A cardiovascular physiological signal feature extraction method applying a pattern recognition algorithm

Abstract

The application relates to the technical field of pattern recognition, in particular to a cardiovascular physiological signal feature extraction method applying a pattern recognition algorithm, which comprises the following steps: collecting original signals through a sensor array and converting the original signals into digital signal streams; synchronously filtering out baseline drift and electromagnetic interference by applying a composite filtering algorithm; determining a key frequency interval by using a discrete orthogonal transform algorithm for multi-scale pattern deconstruction; accurately positioning ECG and pulse wave feature key points by applying a pattern recognition model based on a deep residual network, and marking coordinate sequences; and calculating a multi-dimensional feature vector including a nonlinear dynamics index according to key point attributes. The application improves extraction precision in a complex environment, enhances the generalization ability for different individuals and pathological states, realizes deep quantification of physiological features, and is suitable for mobile monitoring and clinical real-time monitoring.

Description

Technical Field

[0001] This invention relates to the field of pattern recognition technology, and in particular to a method for extracting cardiovascular physiological signal features using pattern recognition algorithms. Background Technology

[0002] With the widespread application of biomedical sensing technology and intelligent medical devices, real-time monitoring of physiological signals of the cardiovascular system has become an important part of clinical diagnosis and daily health management.

[0003] When processing cardiovascular signals with strong noise interference and high dynamic changes, existing technologies struggle to retain weak physiological feature components while removing environmental baseline drift, which can easily lead to irreversible morphological distortion of the signal waveform during feature calculation.

[0004] Meanwhile, conventional feature point recognition strategies rely heavily on fixed empirical thresholds or single time-domain criteria, and often exhibit low generalization performance when faced with significant physiological and anatomical differences between individuals and non-stationary rhythmic fluctuations of the same organism under stress. Summary of the Invention

[0005] The inventors discovered through research that applying pattern recognition algorithms to extract features from cardiovascular physiological signals is a key path to improve monitoring accuracy. The aim is to automatically identify and locate key feature points of the cardiac cycle, such as wave group morphology in electrocardiogram signals or ghost pulse points in pulse waves, by digitally processing and deconstructing the original signals, thereby achieving precise quantification of cardiovascular operating status.

[0006] The purpose of this invention is to provide a method for extracting cardiovascular physiological signal features using a pattern recognition algorithm, in order to solve the problems of the prior art.

[0007] This invention provides a method for extracting cardiovascular physiological signal features using a pattern recognition algorithm, comprising:

[0008] Step 1. Acquire raw cardiovascular physiological signals. Collect raw analog sequences reflecting cardiac electrophysiological activity and circulatory system pressure fluctuations by using a sensor array deployed at specific parts of the organism. Then, use an analog-to-digital conversion module to convert the raw analog sequences into discrete digital signal streams.

[0009] Step 2. Execute the signal preprocessing program, apply a composite filtering algorithm to the digital signal stream, and simultaneously filter out baseline drift interference, power frequency electromagnetic interference and human electromyographic noise in the time and frequency domains to restore the true waveform shape and linear change trend of the cardiovascular physiological signal.

[0010] Step 3. Perform multi-scale pattern deconstruction of the signal. Use the discrete orthogonal transform algorithm to decompose the preprocessed signal into multiple sub-band components with different physical meanings. By analyzing the energy distribution characteristics at different scales, determine the specific frequency range containing key information about the heartbeat.

[0011] Step 4. Identify key physiological features, apply a pre-built pattern recognition model to perform a depth scan on the sub-band components, locate key physiological features, and mark the coordinate sequence of the key points on the time axis;

[0012] Step 5. Extract and quantify physiological feature vectors. Based on the time coordinates and amplitude attributes of the key points, calculate multi-dimensional feature indicators and construct a feature vector set to characterize cardiovascular health status.

[0013] In some embodiments, the signal acquired by the sensor array in step one passes through an input quality assessment stage before entering step two. The input quality assessment stage evaluates the signal-to-noise ratio by calculating the ratio of the root mean square energy value of the signal to the noise floor level within a predetermined observation window, and monitors in real time whether the signal amplitude exceeds the maximum quantization range of the analog-to-digital conversion module to determine the degree of waveform saturation.

[0014] If the current signal-to-noise ratio is detected to be lower than the preset minimum working threshold, or the number of saturation points exceeds the predetermined proportion, the system will automatically generate a prompt command and automatically increase the fault tolerance coefficient of the subsequent filtering algorithm.

[0015] The sensor array includes multiple leads of electrocardiogram electrodes and at least one set of photoplethysmography (PPG) sensors, used to synchronously capture bioelectric potential signals and peripheral blood volume change signals at a preset sampling frequency.

[0016] The analog-to-digital conversion module has multi-channel synchronous acquisition capability and two mirrored ping-pong buffers in memory. When the analog-to-digital conversion module writes a digital signal stream to the first buffer, the processor simultaneously processes the data already stored in the second buffer, ensuring the connection between signal acquisition and feature extraction.

[0017] In some embodiments, the baseline drift removal process in step two employs an algorithm based on a combination of median filtering and moving average, specifically:

[0018] First, the median of the signal within a predetermined length window is calculated, the length of which is set to cover at least 1.5 complete heartbeat cycles, thereby constructing a preliminary baseline estimation curve.

[0019] The curve is then smoothed using a low-pass filter to obtain the final baseline drift component. The baseline drift component is then subtracted from the original digital signal to achieve zero-point calibration of the signal baseline. For power frequency electromagnetic interference, a notch filter bank is used, with its center frequency precisely locked at 50 Hz or 60 Hz and its higher harmonics.

[0020] In some embodiments, the discrete orthogonal transform algorithm in step three involves projecting the preprocessed signal onto a set of orthogonal basis function spaces. Through a multi-level iterative decomposition process, the signal is decomposed into a low-frequency trend part that reflects the slow regulation mechanism of the cardiovascular system and a high-frequency detail part that captures the abrupt information of myocardial depolarization.

[0021] The system dynamically adjusts the number of decomposition layers based on the real-time calculated signal-to-noise ratio. When the ambient noise is high, the decomposition depth is increased to refine the separation of noise and signal. After the decomposition is completed, the system performs energy distribution characteristic analysis on each sub-band component, calculates the proportion of each component in the total energy, and determines the specific frequency range containing key information about the heartbeat by comparing it with the preset typical physiological frequency distribution template.

[0022] The discrete orthogonal transform process is decomposed into a series of addition and shift operations, and the processor automatically determines whether the signal is affected by external high-frequency vibration interference by monitoring the energy entropy of the high-frequency detail coefficients. If the energy entropy exceeds a predetermined threshold, the system automatically increases the attenuation coefficient in that frequency band.

[0023] In some embodiments, the pattern recognition model in step four is based on a deep residual network architecture. The deep residual network is composed of multiple stacked residual blocks. Each residual block includes a convolutional layer, a batch normalization layer, and a nonlinear activation unit. The nonlinear activation unit adopts linear correction logic, that is, when the input value is greater than a preset zero value, the output value is equal to the input value, and when the input value is less than or equal to the preset zero value, the output value is always zero.

[0024] The convolutional layer extracts local morphological features through a sliding weighted window, and the batch normalization layer is used to accelerate network convergence. The pattern recognition model also includes an attention mechanism weight generation branch, which calculates an attention weight distribution map in the time dimension based on the local variance and frequency distribution of the signal, guiding computational resources to focus on regions with drastic waveform variations. For the extraction of atypical cardiac waveforms, the number of branches in the feature extraction path is increased to achieve parallel capture of multi-scale morphological features.

[0025] In some embodiments, the process of identifying key points of physiological features further includes an adaptive threshold determination step; this step dynamically sets a predetermined threshold for detecting feature peaks by statistically analyzing the local energy maxima of the signal within the current processing window and combining the mean information of historical heartbeat cycles; when the waveform amplitude exceeds the predetermined threshold and the sign of the first-order difference value before and after it changes, and the morphological slope requirement is met, it is determined to be a candidate feature point.

[0026] Subsequently, the local windows around the candidate feature points are classified by the pattern recognition model to remove false feature points generated by motion artifacts; when the time span between adjacent feature points is detected to be greater than two seconds or less than 0.3 seconds, it is determined to be an abnormal feature point and the compensation algorithm is automatically started.

[0027] The compensation algorithm is based on the periodicity of historical waveforms. It performs interpolation prediction at the location of suspected lost feature points, or when the electrocardiogram signal is disturbed, it uses the periodicity of the pulse wave signal to infer the heart rate and determines the location of the feature point by calculating the pulse wave propagation time delay.

[0028] In some embodiments, the process of extracting the heartbeat interval in step five involves calculating the time difference between two adjacent feature peak points; the extraction of the waveform slope is achieved by calculating the first-order difference quotient in the neighborhood of the feature point; the calculation of the area under the curve adopts numerical integration logic, that is, dividing the region enclosed by the waveform curve and the time axis into several small trapezoidal units, and calculating the sum of the areas of all trapezoidal units;

[0029] The feature vector set also includes nonlinear dynamic indicators, which are used to describe the changes in the complexity of the cardiovascular system by reconstructing the phase space of the extracted time series, calculating the correlation dimension and information entropy; all extracted feature indicators are normalized by subtracting the minimum value of the sequence from the indicator value and dividing by the range, so that their numerical range is mapped to a predetermined standard interval.

[0030] During the initialization phase, the system also calibrates the anatomical compensation parameters by acquiring a reference signal of a predetermined length. The anatomical compensation parameters are used to correct signal attenuation and phase shift caused by differences in sensor attachment position, skin impedance, and subcutaneous tissue thickness.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] 1. This invention significantly improves the accuracy of cardiovascular physiological signal feature extraction under complex electromagnetic environments and motion artifact interference by applying advanced pattern recognition algorithms and multi-scale decomposition technology. The adaptive filtering and baseline correction logic it employs can effectively eliminate the influence of non-stationary noise on waveform morphology, ensuring that the extracted physical indicators can truly and objectively reflect cardiac electrophysiological activity and the state of the circulatory system.

[0033] 2. The recognition model based on deep residual networks constructed in this invention overcomes the limitations of traditional algorithms that rely on fixed empirical thresholds, and has extremely strong generalization ability. For physiological and anatomical differences among different individuals and signal fluctuations of the same individual under different physiological stress states, this method can maintain extremely high feature localization stability and significantly reduce the false detection rate and false negative rate of atypical waveforms.

[0034] 3. This invention achieves in-depth quantification of multi-dimensional physiological characteristics. In addition to extracting traditional time-domain and amplitude-domain indicators, it also introduces nonlinear dynamic features and multi-scale frequency features to construct a more comprehensive and three-dimensional feature vector set. Attached Figure Description

[0035] 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 these drawings without creative effort.

[0036] Figure 1 This is a flowchart of the cardiovascular physiological signal feature extraction method using the pattern recognition algorithm of the present invention;

[0037] Figure 2 This is a logic diagram of the present invention for locating key points of physiological features based on a deep residual network architecture and adaptive threshold determination. Detailed Implementation

[0038] The following will be based on embodiments of the present invention. Figures 1-2 The technical solutions in the embodiments of the present invention will be clearly and completely described together. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0039] It should be noted that the term "system" in this invention refers to all objects capable of performing the methods of this invention, including but not limited to computers, memory, etc.

[0040] Example

[0041] This embodiment provides a method for extracting cardiovascular physiological signal features using a pattern recognition algorithm, including: Step 1: Obtaining raw cardiovascular physiological signals.

[0042] In this step, synchronous data acquisition is performed using a sensor array deployed at specific locations on the organism. The sensor array is specifically configured with multiple leads of electrocardiogram (ECG) electrodes and at least one set of photoplethysmography (PPG) sensors. The ECG electrodes, made of silver / silver chloride, are attached to specific points on the limbs or chest of the organism to capture the weak bioelectric potential signals generated by the depolarization and repolarization of myocardial cells.

[0043] Photoplethysmography (PPG) sensors include a specific wavelength emission light source and a highly sensitive photodetector. By emitting red or infrared light into the skin tissue of a living organism and receiving reflected or transmitted light after being modulated by peripheral blood vessels, they obtain a simulated sequence reflecting the signal of changes in peripheral blood volume.

[0044] During the acquisition process, the system sets a constant preset sampling frequency, such as, but not limited to, 512 Hz or 1024 Hz, to ensure sufficient resolution of the signal in the time domain. The acquired raw analog sequence is sent to a high-precision analog-to-digital converter module, which has quantization accuracy of more than 24 bits and multi-channel synchronous sample-and-hold capability. Under a unified clock trigger, the analog-to-digital converter module converts multiple analog potential signals into discrete digital signal streams.

[0045] To ensure alignment of signals from different sources, the system embeds a high-precision timestamp in the data packet header, providing a consistent time reference for subsequent multimodal feature fusion. Before the digital signal enters the subsequent processing stage, it undergoes an input quality assessment stage. This stage evaluates the signal-to-noise ratio (SNR) by calculating the ratio of the root mean square energy value to the noise floor level within a predetermined observation window, and monitors in real time whether the signal amplitude exceeds the maximum quantization range of the analog-to-digital conversion module to determine the waveform saturation level. If the current signal SNR is detected to be lower than the preset minimum operating threshold, or the number of saturation points exceeds a predetermined proportion, the system will automatically generate a prompt instruction to guide the user to adjust the sensor attachment position and automatically increase the fault tolerance coefficient of the subsequent filtering algorithm.

[0046] Next, in step two, a signal preprocessing procedure is executed. The first step is to remove baseline drift, which is typically caused by respiratory movements or electrode polarization and manifests as slow fluctuations in the signal at very low frequencies. During processing, the system first calculates the median of the signal within a sliding window of a specific length, set to cover at least 1.5 complete heartbeat cycles. A preliminary baseline estimation curve is constructed by extracting the medians from each window. Then, a low-pass filter with smoothing properties is applied to this curve for secondary processing, removing abrupt transitions to obtain the final baseline drift component. The original digital signal is then subtracted point-by-point from this baseline drift component to achieve zero-point calibration of the signal baseline.

[0047] For power frequency electromagnetic interference, the system employs a set of high-performance notch filters, with their center frequencies precisely locked at 50 Hz or 60 Hz and their second and third harmonics. The bandwidth of this notch filter is set within an extremely narrow range to ensure that physiological characteristic components distributed in adjacent frequency bands are not damaged while suppressing power frequency noise. For human electromyographic noise, the system applies a time-frequency domain synchronous filtering logic to identify and suppress high-frequency, random spikes in the signal, thereby restoring a clear linear trend.

[0048] In addition, the preprocessing process also involves an adaptive adaptation mechanism. The three-axis accelerometer integrated in the sensor array senses the motion intensity of the organism. When the motion intensity exceeds the preset motion state threshold, the system automatically switches to the motion mode parameter package. At this time, the passband width of the filter is appropriately compressed to enhance the resistance to motion artifacts.

[0049] Next, in step three, the preprocessed signal is projected into a set of orthogonal basis function spaces using the discrete orthogonal transform algorithm. Through a multi-level iterative decomposition process, the signal is analyzed layer by layer into low-frequency trend parts and a series of high-frequency detail parts.

[0050] The low-frequency trend portion carries information about the slow regulation of the cardiovascular system, such as the effect of respiratory modulation on heart rate; while the high-frequency detail portion accurately captures rapid abrupt changes in the myocardial depolarization process, such as the spike components in the electrocardiogram waveform. The system dynamically adjusts the number of decomposition layers based on the real-time calculated signal-to-noise ratio. When the ambient noise is high, the decomposition depth is increased to more finely separate noise from the signal.

[0051] After decomposition, the system analyzes the energy distribution characteristics of each sub-band component and calculates the proportion of each component in the total energy. By comparing with a preset typical physiological frequency distribution template, specific frequency intervals containing key information about the heartbeat are determined. This process effectively removes signal redundancy, transforming complex physiological signals into a series of sub-band feature streams with clear physical meanings.

[0052] Then, in step four, the system applies a pre-built deep residual network model to perform a depth scan of the subband components. The deep residual network consists of multiple stacked residual blocks containing convolutional layers, batch normalization layers, and nonlinear activation units.

[0053] Convolutional layers extract local morphological features through sliding weighted windows, batch normalization layers accelerate network convergence and enhance stability, and nonlinear activation units employ linear correction logic, meaning the output equals the input when the input value is greater than zero, and the output is zero otherwise. The design of residual connections allows input information to be passed across layers, effectively avoiding information loss problems in deep models.

[0054] During the recognition process, the system also introduces an attention mechanism, which calculates the importance weight coefficients of different positions of the signal, enabling the model to automatically focus on areas with drastic waveform variations.

[0055] The process of identifying key physiological features also includes an adaptive threshold determination step. This step involves real-time statistical analysis of the local energy maxima of the signal within the current processing window, combined with the average value information of several past cardiac cycles, to dynamically set a predetermined threshold for detecting feature peaks.

[0056] When the waveform amplitude at a certain point exceeds the threshold, and the sign of the first-order difference value before and after it changes (representing a maximum point), and a specific slope threshold requirement is met, it is determined to be a candidate feature point. Subsequently, the pattern recognition model performs fine classification on the local windows around the candidate feature point, and uses its learned deep semantic features to eliminate false feature points caused by electromagnetic interference or motion false alarms.

[0057] For electrocardiogram (ECG) signals, the model accurately locates the peak points of complex waveforms; for pulse wave signals, it utilizes second- and third-order higher-order derivative analysis to amplify subtle inflection points in the waveform, thereby capturing ghost pulse points and notch points. If the system detects an abnormal time span between adjacent feature points, exceeding the normal physiological range, it activates a compensation algorithm to perform interpolation prediction based on the periodicity of historical cycles, ensuring the continuity of the feature sequence.

[0058] Finally, step five obtains the cardiac interval sequence by calculating the time difference between two adjacent characteristic peak points. Next, it calculates the first-order difference quotient within the neighborhood of each characteristic point to characterize the waveform slope, reflecting the rate of pressure rise during cardiac pumping. The peak height is obtained by the difference between the characteristic point amplitude and the baseline value. The area under the curve is calculated using numerical integration logic, dividing the closed region enclosed by the characteristic waveform and the time axis into several trapezoidal units of minimal width, and then summing the areas of all trapezoidal units.

[0059] In some possible implementations, the system also introduces nonlinear dynamic indicators to reconstruct the phase space of the extracted time series, describing the complexity of the cardiovascular system at the nonlinear level by calculating the correlation dimension and information entropy. All extracted feature indicators are normalized by subtracting the sequence minimum value from the indicator value and dividing by the range, mapping them to a standard interval between zero and one.

[0060] The constructed feature vector set is stored in a cyclically updated local buffer. The system aggregates the data in the buffer at predetermined intervals, calculating statistical properties such as mean, standard deviation, and skewness to form a macro-trend descriptor. To reduce redundancy, the system applies principal component analysis to reduce the dimensionality of the feature vectors, selecting the core feature subset with the highest correlation coefficient to physiological state evolution, thus providing efficient and accurate data support for health analysis throughout the entire life cycle.

[0061] Example 2

[0062] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail with reference to specific embodiments.

[0063] This embodiment describes a specific application scheme of the method of the present invention in a long-term static and dynamic mixed monitoring scenario, focusing on the data stream segmentation and processing and the consistency maintenance of feature sequences.

[0064] In step one, when the system performs long-term monitoring lasting for hours or even days, the generated digital signal stream is continuous and massive. To ensure real-time processing, the system divides the continuous signal stream into multiple data frames of predetermined lengths, for example, each frame contains 10 seconds of physiological data. To prevent the loss of key cross-frame features at frame boundaries (such as a heartbeat cycle spanning two frames), the system sets a predetermined overlap region at the end of each data frame, the length of which is typically set to two seconds.

[0065] In step two, each frame of data is independently fed into the composite filter bank. Since the filter produces an initiation effect (i.e., transient response) at the start of processing, the system uses the filter state value of the overlapping area of ​​the previous frame to initialize the filter parameters of the current frame, thereby ensuring that the filtered signal is continuous and smooth at the inter-frame transition. Baseline correction and notch filtering also follow this state continuation mechanism, avoiding traces of artificial splicing in long-term signals.

[0066] In steps three and four, the multi-scale deconstruction and recognition logic operates concurrently on a multi-core processor, with different processor cores handling signal components from different leads. After locating feature points within each frame, the recognition model generates a local time coordinate sequence. The system maintains a global coordinate transformation engine, mapping local coordinates to global timestamps based on the offset of each frame on the overall time axis. For motion monitoring, the adaptive logic fine-tunes the filtering parameters and recognition thresholds for each frame by reading accelerometer data in real time. For example, the system detects a jump in vibration frequency the instant a user transitions from rest to running and immediately shortens the median filter window to track faster baseline fluctuations.

[0067] In step five, feature quantization is no longer limited to instantaneous waveforms but constructs a feature matrix containing temporal evolution information. Extracted indicators such as heart rate intervals and waveform slopes form a long-term sequence in a buffer. The system extracts deep features reflecting the regulatory patterns of the autonomic nervous system by calculating the autocorrelation function and power spectral density of this sequence. To ensure the stability of the overall feature sequence on a macroscopic time scale, the system uses a weighted average method to smooth the overlapping feature parameters extracted from adjacent frames. If the difference in heart rate indicators extracted from the overlapping region between two frames exceeds a preset physiological mutation threshold, the system automatically initiates a verification process, rescanning the original waveform in that region to eliminate recognition errors.

[0068] Ultimately, the constructed feature vector set includes not only instantaneous morphological indicators but also statistics reflecting lifelong cardiac health trends, such as the slope and coefficient of variation of these indicators over time. This high-dimensional data is compressed and stored locally, and periodically synchronized to the central processing unit via Bluetooth Low Energy or Wi-Fi.

[0069] Through this frame-by-frame parallel processing scheme with state continuity, the present invention achieves efficient and stable monitoring of long-term dynamic cardiovascular physiological signals while maintaining high-precision extraction.

[0070] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the scope of the invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0071] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for extracting cardiovascular physiological signal features using a pattern recognition algorithm, characterized in that, include: Step 1. Acquire raw cardiovascular physiological signals. Collect raw analog sequences reflecting cardiac electrophysiological activity and circulatory system pressure fluctuations by using a sensor array deployed at specific parts of the organism. Then, use an analog-to-digital conversion module to convert the raw analog sequences into discrete digital signal streams. Step 2. Execute the signal preprocessing program, apply a composite filtering algorithm to the digital signal stream, and simultaneously filter out baseline drift interference, power frequency electromagnetic interference and human electromyographic noise in the time and frequency domains to restore the true waveform shape and linear change trend of the cardiovascular physiological signal. Step 3. Perform multi-scale pattern deconstruction of the signal. Use the discrete orthogonal transform algorithm to decompose the preprocessed signal into multiple sub-band components with different physical meanings. By analyzing the energy distribution characteristics at different scales, determine the specific frequency range containing key information about the heartbeat. Step 4. Identify key physiological features, apply a pre-built pattern recognition model to perform a depth scan on the sub-band components, locate key physiological features, and mark the coordinate sequence of the key points on the time axis; Step 5. Extract and quantify physiological feature vectors. Based on the time coordinates and amplitude attributes of the key points, calculate multi-dimensional feature indicators and construct a feature vector set to characterize cardiovascular health status.

2. The method according to claim 1, characterized in that, Before entering step two, the signal acquired by the sensor array in step one goes through an input quality assessment stage. The input quality assessment stage evaluates the signal-to-noise ratio by calculating the ratio of the root mean square energy value of the signal to the noise floor level within a predetermined observation window, and monitors in real time whether the signal amplitude exceeds the maximum quantization range of the analog-to-digital conversion module to determine the degree of waveform saturation. If the current signal-to-noise ratio is detected to be lower than the preset minimum working threshold, or the number of saturation points exceeds the predetermined proportion, the system will automatically generate a prompt command and automatically increase the fault tolerance coefficient of the subsequent filtering algorithm. The sensor array includes multiple leads of electrocardiogram electrodes and at least one set of photoplethysmography (PPG) sensors, used to synchronously capture bioelectric potential signals and peripheral blood volume change signals at a preset sampling frequency. The analog-to-digital conversion module has multi-channel synchronous acquisition capability and two mirrored ping-pong buffers in memory. When the analog-to-digital conversion module writes a digital signal stream to the first buffer, the processor simultaneously processes the data already stored in the second buffer, ensuring the connection between signal acquisition and feature extraction.

3. The method according to claim 1, characterized in that, The baseline drift removal process in step two employs an algorithm based on a combination of median filtering and moving average, specifically as follows: First, the median of the signal within a predetermined length window is calculated, the length of which is set to cover at least 1.5 complete heartbeat cycles, thereby constructing a preliminary baseline estimation curve. The curve was then smoothed using a low-pass filter to obtain the final baseline drift component. The baseline drift component is subtracted from the original digital signal to achieve zero-point calibration of the signal baseline; for power frequency electromagnetic interference, a notch filter bank is used, and its center frequency is precisely locked at 50 Hz or 60 Hz and its higher harmonics.

4. The method according to claim 1, characterized in that, The discrete orthogonal transform algorithm in step three involves projecting the preprocessed signal onto a set of orthogonal basis function spaces. Through a multi-level iterative decomposition process, the signal is decomposed into a low-frequency trend part that reflects the slow regulation mechanism of the cardiovascular system and a high-frequency detail part that captures the abrupt information of myocardial depolarization. The system dynamically adjusts the number of decomposition layers based on the real-time calculated signal-to-noise ratio. When the ambient noise is high, the decomposition depth is increased to refine the separation of noise and signal. After decomposition, the system performs energy distribution characteristic analysis on each sub-band component, calculates the proportion of each component in the total energy, and determines the specific frequency range containing key information about heartbeat by comparing it with the preset typical physiological frequency distribution template. The discrete orthogonal transform process is decomposed into a series of addition and shift operations, and the processor automatically determines whether the signal is affected by external high-frequency vibration interference by monitoring the energy entropy of the high-frequency detail coefficients. If the energy entropy exceeds a predetermined threshold, the system automatically increases the attenuation coefficient in that frequency band.

5. The method according to claim 1, characterized in that, In step four, the pattern recognition model is based on a deep residual network architecture, which consists of multiple stacked residual blocks. Each residual block contains a convolutional layer, a batch normalization layer, and a nonlinear activation unit. The nonlinear activation unit adopts linear correction logic, that is, when the input value is greater than the preset zero value, the output value is equal to the input value, and when the input value is less than or equal to the preset zero value, the output value is always zero. The convolutional layer extracts local morphological features through a sliding weighted window, and the batch normalization layer is used to accelerate network convergence. The pattern recognition model also includes an attention mechanism weight generation branch, which calculates an attention weight distribution map in the time dimension based on the local variance and frequency distribution of the signal, guiding computational resources to focus on regions with drastic waveform variations. For the extraction of atypical cardiac waveforms, the number of branches in the feature extraction path is increased to achieve parallel capture of multi-scale morphological features.

6. The method according to claim 1, characterized in that, The process of identifying key points of physiological features also includes an adaptive threshold determination step; this step dynamically sets a predetermined threshold for detecting feature peaks by statistically analyzing the local energy maxima of the signal within the current processing window and combining the mean information of historical heartbeat cycles; when the waveform amplitude exceeds the predetermined threshold and the sign of the first-order difference value before and after it changes, and the morphological slope requirement is met, it is determined to be a candidate feature point. Subsequently, the local windows around the candidate feature points are classified by the pattern recognition model to remove false feature points generated by motion artifacts; when the time span between adjacent feature points is detected to be greater than two seconds or less than 0.3 seconds, it is determined to be an abnormal feature point and the compensation algorithm is automatically started. The compensation algorithm is based on the periodicity of historical waveforms. It performs interpolation prediction at the location of suspected lost feature points, or when the electrocardiogram signal is disturbed, it uses the periodicity of the pulse wave signal to infer the heart rate and determines the location of the feature point by calculating the pulse wave propagation time delay.

7. The method according to claim 1, characterized in that, The process of extracting the heartbeat interval in step five involves calculating the time difference between two adjacent feature peak points; the extraction of the waveform slope is achieved by calculating the first-order difference quotient in the neighborhood of the feature point; the calculation of the area under the curve adopts numerical integration logic, that is, dividing the region enclosed by the waveform curve and the time axis into several small trapezoidal units, and calculating the sum of the areas of all trapezoidal units. The feature vector set also includes nonlinear dynamic indicators, which are used to describe the changes in the complexity of the cardiovascular system by reconstructing the phase space of the extracted time series, calculating the correlation dimension and information entropy; all extracted feature indicators are normalized by subtracting the minimum value of the sequence from the indicator value and dividing by the range, so that their numerical range is mapped to a predetermined standard interval. During the initialization phase, the system also calibrates the anatomical compensation parameters by acquiring a reference signal of a predetermined length. The anatomical compensation parameters are used to correct signal attenuation and phase shift caused by differences in sensor attachment position, skin impedance, and subcutaneous tissue thickness.

Images