calculations in the time dimension and power analysis in the frequency dimension, thereby generating a comprehensive index reflecting the regulatory state of the target autonomic nervous system. This transforms the original cycle sequence into primary medical diagnostic features that can be directly read by deep learning networks.

[0131] Specifically, the time-domain statistical characteristics include at least one of the following: the proportion of adjacent period differences exceeding the limit, the standard deviation of the periodic series, and the root mean square of adjacent period differences.

[0132] Specifically, the underlying mechanism for extracting the aforementioned specific time-domain statistical features lies in quantifying and distinguishing the discrete fluctuation patterns of normal heart rate and sinus arrhythmia on the time axis. Under physiologically normal resting heart rate, the human heartbeat exhibits extremely high stability, with minimal fluctuations in the time interval between two adjacent heartbeat cycles (equivalent to the RR interval on an electrocardiogram). Specifically, the percentage of consecutive cycle differences exceeding 50 milliseconds (pNN50) is low and stable, the standard deviation (SDNN) of all normal sinus intervals inevitably falls within a stable range of 100 to 200 milliseconds, and the root mean square (RMSSD) of the difference between adjacent heartbeat intervals is also highly concentrated in the 20 to 50 millisecond range. Conversely, when sinus arrhythmia occurs, the time-domain interval fluctuations become abnormally significant, directly leading to a surge in the proportion of adjacent cycle differences exceeding the limit or abnormal fluctuations. Furthermore, the calculated standard deviation of the cycle sequence and the root mean square of the difference between adjacent cycles will deviate significantly from the aforementioned normal baseline range.

[0133] In some embodiments, calculating time-domain statistical features based on the heartbeat cycle sequence includes: S331. Calculate the percentage of adjacent heartbeat cycle differences whose absolute value is greater than the preset time difference threshold in the total number of differences in the heartbeat cycle sequence, and use this as the percentage of adjacent cycle differences exceeding the limit.

[0134] Specifically, the processor processes the obtained heartbeat cycle sequence (in (Indicates the number of periods), calculate the absolute value of the difference between each group of consecutive periods. (in From 1 to Subsequently, the processor sets the preset time difference threshold to fifty milliseconds (50ms) and performs a counting action to count the number of absolute values ​​of the above differences that are greater than this threshold, denoted as . .

[0135] Furthermore, the processor uses the formula: Calculate its percentage in the total number of differences to generate the percentage of adjacent period differences exceeding the limit (pNN50).

[0136] Based on this statistical mechanism of exceeding the limit, the system accurately quantifies the short-term drastic fluctuations between adjacent heartbeat cycles. Under normal sinus rhythm, this value is usually low and stable; however, when sinus arrhythmia occurs, due to significant cycle fluctuations, the proportion of this exceeding the limit will increase significantly or exhibit abnormal fluctuations.

[0137] S332. Calculate the standard deviation of the heartbeat cycle sequence as the standard deviation of the cycle sequence.

[0138] Specifically, the processor uses statistical analysis of variance to evaluate the global dispersion of the heartbeat cycle sequence. The specific mathematical formula is as follows: ; in , which is the average value across all heartbeat cycles.

[0139] Using this global standard deviation calculation mechanism, the system accurately extracts the periodic sequence standard deviation (SDNN), which reflects the overall variability of heart rate. Under normal physiological conditions, this value is usually distributed in the range of 100 to 200 milliseconds. If the value deviates from this range, it indicates a potential abnormality in heart rhythm.

[0140] S333. Calculate the root mean square of the difference between adjacent cycles in the heartbeat cycle sequence, and use it as the root mean square of the difference between adjacent cycles.

[0141] Specifically, the processor further performs root mean square (RMS) calculations on the short-term heart rate variability components. The specific calculation formula is as follows: ; Based on the processing mechanism of summing the squares of adjacent differences and then taking the square root, the system amplifies the instantaneous change characteristics between heartbeat cycles and extracts the root mean square difference (RMSSD) of adjacent cycles. Under normal physiological conditions, this feature value mostly falls within the range of 20 to 50 milliseconds, thus providing another core time-domain quantitative reference for the system to determine arrhythmias.

[0142] Specifically, the frequency domain power characteristics include at least one of the following: low-frequency band power, high-frequency band power, and the ratio of low-frequency to high-frequency power.

[0143] Specifically, this frequency domain power calculation process aims to capture the redistribution of cardiac physiological energy caused by abnormal autonomic nervous system regulation. Under normal heart rate conditions, the calculated heartbeat energy is highly concentrated at the fundamental frequency, and the ratio of the low-frequency component (LF, range 0.04-0.15Hz) to the high-frequency component (HF, range 0.15-0.4Hz), reflecting the balance between sympathetic and parasympathetic nerve activity, is strictly constrained within the normal range of 1.5 to 2.0. However, the onset of sinus arrhythmia disrupts this steady-state balance, causing the frequency domain energy distribution to shift abnormally, specifically manifested as an abnormal low-frequency to high-frequency power ratio (e.g., a calculated result less than 1.5 or greater than 3). Based on this frequency domain transformation and power ratio solution mechanism, the system extracts another layer of key pathological characteristics completely independent of the time dimension, thoroughly covering the hidden variation characteristics of arrhythmias in the frequency domain dimension.

[0144] In some embodiments, calculating the time-frequency power characteristics based on the heartbeat cycle sequence includes the following steps S334 to S336: S334. Calculate the integral value of the power spectral density of the heartbeat cycle sequence in the preset low-frequency range, and use it as the low-frequency power.

[0145] Specifically, the processor first calculates the power spectral density (PSD) corresponding to the time-domain sequence using discrete Fourier transform. Then, the processor sets a preset low-frequency range of 0.04 Hz to 0.15 Hz and performs an integral operation on the power spectral density within this frequency band, mathematically expressed as: ; By utilizing the energy integration mechanism of this frequency band, the system accurately quantifies the low-frequency energy components that are mainly regulated by the sympathetic and parasympathetic nervous systems, generating low-frequency power (LF).

[0146] S335. Calculate the integral value of the power spectral density in the preset high-frequency range, and use it as the power in the high-frequency band.

[0147] Similarly, the processor sets the preset high-frequency range to 0.15 Hz to 0.4 Hz, and integrates the power spectral density within this range. The specific calculation formula is as follows: ; By targeting and extracting the isolation mechanism of this high-frequency energy, the high-frequency power (HF) that mainly reflects the changes in parasympathetic (vagus nerve) tension was quantified.

[0148] S336. Calculate the ratio of low-frequency band power to high-frequency band power, and use it as the low-frequency to high-frequency power ratio.

[0149] After obtaining the power values ​​for the two frequency bands mentioned above, the processor performs an arithmetic division operation to calculate the quotient of the two values, using the following formula: ; Based on this power ratio calculation principle, the system intuitively and effectively reflects the balance between the sympathetic and vagus nerves in the human body. Under normal heart rate conditions, the ratio of this low-frequency to high-frequency component is concentrated between 1.5 and 2.0; if the system detects an abnormal ratio (such as less than 1.5 or greater than 3), it will be used as an important frequency domain criterion for identifying sinus arrhythmia.

[0150] In some embodiments, calculating time-domain statistical features and / or frequency-domain power features based on the heartbeat cycle sequence, and generating heart rhythm state features based on the time-domain statistical features and / or frequency-domain power features includes: calculating time-domain statistical features and frequency-domain power features based on the heartbeat cycle sequence, and concatenating the extracted time-domain statistical features and frequency-domain power features by vector dimension to generate heart rhythm state features.

[0151] Specifically, this step is a transformation process from underlying medical and physiological indicators to a data array usable for machine learning. After completing the calculations of the aforementioned time-domain and frequency-domain modules, the processor extracts three time-domain indicators: the proportion of adjacent cycle differences exceeding the limit, the standard deviation of the cycle sequence, and the root mean square of adjacent cycle differences; and three frequency-domain indicators: low-frequency band power, high-frequency band power, and the ratio of low-frequency to high-frequency power. At the data structure level, a one-dimensional vector concatenation operation is performed. Leveraging this multi-modal feature fusion mechanism, the system reconstructs six independent static heart rate features extracted from the perspectives of time-discrete fluctuations and frequency band energy distribution into a high-dimensional combined feature vector containing comprehensive physiological representation information. This vector constitutes the final output heart rate state feature. Through this cross-domain fusion mechanism, the system reconstructs a unified high-dimensional static feature vector from six single-dimensional indicators representing time-discrete fluctuations and frequency energy distribution, thereby providing a decision-making foundation with maximized information and mutual corroboration for the classification and inference network.

[0152] Alternatively, in other embodiments, calculating time-domain statistical features and / or frequency-domain power features based on the heartbeat cycle sequence, and generating heart rhythm state features based on the time-domain statistical features and / or frequency-domain power features includes: in response to extracting only time-domain statistical features or frequency-domain power features, vectorizing the extracted features (time-domain statistical features or frequency-domain power features) to generate heart rhythm state features.

[0153] Specifically, this feature generation step is a process of encapsulating underlying medical indicators into usable input data for the classification network. Based on prior pathological knowledge, sinus arrhythmia physiologically manifests as abnormalities in the time domain, frequency domain, or both. Therefore, the system is endowed with a highly flexible feature processing architecture. In some implementation scenarios where computing resources are limited or specific signals are disturbed, the system can be configured in a single-dimensional detection mode. In this case, the processor only calculates and extracts time-domain statistical features (including the aforementioned pNN50, SDNN, and RMSSD) or only calculates and extracts frequency-domain power features (including the aforementioned LF, HF, and LF / HF ratio), and directly performs vectorization transformation on the obtained single-dimensional feature set to independently generate cardiac rhythm state features. With this single-track execution mechanism, the system utilizes the physiological characteristic that "abnormality in any one aspect is sufficient to characterize" sinus arrhythmia, and can still keenly capture single-dimensional variation information to maintain cardiac rhythm screening even in a low-power state.

[0154] S40. Calculate the temporal correlation between respiratory signals and heart rate cycle sequences to generate respiratory-heart rate correlation features.

[0155] Specifically, step S40 is the quantification and feature extraction process of cardiopulmonary coupling (the physiological mechanism of respiratory sinus arrhythmia). In this step, the processor receives continuous respiratory signals separated in the frequency domain and discrete heartbeat cycle sequences, performs time-dimensional length alignment and normalized cross-correlation analysis, thereby generating correlation features reflecting the degree of synchronization and coordination of the cardiopulmonary system. The underlying physiological mechanism for extracting this respiratory-heartbeat correlation feature is that respiratory sinus arrhythmia has typical cardiopulmonary coupling characteristics, that is, the heart rate changes periodically with respiration (the heart rate increases during inspiration, leading to a shorter cycle, and the heart rate decreases during expiration, leading to a longer cycle).

[0156] Therefore, during respiratory sinus arrhythmia, there is a fixed "one rises, the other falls" phase relationship between the respiratory signal waveform and the heart rate cycle sequence waveform, exhibiting a very strong cross-correlation. However, during non-respiratory sinus arrhythmia, heart rate fluctuations do not change with respiration, and the two lack a fixed period, exhibiting a weak or zero cross-correlation. By calculating the temporal correlation between these two factors, the system can directly reveal the underlying driving force of arrhythmias, providing a core classification dimension for subsequent abnormality determination.

[0157] In some embodiments, step S40 can be achieved by the following steps S41 to S45: S41. Perform time length alignment processing on the heartbeat cycle sequence to align the data length of the heartbeat cycle sequence with that of the respiratory signal in the time dimension, and use the aligned heartbeat cycle sequence as the target cycle sequence.

[0158] Specifically, since the separated respiratory signals are continuous-time signals with a fixed sampling rate, while the heart rate cycle sequence (RR interval) is a sequence generated based on discrete heart rate event points, the two are different in terms of data length and temporal resolution. In order to perform pointwise correlation calculations on the same time scale while preserving the gradual trend of dynamic changes in heart rate to adapt to the periodic characteristics of respiratory abnormalities, the processor performs a time-domain length interpolation alignment operation.

[0159] Specifically, step S41 includes: using linear interpolation to fill in insufficient data between adjacent data points in the heartbeat cycle sequence, generating a target cycle sequence with the same data length as the respiratory signal. The specific calculation formula for this linear interpolation mechanism is expressed as follows: ; in and This is the timestamp of adjacent valid heartbeat peaks in the original heartbeat signal.

[0160] Through this interpolation and reconstruction mechanism, the discrete periodic sequence is smoothly broadened, generating a target periodic sequence with a length completely consistent with the respiratory signal. .

[0161] Furthermore, in other implementation scenarios, to adapt to the nonlinear gradual change characteristics of different physiological signals or to accommodate the computational resource limitations of the underlying hardware, the processor can also be configured to perform the time-series length alignment processing using cubic spline interpolation, zero-order hold interpolation, or a multiphase resampling algorithm. Based on these equivalent data alignment mechanisms, the system can seamlessly fill the physical gaps in the time axis of discrete periodic sequences, thus laying a unified data structure foundation for the subsequent algebraic multiplication of two physiological signals at the same frequency.

[0162] S42. Perform standardization processing on the respiratory signal and the target periodic sequence respectively to generate standard respiratory signal and standard periodic sequence that conform to the preset distribution state.

[0163] Specifically, because there is a significant dimensional difference between the displacement amplitude of thoracic mechanical respiration and the time amplitude of the cardiac interval, directly calculating the correlation would result in severe interference from the absolute amplitude. Therefore, the processor performs dimensionless normalization scaling on the two aligned signals. Specifically, step S42 includes the following steps S421 to S422: S421. Calculate the sequence mean and standard deviation of the respiratory signal and the target cycle sequence, respectively.

[0164] The processor extracts the respiratory signals within the current analysis window through statistical calculations. sequence mean with sequence standard deviation Simultaneously calculate the target periodic sequence sequence mean with sequence standard deviation .

[0165] S422. Subtract the mean from the corresponding signal sequence and divide by the standard deviation of the sequence to convert them into standard normal distribution signals with a mean of zero and a standard deviation of one, which are used as standard respiratory signals and standard periodic sequences.

[0166] After obtaining the above statistical parameters, the processor performs Z-score normalized algebra operations. The specific implementation mechanism is as follows: Through the formula: ; Generate a standard respiratory signal.

[0167] And through the formula: ; Generate a standard periodic sequence.

[0168] Based on this standardized conversion mechanism, the system completely eliminates the difference in absolute values ​​between breathing depth and heart rate, forcing subsequent correlation calculations to focus only on the phase and waveform similarity of the two in their periodic fluctuations.

[0169] S43. Calculate the cross-correlation function between the standard respiratory signal and the standard periodic sequence, and extract the peak value of the cross-correlation coefficient.

[0170] Specifically, the processor performs a sliding inner product operation on the two standardized signals in the time domain.

[0171] First, the system calculates the time lag between the two signals as follows: The linear cross-correlation function at time t is given by the formula: ; The range of this function reflects the hysteresis of respiratory signals. The degree of overlap between the signal and the heartbeat cycle at each time point.

[0172] Subsequently, the system normalizes the results of the cross-correlation function, takes the absolute value of the normalized cross-correlation sequence, and extracts the maximum value as the peak value of the cross-correlation coefficient. During this process, to facilitate the establishment of a global evaluation criterion, the processor uses the formula: ; The cross-correlation results are strictly mapped to the interval [-1, 1].

[0173] Because typical respiratory arrhythmias exhibit a negative correlation (coefficient close to -1) between inspiration (increased waveform) and heart rate (decreased RR interval), the system further... The absolute value of the sequence is taken to eliminate interference from polarity direction, and the maximum value in the absolute value sequence is extracted as the peak value of the cross-correlation coefficient. This absolute peak mechanism precisely pinpoints the strongest instantaneous coupling strength between the cardiopulmonary system that is unaffected by phase reversal.

[0174] Furthermore, in complex monitoring scenarios with severe local burst noise or the need to handle periodic boundary effects, the processor can be configured to perform the aforementioned correlation calculations using a cyclic cross-correlation function or a generalized cross-correlation function with phase transformation weights. By extending these equivalent cross-correlation operation mechanisms, the system effectively broadens the applicability of the algorithm in complex low-noise environments, ensuring the robustness of time delay estimation.

[0175] S44. Calculate the proportion of the target duration that is greater than the preset correlation threshold in the result of the statistical cross-correlation function, and generate the cross-correlation compliance duration percentage.

[0176] In addition to assessing the instantaneous maximum coupling strength, the processor further evaluates the persistence and stability of cardiopulmonary coordination. The processor is configured to set a preset correlation threshold (e.g., 0.6, an empirical value that effectively distinguishes between strong respiratory characteristics) and, along the time axis, calculate the total duration of achievement exceeding this threshold in the aforementioned cross-correlation results (denoted as the target duration). The processor then calculates using the formula: ; Calculate the target duration within the overall analysis of the total monitoring duration. The percentage it accounts for.

[0177] Based on this proportion assessment mechanism, the system effectively filters out localized spurious high correlations caused by accidental bodily movements or brief sighs, and generates a proportion of cross-correlation duration that reflects long-term coupling stability. .

[0178] S45. Combine the peak value of the cross-correlation number with the proportion of cross-correlation achievement time to generate respiratory-heartbeat correlation features.

[0179] In this step, the processor performs a vector-level fusion combination of the peak cross-correlation coefficient, which represents the extreme values ​​of coupling strength, and the proportion of cross-correlation duration that represents the stability of coupling maintenance. Through this dimensional splicing action, the system constructs a comprehensive feature vector that includes both instantaneous deep collaborative information and long-term macroscopic collaborative statistical information, thereby generating the final respiratory-heartbeat correlation features. This provides an impeccable data-driven basis for the subsequent classification network to perform secondary discrimination (distinguishing between respiratory and non-respiratory sinus arrhythmias).

[0180] S50. Input the heart rhythm state features and breathing-heartbeat correlation features into the preset classification network to perform feature fusion and inference operations, and output the sinus rhythm type.

[0181] Specifically, step S50 is the process of deep fusion of multidimensional physiological features and output of medical-grade diagnostic results. In this step, the processor receives the static discrete features and continuous time-series sequences extracted by the front-end module, and performs feature mapping, self-attention dimension interleaving, and classification discrimination based on a multi-task Transformer architecture to generate a final diagnostic label reflecting the target cardiac electrophysiological mechanism. This enables high-precision and automated hierarchical inference of three physiological states: normal sinus rhythm, respiratory sinus arrhythmia, and non-respiratory sinus arrhythmia.

[0182] In some embodiments, step S50 can be achieved by the following steps S51 to S54: S51. Combine the heart rhythm state features with the breathing-heartbeat correlation features into a static feature vector, perform feature embedding processing on the static feature vector, and generate an embedded static vector.

[0183] Specifically, step S51 is the structured encapsulation and dimensional projection process of heterogeneous medical features. In this step, the processor extracts the six static heartbeat features generated in the previous steps (i.e., the proportion of adjacent cycle difference exceeding the limit pNN50, the standard deviation of the cycle sequence SDNN, the root mean square of the adjacent cycle difference RMSSD, the low-frequency power LF, the high-frequency power HF, and the ratio of low-frequency to high-frequency power LF / HF) and two respiratory-heartbeat synchronization features (i.e., the peak value of the cross-correlation coefficient). Percentage of time spent meeting relevant standards At the data structure level, these eight discrete values ​​with different physical dimensions are concatenated and combined to form an 8-dimensional hybrid static feature vector. Subsequently, the processor performs high-dimensional latent space feature embedding processing on this static feature vector, thereby generating an embedded static vector with rich underlying logical representation capabilities. This solves the architectural bottleneck that physiological features of different properties cannot be directly and equally treated by deep learning networks.

[0184] In one embodiment, performing feature embedding processing on a static feature vector to generate an embedded static vector includes: performing standardization processing on the static feature vector and then inputting it into a fully connected embedding layer to output an embedded static vector of uniform dimension.

[0185] Specifically, in order to eliminate the huge absolute numerical differences and dimensional influences among various physiological indicators in the above 8-dimensional static features (for example, the LF / HF ratio is in the single digits, while in SDNN it may reach the hundreds), the processor first calculates using the formula: ; Z-score standardization is performed on the static feature vectors. Specifically, and These represent the mean and standard deviation of each feature to be standardized in the baseline data distribution, respectively.

[0186] Based on this standardization transformation mechanism, the original physical data is strictly stretched and mapped to a standard distribution space with zero mean and unit variance, effectively preventing the network from tilting towards large numerical features during backpropagation. Then, the processor inputs the standardized 8-dimensional static feature vector into the pre-set fully connected embedding layer of the classification network (e.g., a linear layer matrix with configuration parameters of Linear(8,64)).

[0187] Based on this linear projection and dimensional expansion mechanism, the originally low-dimensional and fragmented statistical indicators are cross-crossed with cross-dimensional features, ultimately outputting a statically embedded vector with a unified dimension (e.g., 64 dimensions). Through this embedding transformation, the system completely removes data format barriers for subsequent equal-dimensional concatenation and deep feature fusion of static vectors and temporal sequence vectors in a shared self-attention encoder.

[0188] S52. Obtain the temporal feature sequence corresponding to the static feature vector, perform position encoding and dimension mapping on the temporal feature sequence, and generate a temporal feature encoding vector.

[0189] Specifically, step S52 is the preprocessing and feature projection process of the temporal physiological waveform entering the multi-task Transformer network. In this step, the processor extracts continuous waveform data corresponding to the time period of the static feature extraction mentioned above. By introducing absolute time position information and dimensionality-up transformation of the feature space, a high-dimensional feature representation carrying sequential temporal logic and highly aligned with the static feature dimension is generated. This solves the inherent architectural defect of the self-attention mechanism itself, which lacks the ability to perceive the sequence position, thereby achieving accurate analysis of the dynamic process of cardiopulmonary coupling.

[0190] In one embodiment, performing position encoding and dimension mapping on the temporal feature sequence to generate a temporal feature encoding vector includes: using sine-cosine position encoding to mark the temporal position of the temporal feature sequence containing respiratory signals and target periodic sequences, and then converting it into a temporal feature encoding vector of uniform dimension through a fully connected layer after standardization.

[0191] Specifically, the processor first acquires the separated continuous breathing signals. Target periodic sequence after interpolation alignment The two are combined to form the temporal feature sequence. To fully represent the periodic gradual fluctuation of "inhalation-exhalation" and ensure that the input data segment contains 2 to 3 complete respiratory cycles, the processor uses a data window containing 200 frames (e.g., corresponding to a duration of 10 seconds) to extract the signal. At the same time, the processor translates this window in 2-second sliding steps according to the time sequence. Relying on this time window overlapping and advancement mechanism, the system effectively avoids the loss of key physiological gradual information caused by boundary sequence cutting.

[0192] For the truncated time sequence, the processor uses an internal algorithm module to introduce a sine-cosine position coding mechanism to perform time position marking. The underlying coding mechanism of this action is implemented through the following mathematical formula: ; ; in, This represents the absolute position index of the time step in the sequence. The dimension index representing the encoded vector (i.e., half the number of features), while This represents the hidden layer dimension of the model (corresponding to the length of the input sequence, set to 200 in this embodiment). Leveraging this alternating trigonometric function encoding mechanism, the system injects a unique and smooth, continuous relative position label into the data at each discrete time point, enabling the subsequent attention calculation module to accurately perceive the sequential coupling pattern between respiratory fluctuations and heart rate intervals on the time axis.

[0193] After the location markers are fused, the processor performs Z-score normalization on each of the temporal feature sequences. This sequence-by-sequence normalization mechanism effectively converges outlier values ​​while perfectly preserving the original dynamic fluctuation trend within the cardiopulmonary signal, thus generating a stable input sequence. Subsequently, the processor inputs the normalized sequence into a pre-configured fully connected layer of the classification network (e.g., a linear layer matrix with parameters configured as Linear(200,64)) to perform dimensionality mapping. Through the linear projection mechanism of this fully connected layer, the original 200-dimensional temporal sequence is transformed into a 64-dimensional dual-sequence feature representation. This mapping transformation not only completes the deep abstraction of temporal features into the latent space but also ensures that the final output temporal feature encoding vector is absolutely unified with the 64-dimensional embedded static vector generated in the previous steps in terms of feature dimension, thereby clearing all dimensional barriers for the subsequent deep splicing and fusion of static structured indicators and dynamic temporal waveforms.

[0194] S53. The embedded static vector and the temporal feature encoding vector are fused together and input into the shared encoder in the classification network to perform self-attention feature extraction, generating a deep feature representation.

[0195] Specifically, step S53 is a deep interaction and high-dimensional mapping process for multimodal heterogeneous data. In this step, the processor receives the aforementioned independent static structured indicators and dynamic temporal encoded waveforms, performs dimensional concatenation of the feature space and self-attention mechanism operations based on the Transformer architecture, thereby generating a deep feature representation containing global physiological correlation information. This breaks down the physical isolation between time domain and frequency domain, and between static and dynamic features, providing a unified feature base rich in deep semantics for subsequent multi-level classification inference.

[0196] In one embodiment, the embedded static vector and the temporal feature encoding vector are fused, and the result is input into the shared encoder in the classification network to perform self-attention feature extraction, generating a deep feature representation, including the following steps S531 to S532: S531. Concatenate the embedded static vector and the temporal feature encoding vector according to their dimensions to generate a fused feature vector.

[0197] Specifically, the processor performs a direct concatenation operation on the feature dimensions of the data structure between the uniformly dimensional embedded static vector (e.g., a 64-dimensional static feature vector) and the temporal feature encoding vector (e.g., a 64-dimensional temporal feature encoding vector). Through this direct dimensional concatenation mechanism, the system generates a fused feature vector with horizontally expanded dimensions (e.g., concatenated and merged into 128 dimensions). This computational operation, without destroying their independent physical representation information, forcibly places the static statistical regularity representing heart rhythm and the dynamic change regularity representing cardiopulmonary coupling into the same high-dimensional vector space, which is then used as the standardized input data for the subsequent shared encoder.

[0198] S532. Input the fused feature vector into the shared encoder containing the multi-head self-attention layer and the feedforward network layer, extract features through attention weight calculation and nonlinear fitting, and generate deep feature representation.

[0199] The processor feeds the aforementioned 128-dimensional fused feature vector into a shared encoder designed with multiple layers (e.g., configured with 4 layers) for deep parsing. Within each layer, the processor first uses a multi-head self-attention layer to perform the capture of dynamic correlations between features.

[0200] Specifically, the processor generates the query matrix, key matrix, and value matrix through linear projection, and then calls the self-attention weight calculation formula: ; To calculate and assign attention weights to each feature, where Indicates matrix transpose. This represents the dimension of the hidden layer of the matrix.

[0201] During the calculation process, the processor uses the Softmax normalization function to process the results, and its calculation formula is as follows: ; in , They represent vectors respectively The first in , Each element. The processor uses this normalization function to transform the raw relevance score into a standard probability distribution output. After attention weighting, the processor passes the output of the attention layer to the feedforward network layer (e.g., configuring the hidden layer dimension to 256 and using the GELU activation function) to perform nonlinear fitting and transformation.

[0202] Relying on the alternating deep processing mechanism of multi-head self-attention and feedforward network, the system perfectly captures the long-range dependencies within the cardiopulmonary waveform and the deep nonlinear correlation between static statistics and dynamic waveforms, ultimately generating a highly abstract deep feature representation at the end of the shared encoder.

[0203] S54. Perform multi-task inference operations based on deep feature representation and output the sinus rhythm type.

[0204] Specifically, step S54 is the hierarchical output and risk warning process of medical diagnostic conclusions. In this step, the processor receives the highly abstract deep feature representation mentioned above, and uses a classifier with mutually independent parameters to perform a category mapping inference action with a logical sequence, thereby generating a specific diagnostic label for the target cardiac electrophysiological state, thus achieving high-precision and automated hierarchical screening of normal heart rhythm and arrhythmias of different pathological levels.

[0205] In some embodiments, step S54 can be implemented by the following steps S541 to S544: S541. Input the deep feature representation into the first classification head and output the first state probability vector.

[0206] Specifically, the processor imports the aforementioned 128-dimensional deep feature representation output by the shared encoder into a first classification head with independently independent parameters, specifically adapted for the first-level classification task. Inside the first classification head, the processor performs probability inference through linear mapping of fully connected layers and classification activation functions, thereby outputting a first state probability vector containing two independent classification probability components (for example, this vector is represented as [...]). , Based on this dimensionality reduction mapping and probabilistic computing theory, the system completed a preliminary probabilistic quantification of whether an arrhythmia event has occurred at the target.

[0207] S542. Responding to the first state probability vector, which is represented as normal sinus rhythm, output normal sinus rhythm as the sinus rhythm type.

[0208] The processor performs a comparison and judgment action on the numerical distribution of the first state probability vector. In response to the dominance of the probability component representing "normal sinus rhythm" in the vector (e.g., its value is the largest), the system determines that the current target heart physiological state is healthy, and the processor then cuts off the signal transmission to subsequent network layers, directly outputting the normal sinus rhythm as the sinus rhythm type finally determined by the classification network.

[0209] By utilizing this short-circuit decision mechanism, the system effectively blocks redundant calculations when it confirms that the target's physiological characteristics are normal, thus saving the overall system's computing power and energy consumption.

[0210] S543. In response to the first state probability vector representing sinus arrhythmia, activate the second classification head, input the deep feature representation into the second classification head, and output the second state probability vector.

[0211] Correspondingly, in response to the first state probability vector being determined as "sinus arrhythmia" (i.e., the abnormal probability is dominant), the processor executes a conditional trigger action to activate the second classification head, which has independent parameters and is specifically adapted to the secondary classification task.

[0212] In the active state, the processor again directly inputs the same 128-dimensional deep feature representation output by the shared encoder into the second classification head to perform further depth mapping classification calculations, thereby outputting another two-dimensional second state probability vector (for example, this vector is represented as [...]). , ]).

[0213] Through this conditional activation mechanism, the multi-task network only delves deeper into the specific physiological causes of the pathology after a heart rhythm disorder has been confirmed at the first level. This strictly ensures the progressive rigor of the medical diagnostic logic while also saving the system's computational resources.

[0214] S544. Based on the second state probability vector, determine the target abnormality type in respiratory sinus arrhythmia and non-respiratory sinus arrhythmia, and output the target abnormality type as the sinus rhythm type.

[0215] Specifically, the processor further compares the two numerical components in the second state probability vector, selects the category with the highest probability value, determines the final specific pathological subtype as the target abnormal type between respiratory sinus arrhythmia and non-respiratory sinus arrhythmia, and outputs the target abnormal type as the sinus rhythm type inferred by the system as a whole.

[0216] Furthermore, in order to achieve a complete closed-loop health monitoring system, the processor is configured to transmit the obtained classification results (i.e., the result with the highest probability among the three types of normal sinus rhythm, respiratory sinus arrhythmia, and non-respiratory sinus arrhythmia) to the terminal's display module for user viewing via a built-in communication module (such as WIFI protocol or Bluetooth transmission protocol).

[0217] Since respiratory sinus arrhythmia is generally considered a normal physiological compensatory phenomenon in medical diagnosis, the system is configured to implement a hierarchical feedback mechanism when executing its early warning strategy: the processor only triggers a safety warning signal for non-respiratory sinus arrhythmias that are determined to be pathological. With this targeted early warning mechanism, the system can not only promptly and accurately alert users to potential cardiac health risks, but also perfectly avoid unnecessary medical anxiety caused by frequent alarms due to harmless physiological rhythm fluctuations.

[0218] In summary, the physiological characteristic monitoring data processing of this application has the following beneficial effects: 1. Achieved high-fidelity and robust contactless extraction of low-level physiological signals; This application acquires radar in-phase orthogonal signals and integrates dynamic environmental background cancellation, target space locking based on constant false alarm rate (CA-CFAR), and temporal phase differential compensation mechanism. It not only completely eliminates the physical constraints of traditional contact devices such as electrocardiograms (ECG) on users, but also adaptively eliminates static clutter and phase entanglement distortion in complex and dynamic sleep environments, thereby accurately restoring the real chest cavity micro-displacement signal.

[0219] 2. Achieved cardiopulmonary decoupling of mixed micro-motion signals, solving the problem of modal aliasing; This application employs the Variable Mode Decomposition (VMD) algorithm, combined with a preset penalty factor and physiological frequency bands (0.1-0.6Hz respiratory band, 0.6-2Hz heartbeat band) to perform signal separation. Compared to traditional Empirical Mode Decomposition (EMD), it not only achieves non-recursive adaptive decoupling of respiratory, heartbeat, and noise components in the frequency domain, completely solving the mode aliasing problem, but also losslessly removes out-of-band clutter while preserving the physical meaning of the waveform.

[0220] 3. A bridge for mapping heart rate variability characteristics to connect with mature medical evaluation systems was constructed; This application employs a dynamic sliding window mechanism to robustly extract heartbeat peaks and maps the extracted continuous heartbeat cycle sequence to the standard medical "RR interval sequence." Based on this, the system further calculates time-domain statistical features such as pNN50, SDNN, and RMSSD, as well as frequency-domain power features such as LF, HF, and LF / HF. This design enables seamless integration of radar wave-based physical displacement data into a mature clinical heart rate variability diagnostic system, endowing non-contact radar signals with profound medical characterization value.

[0221] 4. A quantitative identification method for respiratory / non-respiratory arrhythmias based on the cardiopulmonary coupling mechanism was realized; Traditional electrocardiogram (ECG) monitoring often struggles to distinguish between harmless physiological arrhythmias and harmful pathological arrhythmias. This application, based on the medical principle of "heart rate fluctuating periodically with respiration" in respiratory sinus arrhythmia, extracts the degree of cardiopulmonary synchrony as a quantifiable mathematical feature by aligning the data lengths of respiratory signals and heartbeat cycle sequences, and calculating the normalized cross-correlation coefficient and the percentage of time the target was achieved. This provides a crucial analytical dimension for accurately blocking normal physiological warnings and targeting the identification of pathological arrhythmias.

[0222] 5. Based on the multi-task Transformer architecture, a hierarchical and accurate diagnosis and false alarm prevention and early warning system were implemented to reduce costs and increase efficiency; The classification network in this application adopts an advanced architecture of "static feature embedding + temporal feature encoding + shared encoder + dual-task output head". On the one hand, it perfectly realizes the self-attention deep fusion of multimodal heterogeneous data (discrete static statistics and continuous dynamic waveforms); on the other hand, its condition-triggered dual-classification head design of "first-level judgment of abnormality, second-level judgment of etiology" can "short-circuit" the calculation in time when a normal heart rhythm is diagnosed, which greatly saves the computing power on the end side. At the same time, the system only triggers pathological warnings for the most probable "non-respiratory sinus arrhythmia", completely shielding physiological alarms caused by normal respiratory compensation, perfectly eliminating the user's "electronic medical anxiety", and forming a perfect health monitoring closed loop.

[0223] 6. It enables the intuitive output of complex medical data, breaking down the cognitive barriers of professional medical indicators; This application achieves accurate qualitative differentiation between respiratory and non-respiratory sinus arrhythmias through deep fusion and classification network inference of multidimensional features (static heart rate features and dynamic respiratory-heartbeat correlation features). Compared to traditional heart rate variability (HRV) monitoring devices, which can only output obscure low-level statistical parameters, resulting in the inherent limitation that requires physicians with professional medical theoretical knowledge to interpret them, this solution can directly output highly intuitive final heart rhythm type results at the device. This design allows ordinary users, without any professional medical knowledge, to clearly and accurately understand their own cardiac health status, greatly improving the universality and interactive experience of home health monitoring devices.

[0224] Furthermore, embodiments of this application also provide a computer-readable storage medium (or a non-volatile computer-readable storage medium) storing a computer program (or instructions). When the computer program is executed by a processor, it implements the various steps in the embodiments of the above-described physiological characteristic monitoring data processing method. The computer-readable storage medium may include any medium capable of storing program code, including but not limited to: read-only memory (ROM), random access memory (RAM), magnetic disk, optical disk, flash memory, hard disk (HDD), or solid-state drive (SSD). This storage medium may exist independently or be integrated into a processor or server.

[0225] The embodiments described herein may be provided as methods, systems, or computer program products. Therefore, this application may be implemented entirely in hardware, entirely in software, or a combination of hardware and software. Furthermore, this application may also be embodied as a computer program product implemented on one or more computer-readable storage media (including but not limited to disk storage, optical storage, flash memory, etc.).

[0226] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to this embodiment. It should be understood that each flow, block, and combination thereof in the flowchart illustrations and / or block diagrams can be implemented by computer program instructions. These instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device for execution, thereby producing a machine for implementing a specified function. Simultaneously, these instructions can also be stored in a computer-readable storage medium or loaded onto a computer device, causing the device to perform a series of operational steps to produce a computer-implemented process.

[0227] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. If such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for processing physiological characteristic monitoring data, characterized in that, include: Acquire radar in-phase orthogonal signals, locate the target's thoracic cavity range cell based on the radar in-phase orthogonal signals and extract phase information to generate micro-motion signals; Based on a preset frequency range, the micro-motion signal is separated to extract the respiratory and heartbeat signals; Extract the heartbeat cycle sequence from the heartbeat signal, and generate heart rhythm state features based on the heartbeat cycle sequence; Calculate the temporal correlation between respiratory signals and heart rate cycle sequences to generate respiratory-heart rate correlation features; The heart rhythm characteristics and breathing-heartbeat correlation characteristics are input into a preset classification network to perform feature fusion and inference operations, and output the sinus rhythm type.

2. The physiological characteristic monitoring data processing method as described in claim 1, characterized in that, Acquire radar in-phase orthogonal signals, locate the target's thoracic cavity range cell based on the radar in-phase orthogonal signals and extract phase information to generate micro-motion signals, including: Extract the environmental background reference signal within a preset time window, and perform static clutter filtering on the radar in-phase quadrature signal based on the environmental background reference signal to generate a denoised signal; A distance fast Fourier transform is performed on the denoised signal to generate a distance spectrum matrix. Based on the constant false alarm rate detection algorithm, the distance spectrum matrix is ​​traversed, and the distance cell corresponding to the maximum signal energy is locked as the target chest cavity distance cell. The initial phase information of the target thoracic cavity distance unit is extracted, and the phase entanglement in the initial phase information is eliminated based on the temporal phase difference rule to generate a micro-motion signal.

3. The physiological characteristic monitoring data processing method as described in claim 1, characterized in that, Based on a preset frequency range, signal separation is performed on the micro-motion signal to extract respiratory and heartbeat signals, including: The micro-motion signal is subjected to variable mode empirical decomposition to generate multiple eigenmode functions with independent center frequencies; Based on preset heart rate frequency ranges and preset respiratory rate frequency ranges, multiple intrinsic mode functions with independent center frequencies are filtered to extract target heart rate mode functions and target respiratory rate mode functions; Signal reconstruction is performed on the target heart rate mode function and the target respiratory rate mode function respectively to generate heartbeat signals and respiratory signals.

4. The physiological characteristic monitoring data processing method as described in claim 1, characterized in that, Extract the heartbeat cycle sequence from the heartbeat signal, and generate cardiac rhythm state features based on the heartbeat cycle sequence, including: Peak detection is performed on the heartbeat signal based on the statistical characteristics of a preset sliding window, and the peak timestamp sequence is extracted. Calculate the time difference between adjacent peak timestamps to generate a heartbeat cycle sequence; Calculate time-domain statistical features and / or frequency-domain power features based on the heartbeat cycle sequence, and generate heart rhythm state features based on the time-domain statistical features and / or frequency-domain power features.

5. The physiological characteristic monitoring data processing method as described in claim 4, characterized in that, Time-domain statistical characteristics include at least one of the following: the proportion of adjacent period differences exceeding the limit, the standard deviation of the periodic series, and the root mean square of the adjacent period differences; Frequency domain power characteristics include at least one of low-frequency power, high-frequency power, and the ratio of low-frequency to high-frequency power. Based on the heartbeat cycle sequence, time-domain statistical features and / or frequency-domain power features are calculated, and based on the time-domain statistical features and / or frequency-domain power features, cardiac rhythm state features are generated, including: The time-domain statistical features and frequency-domain power features are calculated based on the heartbeat cycle sequence, and the extracted time-domain statistical features and frequency-domain power features are concatenated by vector dimension to generate heart rhythm state features.

6. The physiological characteristic monitoring data processing method as described in claim 1, characterized in that, Calculate the temporal correlation between respiratory signals and heart rate cycle sequences to generate respiratory-heart rate correlation features, including: Perform time-length alignment processing on the heartbeat cycle sequence to align the data length of the heartbeat cycle sequence with that of the respiratory signal in the time dimension, and use the aligned heartbeat cycle sequence as the target cycle sequence. Standardization processing is performed on the respiratory signal and the target periodic sequence respectively to generate standard respiratory signal and standard periodic sequence that conform to the preset distribution state; Calculate the cross-correlation function between the standard respiratory signal and the standard periodic sequence, and extract the peak value of the cross-correlation coefficient; The target duration that is greater than the preset correlation threshold in the result of the statistical cross-correlation function is calculated as the proportion of the target duration in the total monitoring duration, and the cross-correlation compliance duration percentage is generated. By combining the peak value of the cross-correlation number with the proportion of cross-correlation duration meeting the target, a respiratory-heartbeat correlation feature is generated.

7. The physiological characteristic monitoring data processing method as described in claim 6, characterized in that, The heart rhythm characteristics and respiration-heartbeat correlation characteristics are input into a pre-defined classification network to perform feature fusion and inference operations, and output the sinus rhythm type, including: The heart rhythm state features are combined with the breathing-heartbeat correlation features into a static feature vector. Feature embedding processing is performed on the static feature vector to generate an embedded static vector. Obtain the temporal feature sequence corresponding to the static feature vector, perform position encoding and dimension mapping on the temporal feature sequence, and generate a temporal feature encoding vector; The embedded static vector and the temporal feature encoding vector are fused together and input into the shared encoder in the classification network to perform self-attention feature extraction, generating a deep feature representation. Multi-task inference operations are performed based on deep feature representation to output the sinus rhythm type.

8. The physiological characteristic monitoring data processing method as described in claim 7, characterized in that, The classification network includes a first classification head and a second classification head; Multi-task inference operations are performed based on deep feature representation to output sinus rhythm type, including: Input the deep feature representation into the first classification head and output the first state probability vector; In response to the first state probability vector being represented as normal sinus rhythm, normal sinus rhythm is output as the sinus rhythm type; In response to the first state probability vector representing sinus arrhythmia, the second classifier is activated, the deep feature representation is input into the second classifier, and the second state probability vector is output. Based on the second state probability vector, the target abnormality type is determined in respiratory sinus arrhythmia and non-respiratory sinus arrhythmia, and the target abnormality type is output as the sinus rhythm type.

9. A physiological characteristic monitoring data processing device, characterized in that, It includes a memory, a processor, and a physiological characteristic monitoring data processing program stored in the memory and executable on the processor. When the processor executes the physiological characteristic monitoring data processing program, it implements the physiological characteristic monitoring data processing method as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, A physiological characteristic monitoring data processing program is stored on a computer-readable storage medium, and when executed by a processor, the physiological characteristic monitoring data processing program implements the physiological characteristic monitoring data processing method as described in any one of claims 1-8.

Images