Electrocardiogram intelligent diagnosis and analysis system based on multi-mode fusion

Abstract

The invention discloses an intelligent electrocardiogram diagnosis and analysis system based on multi-modal fusion. The system comprises a multi-modal acquisition module, a time-frequency decoupling module, a feature reconstruction module, a geometric calibration module and an intelligent diagnosis module. According to the system, self-adaptive synchronous extrusion transformation is used for processing synchronously-collected electrocardio and cardiac seismic signals, and window function parameters are dynamically optimized by minimizing time-frequency ridge curvature entropy; constructing an electro-seismic dynamic coupling hysteresis ring, and extracting physical characteristics such as a ring area and a main shaft deflection angle to quantify an electro-mechanical coupling state; adopting a Riemannian manifold parallel movement algorithm to carry out breathing layered calibration on the characteristic covariance matrix, and eliminating nonlinear interference caused by breathing movement; and finally, projecting the calibrated features to an Euclidean tangent space for classification evaluation. According to the method, deep feature extraction of an electro-mechanical coupling mechanism and effective elimination of respiratory interference are realized, and the accuracy and robustness of heart function diagnosis are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of biomedical signal processing and intelligent assisted diagnosis technology, specifically to an intelligent electrocardiogram diagnosis and analysis system based on multimodal fusion. Background Technology

[0002] Early detection and accurate diagnosis of cardiovascular diseases are crucial for reducing patient mortality. While surface electrocardiography (ECG), a commonly used clinical method, reflects the electrophysiological activity of the heart, it has limitations in assessing mechanical functions such as myocardial contractility and pumping efficiency. Seismographing, by recording the mechanical vibrations of the chest wall caused by the heart's pulsation, can supplement information about cardiac mechanical activity. Therefore, multimodal analysis techniques that integrate ECG and seismographing signals are gradually becoming an important research direction for comprehensive cardiac function assessment.

[0003] However, existing multimodal cardiac signal analysis techniques still have limitations in practical applications. First, electrocardiogram (ECG) and cardiac oscillation signals are typical non-stationary signals, containing rich transient pathological features. Traditional time-frequency analysis methods are limited by the uncertainty principle, making it difficult to simultaneously achieve both time and frequency resolution. During transformations using fixed window functions, rapid changes in the signal often lead to energy divergence and ridge blurring in the time-frequency distribution, making it difficult to accurately extract the instantaneous frequency and energy features reflecting pathological details.

[0004] Secondly, existing fusion methods mostly employ simple concatenation or weighted summation of feature vectors, failing to fully reveal the intrinsic physical coupling mechanism between cardiac electroactivity and mechanical contraction. The electrical excitation of the heart triggering mechanical contraction is a complex energy conversion process, with a specific nonlinear dynamic relationship between the two in terms of phase and amplitude. Feature fusion methods that ignore this physical coupling mechanism struggle to effectively identify early, subtle lesions such as electromechanical decoupling.

[0005] Furthermore, cardiac vibration signals acquired through the body surface are highly susceptible to interference from respiratory movements. Respiration not only causes low-frequency drift in the signal baseline but also induces nonlinear modulation of the signal amplitude due to changes in thoracic cavity volume. Conventional linear filtering methods struggle to remove respiratory interference while preserving complete pathological waveform information and cannot address the non-rigid drift in data distribution caused by respiration, resulting in poor stability of the diagnostic model under different respiratory states. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides an intelligent electrocardiogram diagnosis and analysis system based on multimodal fusion. This system solves the problems of traditional time-frequency analysis methods being unable to balance the time and frequency resolution of non-stationary signals, single-modal or simple feature splicing being unable to reveal the intrinsic physical mechanism of cardiac electromechanical coupling, and conventional filtering methods being unable to eliminate non-rigid geometric interference caused by respiratory motion.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] The first aspect of this invention provides an intelligent electrocardiogram diagnosis and analysis system based on multimodal fusion. This system addresses the technical problems of existing cardiac monitoring technologies, such as the lack of deep coupling analysis between electrical and mechanical signals and susceptibility to nonlinear interference from respiratory motion, through synchronous acquisition at the hardware level and manifold geometry calculation at the software level.

[0009] The system includes a multimodal signal acquisition module, an adaptive time-frequency decoupling module, a phase space feature reconstruction module, a manifold geometry calibration module, and an intelligent diagnostic module.

[0010] The multimodal signal acquisition module is used to simultaneously acquire the subject's surface electrocardiogram (ECG) and cardiac vibration signals, and extract respiratory signals from the acquired signals. This module achieves homogeneous acquisition at physical locations through a stacked sensor assembly, and uses the same clock source to trigger analog-to-digital conversion, ensuring precise alignment of the ECG and cardiac vibration signals in the time domain.

[0011] The adaptive time-frequency decoupling module addresses the time-frequency ambiguity problem of non-stationary biological signals. This module performs synchronous squeeze wavelet transform on both electrocardiogram (ECG) and cardiac vibration (ECG) signals, introducing a feedback mechanism based on topological curvature during the transform process. Specifically, the module extracts the path with the highest energy from the time-frequency distribution as a time-frequency ridge, and calculates the topological curvature sequence of this ridge and its corresponding curvature entropy. The system uses curvature entropy as the optimization objective and adaptively adjusts the window function parameters using gradient descent. When the curvature entropy reaches convergence, it indicates that the complexity of the time-frequency ridge is minimized, at which point the output instantaneous electrical energy sequence and instantaneous mechanical momentum sequence have optimal time-frequency resolution.

[0012] The phase space feature reconstruction module is used to explore the dynamic coupling relationship between cardiac electrical activity and mechanical contraction. This module abandons traditional time series analysis methods, instead mapping instantaneous electrical energy and mechanical momentum sequences at the same moment to a two-dimensional physical phase space, constructing an electro-seismic dynamic coupling hysteresis loop. The geometry of this hysteresis loop directly reflects the heart's pumping efficiency and electrical conduction synergy. The module further extracts the physical feature vectors of the hysteresis loop, including the hysteresis loop area representing the work done in the electro-mechanical energy conversion of a single heartbeat, the principal axis deflection angle of the fitted ellipse representing phase synchronization, and the maximum curvature value of the trajectory representing the peak torque of myocardial contraction.

[0013] The manifold geometry calibration module is used to eliminate nonlinear interference caused by respiratory motion on physiological signals. Considering that the sample covariance matrix constructed from physical feature vectors lies on a symmetric positive definite matrix manifold, rather than a flat Euclidean space, this module employs a Riemannian manifold geometry method. The module first uses the phase information of the respiratory signal to divide the sample covariance matrix into different respiratory layer subsets. Then, it calculates the Fréchet mean of each respiratory layer subset on the Riemannian manifold as the geometric center, and uses the respiratory layer subset corresponding to the end of expiration as the reference layer. The module constructs a Riemann geodesic connecting the geometric center of the layer to be calibrated and the geometric center of the reference layer, and uses a parallel translation operator to move the sample covariance matrix in the layer to be calibrated along the geodesic to the tangent space position of the reference layer. This operation eliminates the overall distribution shift caused by respiration while maintaining the Riemannian geometric structure of the sample points relative to their geometric centers, thus preserving local features reflecting cardiac pathology.

[0014] The intelligent diagnostic module is used to achieve automatic classification based on manifold features. This module selects the geometric center of the baseline layer as the tangent point to construct a Euclidean tangent space, and uses a log-Euclidean mapping to project the calibrated covariance matrix into a symmetric matrix form of the tangent vector. Subsequently, the tangent vector is semi-vectorized, removing duplicate elements and weighting the off-diagonal elements to generate a one-dimensional feature vector. This module uses a pre-trained support vector machine classifier to analyze the one-dimensional feature vector, calculates its distance relative to the classification hyperplane and maps it to the disease probability, and finally combines the geometric features of the hysteresis loop to output the cardiac function assessment result.

[0015] A second aspect of this invention provides a method for intelligent diagnosis and analysis of electrocardiograms based on multimodal fusion. This method, implemented based on the aforementioned system, mainly includes the following steps:

[0016] First, the electrocardiogram and cardiac vibration signals of the subjects were acquired simultaneously, and the respiratory signals were separated to ensure the consistency of data from different modalities in terms of time and space.

[0017] Secondly, adaptive time-frequency decoupling is performed on the electrocardiogram (ECG) and cardiac vibration (CVR) signals. By calculating the topological curvature entropy of the time-frequency ridges in the synchronous squeeze wavelet transform, the degree of clustering of the time-frequency distribution is quantified, and the window function width is iteratively updated using the gradient descent method until the optimal time-frequency distribution that can accurately characterize instantaneous electrical energy and instantaneous mechanical momentum is obtained.

[0018] Next, an electro-seismic dynamic coupling hysteresis loop is constructed and its features are extracted. Instantaneous electrical energy and instantaneous mechanical momentum are mapped to a two-dimensional phase space, and the electro-mechanical coupling performance of the heart is quantified using the geometric properties of the closed trajectory.

[0019] Subsequently, breathing interference calibration based on manifold geometry is performed. The extracted physical feature vectors are transformed into covariance matrices and placed in Riemannian manifold space. The matrix is ​​layered according to the breathing phase, and data from different breathing states are aligned to the baseline breathing layer by parallel shifting along geodesics, thus eliminating global geometric deformation caused by breathing motion.

[0020] Finally, tangent space projection and intelligent diagnosis are performed. The calibrated manifold data is projected onto the Euclidean tangent space and quantized. A classification model is used to identify the distribution pattern of feature vectors, and the output is a cardiac function assessment result including the probability of disease and the electromechanical coupling comprehensive index.

[0021] This invention provides an intelligent electrocardiogram diagnosis and analysis system based on multimodal fusion. It has the following beneficial effects:

[0022] 1. This invention employs an adaptive synchronous squeezing transform based on topological curvature feedback, solving the problem of mutual constraints between time and frequency resolution in non-stationary biological signal processing. By monitoring the topological curvature entropy of time-frequency ridges and dynamically optimizing window function parameters using gradient descent, the system can automatically adjust the analysis window at points of drastic signal change, effectively suppressing ambiguity and artifacts in the time-frequency distribution, thereby ensuring the extraction accuracy of instantaneous electrical energy and mechanical momentum sequences.

[0023] 2. This invention utilizes phase space reconstruction technology to construct an electro-seismic dynamic coupling hysteresis loop, achieving physical-level fusion of electrocardiogram (ECG) and cardiac seismic signals. This method maps a one-dimensional time series to a two-dimensional phase space trajectory, directly quantifying the energy conversion efficiency and phase synchronization relationship between cardiac electrical excitation and mechanical contraction using geometric features such as the area and principal axis deflection angle of the hysteresis loop. It reveals an electro-mechanical coupling mechanism that cannot be reflected by single-modal analysis, providing more physiologically meaningful diagnostic indicators.

[0024] 3. This invention introduces a Riemannian manifold calibration algorithm based on parallel translation, which effectively eliminates the nonlinear interference of respiratory motion on the multimodal feature distribution. By performing respiratory layering and geometric alignment of the covariance matrix in the Riemannian manifold space, this method can standardize respiratory-modulated feature data to the same reference space, removing the influence of respiration while maintaining the geometric structure of the data that reflects cardiac pathology, significantly improving the robustness of the diagnostic model under different respiratory modes. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the overall system architecture according to an embodiment of the present invention;

[0026] Figure 2 This is a flowchart illustrating the overall process of the cardiac function assessment method based on manifold geometry calibration according to an embodiment of the present invention.

[0027] Figure 3This is a timing diagram of the stacked sensor structure and multimodal signal synchronous acquisition in an embodiment of the present invention;

[0028] Figure 4 This is a schematic diagram illustrating the adaptive synchronous squeezing wavelet transform (SST) and window function feedback adjustment principle of an embodiment of the present invention.

[0029] Figure 5 This is a schematic diagram illustrating the construction and geometric feature definition of a two-dimensional electro-seismic dynamic coupling hysteresis loop (ESCHL) according to an embodiment of the present invention.

[0030] Figure 6 This is a schematic diagram of the Riemannian manifold spatial mapping and the parallel movement calibration geometry based on respiratory stratification, which is an embodiment of the present invention.

[0031] Figure 7 This is a schematic diagram of the logarithmic projection of the tangent space and the eigenvectorization process in an embodiment of the present invention.

[0032] Among them, 110 is the multimodal signal acquisition module; 120 is the adaptive time-frequency decoupling module; 130 is the phase space feature reconstruction module; 140 is the manifold geometry calibration module; and 150 is the intelligent diagnostic module. Detailed Implementation

[0033] 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.

[0034] Example:

[0035] Please see the appendix Figure 1 - Appendix Figure 7 This invention provides an intelligent electrocardiogram diagnosis and analysis system based on multimodal fusion, comprising:

[0036] Multimodal signal acquisition module 110:

[0037] The multimodal signal acquisition module 110 performs the following sub-steps to obtain high-quality synchronous physiological data.

[0038] Step S1011, Sensor Array Layout and Heterogeneous Signal Acquisition: This embodiment employs a stacked composite sensor structure, in which a microelectromechanical system (MEMS) triaxial accelerometer is directly mounted on a rigid printed circuit board on the back of a single-lead ECG electrode. This structure ensures that the origin of the accelerometer's sensitive axis coincides with the contact center of the ECG electrode in a direction perpendicular to the skin surface, thereby guaranteeing that the ECG signal and mechanical vibration signal are acquired from the same anatomical location. When the sensor is attached to the fourth intercostal space at the left sternal border of the subject or at the apex of the heartbeat, the analog front-end circuit simultaneously activates two acquisition channels. The first channel acquires surface potential changes with a frequency range of 0.05Hz to 100Hz, generating a raw simulated ECG signal; the second channel acquires chest wall vibration acceleration with a frequency range of 0.1Hz to 500Hz, generating a raw simulated cardiac vibration signal.

[0039] Step S1012, hardware-level time synchronization and digitization. To eliminate time delay errors between heterogeneous signals, the multimodal signal acquisition module 110 is configured with a unified clock source. The microcontroller (MCU) uses the trigger pulse generated by this unified clock source to synchronously trigger the analog-to-digital converter (ADC) to sample the first and second channels. In this embodiment, the sampling rate is set to 1000Hz and the quantization precision is 24 bits. Through the shared clock and synchronous triggering mechanism, the ECG signal sequence x e [n] and the cardiac signal sequence x s [n] implements strict point-to-point alignment on the time axis, where n is the discrete-time index. For any fixed phase delay that may occur during sampling, the system compensates using pre-calibrated hardware delay parameters.

[0040] Step S1013: Baseline Drift Removal and Power Frequency Notch Filtering. For the acquired raw digital signal, the system first applies a linear phase finite impulse response (FIR) high-pass filter to remove low-frequency baseline drift caused by breathing or body movement. The cutoff frequency of this high-pass filter is set to 0.5Hz. Subsequently, an adaptive notch filter is used to filter out 50Hz or 60Hz power frequency interference. The adaptive notch filter uses a least mean square (LMS) algorithm to dynamically adjust the filter coefficients to suppress narrowband power frequency noise to the greatest extent possible without compromising the signal's spectral characteristics. After the above processing, the preprocessed ECG signal x is obtained. e (t) and cardiac signal x s (t), for subsequent module calls. Regarding the specific design and parameter selection of the aforementioned FIR filter and adaptive notch filter, those skilled in the art can make conventional settings based on the noise characteristics in the actual application scenario; these are well-known technologies in the field and will not be elaborated further here.

[0041] After completing basic signal acquisition and preprocessing, the system needs to further extract respiratory information hidden in physiological signals to provide a phase reference for subsequent manifold geometry calibration. The multimodal signal acquisition module 110 performs the following sub-steps to separate the respiratory components and determine the respiratory phase.

[0042] Step S1014, respiratory modulation signal separation. This embodiment utilizes the characteristic that the amplitude of the R-wave peak of the electrocardiogram (ECG) signal is modulated by changes in thoracic impedance (i.e., the principle of ECG-derived respiratory modulation (EDR)) to separate the preprocessed ECG signal x. e The system first extracts the respiratory signal from the ECG signal r(t). The system first uses the Pan-Tompkins algorithm to detect the peak position of the R-wave in the ECG signal and constructs an R-wave peak amplitude sequence. Then, cubic spline interpolation is performed on this non-uniformly sampled amplitude sequence, and resampling is performed to obtain a uniformly sampled respiratory envelope signal r(t). To eliminate high-frequency noise and extremely low-frequency trend terms introduced by interpolation, the system applies a bandpass filter with a passband range of 0.1Hz to 0.5Hz to r(t), outputting a clean respiratory modulation signal r(t). resp (t).

[0043] Step S1015, Instantaneous respiratory phase calculation. To accurately describe the respiratory state at each moment, the system uses the Hilbert Transform to analyze the instantaneous phase of the respiratory modulation signal. For the respiratory modulation signal r... resp (t), whose analytic signal z(t) is defined as:

[0044]

[0045] Where j is the imaginary unit, The Hilbert transform operator is defined as follows:

[0046]

[0047] Calculate the instantaneous respiratory phase φ based on the analytic signal z(t). resp (t):

[0048]

[0049] The calculated phase φ resp (t) is mapped to the interval [-π,π], where -π corresponds to the end of expiration and π corresponds to the end of inspiration. The continuous change in phase reflects the progress of the respiratory cycle.

[0050] Step S1016, respiratory phase layer division. Based on the accuracy requirements of subsequent manifold calibration, the system pre-determines that a complete respiratory cycle is divided into N phase layers, denoted as {L1, L2, ..., L...}. N Each phase layer L nCorresponding phase interval [θ n-1 ,θ n ), where θ0=-π, θ N =π, and For any time t, if the calculated instantaneous respiratory phase φ resp (t) falls within the interval [θ n-1 ,θ n If the data falls within the respiratory phase layer (L), then the physiological data at that moment is determined to belong to the respiratory phase layer (L). n In this embodiment, N is set to 8 to balance computational complexity and phase resolution. The phase layer index n will serve as a metadata tag, associated with the synchronously acquired ECG and MRI signal segments, and used for covariance matrix classification and calibration in subsequent steps.

[0051] Adaptive Time-Frequency Decoupling Module 120:

[0052] The adaptive time-frequency decoupling module 120 uses synchronous squeeze wavelet transform (SST) to process the electrocardiogram signal x. e (t) and cardiac signal x s (t) Perform high-resolution time-frequency analysis. The specific steps are as follows.

[0053] Step S1021, continuous wavelet transform calculation. The adaptive time-frequency decoupling module 120 performs calculations on the preprocessed ECG signal x. e (t) and cardiac signal x s (t) Perform continuous wavelet transform (CWT). Let the signal to be processed be x(t) (i.e., x e (t) or x s (t)), its continuous wavelet coefficients W x The formula for calculating (a, b) is as follows:

[0054]

[0055] Where 'a' is the scaling factor, used to control the frequency range of the analysis window, a>0; and 'b' is the time shift factor, used to locate the time window for analysis. ψ(t) is the selected mother wavelet function. In this embodiment, the complex Morlet wavelet is chosen as the mother wavelet because it has good time-frequency localization characteristics; ψ * This represents the complex conjugate of the mother wavelet function. This step maps the one-dimensional time signal to a coefficient distribution on a two-dimensional time-scale plane.

[0056] Step S1022, Instantaneous Frequency Estimation. To resolve the energy ambiguity problem of wavelet transform in the time-frequency plane, the system needs to calculate the instantaneous frequency ω at each time-frequency point (a,b). x (a, b). The instantaneous frequency is obtained by calculating the partial derivatives of the wavelet coefficients with respect to the time shift factor b. The specific calculation formula is as follows:

[0057]

[0058] Where i is the imaginary unit, Let W be the partial derivative of the wavelet coefficients with respect to time. x For the point (a,b) = 0, set its instantaneous frequency ω. x (a,b)=0. The instantaneous frequency ω x (a,b)=0 represents the true frequency location of the signal energy concentration at a specific scale a and time b.

[0059] Step S1023, Synchronous Squeezing and Rearrangement. Using the calculated instantaneous frequency information, the wavelet coefficients are redistributed from the time-scale plane (b,a) to the time-frequency plane (b,ω). Specifically, for any given discrete frequency ω... l The system will satisfy |ω x (a,b)-ω l The wavelet coefficients corresponding to scale a with a value less than Δω / 2 are accumulated, where Δω is the frequency resolution. The time-frequency distribution T after synchronous compression transform is shown. x (ω l b) The calculation is as follows:

[0060]

[0061] Among them, a k For a discrete-scale sequence, (Δa) k =a k -a k-1 Through the above-mentioned squeezing operation, the signal energy is focused near the instantaneous frequency trajectory, significantly sharpening the time-frequency representation, thereby enabling precise separation of the electrical excitation energy component in the electrocardiogram signal and the mechanical vibration energy component in the cardiac vibration signal.

[0062] Step S1024, instantaneous envelope extraction. Based on the sharpened time-frequency distribution T x (ω,b), the system extracts the instantaneous energy envelope of the signal by integrating along the frequency axis. For electrocardiogram signals, the instantaneous electrical energy is extracted. For cardiac vibration signals, extract instantaneous mechanical momentum. These two physical quantities will serve as the basic variables for the subsequent construction of the electro-seismic dynamic coupling hysteresis loop.

[0063] To address the issue that fixed window function parameters in traditional time-frequency analysis cannot adapt to non-stationary abrupt changes in cardiac signals, the adaptive time-frequency decoupling module 120 introduces a feedback control mechanism based on hysteresis loop topology. This mechanism inversely corrects the window function parameters in the SST algorithm by evaluating the smoothness of the phase space trajectory. This process includes the following specific sub-steps.

[0064] Step S1025: Initial hysteresis loop construction and curvature calculation. The system first executes steps S1021 to S1024 using the preset initial Gaussian window function width parameter σ0 to obtain the initial instantaneous electrical energy. With instantaneous mechanical momentum Subsequently, an initial two-dimensional state vector is constructed. And an exploratory hysteresis loop trajectory is generated in phase space. For this trajectory, the local curvature sequence κ[n] of its discrete points is calculated:

[0065]

[0066] in, and κ[n] represents the first and second time derivatives calculated using the central difference method, respectively, where n is the index of the discrete sampling point. The curvature κ[n] quantifies the degree of local curvature of the trajectory in phase space.

[0067] Step S1026, Definition of curvature entropy and topological loss function. To comprehensively evaluate the overall morphological quality of the hysteresis loop, the system defines curvature entropy H. κ As the topological loss function, the curvature sequence κ[n] is first normalized to a probability distribution p. n =|κ[n]| / ∑ m |κ[m]|, then calculate the Shannon entropy:

[0068] H κ =-∑ n p n ln(p n );

[0069] The curvature entropy H κ This reflects the statistical characteristics of the hysteresis loop trajectory complexity. When signal processing parameter mismatch leads to non-physiological high-frequency jitter, spikes, or self-intersections in the trajectory, the curvature distribution will become disordered, resulting in H... κ The value increases significantly; conversely, for smooth and physiologically significant quasi-elliptical loci, H... κ The value is relatively small.

[0070] Step S1027, reverse correction of window function parameters. The system sets the target threshold H for curvature entropy. target Compare the currently calculated H κ With H target The difference ΔH = H κ -H target If ΔH > ε (ε is the tolerance limit), then the parameter correction logic is triggered. The correction algorithm uses a variant of the gradient descent method, updating the width parameter σ of the Gaussian window function according to the following rules:

[0071]

[0072] Where η is the learning rate and sgn(·) is the sign function. In practical engineering implementations, due to the analytical gradient... Since the gradient direction is difficult to obtain directly, the system uses a perturbation method to estimate it, that is, by slightly changing the value of σ and observing H. κ The direction of the next adjustment will be determined by the changing trend.

[0073] Step S1028, Iterative Optimization and Final Output. The system repeats steps S1025 to S1027 until the convergence condition |ΔH|≤∈ is met or the preset maximum number of iterations is reached. The corresponding parameter σ at this point... opt This is the optimal window function width for that cardiac cycle. The adaptive time-frequency decoupling module 120 utilizes σ opt Recalculate SST and output the final optimized instantaneous electrical energy E. e (t) and instantaneous mechanical momentum E s (t). This feedback mechanism uses high-level physical features (the topology of the ring) to constrain the low-level signal processing parameters, ensuring that the feature extraction process conforms to the continuous physiological laws of cardiac electromechanical coupling.

[0074] Phase space feature reconstruction module 130:

[0075] The phase space feature reconstruction module 130 is responsible for mapping the adaptively optimized one-dimensional time-series signal to the two-dimensional physical phase space to reveal the dynamic coupling relationship between cardiac electrical activity and mechanical contraction. The specific steps are as follows.

[0076] Step S1031, cardiac cycle segmentation and normalization. The phase space feature reconstruction module 130 first uses the optimized instantaneous electrical energy signal E e (t) Locate the boundaries of the cardiac cycle. By detecting E e The local maxima of (t) are used to identify the reference point corresponding to the time of R-wave occurrence. Where k represents the k-th cardiac cycle. The analysis time window for the k-th cycle is defined as... Where δ pre and δ post These represent the time offsets before and after the R wave, set to 200ms and 600ms respectively, based on typical heart rate ranges. To eliminate the amplitude effects caused by individual differences and signal acquisition gain, the system adjusts the time window τ. k The signal within is normalized to its maximum value:

[0077]

[0078] Normalized dimensionless signal and The range of values ​​is mapped to the interval [0,1].

[0079] Step S1032, Definition of Two-Dimensional Physical Phase Space. Unlike traditional electrocardiogram waveform analysis, this invention constructs an orthogonal two-dimensional Cartesian coordinate system as the physical phase space. The horizontal axis (X-axis) of this coordinate system is defined as the normalized instantaneous electrical energy. The vertical axis (Y-axis) is defined as the normalized instantaneous mechanical momentum. Each point (x, y) in this phase space represents the simultaneous electrical excitation intensity and mechanical contractile intensity of the heart at a specific moment. Therefore, the heart's beating process no longer manifests as a waveform extending over time, but rather as the movement of state points on this two-dimensional plane.

[0080] Step S1033, hysteresis loop trajectory plotting and reconstruction. Within the defined time window τ... k Within, as time t from Increase to State vector A continuous evolution curve is plotted in phase space. Due to the inherent electromechanical delay (EMD) between the reception of electrical signals and the generation of mechanical contraction in cardiomyocytes, and the fact that the duration of mechanical contraction is usually longer than the electrical excitation time, this evolution curve exhibits a significant hysteresis effect. That is, the ascending and descending paths do not coincide, thus forming a closed or nearly closed loop structure, namely the electro-vibrational dynamic coupling hysteresis loop. The hysteresis loop The trajectory equation can be parameterized as follows:

[0081]

[0082] This trajectory visually reflects the dynamic process of electrical energy conversion into mechanical energy during the heart's pumping action. If myocardial contractility weakens or conduction is blocked, the morphology of the hysteresis loop will become flattened, distorted, or reduced in area. The phase space feature reconstruction module 130 stores this trajectory data as a discrete point set sequence for subsequent feature quantization calculations.

[0083] To make the hysteresis loop The visual form is transformed into a computer-processable quantitative indicator. The phase space feature reconstruction module 130 extracts feature vectors f from two dimensions: geometry and physics. k The specific execution steps are as follows.

[0084] Step S1034, calculate the coupling efficiency area A kThe area enclosed by the hysteresis loop characterizes the effective work or energy conversion of cardiac electrical activity driving mechanical contraction within a single cardiac cycle. According to Green's Theorem, the closed-path integral is converted into an area calculation. For a discretized sequence of trajectory points (x[n], y[n]), where n = 1, ..., M, the polygon area formula is used for calculation:

[0085]

[0086] The area A k It is a dimensionless scalar value. In pathological conditions, such as electromechanical decoupling caused by myocardial ischemia, A k The value will decrease significantly; however, during the compensatory phase of heart failure, A k An abnormal increase may occur.

[0087] Step S1035: Calculate the principal axis deflection angle θ of the trajectory. k The system uses principal component analysis (PCA) to determine the predominant distribution direction of hysteresis loops in phase space. First, the covariance matrix Σ of the trajectory point set is calculated. xy Eigenvalue decomposition yields the eigenvector v corresponding to the largest eigenvalue. max =[v x ,v y ] T The principal axis deflection angle is defined as the angle between this eigenvector and the X-axis (electric energy axis):

[0088]

[0089] The angle θ k This reflects the synchronicity and relative intensity ratio of the changes in the two physical quantities, electricity and mechanics. If θ k A bias towards the Y-axis indicates that the mechanical response is more significant or lags behind the electrical excitation; conversely, a bias away from the Y-axis indicates that electrical activity is dominant.

[0090] Step S1036, calculate the maximum topological curvature κ. max To quantify the smoothness of the trajectory and identify abrupt changes, the system reuses the curvature sequence κ[n] calculated in step S1025 and extracts its maximum value:

[0091]

[0092] Maximum topological curvature κ max It usually appears at the top or turning point of the hysteresis loop, and its magnitude is related to the nonlinear dynamic characteristics of the rapid change phase of ventricular ejection.

[0093] Step S1037: Construct a comprehensive feature vector. In addition to the three core features mentioned above, the system also calculates the trajectory perimeter. and the coordinates of the centroid of the trajectory (x) c ,y c Combine all extracted scalar features to generate the initial feature vector for the k-th cardiac cycle:

[0094] f k =[A k ,θ k ,κ max ,L k ,x c ,y c ] T ;

[0095] The eigenvector f k As a digital fingerprint of this cardiac cycle, it fully preserves the physical and geometric information of the electro-vibration dynamic coupling, which can be used by the subsequent manifold calibration module.

[0096] Manifold geometry calibration module 140:

[0097] The manifold geometry calibration module 140 first converts the extracted one-dimensional feature vector into a covariance matrix with rich statistical information, thereby mapping it to the Riemannian manifold space. The specific steps are as follows.

[0098] Step S1041, temporal delay embedding of feature vectors. Since the state evolution of the cardiac system has memory, a feature vector at a single moment is insufficient to fully describe its dynamic characteristics. The manifold geometry calibration module 140 uses temporal delay embedding to expand the feature space. For the feature vector of the k-th cardiac cycle... (where d is the feature dimension), the system selects a window containing several preceding and following periods to construct an expanded data matrix X. k :

[0099] X k =[f k-m ,…,f k ,…,f k+m ];

[0100] Where m is the radius of the embedded window, and in this embodiment m is 2, therefore X k It contains 2m+1 sample vectors. This operation introduces short-term temporal context information, enabling the subsequently constructed covariance matrix to capture the evolution of features over time.

[0101] Step S1042, calculate the sample covariance matrix. For the expanded data matrix X... k The system calculates its sample covariance matrix C. kTo enhance the robustness of covariance estimation, shrinkage estimation or maximum likelihood estimation is employed. This embodiment uses the standard unbiased estimation formula:

[0102]

[0103] in, Let I be the mean vector of the feature vectors within the window, λ be the identity matrix, and λ be a small regularization parameter (e.g., 10). -6 ), used to ensure matrix C k Strictly positive definite (i.e., all eigenvalues ​​are greater than zero). The calculated C... k It belongs to the set of d×d dimensional symmetric positive definite matrices (SPD).

[0104] Step S1043, Riemannian manifold space mapping. In mathematical geometry, all d×d-dimensional symmetric positive definite matrices do not form a flat Euclidean space, but rather a curved Riemannian manifold. Therefore, the system treats the physical state corresponding to each cardiac cycle k as a manifold. A point P on k ≡C k On this manifold, the distance between two points P1 and P2 is no longer the Euclidean distance, but the geodesic distance δ defined by the affine invariant Riemannian metric (AIRM). R (P1,P2):

[0105]

[0106] Among them, ||·|| F Let λ denote the Frobenius norm, log(·) be the matrix logarithm operation, and λ be the matrix norm. i For matrix The i-th generalized eigenvalue. Through this mapping, the system transforms the statistical distribution characteristics of the signal features into geometric points on the manifold, laying the mathematical foundation for subsequent breathing interference calibration based on geometric operations.

[0107] To eliminate the nonlinear interference of heart position shift and thoracic impedance changes caused by respiratory motion on the feature matrix, the manifold geometry calibration module 140 uses the parallel transport technique in Riemannian geometry to align data under different respiratory phases to the same reference space. The specific execution process is as follows.

[0108] Step S1044: Manifold data stratification and mean calculation. Based on the respiratory phase labels n∈{1,…,N} extracted in step S1016, the system stratifies the set of all sample points {P} on the manifold. k Divide into N subsets (i.e., respiratory layers) Among them, subset It contains the covariance matrix of all phases in the nth respiratory phase. Then, for each respiratory layer... The system calculates its Riemannian geometric center (i.e., Fréchet mean) M. n Fréchet mean M n Defined as the point in this layer that minimizes the sum of the squared distances from all matrices to their geodesics:

[0109]

[0110] This optimization problem has no closed-form solution. The system uses the gradient descent algorithm to iteratively solve the problem on the manifold until it converges.

[0111] Step S1045, Reference Layer Selection and Transport Operator Construction. The system selects one of the respiratory layers as the reference layer. (The layer corresponding to the end of expiration is usually selected because the heart's position is relatively stable at this time), and its mean value is M. ref In order to obtain any sample point P in the nth layer n,i Calibrate to the baseline layer, the system is built from the source mean M n To the target mean M ref Parallel translation operator On symmetric positive definite matrix manifolds, this operator is constructed based on Schild's Ladder algorithm or directly using geometric properties, and its explicit expression is as follows:

[0112]

[0113] in, This is the transformation matrix. Geometrically, this operation is equivalent to transforming the distribution center M of the nth layer... n Along connection M n With M ref The geodesic line was "pushed" to M ref At the same time, the relative geometric structure of the sample points within the layer (i.e., the distribution divergence of each sample point relative to the mean) remains unchanged.

[0114] Step S1046, Geometric calibration of the entire dataset. The system iterates through all non-reference respiratory layers n≠ref, and performs geometric calibration on each sample matrix P within the layer. n,i Applying the parallel shift operator described above, the calibrated matrix is ​​obtained. For the base layer The samples within the range have calibration values ​​equal to their original values. After this step, the distribution centers of the covariance matrices of all cardiac cycles on the manifold space are forced to coincide at M. ref This effectively removes the global non-rigid deformation components caused by respiration, retaining only the local differences reflecting the pathological state of the heart. These calibrated matrix sets {P} cal This will be used as clean input data for the next stage of processing.

[0115] The covariance matrix after parallel translation calibration still lies on the curved Riemannian manifold space, while most existing standard classification algorithms are designed based on flat Euclidean space. Therefore, the manifold geometry calibration module 140 needs to perform tangent space logarithmic mapping and vectorization operations to project the points on the manifold onto the tangent plane without loss. The specific steps are as follows.

[0116] Step S1047, tangent space logarithmic mapping. The system selects the Fréchet mean M of the baseline respiratory layer. ref As the tangent point, construct the tangent space. The tangent space is a space that is connected to the manifold. In M ref The Euclidean vector space tangent at each point has a dimension of d(d+1) / 2. Using the Riemann logarithm map operator, each calibrated covariance matrix is... The projection is the tangent vector (in the form of a symmetric matrix) S in the tangent space. i The mapping formula is as follows:

[0117]

[0118] Here, log(·) represents the matrix logarithm operation. This operation is implemented for symmetric positive definite matrices through eigenvalue decomposition: if matrix A = UΛU T Then log(A) = Uln(Λ)U T , where ln(Λ) is the natural logarithm of each diagonal element of the diagonal matrix Λ. Through this mapping, the geodesic distance between two points on a manifold is transformed into the Euclidean distance between tangent vectors in the tangent space, i.e.

[0119] Step S1048, semi-vectorization dimensionality reduction. Because the tangent vector S... i It retains its d×d symmetric matrix form and contains redundant information. The system performs a half-vectorization operation on it to generate the final one-dimensional feature vector z. i To preserve the inner product structure and norm in the tangent space during vectorization (i.e., isomorphism), for S... i The diagonal element S in j,jremain unchanged, while for non - diagonal elements S j,k (j < k) is multiplied by a coefficient The generated eigenvector z i is expressed as:

[0120]

[0121] This vector z i has a dimension of D = d(d + 1) / 2. After the above - mentioned processing, the original complex non - linear manifold data is transformed into a high - dimensional eigenvector z that conforms to the characteristics of Euclidean geometry i , which can be directly adapted to standard classifiers such as Support Vector Machine (SVM) or Linear Discriminant Analysis (LDA), thus completing the transformation from the geometric space to the algebraic space.

[0122] Intelligent diagnosis module 150:

[0123] The intelligent diagnosis module 150 uses the eigenvector z projected onto the Euclidean tangent space i to construct a classification model to identify the cardiac electro - mechanical coupling functional state. The specific implementation steps are as follows.

[0124] Step S1051, classification model training and hyperplane construction. In this embodiment, Support Vector Machine (SVM) is used as the core of the classifier. In the training stage, the system receives a set of sample sets with expert - annotated labels where y m ∈{-1, +1} respectively represent normal and abnormal (such as heart failure or myocardial ischemia) categories, and K is the total number of samples. Since the eigenvector z has been transformed into a flat Euclidean space through the manifold logarithmic mapping and the respiratory interference has been eliminated, the data shows good linear separability. Therefore, a linear kernel function The classifier aims to find an optimal hyperplane w T z + b = 0, so as to maximize the margin between the two types of samples. This optimization problem is expressed as:

[0125]

[0126] where w is the normal vector, b is the bias term, ξ m is the slack variable used to handle the case of non - completely linearly separable, and C is the penalty coefficient. By solving the above quadratic programming problem using the Lagrangian dual method, the optimal weight vector w * and the bias b * are obtained.

[0127] Step S1052, online diagnosis and probability estimation. In the real - time monitoring stage, for the new eigenvector z input in the current cardiac cycle newThe classifier calculates its decision function value f(z). new )=(w * ) T z new +b * To provide a more intuitive clinical reference, the system uses the Platt Scaling method to map the decision function value to the posterior probability value P(y=1|z). new ), that is, the probability of developing the disease:

[0128]

[0129] Here, A and B are the Sigmoid fitting parameters obtained by maximum likelihood estimation on the cross-validation set.

[0130] Step S1053, Multi-classification Expansion Strategy. For various cardiac pathological types (such as myocardial infarction, heart failure, and valvular disease), the system employs a "one-vs-rest" strategy to construct a multi-classifier architecture. That is, for each specific pathological category c... j A binary classifier is trained to distinguish between "this pathology" and "all other states". The final diagnosis is determined by the category with the highest probability output value.

[0131]

[0132] Through the linear discrimination in Euclidean space described above, the system achieves fast and accurate decoding of high-dimensional physical features after manifold geometric calibration.

[0133] In order to provide clinicians or users with diagnostic evidence that is physically meaningful and easy to interpret, the intelligent diagnostic module 150 converts the mathematical results of the classifier into specific physiological quantitative indicators and outputs them in a visual manner. The specific execution steps are as follows.

[0134] Step S1054, Electro-Mechanical Coupling Index (EMCI) Calculation. The system calculates the electro-mechanical coupling index I based on the statistical mean of the physical characteristics of the hysteresis loop extracted within a time window (e.g., the most recent 30 seconds). EMC This index aims to quantify the efficiency with which the myocardium converts electrical excitation into mechanical pumping of blood. The calculation formula is defined as follows:

[0135]

[0136] in, Let A be the mean area of ​​the hysteresis loop. min With A max This refers to the preset statistical extreme values ​​of the population. Mean of main axis deflection angle, θ opt The optimal deflection angle under ideal healthy conditions (typically 45 degrees); Prisk The disease probability value output in step S1052; w1, w2, w3 are the weight coefficients of each item, satisfying w1 + w2 + w3 = 1. The calculated I EMC Normalized to values ​​between 0 and 100, the higher the value, the more coordinated the electromechanical coupling function of the heart.

[0137] Step S1055: Visual reconstruction of hysteresis loop morphology. The system plots the currently detected average hysteresis loop trajectory on the display terminal and overlays a standard healthy reference loop with a semi-transparent shadow. By intuitively comparing the geometric differences between the two (such as the width, tilt direction, and closure of the loop), users can directly observe changes in myocardial contractility or delays in electrical conduction. Furthermore, the system uses different colors to mark specific time periods on the trajectory; for example, red lines mark the isovolumetric contraction period (the stage where electrical activity has occurred but mechanical momentum has not yet increased significantly), visually demonstrating the electromechanical delay time (EMD).

[0138] Step S1056, Tiered Early Warning and Data Transmission. The system will calculate I... EMC The index is compared with a preset clinical threshold. If I EMC <T warn (Warning threshold) The system triggers a Level 1 warning, indicating the need to monitor cardiac load; if I EMC <T crit (Critical threshold) triggers a level two alarm, recommending immediate medical attention. The final diagnostic report includes a classification result (e.g., "normal," "suspected heart failure"), I EMC Numerical values, average electromechanical delay time, and hysteresis loop morphology diagrams are collected and transmitted via Bluetooth or Wi-Fi communication modules to external smart terminals or medical cloud platforms. The data packets conform to medical data exchange standards such as HL7 or FHIR to ensure compatibility with hospital information systems.

Claims

1. A multi-modal fusion-based electrocardiogram intelligent diagnosis and analysis system, characterized in that, The method comprises the following steps: A multi-modal signal acquisition module is used to synchronously acquire the body surface electrocardio signal and the heart shock signal of a subject, and extract the respiratory signal from the acquired signals; An adaptive time-frequency decoupling module is used to perform synchronous squeezing wavelet transformation on the electrocardio signal and the heart shock signal respectively, and adaptively adjust the window function parameter in the transformation process according to the topological curvature feature of the time-frequency ridge line in the transformation result, so as to extract the instantaneous electrical energy sequence and the instantaneous mechanical momentum sequence; A phase space feature reconstruction module is used to map the instantaneous electrical energy sequence and the instantaneous mechanical momentum sequence to a two-dimensional physical phase space, construct an electro-seismic dynamic coupling hysteresis loop, and extract the physical feature vector of the electro-seismic dynamic coupling hysteresis loop; A manifold geometry calibration module is used to construct the physical feature vector into a covariance matrix and map it to a Riemannian manifold space, stratify the covariance matrix by using the respiratory signal, and eliminate the influence of the respiratory on the distribution of the covariance matrix by parallel moving operation on the Riemannian manifold, to obtain the calibrated covariance matrix; An intelligent diagnosis module is used to project the calibrated covariance matrix to the Euclidean tangent space, analyze the projected features by using a classification model, and output the heart function evaluation result. 2.The multi-modal fusion based electrocardiogram intelligent diagnosis and analysis system according to claim 1, characterized in that, The multi-modal signal acquisition module comprises a stacked sensor assembly and a multi-channel analog front end; The stacked sensor assembly comprises a silver / silver chloride electrode layer, an insulating shielding layer and a MEMS accelerometer arranged in sequence from bottom to top; the silver / silver chloride electrode layer is used to adhere to the chest wall of the subject to acquire the electrocardio signal, and the MEMS accelerometer is used to acquire the mechanical vibration of the chest wall at the same position as the heart shock signal; The multi-channel analog front end is connected to the silver / silver chloride electrode layer and the MEMS accelerometer at the same time, triggers analog-to-digital conversion by using the same clock source, and realizes synchronous digital acquisition of the electrocardio signal and the heart shock signal. 3.The multi-modal fusion based electrocardiogram intelligent diagnosis and analysis system according to claim 1, characterized in that, The adaptive time-frequency decoupling module is configured to perform the following operations to realize adaptive adjustment: Initialize the width parameter of the Gaussian window function; Perform short-time Fourier transform on the input signal by using the window function with the current width, and rearrange the transformed time-frequency coefficients to the time-frequency plane by using the synchronous squeezing operator to obtain the time-frequency distribution; Extract the path with the maximum energy from the time-frequency distribution as the time-frequency ridge line, and calculate the topological curvature sequence of the time-frequency ridge line; Calculate the curvature entropy of the topological curvature sequence, if the curvature entropy does not reach the convergence condition, update the width parameter according to the gradient direction of the curvature entropy relative to the width parameter and repeat the transformation process until the curvature entropy converges, and output the corresponding optimal time-frequency distribution.

4. The multi-modal fusion-based electrocardiogram intelligent diagnosis and analysis system according to claim 1, characterized in that, When updating the width parameter, the adaptive time-frequency decoupling module calculates the partial derivative of the curvature entropy with respect to the width parameter, and iteratively adjusts the width parameter in the direction that reduces the curvature entropy by using the gradient descent method, so as to minimize the complexity of the extracted time-frequency ridge line on the time-frequency plane.

5. The multi-modal fusion-based electrocardiogram intelligent diagnosis and analysis system according to claim 1, characterized in that, The phase space feature reconstruction module is configured to: Draw a closed trajectory evolving over time in a two-dimensional Cartesian coordinate system by taking the instantaneous electrical energy sequence value at the same time as the horizontal coordinate and the instantaneous mechanical momentum sequence value as the vertical coordinate, to form the electro-seismic dynamic coupling hysteresis loop; extracting geometric properties of the electro-seismic mechanical coupling hysteresis loop as the physical feature vector, the geometric properties at least including: an area enclosed by the hysteresis loop, for representing electro-mechanical energy conversion work of a single heart beat; a major axis deflection angle of a fitted ellipse, for representing phase synchronism between electrical excitation and mechanical contraction; and a maximum curvature value on the trajectory, for representing peak torque of myocardial contraction. 6.The multi-modal fusion based electrocardiogram intelligent diagnosis and analysis system according to claim 1, characterized in that, The manifold geometry calibration module is configured to: extracting geometric properties of the electro-seismic mechanical coupling hysteresis loop as the physical feature vector, the geometric properties at least including: an area enclosed by the hysteresis loop, for representing electro-mechanical energy conversion work of a single heart beat; a major axis deflection angle of a fitted ellipse, for representing phase synchronism between electrical excitation and mechanical contraction; and a maximum curvature value on the trajectory, for representing peak torque of myocardial contraction. The manifold geometry calibration module is configured to: calculate a sample covariance matrix of the physical feature vector of a plurality of continuous heart cycles using a sliding time window; 7. The multi-modal fusion-based electrocardiogram intelligent diagnosis and analysis system according to claim 6, characterized in that, regard the sample covariance matrix as a point on a symmetric positive definite matrix manifold; divide the sample covariance matrix into different respiratory layer subsets according to phases of the respiratory signal, each respiratory layer subset containing sample covariance matrices within a specific respiratory phase interval. The manifold geometry calibration module performs parallel translation operations in the following manner: calculate a Fréchet mean of each respiratory layer subset on the Riemannian manifold as a geometric center of the respiratory layer subset; select a respiratory layer subset corresponding to an end-expiratory phase as a reference layer, and the remaining respiratory layer subsets as calibration layers; 8.The multi-modal fusion based electrocardiogram intelligent diagnosis and analysis system according to claim 7, characterized in that, for any calibration layer, construct a Riemannian geodesic connecting its geometric center and the geometric center of the reference layer; use a parallel translation operator to move all sample covariance matrices in the calibration layer along the Riemannian geodesic to the tangent space position of the reference layer, so that the Riemannian geometric structure of the moved sample covariance matrices relative to the geometric center of the reference layer remains consistent with the Riemannian geometric structure before movement relative to the geometric center of the calibration layer. The intelligent diagnosis module is configured to: select the geometric center of the reference layer as a tangent point to construct an Euclidean tangent space; 9. The multi-modal fusion-based electrocardiogram intelligent diagnosis and analysis system according to claim 8, characterized in that, project the calibrated covariance matrix into the Euclidean tangent space using a logarithmic Euclidean mapping to obtain a tangent vector in symmetric matrix form; perform semi-vectorization processing on the tangent vector to remove duplicate elements and weight non-diagonal elements, generating a one-dimensional feature vector. The intelligent diagnosis module is further configured to:

10. The multi-modal fusion based electrocardiogram intelligent diagnosis and analysis system according to claim 9, characterized in that, calculate the distance of the one-dimensional feature vector relative to a classification hyperplane using a pre-trained support vector machine classifier, and map it to a disease probability; calculate an electro-mechanical coupling comprehensive index, which is a weighted combination based on the area mean, major axis deflection angle mean of the electro-seismic mechanical coupling hysteresis loop, and the disease probability, for quantifying the coordination degree between cardiac pumping efficiency and electrical conduction. The system further comprises a visual display terminal for displaying the real-time morphology of the electro-seismic mechanical coupling hysteresis loop, superimposed with a standard healthy reference loop for comparison, as well as displaying the electro-mechanical coupling comprehensive index and classification results.

Images