A model algorithm for estimating metabolic syndrome based on cardiac deceleration force and heart rate variability.

Abstract

This application relates to the field of biomedical engineering technology and discloses a model algorithm for predicting metabolic syndrome based on cardiac deceleration force and heart rate variability. The algorithm includes: acquiring the electrocardiogram (ECG) and mechanical waveform signal sequences of the target object, and extracting the RR interval time series using adaptive anti-interference processing; extracting heart rate variability and cardiac deceleration force features characterizing vagal tone; locating key physiological markers and calculating underlying hemodynamic features such as ventricular ejection time, ventricular isovolumic contraction period, and electro-mechanical coupling delay; performing dimensionless mapping and feature fusion on the above heterogeneous feature set to generate a one-dimensional vector; inputting the data into a state assessment model, obtaining a weighted context vector through nonlinear high-dimensional mapping and dynamic weight allocation, calculating a comprehensive risk index, and outputting a state classification. This invention overcomes the bottleneck of single-signal assessment, taking into account both autonomic nervous system regulation and myocardial mechanical efficacy, and improving the algorithm's robustness and clinical prediction accuracy in complex monitoring scenarios.

Description

Technical Field

[0001] This invention relates to the fields of biomedical engineering and medical health data analysis technology, specifically a model algorithm for estimating metabolic syndrome based on cardiac deceleration force and heart rate variability. Background Technology

[0002] Metabolic syndrome is a complex metabolic disorder closely related to the occurrence and development of cardiovascular disease. Currently, non-invasive assessment methods for metabolic syndrome largely rely on a single electrocardiogram (ECG) signal, extracting parameters such as heart rate variability for risk prediction. However, a single ECG signal only reflects the electrophysiological activity of the heart and cannot capture the mechanical pulsation state and contractile efficacy of the myocardium. This reliance on a single signal means that the assessment of metabolic syndrome cannot combine the autonomic nervous system's regulatory mechanisms on the heart with the mechanical contractile efficacy of the underlying myocardium, resulting in insufficient acquisition and integration of physiological characteristics and limiting the overall accuracy of model predictions.

[0003] To obtain more comprehensive information on cardiovascular status, the joint analysis of multi-source physiological signals has become a trend. However, this has led to errors in the extraction and alignment of underlying physiological parameters. Features extracted from heterogeneous multi-source signals differ significantly in physical dimensions and numerical amplitudes. Furthermore, macroscopic sequence features such as cardiac deceleration force and heart rate variability are difficult to match with beat-by-beat cardiac cycle features such as ventricular isovolumetric contraction and electro-mechanical coupling delay on a temporal scale. Existing data processing workflows lack dimensionless processing for heterogeneous features and strict alignment mechanisms in the time and matrix row dimensions. This makes the model susceptible to large-scale numerical interference during computation, resulting in feature weight skew and making it difficult for the model to achieve stable convergence during training.

[0004] Existing metabolic syndrome risk estimation models typically employ static weight allocation when handling multidimensional fusion features, lacking the ability to dynamically evaluate deeper features. Faced with complex individual physiological differences, existing models cannot adaptively quantify and allocate the contribution of different dimensional features to the classification results based on the actual state of the target individual. Because they cannot selectively strengthen key pathogenic features and suppress irrelevant background noise, the reliability and accuracy of the final comprehensive risk grading results are limited in clinical applications. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a model algorithm for estimating metabolic syndrome based on cardiac deceleration force and heart rate variability. This algorithm solves the problems of insufficient feature acquisition and fusion of multi-source physiological signals in metabolic syndrome risk assessment, large alignment errors of underlying heterogeneous physiological parameters, and the lack of deep feature dynamic allocation capability in state assessment models.

[0006] To achieve the above objectives, the present invention provides the following technical solution: This invention provides a model algorithm for estimating metabolic syndrome based on cardiac deceleration force and heart rate variability, comprising: Obtain the electrocardiogram signal sequence and mechanical waveform signal sequence of the target object, and extract the corresponding RR interval time series; Heart rate variability and cardiac deceleration force features were extracted based on the RR interval time series. Based on the electrocardiogram signal sequence and mechanical waveform signal sequence, key physiological markers are located, and hemodynamic features reflecting the mechanical contractile efficacy of the myocardium are extracted. The hemodynamic features include ventricular ejection time, ventricular isovolumetric contraction period, and electro-mechanical coupling delay. The heart rate variability feature, the cardiac deceleration force feature and the hemodynamic feature are combined into a heterogeneous feature set, and dimensionless mapping and feature fusion are performed to generate a one-dimensional fused feature vector. The one-dimensional fused feature vector is input into the state evaluation model, and a weighted context vector is obtained through nonlinear high-dimensional mapping and dynamic weight allocation; The comprehensive risk index of the target object is calculated based on the weighted context vector, and the risk status classification result of metabolic syndrome is output.

[0007] Further, the step of acquiring the electrocardiogram signal sequence and mechanical waveform signal sequence of the target object, and extracting the corresponding RR interval time series, includes: A median sliding window filter is used to eliminate high-frequency noise and baseline drift in the ECG signal sequence and mechanical waveform signal sequence. The difference between the peak timestamps of adjacent R waves is calculated to generate an initial cardiac cycle time interval sequence, and its arithmetic mean and standard deviation are calculated. Abnormal data are identified and marked based on the 3σ criterion. The arithmetic mean of the normal time intervals before and after the abnormal data is extracted as the replacement value to cover the abnormal data, and the corrected RR interval time series is obtained.

[0008] Furthermore, the extraction of heart rate variability features and cardiac deceleration force features includes: The standard deviation, root mean square of adjacent differences, and percentage of adjacent differences exceeding a preset time difference threshold of the RR interval time series are calculated as time-domain heart rate variability features. A fast Fourier transform is performed on the equal-interval series to obtain the power spectral density function, and the ratio of the total energy in the low-frequency band to the total energy in the high-frequency band is calculated as a frequency-domain heart rate variability feature. The RR interval time series is constructed as a two-dimensional delayed reconstruction scatter plot, and the short-axis and long-axis standard deviations are calculated as nonlinear dynamic features. The position indices in the RR interval time series that satisfy the heart rate deceleration condition are identified as anchor points. Signal segments are extracted centered on the anchor points, phase-aligned, and their arithmetic mean is calculated to generate a reconstructed sequence. The deceleration force features are obtained using differential calculation.

[0009] Furthermore, the method of locating key physiological markers based on the electrocardiogram signal sequence and the mechanical waveform signal sequence includes: The first and second derivative sequences of the mechanical waveform signal sequence are obtained through differential operations; the zero-crossing point in the first derivative sequence that changes from positive to negative and whose original amplitude is greater than the dynamic threshold is identified as the main peak; the local maximum point of the second derivative sequence is searched within a set time interval before the main peak and identified as the trough start point; the local minimum point of the first derivative sequence is searched between the current main peak and the trough start point of the next cardiac cycle and identified as the diphtheria wave notch point.

[0010] Furthermore, the extraction of ventricular ejection time, ventricular isovolumetric contraction period, and electromechanical coupling delay includes: The length of the discrete sampling point interval between the trough start point and the diphtheria notch point is converted into absolute time to obtain the ventricular ejection time; the time span between the peak point of the ECG signal R wave and the trough start point of the mechanical waveform is calculated to obtain the pre-ejection time of the heart, and the preset electromechanical delay time constant is subtracted to obtain the isovolumetric contraction period of the ventricle; based on the synchronous calibration pulse segment corresponding to the hardware synchronous level jump signal, the peak offset is extracted by cross-correlation function for phase shift compensation to achieve fine alignment of the time axis of the ECG and mechanical signals; the ventricular depolarization electrical start point is located in the aligned ECG signal sequence, and the time span between it and the trough start point is calculated as the electromechanical coupling delay.

[0011] Further, the step of performing dimensionless mapping and feature fusion to generate a one-dimensional fused feature vector includes: The features in the heterogeneous feature set are arranged in columns according to the cardiac cycle to construct an initial feature matrix, and the statistical mean and standard deviation of each feature column are calculated. The initial feature matrix is ​​transformed into a dimensionless standardized feature matrix with a mean of 0 and a variance of 1 using the mean and standard deviation. The standardized feature matrix is ​​flattened and arranged in ascending order of feature column index, and spliced ​​to form a one-dimensional fused feature vector.

[0012] Furthermore, the step of obtaining the weighted context vector through nonlinear high-dimensional mapping and dynamic weight allocation includes: Calculate the linear combination of the one-dimensional fused feature vector at the hidden layer nodes, and perform a modified linear unit activation operation with leakage to obtain the hidden layer output vector; calculate the unnormalized attention score vector through a single-layer feedforward attention mechanism network, and obtain the adaptive weight vector by processing it through a normalization function; calculate the Hadamard product of the adaptive weight vector and the hidden layer output vector to obtain the weighted context vector.

[0013] Furthermore, the calculation of the comprehensive risk index of the target object and the output of the risk status classification result of metabolic syndrome include: Linear dimensionality reduction is performed on the weighted context vector, and the scalar output value is mapped to an initial probability value using the Sigmoid activation function. The initial probability value is then combined with the full range to convert it into a comprehensive risk index with a value between 0 and 100. A low-risk state is determined when the comprehensive risk index is less than a first risk threshold; a medium-risk state is determined when it is greater than or equal to the first risk threshold and less than a second risk threshold; and a high-risk state is determined when it is greater than or equal to the second risk threshold.

[0014] Furthermore, the hardware architecture for acquiring the electrocardiogram signal sequence and the mechanical waveform signal sequence includes: The system includes an electrocardiogram (ECG) signal acquisition channel and a mechanical waveform signal acquisition channel. The mechanical waveform signal acquisition channel includes a microelectromechanical system (MEMS) accelerometer or a microelectromechanical system (MEMS) gyroscope sensor. The main control microprocessor is used to synchronously inject global hardware trigger pulses into the ECG signal acquisition channel and the mechanical waveform signal acquisition channel to configure a global timestamp and control the synchronization time error to be less than a preset tolerance threshold.

[0015] This invention provides a model algorithm for predicting metabolic syndrome based on cardiac deceleration force and heart rate variability. It has the following beneficial effects: 1. This invention simultaneously acquires electrocardiogram (ECG) signal sequences and mechanical waveform signal sequences, and comprehensively extracts time-frequency domain heart rate variability features, nonlinear dynamic features, myocardial deceleration force features, and underlying hemodynamic parameters such as ventricular isovolumetric contraction period and motor coupling delay. This joint feature extraction method using multi-source heterogeneous signals overcomes the limitations of single ECG signals in metabolic syndrome risk assessment. It can quantify the cardiovascular status of the target subject from two dimensions: autonomic vagal tone regulation mechanism and myocardial mechanical contractile efficacy, thereby significantly improving the prediction accuracy of the risk estimation model.

[0016] 2. This invention addresses the issues of inconsistent dimensions and large differences in physical amplitude of multi-source physiological signals by proposing a dimensionless mapping and feature vector concatenation for a heterogeneous feature set composed of heart rate variability, hemodynamic parameters, etc. Combined with a feature local sliding window truncation and alignment mechanism, it effectively eliminates the feature weight skew caused by differences in numerical scales between different modalities of physiological information. This ensures strict alignment between heterogeneous macroscopic features and beat-by-beat cardiac cycle features in both the time and matrix row dimensions, providing standardized input data support for the stable convergence of subsequent state assessment models.

[0017] 3. This invention constructs a multilayer perceptron topology and adaptive weight allocation logic within the state assessment model. It calculates the normalized weight coefficients of the output vectors of each hidden layer node through a feedforward attention mechanism network and performs weighted aggregation to obtain a weighted context vector. This network topology mechanism can dynamically quantify and allocate the contribution of different dimensions of deep features to the final metabolic syndrome risk classification result according to the individual physiological state of the target object. This achieves targeted enhancement of key pathogenic features and effective suppression of background noise interference, improving the clinical reliability of the comprehensive risk index output by the fully connected classification layer. Attached Figure Description

[0018] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a system architecture diagram of the present invention. Detailed Implementation

[0019] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only 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 are within the scope of protection of the present invention.

[0020] Please see the appendix Figure 1 -Appendix Figure 2 This invention provides a model algorithm for estimating metabolic syndrome based on cardiac deceleration force and heart rate variability.

[0021] A cardiovascular condition assessment system according to an embodiment of the present invention.

[0022] The cardiovascular status assessment system provides the hardware and software foundation for signal acquisition, data analysis, and status prediction. The system includes a signal acquisition module, a data preprocessing module, a core processing unit, a storage module, and a status output terminal. These modules interact via an internal system bus or wireless communication interface.

[0023] The signal acquisition module is used to synchronously acquire multi-source physiological signals from the target object. Lower-level features of the signal acquisition module include ECG electrode patches, photoplethysmography (PPG) pulse wave sensors, cardiac impulse-measuring piezoelectric sensors, or seismic electrocardiogram (MEMS) accelerometers. The signal acquisition module continuously acquires the target object's cardiac electrophysiological activity and vascular volume or myocardial mechanical pulsation state through physical contact or non-contact optical detection methods, and outputs raw ECG signal sequences and raw mechanical waveform signal sequences. The signal acquisition module incorporates a built-in clock synchronization generator in its hardware circuitry, synchronously injecting hardware synchronization level transition signals into the ECG signal acquisition channel and the mechanical waveform signal acquisition channel to ensure strict alignment of the multi-source physiological signals at the start time.

[0024] The data preprocessing module is physically or communicatively connected to the signal acquisition module. The data preprocessing module receives the raw ECG signal sequence and the raw mechanical waveform signal sequence, and performs baseline drift removal and high-frequency noise filtering operations on them. For the digital filter design logic of baseline drift removal and high-frequency noise filtering, those skilled in the art can implement it using a finite-length impulse response (FIR) filter or an infinite-length impulse response (IIR) filter. The underlying digital filtering algorithms of finite-length impulse response filters or infinite-length impulse response filters are well-known technologies in the field and will not be elaborated upon here.

[0025] The core processing unit receives the physiological signal sequence output by the data preprocessing module and executes feature extraction and state assessment logic. The storage module stores intermediate physiological variables, feature matrices, and network weight parameters of the state assessment model generated by the core processing unit during the calculation process. The state output terminal receives the final risk level instruction output by the core processing unit and executes the visualization rendering of the warning interface.

[0026] The macro-workflow defines the complete computational chain of the cardiovascular status assessment system from signal input to risk output. The macro-workflow includes an ordered sequence of execution from step S100 to step S900.

[0027] In step S100, the signal acquisition module acquires the electrocardiogram (ECG) signal sequence and mechanical waveform signal sequence of the target object. The signal acquisition module converts the analog physiological signals into discrete digital signal sequences through an internal analog-to-digital converter, and sends the ECG signal sequence and mechanical waveform signal sequence to the data preprocessing module.

[0028] In step S200, the data preprocessing module performs filtering and noise reduction operations on the ECG signal sequence and the mechanical waveform signal sequence. The data preprocessing module outputs the filtered and denoised ECG signal sequence and the filtered and denoised mechanical waveform signal sequence, and extracts the RR interval time series reflecting heart rate variability from the filtered and denoised ECG signal sequence.

[0029] In step S300, the core processing unit receives the RR interval time series, extracts the nonlinear dynamic characteristics of the RR interval time series, and uses the phase space reconstruction signal averaging algorithm to calculate the cardiac deceleration force characteristics and obtain the quantified vagal nerve tension index.

[0030] In step S400, the core processing unit receives the filtered and denoised mechanical waveform signal sequence. The core processing unit uses first-order and second-order difference operations to locate the time index of the trough start point and the time index of the dicrotic notch point corresponding to each cardiac cycle in the mechanical waveform signal sequence.

[0031] In step S500, the core processing unit calculates the ventricular ejection time and ventricular isovolumetric contraction period of the target object based on the time index of the trough start point, the time index of the diphthora notch point, and the feature points in the filtered and denoised ECG signal sequence.

[0032] In step S600, the core processing unit performs fine-tuned timing alignment between heterogeneous sensors based on the filtered and denoised ECG signal sequence and the filtered and denoised mechanical waveform signal sequence. The core processing unit calculates the electromechanical coupling delay of the cardiac cycle and obtains an independent evaluation index reflecting the excitation-contraction coupling efficiency of the myocardium.

[0033] In step S700, the core processing unit extracts nonlinear dynamic features, myocardial deceleration force, ventricular ejection time, ventricular isovolumetric contraction period, and electromechanical coupling delay to form a heterogeneous feature set. The core processing unit performs dimensionless mapping and dimensionality reduction flattening and splicing on the heterogeneous feature set to generate a unified one-dimensional fused feature vector on a numerical scale.

[0034] In step S800, the core processing unit inputs the one-dimensional fused feature vector into the pre-trained state evaluation model. The core processing unit performs nonlinear high-dimensional mapping through the multilayer perceptron topology inherent in the state evaluation model and obtains the weighted context vector based on adaptive weight allocation logic.

[0035] In step S900, the core processing unit inputs the weighted context vector into the fully connected classification layer to calculate the comprehensive risk index of the target object. Based on the comprehensive risk index, the core processing unit performs a risk status classification comparison and drives the status output terminal to output the final cardiovascular risk level and early warning prompt.

[0036] The execution step S100 involves controlling the multi-source signal acquisition module to simultaneously acquire the electrocardiogram signal and chest cavity physical motion signal of the subject in a resting state. Step S100 specifically includes sub-steps S110 to S130.

[0037] S110: The multi-source signal acquisition module acquires electrocardiogram (ECG) signals via surface ECG electrodes. The module attaches single-lead surface ECG electrodes to the subject's skin to measure changes in electrical signals generated during myocardial cell depolarization and repolarization. The lower limit threshold for ECG signal sampling frequency is set to 250Hz. For analog amplification and analog-to-digital conversion of the ECG signals, those skilled in the art can implement this using an integrated ECG front-end chip and supporting peripheral circuitry. The construction of such an integrated ECG front-end chip and supporting peripheral circuitry is well-known in the field and will not be elaborated upon here.

[0038] The multi-source signal acquisition module described in step S120 acquires the physical motion signal of the thoracic cavity through the microelectromechanical system sensor. Step S120 specifically includes sub-steps S121 to S122.

[0039] S121: The multi-source signal acquisition module uses a high-sensitivity accelerometer to acquire seismic echocardiogram signals. The microelectromechanical system (MEMS) sensor is attached to the surface of the subject's sternum. The seismic echocardiogram signals are used to capture micro-vibrations in the chest cavity caused by cardiac contraction, valve opening and closing, and blood flow. The high-sensitivity accelerometer measures linear acceleration, with units of [unit missing]. The detection frequency range of the high-sensitivity accelerometer is set to 1Hz to 50Hz.

[0040] S122: The multi-source signal acquisition module calls the microelectromechanical gyroscope sensor to acquire gyrocardiogram signals. The gyrocardiogram signals are used to measure the angular velocity of the thoracic cavity rotation caused by the three-dimensional rotational motion of the heart. The physical quantity measured by the microelectromechanical gyroscope sensor is angular velocity, and the unit of this physical quantity is angular velocity. The angular velocity range of the microelectromechanical gyroscope sensor is set to 0.1. Up to 20 Microelectromechanical gyroscope sensors are insensitive to linear translational motion, thereby reducing motion artifact interference.

[0041] S130: The multi-source signal acquisition module establishes a synchronized time series between the ECG signal and the chest cavity physical motion signal. The main control microprocessor inside the multi-source signal acquisition module sends a global hardware trigger pulse to the integrated ECG front-end chip and the microelectromechanical system (MEMS) sensor. The integrated ECG front-end chip and the MEMS sensor initiate sampling at the physical moment they receive the global hardware trigger pulse. The main control microprocessor configures a global timestamp for the ECG signal data and the chest cavity physical motion signal data based on its internal hardware oscillation clock. The timestamp sequence of the ECG signal is defined as follows: The timestamp sequence of the physical motion signal in the chest cavity is Synchronization time error The calculation formula is: In the formula, For sampling point index, For the first Global timestamps of each ECG signal sampling point For the first Global timestamps for each sampling point of the chest cavity physical motion signal. Controlling synchronization time error. The time tolerance threshold is less than the preset time tolerance threshold. For hardware clock synchronization of multi-sensor arrays, those skilled in the art can use existing synchronous serial communication protocol mechanisms. The underlying communication timing control is a well-known technology in the field and will not be described in detail here.

[0042] Receive the synchronous time series output by the aforementioned multi-source signal acquisition module, and execute the data preprocessing module described in step S200 to perform digital denoising and cardiac cycle time series reconstruction and extraction on the synchronously acquired electrocardiogram signal. Step S200 specifically includes sub-steps S210 to S230.

[0043] S210: The data preprocessing module applies a sliding window filter to the ECG signal to eliminate high-frequency noise and baseline drift. The data preprocessing module uses a median sliding window filtering algorithm, with the window length set to 3 to 5 sampling points. The data preprocessing module identifies the R-wave peak in the ECG signal sequence using a localization algorithm. For the R-wave peak identification and localization algorithm, those skilled in the art can implement it using existing Pan-Tompkins algorithms or wavelet transform modulus maxima methods. These existing Pan-Tompkins algorithms or wavelet transform modulus maxima methods are well-known technologies in the field and will not be elaborated upon here. The data preprocessing module calculates the difference between the timestamps of two adjacent R-wave peaks to generate an initial cardiac cycle time interval sequence. The initial cardiac cycle time interval sequence is set to include... The initial cardiac cycle time interval sequence is denoted as a set of time interval data. ,in Represents a set The Middle The time interval of one cardiac cycle.

[0044] S220: Data preprocessing module based on The criteria detect and remove outliers from the initial cardiac cycle time interval sequence. The data preprocessing module calculates the initial set of cardiac cycle time interval sequences. arithmetic mean with standard deviation .

[0045] Arithmetic mean The calculation formula is: Standard deviation The calculation formula is: The data preprocessing module will result in values ​​smaller than the arithmetic mean. Subtract three standard deviations or greater than the arithmetic mean Add three standard deviations The time interval data is determined to be abnormal. The determination logic formula is as follows: If time interval data If the conditions of the judgment logic formula are met, the data preprocessing module will process the time interval data. Mark as abnormal data.

[0046] S230: The data preprocessing module uses linear interpolation to replace and correct the marked outlier data. The module extracts the normal time intervals preceding and following each of the marked outlier data. It then calculates the arithmetic mean of these two normal time intervals. This arithmetic mean is input as the replacement value into the initial cardiac cycle time interval sequence, overwriting the marked outlier data. The new value after the outlier data replacement is denoted as... The replacement calculation formula is: In the formula, Time interval data The normal time interval of the former, Time interval data The normal time interval of the latter item. The data preprocessing module outputs the corrected cardiac cycle time interval sequence. The corrected cardiac cycle time interval sequence serves as the basic input data for electrophysiological feature extraction.

[0047] The corrected cardiac cycle time interval sequence (i.e., the preprocessed RR interval time sequence) output by the data preprocessing module is used by the feature engineering extraction module to extract multidimensional feature parameters based on this sequence. Step S300 specifically includes time domain feature extraction sub-step S310, frequency domain feature extraction sub-step S320, nonlinear dynamics feature extraction sub-step S330, and cardiac deceleration force feature extraction sub-step S340.

[0048] In step S310, the feature engineering extraction module calculates the time-domain heart rate variability features. Step S310 further includes sub-steps S311 and S312.

[0049] In step S311, the feature engineering extraction module calculates the standard deviation of the mean of the corrected cardiac cycle time intervals and the standard deviation of all normal sinus beat time intervals. The total number of data points in the corrected cardiac cycle time interval sequence is set to... , No. The corrected time interval data is denoted as The feature engineering extraction module calculates the arithmetic mean of the corrected cardiac cycle time interval sequence. The feature engineering extraction module calculates the standard deviation of all normal sinus beat intervals. Standard deviation The calculation formula is as follows: in, This represents the total number of cardiac cycle time interval data. For the first The corrected time interval data, This is the arithmetic mean of the corrected cardiac cycle time interval sequence.

[0050] In step S312, the feature engineering extraction module calculates the root mean square of the differences between adjacent time intervals and the percentage of absolute values ​​of adjacent differences that are greater than a preset time difference threshold. The feature engineering extraction module calculates the root mean square of the differences between adjacent time intervals. Root mean square The calculation formula is as follows: The feature engineering extraction module calculates the percentage of adjacent time interval differences with an absolute value greater than 50ms. The preset time difference threshold is fixed at 50ms. (Percentage) The calculation formula is as follows: in, This represents the number of differences in the corrected cardiac cycle time interval sequence that satisfy the absolute value condition. This represents the total number of differences between adjacent time intervals.

[0051] In step S320, the feature engineering extraction module calculates the frequency domain heart rate variability features. Step S320 further includes sub-steps S321 to S323.

[0052] In step S321, the feature engineering extraction module transforms the corrected cardiac cycle time interval sequence into an isochronous interval sequence. The feature engineering extraction module constructs a non-uniform time series using the cumulative cardiac cycle time as the horizontal axis time index and the corrected time interval data as the vertical axis amplitude. The feature engineering extraction module resamples the non-uniform time series using a linear interpolation algorithm, with the resampling frequency set to 4Hz. The feature engineering extraction module outputs the isochronous interval sequence.

[0053] In step S322, the feature engineering extraction module performs spectral estimation and calculates the power spectral density function on the equal-time-interval sequence. The feature engineering extraction module performs a Fast Fourier Transform (FFT) on the equal-time-interval sequence to obtain the complex spectrum. For the FFT operation of the equal-time-interval sequence, those skilled in the art can implement it using the radix-2 decimation-time FFT algorithm. Existing radix-2 decimation-time FFT algorithms are well-known in the field and will not be elaborated upon here. The feature engineering extraction module calculates the square of the modulus of the complex spectrum, divides it by the sampling frequency, and obtains the power spectral density function. .

[0054] In step S323, the feature engineering extraction module calculates the energy-to-frequency-power ratio of each preset frequency band based on the power spectral density function. The preset frequency bands include extremely low frequency, low frequency, and high frequency bands. The frequency range of the extremely low frequency band is set to 0.0033Hz to 0.04Hz. The frequency range of the low frequency band is set to 0.04Hz to 0.15Hz. The frequency range of the high frequency band is set to 0.15Hz to 0.4Hz. The feature engineering extraction module integrates the power spectral density function within the corresponding frequency band range to calculate the total energy of the corresponding frequency band. The high frequency band power... For example, high-frequency power The calculation formula is as follows: in, For frequency parameters, Let be the power spectral density function. The lower limit of integration in the high-frequency band is 0.15Hz. The upper limit of the high-frequency integration is 0.4Hz. Similarly, the feature engineering extraction module calculates the low-frequency power. Low-frequency power The integration interval is from the lower limit of integration frequency (0.04Hz) to the upper limit of integration frequency (0.15Hz) in the low-frequency band. The feature engineering extraction module calculates the low-frequency power. With high-frequency power The ratio of low frequency to high frequency ratio is used to generate characteristic parameters of low frequency to high frequency ratio.

[0055] In step S330, the feature engineering extraction module extracts the nonlinear dynamic features of the RR interval time series. Step S330 further includes a Poincaré scatter plot feature extraction sub-step S331 and a sample entropy feature extraction sub-step S332.

[0056] In step S331, the feature engineering extraction module calculates the spatial distribution parameters of the Poincaré scatter plot. The feature engineering extraction module constructs a two-dimensional delayed reconstruction scatter plot from the RR interval sequence, setting the horizontal axis to the first... The RR interval is denoted as . The vertical axis is the first The RR interval is denoted as . ,in The range of values ​​is to Positive integers. Based on the two-dimensional delayed reconstruction scatter plot, the feature engineering extraction module calculates values ​​perpendicular to... Standard deviation of the minor axis of the identity lines And parallel to Standard deviation of the major axis of the identity lines The specific calculation formula is as follows: in, This represents the statistical variance of the calculated sequence.

[0057] In step S332, the feature engineering extraction module calculates the sample entropy of the RR interval sequence. For the phase space reconstruction, template vector distance calculation, and logarithmic ratio solution of the sample entropy, those skilled in the art can consult relevant nonlinear time series analysis literature. The basic algorithm implementation of sample entropy is a well-known technique in this field and will not be elaborated here. The feature engineering extraction module outputs the calculated sample entropy value as a quantitative indicator characterizing the complexity of heart rate variability.

[0058] In step S340, the feature engineering extraction module uses a phase space reconstruction signal averaging algorithm to process the RR interval sequence to obtain cardiac deceleration force features. Step S340 includes multiple sub-steps from S341 to S344.

[0059] In step S341, the feature engineering extraction module determines anchor points in the RR interval sequence. The feature engineering extraction module traverses the entire RR interval sequence, identifying points that satisfy the heart rate deceleration condition. When two adjacent RR intervals satisfy the relationship... Time (of which) For traversal index, The range of values ​​is to The feature engineering extraction module will extract the location index. This is marked as an anchor point location. Assume that a common feature is identified in the RR interval series. The feature engineering extraction module will use the anchor point as the first anchor point. The position index of each anchor point in the RR interval sequence is denoted as . ,in For anchor point number, .

[0060] In step S342, the feature engineering extraction module extracts signal segments based on the determined anchor point positions. The feature engineering extraction module extracts signal segments based on each anchor point position. Extract forward from the center. Extracting data points backwards Data points, forming a length of The data window. Let the captured data be the first... The segment is The feature engineering extraction module will extract the first... The first segment Each data point is defined as: in, The set window half-length constant, This is the relative position index within the data window, with a value range of [value range missing]. During the extraction process, if the anchor point position... satisfy or The feature engineering extraction module then removes anchor point locations. The corresponding anchor points will not be subject to signal segmentation.

[0061] In step S343, the feature engineering extraction module performs signal alignment and averaging. The feature engineering extraction module then processes all the extracted segments... Phase alignment is performed with the anchor point as the center, and the relative position indices are calculated. The arithmetic mean of the data points is used to generate the reconstructed average signal sequence. : in, This represents the total number of effective anchor points actually used for segment extraction after removing boundary anchor points. In the reconstructed average signal sequence, This represents the average value of all valid anchor point location data.

[0062] In step S344, the feature engineering extraction module calculates the cardiac deceleration force based on the reconstructed average signal sequence. The specific numerical value. The feature engineering extraction module uses the data of the center point and neighboring points of the reconstructed average signal sequence to perform difference calculation, the formula is: in, The average RR interval for the effective anchor point location, The average RR interval at the location following the effective anchor point; and These are the average RR intervals at the positions preceding and two preceding the effective anchor point, respectively. The feature engineering extraction module outputs the cardiac deceleration force. The specific numerical value is used as a quantitative feature.

[0063] In the previous steps, the processing unit has acquired a filtered and denoised mechanical waveform signal sequence. In step S400, the processing unit receives the filtered and denoised mechanical waveform signal sequence and performs automated localization of key physiological markers on it. The filtered and denoised mechanical waveform signal sequence includes any one of photoplethysmography (PPG) signals, impact graph (IPG) signals, or seismic electrocardiogram (ECG) signals. The automated localization process of key physiological markers includes a signal derivative calculation sub-step S410, a main peak identification sub-step S420, and an auxiliary marker extraction sub-step S430.

[0064] In step S410, the processing unit calculates the first and second derivative sequences of the mechanical waveform signal sequence. Let the discretized mechanical waveform signal sequence be... ,in For sampling time index, This represents the total number of sampling points in the mechanical waveform signal sequence. The processing unit obtains the first-order derivative sequence through first-order difference operations. The specific calculation formula is as follows: Among them, the first derivative sequence is calculated. Time index The range of values ​​is to Positive integers. The processing unit processes the sequence of first derivatives. Perform the same difference operation again to obtain the second derivative sequence. Calculate the second derivative sequence Time index The range of values ​​is to Positive integers. First derivative sequence. Characterizing the instantaneous rate of change of a mechanical waveform signal sequence, the second derivative sequence Characterizes the acceleration changes of a mechanical waveform signal sequence.

[0065] In step S420, the processing unit is based on the first derivative sequence Identify the main peak of the mechanical waveform signal sequence. Step S420 further includes a zero-crossing detection sub-step S421 and a peak confirmation sub-step S422.

[0066] In step S421, the processing unit processes the first derivative sequence. Find the zero-crossing point where the value changes from positive to negative. (When the time index...) satisfy and At that time, the processing unit will index the time. Recorded as candidate peak points.

[0067] In step S422, the processing unit performs amplitude threshold filtering on the candidate peak points. The processing unit calculates the average amplitude of the mechanical waveform signal sequence within a global time window as the dynamic threshold. For any candidate peak point If candidate peak points The corresponding original signal amplitude satisfies The processing unit will index the time. The main peak point has been identified. The processing unit will then... The time index corresponding to the main peak of each cardiac cycle is denoted as . ,in This is the cardiac cycle number.

[0068] In step S430, the processing unit is based on the main wave peak point First derivative sequence and second derivative sequence Extract the trough start point and diphtheria notch point of the mechanical waveform signal sequence. Step S430 further includes a trough start point location sub-step S431 and a notch point location sub-step S432.

[0069] In step S431, the processing unit locates the starting point of the trough in the mechanical waveform signal sequence. The processing unit also locates the main peak of the current cardiac cycle. The previously set time range Inside, search for the second derivative sequence The local maximum point, where The number of data points in the preset search window is a constant. The processing unit determines the time index corresponding to the local maximum point of the second derivative as the [number]. The beginning of the trough of a cardiac cycle Trough starting point This represents the initial response time of the mechanical pulsation caused by ventricular ejection at the measurement site.

[0070] In step S432, the processing unit locates the diphtheria notch of the mechanical waveform signal sequence. The processing unit locates the main peak of the current cardiac cycle. The beginning of the trough of the next cardiac cycle Search for the first derivative sequence within the time interval between these two points. The local minimum of the first derivative. The processing unit determines the time index corresponding to the local minimum of the first derivative as the [number]. Dipbeat notch point in one cardiac cycle Diptacardial wave notch This represents the physiological characteristic moment when the aortic valve closes. For conventional search and matching algorithms for local extrema and zero-crossing points, those skilled in the art can implement them using sliding window traversal or binary search algorithms. The specific code judgment logic of sliding window traversal or binary search algorithms is well-known in the field and will not be elaborated here.

[0071] In step S440, the processing unit extracts the main wave peak point trough starting point and diabetic wave notch The time index is output to the storage module. The processing unit will output the main peak point. trough starting point and diabetic wave notch As a basic characteristic variable, it is used for subsequent calculation of cardiovascular hemodynamic parameters.

[0072] In the previous steps, the processing unit had already acquired the mechanical waveform signal sequence at the [number]th [step]. Index of the start point time of the trough in each cardiac cycle Time index of diphtheria wave notch Simultaneously, the ECG signal feature points corresponding to the cardiac cycle were acquired. In step S500, the processing unit uses the time index of the trough start point of the extracted mechanical waveform signal sequence. Diptacardiac wave notch time index The calculation of ventricular ejection time and isovolumetric contraction period of the target object is performed using ECG signal feature points. The calculation process of ventricular ejection time and isovolumetric contraction period includes ventricular ejection time calculation sub-step S510 and isovolumetric contraction period calculation sub-step S520.

[0073] In step S510, the processing unit calculates the ventricular ejection time corresponding to the mechanical waveform signal sequence. Step S510 further includes an interval calculation sub-step S511 and a time conversion sub-step S512.

[0074] In step S511, the processing unit extracts the time index of the trough start point of the current cardiac cycle. and diphthous pulsation wave notch time index The processing unit obtains the length of the discrete sampling point interval corresponding to ventricular ejection through differential operations. (Trough start point time index) Time index of diphtheria wave notch corresponding to the physiological characteristic points of aortic valve opening The length of the discrete sampling point interval corresponding to the aortic valve closure and the ventricular ejection represents the data point span of the continuous phase of cardiac ejection.

[0075] In step S512, the processing unit converts the length of the discrete sampling point interval corresponding to ventricular ejection into an absolute time dimension. The uniform sampling frequency of the data acquisition system is set to... The processing unit calculates the first Ventricular ejection time per cardiac cycle The calculation formula is: Among them, unified sampling frequency The unit is Hertz, ventricular ejection time. The unit is seconds. The processing unit will calculate the ventricular ejection time. Output and store in a register.

[0076] In step S520, the processing unit calculates the isovolumetric contraction period of the ventricle. Step S520 further includes a pre-ejection time calculation sub-step S521 and an electromechanical delay deduction sub-step S522.

[0077] In step S521, the processing unit acquires the electrocardiogram signal that is synchronously acquired with the mechanical waveform signal in the [missing information]. Time index corresponding to the peak R-wave point within each cardiac cycle To ensure consistency in time scale, the ECG signal and mechanical waveform signal are synchronized by the underlying hardware clock and aligned by software resampling before being input to the processing unit, so that the ECG signal and mechanical waveform signal have the same unified sampling frequency. The processing unit calculates the time before cardiac ejection. Processing unit calculation time index Time index of the trough start point The span between, and combined with a unified sampling frequency Convert to absolute time to obtain the first... Time before cardiac ejection in one cardiac cycle The calculation formula is: Among them, time before cardiac ejection The unit is seconds.

[0078] In step S522, the processing unit calculates the time before the heart ejects blood. Isolation of the isovolumetric contraction phase of the ventricle Time before the heart ejects blood From the electromechanical delay time constant Isovolumetric contraction of the ventricle It consists of two parts. The processing unit retrieves the electromechanical delay time constant pre-stored in the device memory. Calculate the first The isovolumetric contraction period of the ventricle in one cardiac cycle The calculation formula is: For electromechanical delay time constant The individualized calibration measurement process can be obtained by those skilled in the art using the simultaneous echocardiography measurement method. The specific operating procedures of the simultaneous echocardiography measurement method are well-known in the art and will not be elaborated here.

[0079] In step S530, the processing unit summarizes the first Ventricular ejection time per cardiac cycle Isovolumetric contraction of the ventricle This generates a feature set representing the state of myocardial contractility. The processing unit then sends this feature set to the subsequent hemodynamic model for parameter inversion and state classification.

[0080] In the previous steps, the processing unit had completed the extraction of physiological marker points from the mechanical waveform signal sequence and obtained the time index of the trough start point. In step S600, the processing unit acquires the synchronously acquired electrocardiogram (ECG) signal sequence and mechanical waveform signal sequence, and performs timing alignment and electromechanical coupling delay calculation based on the ECG signal sequence and mechanical waveform signal sequence. The timing alignment and electromechanical coupling delay calculation process includes a multi-source signal timing alignment sub-step S610 and an electromechanical coupling delay parameter calculation sub-step S620.

[0081] In step S610, the processing unit performs time axis synchronization alignment on the input ECG signal sequence and the mechanical waveform signal sequence. Step S610 further includes a hardware timestamp coarse alignment sub-step S611 and a phase fine alignment sub-step S612 based on cross-correlation.

[0082] In step S611, the processing unit extracts the start timestamp of the electrocardiogram signal sequence. and the start timestamp of the mechanical waveform signal sequence The processing unit calculates the difference between the start timestamp of the ECG sequence and the start timestamp of the mechanical sequence. The specific calculation formula is as follows: The processing unit is based on the difference. The mechanical waveform signal sequence is shifted along the entire time axis to achieve coarse alignment of the data sequences of the electrocardiogram signal sequence and the mechanical waveform signal sequence.

[0083] In step S612, the processing unit eliminates the sampling-level phase shift between heterogeneous sensors. During the hardware data acquisition phase, the acquisition system pre-injects hardware synchronization level transition signals into both the ECG sensor and the mechanical waveform sensor. The processing unit extracts synchronization calibration pulse segments corresponding to the hardware synchronization level transition signals from the coarsely aligned ECG signal sequence and mechanical waveform signal sequence. For the calculation of the cross-correlation function and peak offset extraction process of the synchronization calibration pulse segments, those skilled in the art can consult relevant literature on digital signal processing. The specific operating procedures of the sliding window cross-correlation algorithm are well-known in the field and will not be elaborated here. The processing unit performs phase shift compensation on the mechanical waveform signal sequence based on the extracted peak offset to obtain the ECG signal sequence and mechanical waveform signal sequence under the same global time index coordinate system. The processing unit sets the unified sampling frequency under the same global time index coordinate system as... .

[0084] In step S620, the processing unit calculates the electro-mechanical coupling delay of the cardiac cycle based on the aligned ECG signal sequence and the mechanical waveform signal sequence. Step S620 further includes an electrical origin extraction sub-step S621 and a delay calculation sub-step S622.

[0085] In step S621, the processing unit locates the first [electrocardiogram signal] in the aligned electrocardiogram signal sequence. The ventricular depolarization electrical initiation point of each cardiac cycle. The processing unit identifies the start point of the QRS complex in the ECG signal sequence and determines the global time index of the QRS complex initiation point as the electrical initiation point time index. ,in This refers to the cardiac cycle sequence number. Electrical start-time index. This represents the initial moment of electrical excitation of ventricular myocytes.

[0086] In step S622, the processing unit calculates the time span consumed by the ventricular myocardium from electrical excitation to the generation of substantial mechanical contraction. The processing unit retrieves the time elapsed in the previous step... Index of the start point time of the trough in each cardiac cycle Trough Start Point Time Index This represents the initial response time of the mechanical pulsation caused by ventricular ejection at the measurement site. The processing unit obtains the first response time through differential operations. Electromechanical coupling delay per cardiac cycle The calculation formula is: Among them, the electromechanical coupling delay The unit is seconds. The processing unit will calculate the electromechanical coupling delay. As an independent evaluation index reflecting the efficiency of myocardial excitation-contraction coupling, and considering the electromechanical coupling delay... Stored in a local register for subsequent hemodynamic state classification modeling.

[0087] In previous steps, the processing unit has acquired multiple sets of nonlinear dynamic features, myocardial deceleration force, ventricular ejection time, ventricular isovolumetric contraction period, and electromechanical coupling delay. In step S700, the processing unit extracts the nonlinear dynamic features, myocardial deceleration force, ventricular ejection time, ventricular isovolumetric contraction period, and electromechanical coupling delay to form a heterogeneous feature set, and performs dimensionless processing and vector concatenation on the heterogeneous feature set. The heterogeneous feature set contains physiological parameters with different dimensions and physical amplitudes. The dimensionless processing and vector concatenation process includes a feature matrix construction sub-step S710, a dimensionless mapping sub-step S720, and a high-dimensional vector fusion sub-step S730.

[0088] In step S710, the processing unit constructs an initial feature matrix based on the heterogeneous feature set. The processing unit then arranges the features of multiple cardiac cycles extracted within a set observation time window in a column-wise order. Let the observation time window contain... The processing unit will process the first effective cardiac cycle. The nonlinear short-axis standard deviation corresponding to each cardiac cycle Nonlinear major axis standard deviation Sample entropy Cardiac deceleration force Ventricular ejection time Isovolumetric contraction of the ventricle and electromechanical coupling delay Arranged sequentially, the construction dimension is initial characteristic matrix .in, This is the index number of the effective cardiac cycle within the observation time window, with a value range of [value range missing]. to Positive integers. For macroscopic features requiring calculation from sequence data, such as nonlinear minor axis standard deviation, nonlinear major axis standard deviation, sample entropy, and cardiac deceleration force, the processing unit uses the first... Using the first cardiac cycle as a baseline, a fixed-length segment of the historical cardiac cycle sequence is extracted as a local sliding window to calculate the value corresponding to the first cardiac cycle. Nonlinear short-axis standard deviation of one cardiac cycle Nonlinear major axis standard deviation Sample entropy and cardiac deceleration force This achieves alignment between macroscopic features and beat-by-beat cardiac cycle features in both the time dimension and the matrix row dimension.

[0089] In step S720, the processing unit processes the initial feature matrix. Perform a dimensionless mapping. Step S720 further includes a statistical parameter calculation sub-step S721 and a standardization transformation sub-step S722.

[0090] In step S721, the processing unit calculates the initial feature matrix. The statistical mean and standard deviation of each feature column in the matrix. Let the initial feature matrix be... The The column feature vector is ,in This is the index for the feature column, and its value range is... to The processing unit calculates the positive integer. Column feature vector mean with standard deviation The calculation formula is: in, Represents the initial characteristic matrix The Middle The first column of the eigenvector Each element.

[0091] In step S722, the processing unit utilizes the mean. with standard deviation For the initial characteristic matrix Z-score normalization is performed. The processing unit initializes the feature matrix. Transform into a dimensionless standardized feature matrix Standardized feature matrix The Middle The first column of the eigenvector element The calculation formula is: After Z-score standardization, the standardized feature matrix Various physiological characteristic data are uniformly mapped to a unified numerical scale with a mean of 0 and a variance of 1, eliminating the numerical scale differences caused by different units of measurement in the original heterogeneous feature set. For other types of dimensionless mapping algorithms, those skilled in the art can use the Min-Max normalization algorithm as a replacement. The specific calculation logic of the Min-Max normalization algorithm is well-known in the field and will not be elaborated here.

[0092] In step S730, the processing unit normalizes the feature matrix. Perform vector concatenation operation. Step S730 further includes the dimensionality reduction and flattening sub-step S731.

[0093] In step S731, the processing unit processes the two-dimensional normalized feature matrix. Flattened into a one-dimensional fused feature vector The processing unit is indexed according to the feature column. The increasing priority order will standardize the feature matrix. The elements in the array are arranged and combined sequentially to form a structure with a total length of [length missing]. One-dimensional fusion feature vector One-dimensional fused feature vector As a global state representation incorporating multimodal physiological information, it integrates comprehensive indices of heart rate variability, vagal tone, and electro-mechanical coupling hemodynamics extracted from electrocardiogram and mechanical waveform signal sequences. The processing unit fuses the one-dimensional feature vector. Output and write to device memory, one-dimensional fused feature vector Used as standardized input data for subsequent machine learning classifier models or regression analysis models.

[0094] In the previous steps, the processing unit has obtained a one-dimensional fused feature vector formed by dimensionless and concatenated processing of multimodal physiological features. In step S800, the processing unit fuses the one-dimensional feature vector. The data is input into a pre-built state assessment model, and the cardiovascular state assessment results of the target object are output through the intrinsic network topology and adaptive weight allocation logic of the state assessment model. The processing of the intrinsic network topology and adaptive weight allocation logic of the state assessment model includes a high-dimensional feature mapping sub-step S810 and an adaptive weight allocation and feature aggregation sub-step S820.

[0095] In step S810, the processing unit processes the one-dimensional fused feature vector through the multilayer perceptron topology inherent in the state evaluation model. A nonlinear high-dimensional mapping is performed. The multilayer perceptron topology is used to transform the original feature space to a higher-dimensional latent feature space. Step S810 further includes a linear transformation sub-step S811 and an activation function mapping sub-step S812.

[0096] In step S811, the processing unit calculates the one-dimensional fused feature vector. A linear combination of hidden layer nodes. Let there be a one-dimensional fused feature vector. For column vectors, one-dimensional fused feature vectors The vector dimension is Set the number of nodes in the hidden layer to The processing unit retrieves the weight matrix stored in the model parameter memory. and bias vector Calculate the linear output vector The specific calculation formula is as follows: Among them, the weight matrix The dimension is Bias vector The dimension is Linear output vector The dimension is .

[0097] In step S812, the processing unit processes the linear output vector. Perform nonlinear activation operations. The processing unit uses a leaky modified linear unit (LeakyReLU) as the activation function to obtain the hidden layer output vector. The processing unit will process the hidden layer output vector. The first in element The calculation process is defined as follows: in, For linear output vectors The One element, The default non-negative leakage coefficient constant is used. This is the index of the hidden layer node, and its value range is... to Positive integers. Hidden layer output vector. This constitutes the deep feature representation within the state evaluation model. For the parameter update process of backpropagation and gradient descent in a multilayer perceptron, those skilled in the art can consult relevant deep learning literature. The basic network training logic of the parameter update process of backpropagation and gradient descent in a multilayer perceptron is a well-known technique in this field and will not be elaborated upon here.

[0098] In step S820, the processing unit outputs a vector to the hidden layer. Perform adaptive weight allocation. Adaptive weight allocation is used to quantify the contribution of deep features of different dimensions to the final classification result, strengthen the expression of key physiological features, and suppress the interference of noisy features. Step S820 further includes attention score calculation sub-step S821 and context vector fusion sub-step S822.

[0099] In step S821, the processing unit calculates the hidden layer output vector. The adaptive weight coefficients are determined. The processing unit constructs a single-layer feedforward attention mechanism network and retrieves the attention weight matrix. and attention bias vector Calculate the unnormalized attention score vector The calculation formula is: Among them, the attention weight matrix The dimension is Attention bias vector The dimension is Unnormalized attention score vector The dimension is The processing unit applies the Softmax function to the unnormalized attention score vector. Perform normalization to obtain the adaptive weight vector. Adaptive weight vector The Middle The weight coefficient of each element The calculation formula is: in, Unnormalized attention score vector The One element, Unnormalized attention score vector The traversing elements, and All are index values, and their range is [value range missing]. to Positive integers. Weighting coefficients. Characterized the first The relative importance of each deep feature dimension for assessing the current cardiovascular status.

[0100] In step S822, the processing unit is based on the adaptive weight vector Output vector of hidden layer Perform weighted aggregation. The processing unit calculates the adaptive weight vector. With hidden layer output vector The Hadamard product is used to obtain the weighted context vector. The calculation formula is: in, This represents the operation of multiplying corresponding elements one by one, using a weighted context vector. The dimension and hidden layer output vector The dimensions remain consistent. The processing unit will weight the context vector. The input is fed into a fully connected output layer, and the classifier function outputs the predicted probability values ​​of the target object corresponding to each state category.

[0101] In practical applications, the adaptive weight allocation logic is not limited to the attention mechanism described above; it can also be implemented using a weight scaling module based on a gating mechanism. Specifically, the processing unit calculates a scaling factor between 0 and 1 based on the input features themselves through a gating mechanism and applies it to the corresponding feature channels for dynamic weighting. Both the attention mechanism and the gating mechanism can achieve the technical effect of strengthening effective features and suppressing ineffective interference, and they are equivalent implementations in this field.

[0102] In the previous steps, the processing unit has obtained the weighted context vector that has been weighted and aggregated by the attention mechanism. In step S900, the processing unit receives the weighted context vector. And based on weighted context vectors Perform the mapping and state output of the comprehensive risk index. The mapping and state output process of the comprehensive risk index includes the fully connected mapping sub-step S910, the index normalization sub-step S920, and the risk classification output sub-step S930.

[0103] In step S910, the processing unit weights the context vector. The input is fed into the fully connected classification layer to calculate the initial probability value of the target object being in a cardiovascular abnormal state. The fully connected mapping sub-step S910 further includes a linear dimensionality reduction sub-step S911 and a probability mapping sub-step S912.

[0104] In step S911, the processing unit processes the weighted context vector. Perform linear dimensionality reduction. Let the weighted context vector be... For column vectors, weighted context vectors The dimension is The processing unit retrieves the weight matrix of the fully connected classification layer. and bias scalar Calculate scalar output value The calculation formula is: Among them, the weight matrix The dimension is Bias scalar The dimension is .

[0105] In step S912, the processing unit outputs the scalar value. Mapped to initial probability values The processing unit uses the Sigmoid activation function to process scalar output values. The calculation formula is: Initial probability value The range of values ​​is to Initial probability value The statistical tendency to cause adverse cardiovascular events in the target population.

[0106] In step S920, the processing unit sets the initial probability value Converted into a comprehensive risk index that aligns with clinical reading habits. The processing unit sets the full-scale range of the comprehensive risk index to be... By using the initial probability value Multiply Calculate the comprehensive risk index Overall Risk Index For the range of values ​​within to The processing unit will integrate the risk index with the real numbers between these values. Write to the specified register address in the device memory.

[0107] In step S930, the processing unit bases the comprehensive risk index on... The risk status is classified and the results are output. The risk classification output sub-step S930 further includes the threshold comparison sub-step S931 and the early warning generation sub-step S932.

[0108] In step S931, the processing unit will integrate the risk index. The data is compared numerically with a preset set of thresholds. This preset set of thresholds is determined in advance based on the statistical distribution characteristics of historical cardiovascular clinical sample data. The processing unit retrieves a first risk threshold stored in the device's memory. Second risk threshold The first risk threshold Less than the second risk threshold When the comprehensive risk index Less than the first risk threshold When the comprehensive risk index is high, the processing unit determines that the target object is in a low-risk state; when the comprehensive risk index is low, the processing unit determines that the target object is in a low-risk state. Greater than or equal to the first risk threshold And less than the second risk threshold When the processing unit determines that the target object is in a medium-risk state, the comprehensive risk index... Greater than or equal to the second risk threshold At that time, the processing unit determines that the target object is in a high-risk state.

[0109] In step S932, the processing unit generates a corresponding output instruction based on the determined risk status of the target object. The processing unit will then integrate the risk index... The cardiovascular risk status level determined and the various physiological characteristic parameters extracted in the previous steps are spliced ​​together to form a standardized data frame, which is then sent to a display terminal for visualization. For the communication transmission encapsulation and parsing extraction process of the standardized data frame, those skilled in the art can use TCP / IP or Bluetooth BLE protocols for code implementation. The underlying communication interaction process of TCP / IP or Bluetooth BLE protocols is well-known in the field and will not be elaborated here.

[0110] Besides using a fully connected classification layer combined with a sigmoid activation function to map the comprehensive risk index, those skilled in the art can also use a support vector machine (SVM) algorithm as a replacement. Specifically, the processing unit calculates the geometric distance between the high-dimensional feature vector and the classification boundary using an SVM algorithm, and then converts the calculated geometric distance into a probability index reflecting the degree of risk. Both fully connected classification layer mapping and SVM mapping can achieve the technical function of converting high-dimensional features into target risk probabilities; they are equivalent alternative implementation methods in this field.

[0111] This invention provides an electronic device for providing a low-level hardware operating environment for executing the cardiovascular status assessment logic disclosed in the foregoing embodiments. The electronic device includes a processor, a memory, and a communication bus.

[0112] The processor and memory are physically connected and interact with each other via a communication bus. The communication bus is used to transmit address signals, data signals, and control signals between the processor and memory of an electronic device. Communication buses include industry-standard architecture buses, peripheral component interconnection standard buses, or extended industry-standard architecture buses. For the underlying data transmission control logic of the communication bus, those skilled in the art can use existing bus arbitration protocols for implementation. The underlying data transmission control logic of the communication bus is well-known in the art and will not be described in detail here.

[0113] The memory is used to store computer-executable instructions and physiological characteristic data generated during runtime. The memory includes volatile memory and non-volatile memory. Lower-level implementations of volatile memory include random access memory or static random access memory, used to provide data read / write cache for the processor. Lower-level implementations of non-volatile memory include read-only memory, erasable programmable read-only memory, electrically erasable programmable read-only memory, or solid-state drive. The memory pre-stores the operating system environment, data processing applications, and various weight parameters of the cardiovascular status assessment model. Specifically, the memory stores a computer program; when the processor executes the computer program, it implements the steps included in the aforementioned embodiments: nonlinear feature and cardiac deceleration force extraction steps, mechanical waveform key physiological marker point localization steps, ventricular ejection time and isovolumetric contraction period calculation steps, timing alignment and electromechanical coupling delay calculation steps, heterogeneous feature dimensionless processing and vector concatenation steps, model mapping calculation steps, and comprehensive risk index mapping and output steps.

[0114] A processor is the core component of an electronic device, responsible for computation and control. It executes specific logical operations based on computer-executable instructions stored in memory. The underlying features of a processor are not limited to general-purpose central processing units (CPUs); in specific low-level circuit designs, they can be equivalently replaced by digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or graphics processing units (GPUs). General-purpose CPUs and DSPs or FPGAs are completely interchangeable in their technical function of parsing instructions and driving hardware circuits to perform arithmetic and logical operations. The aforementioned description of equivalent substitution of hardware components provides clear technical support for the generalization of processors in the claims, avoiding limiting the claims to a single hardware chip form.

[0115] The specific hardware interaction process for electronic devices to execute various computational logics includes the data interaction stage and the instruction decoding and execution stage.

[0116] During the data interaction phase, the processor sends an operation request to the memory via the communication bus to obtain the underlying data required for performing the aforementioned cardiac deceleration force calculation, isochoric contraction period extraction, or multilayer perceptron mapping operations. Specifically, the processor calculates the physical address of the target data in memory and places the calculated physical address on the address bus channel of the communication bus. The physical address value issued by the processor is within the effective range of the maximum physical addressing space limited by the hardware bit width of the address bus of the communication bus. Subsequently, the memory controller inside the memory receives the physical address on the address bus channel and extracts the binary target data from the corresponding memory cell. The memory controller places the binary target data on the data bus channel of the communication bus and sends it back to the processor's internal general-purpose register.

[0117] During the instruction decoding and execution phase, the processor performs specific operations on the binary target data returned to the general-purpose registers. The processor's internal instruction decoder parses the binary target instruction and drives the internal arithmetic logic unit (ALU) to execute the corresponding hardware circuit operations. The ALU calls the underlying floating-point adder and multiplier arrays to perform the aforementioned differential calculations, activation function mappings, and attention weight normalization operations on the ECG signal time index, mechanical waveform amplitude values, or model weight matrices stored in the general-purpose registers. After the operations are completed, the processor temporarily stores the final result output by the ALU in a buffer register. The processor then rewrites the final result from the buffer register into the pre-allocated static storage area in memory via the communication bus, thus completing a closed loop of the underlying hardware execution for a full operation cycle.

[0118] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A model algorithm for estimating metabolic syndrome based on cardiac deceleration force and heart rate variability, characterized in that, Includes the following steps: Obtain the electrocardiogram signal sequence and mechanical waveform signal sequence of the target object, and extract the corresponding RR interval time series; Heart rate variability and cardiac deceleration force features were extracted based on the RR interval time series. Based on the electrocardiogram signal sequence and mechanical waveform signal sequence, key physiological markers are located, and hemodynamic features reflecting the mechanical contractile efficacy of the myocardium are extracted. The hemodynamic features include ventricular ejection time, ventricular isovolumetric contraction period, and electro-mechanical coupling delay. The heart rate variability feature, the cardiac deceleration force feature and the hemodynamic feature are combined into a heterogeneous feature set, and dimensionless mapping and feature fusion are performed to generate a one-dimensional fused feature vector. The one-dimensional fused feature vector is input into the state evaluation model, and a weighted context vector is obtained through nonlinear high-dimensional mapping and dynamic weight allocation; The comprehensive risk index of the target object is calculated based on the weighted context vector, and the risk status classification result of metabolic syndrome is output.

2. The model algorithm for estimating metabolic syndrome based on cardiac deceleration force and heart rate variability according to claim 1, characterized in that, The process of acquiring the electrocardiogram signal sequence and mechanical waveform signal sequence of the target object, and extracting the corresponding RR interval time series, includes: A median sliding window filter was used to eliminate high-frequency noise and baseline drift in the ECG signal sequence and mechanical waveform signal sequence; The difference between adjacent R-wave peak timestamps is calculated to generate an initial cardiac cycle time interval sequence, and its arithmetic mean and standard deviation are calculated. Outlier data is identified and marked based on the 3σ criterion. The arithmetic mean of the normal time intervals before and after the outlier data is extracted as replacement values ​​to cover the outlier data, thus obtaining the corrected RR interval time series.

3. The model algorithm for estimating metabolic syndrome based on cardiac deceleration force and heart rate variability according to claim 1, characterized in that, The extraction of heart rate variability features and cardiac deceleration force features includes: The standard deviation, root mean square of adjacent differences, and percentage of absolute values ​​of adjacent differences in the RR interval time series are calculated as time-domain heart rate variability features. Perform a fast Fourier transform on the time-interval sequence to obtain the power spectral density function, and calculate the ratio of the total energy in the low-frequency band to the total energy in the high-frequency band as a characteristic of the frequency domain heart rate variability. The RR interval time series is constructed into a two-dimensional delayed reconstruction scatter plot, and the short-axis standard deviation and long-axis standard deviation are calculated as nonlinear dynamic features. The position index that satisfies the heart rate deceleration condition in the RR interval time series is identified as the anchor point. The signal segment is extracted with the anchor point as the center, the phase is aligned, and the arithmetic mean is calculated to generate the reconstructed sequence. The heart rate deceleration force characteristics are obtained by differential calculation.

4. The model algorithm for estimating metabolic syndrome based on cardiac deceleration force and heart rate variability according to claim 1, characterized in that, The method of locating key physiological markers based on the electrocardiogram signal sequence and mechanical waveform signal sequence includes: The first and second derivative sequences of the mechanical waveform signal sequence are obtained through difference operations; Find the zero-crossing point in the first derivative sequence that changes from positive to negative and whose original amplitude is greater than the dynamic threshold, and confirm it as the main wave peak point; Search for local maximum points of the second derivative sequence within a set time interval before the main peak point, and confirm them as the starting points of the troughs; Search for the local minimum of the first derivative sequence between the current main peak and the starting point of the next cardiac cycle trough, and identify it as the diabetic wave notch.

5. The model algorithm for estimating metabolic syndrome based on cardiac deceleration force and heart rate variability according to claim 4, characterized in that, The extraction of ventricular ejection time, ventricular isovolumetric contraction period, and electromechanical coupling delay includes: The length of the discrete sampling point interval between the trough initiation point and the diabetic wave notch point is converted into absolute time to obtain the ventricular ejection time. The time span between the peak point of the R wave in the electrocardiogram signal and the starting point of the trough of the mechanical waveform is calculated to obtain the time before cardiac ejection. The preset electromechanical delay time constant is subtracted to obtain the isovolumetric contraction period of the ventricle. Based on the synchronous calibration pulse segment corresponding to the hardware synchronous level transition signal, the peak offset is extracted by cross-correlation function for phase shift compensation, so as to achieve fine alignment of the time axis of ECG and mechanical signals. Locate the ventricular depolarization electrical initiation point in the aligned ECG signal sequence, and calculate the time span between it and the trough initiation point as the electromechanical coupling delay.

6. The model algorithm for estimating metabolic syndrome based on cardiac deceleration force and heart rate variability according to claim 1, characterized in that, The step of performing dimensionless mapping and feature fusion to generate a one-dimensional fused feature vector includes: The features in the heterogeneous feature set are arranged in columns according to the cardiac cycle to construct an initial feature matrix, and the statistical mean and standard deviation of each feature column are calculated. The initial feature matrix is ​​transformed into a dimensionless standardized feature matrix with a mean of 0 and a variance of 1 using the mean and standard deviation. The standardized feature matrix is ​​flattened and arranged in ascending order of feature column index, and then spliced ​​together to form a one-dimensional fused feature vector.

7. The model algorithm for estimating metabolic syndrome based on cardiac deceleration force and heart rate variability according to claim 1, characterized in that, The process of obtaining the weighted context vector through nonlinear high-dimensional mapping and dynamic weight allocation includes: Calculate the linear combination of the one-dimensional fused feature vector at the hidden layer nodes, and perform a modified linear unit activation operation with leakage to obtain the hidden layer output vector; The unnormalized attention score vector is calculated through a single-layer feedforward attention mechanism network, and the adaptive weight vector is obtained by processing it through a normalization function. Calculate the Hadamard product of the adaptive weight vector and the hidden layer output vector to obtain the weighted context vector.

8. The model algorithm for estimating metabolic syndrome based on cardiac deceleration force and heart rate variability according to claim 1, characterized in that, Linear dimensionality reduction is performed on the weighted context vector, and the scalar output value is mapped to the initial probability value using the Sigmoid activation function; The initial probability value is combined with the full range to convert it into a comprehensive risk index with a value between 0 and 100. When the comprehensive risk index is less than the first risk threshold, it is judged to be in a low-risk state; When the risk level is greater than or equal to the first risk threshold and less than the second risk threshold, it is determined to be a medium-risk state. A state is considered high-risk when the risk level is greater than or equal to the second risk threshold.

9. A model algorithm for estimating metabolic syndrome based on cardiac deceleration force and heart rate variability according to any one of claims 1 to 8, characterized in that, The hardware architecture for acquiring the electrocardiogram signal sequence and the mechanical waveform signal sequence includes: The electrocardiogram signal acquisition channel and the mechanical waveform signal acquisition channel include a microelectromechanical system accelerometer or a microelectromechanical system gyroscope sensor. The main control microprocessor is used to synchronously inject global hardware trigger pulses into the ECG signal acquisition channel and the mechanical waveform signal acquisition channel to configure a global timestamp and control the synchronization time error to be less than a preset tolerance threshold.

Images