Method, device and equipment for locating source of abnormal heart beat and medium

Abstract

The application relates to the technical field of non-invasive detection of cardiac electrophysiology. A cardiac abnormal beat source positioning method, device, equipment and medium are disclosed. The method comprises the following steps: synchronously collecting MCG signals and ECG signals; based on a preset feature extraction model, performing feature extraction on the MCG signals and the ECG signals to obtain an ECG time sequence feature sequence and an MCG space-time feature set; obtaining MCG sensor coordinates and ECG electrode coordinates, and based on a preset space coordinate mapping model, performing space alignment processing on the MCG sensor coordinates and the ECG electrode coordinates; based on a preset cardiac geometry and electrical conduction model, determining an abnormal beat source according to a mapping relationship between the ECG time sequence feature sequence, the MCG space-time feature set, the MCG sensor coordinates and the ECG electrode coordinates by means of a Bayesian framework model, a deep learning model and a sparse inversion model. The application realizes non-invasive cardiac electrical activity source positioning with high space-time resolution, high stability and individualization.

Description

Technical Field

[0001] This application relates to the field of non-invasive cardiac electrophysiology detection technology, and more specifically, to a method, device, equipment, and medium for locating abnormal cardiac pulsation sources. Background Technology

[0002] Precise localization of abnormal cardiac pulsations (such as premature beats, tachycardia, and atrial fibrillation triggers) is crucial for interventional treatment of arrhythmias. Currently, commonly used localization techniques in clinical practice mainly include:

[0003] Solving the inverse problem of surface electrocardiography (ECG): Based on the inversion of cardiac electrical activity by surface potential distribution, it has high temporal resolution but limited spatial resolution. It is easily affected by individual differences in thoracic cavity structure and tissue conductivity, and its positioning stability is insufficient.

[0004] Magnetocardiography (MCG) imaging: This method measures the weak magnetic field generated by the heart's electrical current using a superconducting quantum interference device (SQUID). It has high spatial resolution and is not affected by tissue conductivity, but its temporal resolution is relatively low, and the equipment is expensive and requires high environmental shielding.

[0005] Intracardiac electrophysiological mapping: an invasive procedure with high accuracy but also high risk and cost, making it unsuitable for initial screening or long-term monitoring.

[0006] Existing localization technologies are prone to localization errors under conditions of noise interference and individual anatomical differences, and struggle to achieve both high temporal and high spatial resolution. Therefore, there is a need to find a non-invasive localization method that can combine individualized cardiac structural information to achieve high accuracy and robustness. Summary of the Invention

[0007] In view of the above situation, this application provides a method, device, equipment and medium for locating the source of abnormal cardiac pulsation, which aims to solve the above problems or at least partially solve the above problems.

[0008] In a first aspect, this application provides a method for locating the source of abnormal cardiac pulsation, including:

[0009] Simultaneously acquire and preprocess MCG and ECG signals;

[0010] Based on a pre-set feature extraction model, features are extracted from the pre-processed MCG signal and ECG signal respectively to obtain ECG time-series feature sequences and MCG spatiotemporal feature sets.

[0011] Obtain the coordinates of the MCG sensor and the coordinates of the ECG electrode, and perform spatial alignment processing on the MCG sensor coordinates and the ECG electrode coordinates based on a pre-set spatial coordinate mapping model;

[0012] Based on a pre-set cardiac geometry and conductance model, and according to the mapping relationship between ECG time-series feature sequences, MCG spatiotemporal feature sets, MCG sensor coordinates, and ECG electrode coordinates, the abnormal pulsation source is determined through a Bayesian framework model, a deep learning model, and a sparse inversion model.

[0013] For example, synchronously acquiring MCG and ECG signals and performing preprocessing includes:

[0014] The MCG signal is subjected to bandpass filtering, power frequency notch filtering, spatial filtering, and wavelet denoising.

[0015] The ECG signal is subjected to bandpass filtering, power frequency notch filtering, and electromyography interference suppression processing.

[0016] The low-frequency baseline drift of the MCG signal and the ECG signal is corrected;

[0017] The QRS complex in the MCG and ECG signals is detected and the R-wave peak timestamp is located to generate an anchor time series.

[0018] For example, based on a pre-set feature extraction model, features are extracted from the preprocessed MCG signal and the ECG signal respectively to obtain an ECG time-series feature sequence and an MCG spatiotemporal feature set, including:

[0019] Based on a pre-set feature extraction model, the spatiotemporal magnetic field distribution features in the preprocessed MCG signal are extracted. The spatial magnetic field distribution features include at least one of the following: the maximum magnetic field gradient point, the zero contour line position, the magnetic field polarity reversal time, and the dominant component time series.

[0020] Based on a pre-set feature extraction model, the temporal features in the pre-processed ECG signal are extracted. The temporal features include at least one of the following: QRS start point, R-wave peak value, QRS end point, and T-wave end point.

[0021] The anchor time series obtained after preprocessing is obtained, the ECG time series feature series is obtained based on the anchor time series and the time series features, and the MCG spatiotemporal feature set is obtained based on the anchor time series and the spatial magnetic field distribution features.

[0022] For example, the process involves obtaining the anchor time series after preprocessing, obtaining the ECG time series feature series based on the anchor time series and the time series features, and obtaining the MCG spatiotemporal feature set based on the anchor time series and the spatial magnetic field distribution features, including:

[0023] The MCG signal and the ECG signal are each divided into multiple heartbeat cycles;

[0024] The time anchor point for each heartbeat cycle is determined based on the anchor point time series.

[0025] Using the time anchor point as the origin, based on the temporal characteristics and spatial magnetic field distribution characteristics within each heartbeat cycle, ECG temporal feature subsequence and MCG spatiotemporal feature subset are obtained;

[0026] The ECG time-series feature sequence is obtained based on multiple ECG time-series feature subsequences, and the MCG spatiotemporal feature set is obtained based on multiple MCG spatiotemporal feature subsets.

[0027] For example, the coordinates of the MCG sensor and the coordinates of the ECG electrode are obtained, and based on a pre-set spatial coordinate mapping model, the coordinates of the MCG sensor and the coordinates of the ECG electrode are spatially aligned, including:

[0028] Based on a pre-set three-dimensional anatomical coordinate system, the coordinates of the MCG sensor and the coordinates of the ECG electrode are obtained;

[0029] Based on the pre-set chest wall deformation degree, the target registration model is determined from multiple registration models;

[0030] According to the target registration model, the MCG sensor coordinates and the ECG electrode coordinates are transformed into the same coordinate system, and the MCG sensor coordinates and the ECG electrode coordinates are spatially aligned.

[0031] For example, based on a pre-set cardiac geometry and conductance model, and according to the mapping relationship between ECG time-series feature sequences, MCG spatiotemporal feature sets, the coordinates of the MCG sensor, and the coordinates of the ECG electrode, the abnormal pulsation source is determined through a Bayesian framework model, a deep learning model, and a sparse inversion model, including:

[0032] Based on a pre-set cardiac geometry and conductance model, and according to the mapping relationship between ECG time-series feature sequence, MCG spatiotemporal feature set, MCG sensor coordinates and ECG electrode coordinates, the first abnormal pulsation source is obtained through the Bayesian framework model.

[0033] Based on a pre-set cardiac geometry and conductance model, and according to the mapping relationship between ECG time-series feature sequences, MCG spatiotemporal feature sets, MCG sensor coordinates, and ECG electrode coordinates, the second abnormal pulsation source is obtained through the deep learning model.

[0034] Based on a pre-set cardiac geometry and conductance model, and according to the mapping relationship between ECG time-series feature sequences, MCG spatiotemporal feature sets, MCG sensor coordinates, and ECG electrode coordinates, the third abnormal pulsation source is obtained through the sparse inversion model.

[0035] Based on a pre-set voting mechanism, a target abnormal pulsation source is determined according to the first abnormal pulsation source, the second abnormal pulsation source, and the third abnormal pulsation source.

[0036] For example, the method for obtaining the cardiac geometry and conductance model includes:

[0037] Acquire cardiac imaging data;

[0038] Based on a pre-set deep learning segmentation algorithm, multiple tissue structures are determined according to the cardiac imaging data;

[0039] Based on a pre-set cardiac geometry mesh generation algorithm, a cardiac geometry model is generated according to multiple tissue structures.

[0040] Based on pre-set conductivity parameters, a cardiac conductance model is generated according to multiple tissue structures.

[0041] Secondly, this application provides a device for locating the source of abnormal cardiac pulsation, comprising:

[0042] The acquisition module is used to simultaneously acquire MCG and ECG signals and perform preprocessing.

[0043] The feature extraction module is used to extract features from the preprocessed MCG signal and the ECG signal based on a pre-set feature extraction model, so as to obtain the ECG time series feature sequence and the MCG spatiotemporal feature set.

[0044] The spatiotemporal alignment module is used to obtain the coordinates of the MCG sensor and the coordinates of the ECG electrode, and to perform spatial alignment processing on the coordinates of the MCG sensor and the coordinates of the ECG electrode based on a pre-set spatial coordinate mapping model.

[0045] The fusion inversion module is used to determine the abnormal pulsation source based on a pre-set cardiac geometry and conductance model, according to the mapping relationship between ECG time-series feature sequences, MCG spatiotemporal feature sets, MCG sensor coordinates and ECG electrode coordinates, through a Bayesian framework model, a deep learning model and a sparse inversion model.

[0046] Thirdly, this application provides a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method for locating abnormal cardiac pulsation sources as described in the first aspect.

[0047] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method for locating abnormal cardiac pulsation sources as described in the first aspect.

[0048] The above-described technical solutions adopted in the embodiments of this application can achieve the following beneficial effects:

[0049] This application integrates MCG and ECG signals to locate the source of abnormal cardiac pulsation, thereby solving the problems of low positioning accuracy, poor anti-interference ability, and neglect of individual anatomical differences in the prior art, and achieving high spatiotemporal resolution, high stability and individualized non-invasive cardiac electrical activity source localization. Attached Figure Description

[0050] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0051] Figure 1 This is a schematic diagram of an application environment for a method for locating the source of abnormal cardiac pulsation in one embodiment of the present invention;

[0052] Figure 2 This is a flowchart illustrating a method for locating the source of abnormal cardiac pulsation in one embodiment of the present invention;

[0053] Figure 3 yes Figure 2 A flowchart illustrating a specific implementation method of step S1;

[0054] Figure 4 yes Figure 2 A flowchart illustrating a specific implementation method of step S2;

[0055] Figure 5 yes Figure 2 A flowchart illustrating a specific implementation method of step S3;

[0056] Figure 6 yes Figure 2 A flowchart illustrating a specific implementation of step S4;

[0057] Figure 7 This is a schematic diagram of a process for obtaining a cardiac geometry and conductance model in one embodiment of the present invention;

[0058] Figure 8 This is a schematic diagram of a cardiac abnormal pulsation source localization device according to an embodiment of the present invention;

[0059] Figure 9 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention;

[0060] Figure 10 This is another structural schematic diagram of a computer device according to one embodiment of the present invention. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0062] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such use can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the term "comprising" and its variations should be interpreted as open-ended terms meaning "including but not limited to."

[0063] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0064] As mentioned above, current localization technologies are prone to localization errors under conditions of noise interference and individual anatomical differences, and it is difficult to simultaneously achieve high temporal and high spatial resolution. To address this technical problem, embodiments of this application provide a method for locating the source of abnormal cardiac pulsation.

[0065] The method for locating the source of abnormal cardiac pulsation provided in this embodiment of the invention can be applied to, for example... Figure 1In this application environment, the device communicates with the server via a network. The server can synchronously acquire and preprocess MCG and ECG signals from the device. Based on a pre-set feature extraction model, features are extracted from the pre-processed MCG and ECG signals to obtain ECG temporal feature sequences and MCG spatiotemporal feature sets. MCG sensor coordinates and ECG electrode coordinates are acquired, and spatial alignment is performed on them based on a pre-set spatial coordinate mapping model. Based on a pre-set cardiac geometry and conductance model, the abnormal pulsation source is determined using a Bayesian framework model, a deep learning model, and a sparse inversion model, according to the mapping relationship between the ECG temporal feature sequences, the MCG spatiotemporal feature sets, and the MCG sensor and ECG electrode coordinates. This application integrates MCG and ECG signals to locate abnormal cardiac pulsation sources, addressing the problems of low positioning accuracy, poor anti-interference ability, and neglect of individual anatomical differences in existing technologies. It achieves high spatiotemporal resolution, high stability, and individualized non-invasive cardiac electrical activity source localization. The device side can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server side can be implemented using a standalone server or a server cluster consisting of multiple servers. The invention will now be described in detail through specific embodiments.

[0066] Please see Figure 2 As shown, Figure 2 A flowchart illustrating the method for locating the source of abnormal cardiac pulsation provided in an embodiment of the present invention includes the following steps:

[0067] S1: Synchronously acquire MCG and ECG signals and perform preprocessing.

[0068] In one embodiment, the magnetocardiogram (MCG) signal and electrocardiogram (ECG) signal of the subject are acquired simultaneously, as follows:

[0069] a. Hardware Configuration: Magnetocardiogram (MCG) signals are acquired using a multi-channel optically pumped magnetometer (OPM) system. The MCG sensor array covers the anterior chest area (typically a 5×5 or 8×8 grid, covering an area of ​​approximately 20cm×20cm). Electrocardiogram (ECG) signals are acquired using a multi-lead high-density surface electrode array (e.g., 64 channels). The MCG and ECG signal acquisition devices share the same high-stability clock source (e.g., a 10MHz temperature-compensated crystal oscillator) and sampling is initiated via a TTL synchronous trigger signal to ensure that the start time error between the two signals is less than 1μs.

[0070] b. Sampling parameters: The sampling frequency is uniformly set to 1000Hz or higher (e.g., 2000Hz) to preserve the high-frequency components of the QRS composite wave. The analog-to-digital conversion resolution during the acquisition process is no less than 16 bits, and the dynamic range meets the common-mode rejection requirements of MCG signals (typical amplitude 10-100pT) and ECG signals (typical amplitude 0.5-5mV).

[0071] In one embodiment, such as Figure 3 As shown, the preprocessing of the synchronously acquired MCG and ECG signals includes:

[0072] S11: Perform bandpass filtering, power frequency notch filtering, spatial filtering, and wavelet denoising on the MCG signal.

[0073] In one embodiment, bandpass filtering is employed, using a zero-phase FIR or IIR filter with a passband of 1-100Hz to suppress low-frequency drift below 1Hz and high-frequency electromagnetic interference above 100Hz.

[0074] In one embodiment, power frequency notch filtering: a narrowband notch filter (Q value ≥ 30) is set at 50Hz (or 60Hz) to eliminate power supply interference.

[0075] In one embodiment, spatial filtering is employed: external magnetic field noise (such as that from elevators, vehicles, etc.) is suppressed using a reference sensor and signal spatial projection (SSP) method.

[0076] In one embodiment, wavelet denoising is performed using Daubechies wavelets (such as db4) for multi-scale threshold denoising while preserving the transient features of the QRS segments.

[0077] S12: Perform bandpass filtering, power frequency notch filtering, and electromyography interference suppression on the ECG signal.

[0078] In one embodiment, the bandpass filter is a 0.5-150Hz zero-phase filter that preserves the complete P-QRS-T waveform.

[0079] In one embodiment, power frequency notch filtering: a narrowband notch filter (Q value ≥ 30) is set at 50Hz (or 60Hz) to eliminate power supply interference.

[0080] In one embodiment, electromyographic interference suppression employs empirical mode decomposition (EMD) and adaptive filtering to remove high-frequency electromyographic noise.

[0081] S13: Correct the low-frequency baseline drift of the MCG signal and the ECG signal.

[0082] In one embodiment, low-frequency baseline drift in MCG and ECG signals, primarily caused by respiration or body movement, is corrected. The correction method employs high-pass filtering, polynomial fitting, or morphological filtering, ensuring the corrected signal baseline fluctuation is controlled within 5% of the original signal amplitude. The cutoff frequency for high-pass filtering is 0.5 Hz. Polynomial fitting, such as using a 5th-order polynomial, fits and subtracts the baseline. Morphological filtering uses an opening-closing operation with a structuring element length greater than the QRS duration (typically >150 ms) to estimate and remove the baseline.

[0083] S14: Detect the QRS complex in the MCG signal and the ECG signal and locate the R-wave peak timestamp to generate an anchor time series.

[0084] In one embodiment, such as Figure 3 As shown, an improved Pan-Tompkins algorithm is used for preliminary detection of QRS complexes in ECG signals. This includes sequentially performing differential QRS slope enhancement; adaptive threshold dynamic adjustment; and combining RR interval constraints to eliminate false detections. Differential QRS slope enhancement replaces the squaring operation with a difference, reducing noise amplification; adaptive threshold dynamic adjustment improves robustness to heart rate variability and noisy environments; and historical trend analysis of RR intervals is introduced to reduce the false detection rate.

[0085] In one embodiment, such as Figure 3 As shown, the MCG signal is used to assist in the verification of QRS group detection. This includes selecting the channel with the largest magnetic field gradient (usually the Z component) in the MCG signal, calculating the local maximum of the absolute value of its first derivative, and using it as a candidate point for time R in the MCG domain. When the time difference between the QRS group detection result in the ECG signal and the candidate point in the MCG signal is within a reasonable range (e.g., <20ms), the QRS group detection result in the ECG signal can be confirmed as valid.

[0086] In one embodiment, such as Figure 3 As shown, the peak R-wave timestamp is located using multi-modal fusion for precise positioning. The R-wave position is located within the QRS complex detected by the ECG signal and used as the initial timestamp. Within a ±20ms window of the initial timestamp, candidate timestamps corresponding to the maximum values ​​of magnetic field strength or gradient are searched in the MCG signal. The final R-wave peak timestamp, t, is determined by a weighted average of the initial and candidate timestamps, or by using the one with the higher signal-to-noise ratio. R It is understandable that the above R-wave peak timestamp is obtained based on one heartbeat cycle. ECG and MCG signals include multiple heartbeat cycles, thus obtaining the R-wave peak timestamp t corresponding to each heartbeat cycle. R R-wave peak timestamp t RThis serves as the time anchor point for each heartbeat cycle in subsequent processing. Furthermore, RR interval consistency checks are performed on multiple consecutive heartbeats to eliminate abnormal jump points.

[0087] Furthermore, multiple R-wave peak timestamps t corresponding to multiple heartbeat cycles R This constitutes an anchor point time series.

[0088] S2: Based on the pre-set feature extraction model, feature extraction is performed on the pre-processed MCG signal and the ECG signal respectively to obtain the ECG time series feature sequence and the MCG spatiotemporal feature set.

[0089] In one embodiment, such as Figure 4 As shown, step S2, based on a pre-set feature extraction model, extracts features from the preprocessed MCG signal and the ECG signal respectively, obtaining an ECG time-series feature sequence and an MCG spatiotemporal feature set, including:

[0090] S21: Based on a pre-set feature extraction model, extract the spatiotemporal magnetic field distribution features in the preprocessed MCG signal. The spatial magnetic field distribution features include at least one of the following: maximum magnetic field gradient point, zero contour line position, magnetic field polarity reversal time, and dominant component time series.

[0091] In one embodiment, for a multi-channel MCG signal (typically containing B...) x B y B z Three components, with B z (primarily), extract the following spatiotemporal magnetic field distribution features within each heartbeat cycle. These spatiotemporal magnetic field distribution features are spatial-temporal coupling features.

[0092] In one embodiment, the point of maximum magnetic field gradient (grad_max) is near the R-wave (e.g., t). R Within a ±30ms window, calculate the spatial gradient of the magnetic field signal for each channel (e.g., differential gradient between adjacent sensors); locate the sensor with the largest gradient amplitude (x). g ,y g ) and corresponding time t g This point usually corresponds to the steepest region of the excitation wavefront in the heart.

[0093] In one embodiment, the zero contour location (zero_contour): for B zTwo-dimensional interpolation (such as radial basis function RBF or bilinear interpolation) is performed on the chest plane to generate a continuous magnetic field distribution map; closed or open curves with zero magnetic field value (i.e. zero contour lines) are extracted, and their geometric center or curvature extrema can reflect the direction and depth of the dipole source; the spatial morphological parameters (such as area, eccentricity, principal axis direction) of the zero contour lines are recorded at key time phases (such as the start, peak, and end of QRS).

[0094] In one embodiment, the magnetic field polarity reversal time (PRT) is the time required to detect B in a specific channel (such as the central region). z The zero-crossing point where the signal changes from positive to negative (or vice versa) serves as a marker of the completion of local depolarization; this time point has a physiological correspondence with the termination point of the QRS wave in the ECG signal.

[0095] In one embodiment, principal component analysis (PCA) is used to determine the dominant component time series: PCA is performed on the full array MCG signal, and the first principal component (PC1) is taken as the representative of global magnetic field activity; the peak value, starting point and full width at half maximum (FWHM) of PC1 are extracted for cross-subject standardized comparison.

[0096] The spatial magnetic field distribution characteristics of the above-mentioned MCG signals are all represented by three-dimensional spatial coordinates (x, y, z_sensor) and precise timestamps (t), forming a "spatiotemporal feature point set".

[0097] S22: Based on a pre-set feature extraction model, extract the timing features from the pre-processed ECG signal. The timing features include at least one of the following: QRS start point, R-wave peak value, QRS end point, and T-wave end point.

[0098] In one embodiment, the timing features in the ECG signal are high temporal resolution electrical activity timing features. The preprocessed ECG signal (preferably leads such as II, V1, and V5) is used to extract the following key timing markers using a multi-scale analysis method:

[0099] QRS start point (QRS) onset ): Wavelet transform (such as Mexican Hat wavelet, scale corresponding to 5-30ms) is used to detect abrupt changes at the leading edge of the QRS complex; or the first / second derivative zero-crossing point combined with the amplitude threshold method is used to search for the first significant slope increase point in the forward direction of the R wave; combined with the consistency check of adjacent leads, misjudgments caused by noise are eliminated.

[0100] R-wave peak value (R peak The timestamp t, which serves as the primary time anchor, has been precisely located in the preprocessing stage of step S14 (see above). Here, it is reused. R .

[0101] QRS termination point (QRS)offset ): Search for point J (the boundary between QRS and ST segments) in the backward direction of the R-wave, using template matching or energy attenuation threshold method (e.g., the signal amplitude drops to below 10% of the R-wave peak value and lasts for 10ms).

[0102] T wave endpoint (T end ): Utilizing the assumption of T-wave morphology symmetry, the peak value of the T-wave is backfitted to a Gaussian or double exponential model and extrapolated to the baseline intersection; or the Teager-Kaiser energy operator is used to enhance the tail of the T-wave and the endpoint is determined by combining an adaptive threshold.

[0103] The timing accuracy of all ECG feature points is controlled within ±2ms, meeting the requirements for millisecond-level ECG timing analysis.

[0104] S23: Obtain the anchor time series obtained after preprocessing, obtain the ECG time series feature series based on the anchor time series and the time series features, and obtain the MCG spatiotemporal feature set based on the anchor time series and the spatial magnetic field distribution features.

[0105] In one embodiment, to achieve strict alignment of MCG spatiotemporal features and ECG temporal features in the time dimension, the R-wave peak timestamp t in the anchor time series is used. R A unified time axis for MCG spatiotemporal features and ECG temporal features is established for anchor points.

[0106] In one embodiment, step S23 obtains the anchor time series obtained after preprocessing, obtains the ECG time series feature series based on the anchor time series and the time series features, and obtains the MCG spatiotemporal feature set based on the anchor time series and the spatial magnetic field distribution features, including:

[0107] S231: Divide the MCG signal and the ECG signal into multiple heartbeat cycles respectively.

[0108] In one embodiment, based on the QRS complex detection results in step S14, the MCG signal and the ECG signal are respectively divided into multiple heartbeat cycles.

[0109] S232: Determine the time anchor point for each heartbeat cycle based on the anchor point time series.

[0110] In one embodiment, when obtaining the anchor time series in step S14, a timestamp t of the R-wave peak in the anchor time series is established. R Based on the mapping relationship between the MCG signal and the QRS complex in the ECG signal, and the anchor point time series, the time anchor point of each heartbeat cycle is determined.

[0111] S233: Taking the time anchor point as the origin, based on the temporal characteristics and spatial magnetic field distribution characteristics within each heartbeat cycle, obtain the ECG temporal feature subsequence and the MCG spatiotemporal feature subset.

[0112] In one embodiment, each heartbeat cycle is timestamped with the peak R-wave time stamp t. R Set the origin (i.e., t=0ms); then set all ECG timing feature points (such as QRS) as the origin. onset T end QRS offset (etc.) and characteristic points of MCG spatial magnetic field distribution (such as t) g Convert grad_max, PRT, zero_contour, etc. to relative t R The relative time offset (unit: ms). For example, ECG time series feature subsequence: {(t QRSonset QRS onset ”),(0,“R”),(t Tend ,“T end ”)}, where (0,“R”) represents the peak value of the R wave and its time 0, (t) within a heartbeat cycle. QRSonset QRS onset ") represents the starting point of the QRS complex within a heartbeat cycle. onset and its relative t R Time offset t QRSonset ,(t Tend ,“T end ") indicates the T-wave endpoint within a heartbeat cycle. end and its relative t R Time offset t Tend MCG spatiotemporal feature subset: {(x j ,y j ,t j1 ,“grad_max”),(C j ,t j2 ,“zero_contour”)}, where: (x j ,y j ,t j1 "grad_max" indicates that relative to t R Time offset t j1 At time, at sensor coordinates (x j ,y j The maximum value of the spatial magnetic field gradient was detected at (C). j ,t j "zero_contour" indicates that in relative to t R Time offset t j2 The zero contour features of the magnetic field extracted at time Cj This is a set of characteristic parameters describing the spatial morphology and location of the contour lines, such as their geometric center coordinates, major axis direction angle, and enclosing area. By tracking the morphological evolution of zero_contour at multiple key moments (such as the start, peak, and end of the QRS), the propagation path and source direction of intracardiac depolarization waves can be inferred.

[0113] S234: For each target heartbeat cycle to be analyzed (including suspected abnormal heartbeat cycles), its corresponding ECG time-series feature subsequence and MCG spatiotemporal feature subset are respectively used as elements of the ECG time-series feature sequence and the MCG spatiotemporal feature set. Thus, the ECG time-series feature sequence is a list composed of ECG time-series feature subsequences of multiple heartbeat cycles arranged in chronological order, and the MCG spatiotemporal feature set is a set composed of MCG spatiotemporal feature subsets of multiple heartbeat cycles.

[0114] In another embodiment, to improve the stability of the features and suppress random noise, the following steps can be performed: Based on a pre-constructed individualized normal heartbeat template signal, the ECG signal and the MCG signal are compared with the individualized normal heartbeat template signal; based on a pre-set signal threshold and the comparison result, ECG signals and MCG signals of abnormal heartbeat cycles are selected; based on a pre-set feature extraction model, features are extracted from the ECG signals and MCG signals of the abnormal heartbeat cycles to obtain the ECG temporal feature subsequence and the MCG spatiotemporal feature subset of the abnormal heartbeat cycles, wherein the ECG temporal feature subsequence of the abnormal heartbeat cycles constitutes the ECG temporal feature sequence, and the MCG spatiotemporal feature subset of the abnormal heartbeat cycles constitutes the MCG spatiotemporal feature set.

[0115] Specifically, the construction process of the personalized normal heartbeat template signal includes: acquiring ECG and MCG signals from multiple normal heartbeat cycles; locating the R-wave peak timestamp of each normal heartbeat cycle as the time anchor point for that cycle; performing time alignment on the ECG and MCG signals of each cycle based on their respective time anchor points; averaging the ECG signals from multiple normal heartbeat cycles to obtain the template ECG signal in the personalized normal heartbeat template signal, and averaging the MCG signals from multiple normal heartbeat cycles to obtain the template MCG signal. Specifically, the ECG and MCG signals from multiple normal heartbeat cycles are extracted from a continuous signal segment. The normal heartbeat cycle is a morphologically stable normal sinus heartbeat. The averaged personalized normal heartbeat template signal has a high signal-to-noise ratio.

[0116] Specifically, the ECG and MCG signals compared with the individualized normal heartbeat template signal are real-time or offline heartbeat signals. The comparison algorithm can be to calculate correlation coefficients, morphological distances, etc. When the difference between the ECG and MCG signals of a certain heartbeat cycle and the individualized normal heartbeat template signal exceeds a preset signal threshold, the heartbeat cycle is determined to be an abnormal heartbeat cycle and locked.

[0117] Specifically, the peak R-wave timestamp t of this abnormal heartbeat cycle is located. R Using time anchors, the temporal features and spatiotemporal magnetic field distribution features of the abnormal heartbeat cycle are extracted from the original ECG and MCG signal segments (rather than the averaged signals) based on a pre-set feature extraction model, resulting in the ECG temporal feature subsequence and MCG spatiotemporal feature subset of the abnormal heartbeat cycle.

[0118] Furthermore, optionally, based on the temporal and spatiotemporal magnetic field distribution characteristics of the abnormal heartbeat cycle and the corresponding temporal and spatiotemporal magnetic field distribution characteristics of the individualized normal heartbeat template signal, differential features are obtained using a pre-set feature calculation algorithm to achieve feature enhancement. Specifically, the feature calculation algorithm is either comparative or differential, such as calculating time offset and morphological changes. These differential features better highlight the specificity of the anomaly source and can serve as supplementary inputs for subsequent fusion and inversion modules.

[0119] Ultimately, the input used to locate the source of the abnormal heartbeat is the ECG temporal feature subsequence and MCG spatiotemporal feature subset from the single abnormal heartbeat, which has not been averaged with other heartbeats, and its format is consistent with the single heartbeat cycle feature format defined in S233.

[0120] S3: Obtain the coordinates of the MCG sensor and the coordinates of the ECG electrode, and perform spatial alignment processing on the MCG sensor coordinates and the ECG electrode coordinates based on a pre-set spatial coordinate mapping model.

[0121] In one embodiment, step S3 is used for multimodal time-space alignment: constructing a spatial coordinate mapping relationship between the MCG sensor array coordinate system and the ECG lead system; using a rigid / non-rigid registration algorithm, unifying the MCG sensor coordinates and ECG electrode coordinates to the same three-dimensional spatial coordinate system to achieve time synchronization and spatial coordinate consistency.

[0122] In one embodiment, such as Figure 5 As shown, step S3 obtains the MCG sensor coordinates and ECG electrode coordinates, and based on a pre-set spatial coordinate mapping model, performs spatial alignment processing on the MCG sensor coordinates and ECG electrode coordinates, including:

[0123] S31: Based on a pre-set three-dimensional anatomical coordinate system, obtain the coordinates of the MCG sensor and the coordinates of the ECG electrode.

[0124] In one embodiment, a unified three-dimensional anatomical coordinate system is defined: a right-handed Cartesian coordinate system (X, Y, Z) is established with the center of the subject's thoracic cavity or the centroid of the heart as the origin: the +X axis points to the left side of the subject; the +Y axis points to the back (or, as defined by MRI, the head-to-foot direction); and the +Z axis points to the front of the chest (i.e., the direction the sensor is facing). Specifically, this coordinate system can be determined in the following way: based on surface landmarks: a local coordinate system is defined by fitting the thoracic plane with anatomical landmarks such as the sternal angle (Angle of Louis), xiphoid process, seventh cervical vertebra (C7), and tenth thoracic vertebra (T10).

[0125] In one embodiment, ECG electrode coordinates are obtained as follows: Six chest leads (V1-V6) out of the standard 12-lead ECG system are manually marked on the body surface according to international standard anatomical locations; using a high-precision optical positioning system (such as NDIPolaris, Fastrak) or a structured light 3D scanner, the ECG electrode coordinates p of each ECG electrode attachment point in the world coordinate system are measured. l ECG =(x l ,y l ,z l ), in millimeters, where l For the first l Each electrode. When using a high-density ECG (such as 64 channels), coordinates are acquired for all ECG electrodes; limb leads (I, II, III, etc.) are usually not involved in spatial modeling because they are far from the heart, or are replaced by approximate points on the shoulder / hip.

[0126] In one embodiment, the MCG sensor coordinates are obtained as follows: the fixed coordinates p of each channel sensor relative to the detector housing are pre-calibrated at the factory when the MCG system (OPM array) is shipped. j MCG,local , where j is the j-th sensor. Before measurement, the spatial positions of at least three non-collinear marker points on the detector housing are measured using the same optical positioning system. The local coordinates are transformed to the world coordinate system through rigid body transformation to obtain the MCG sensor coordinates p. j MCG =R MCG ·p j MCG,local +t MCG , where R MCG Let t be a rotation matrix. MCG It is a translation vector.

[0127] In one embodiment, the measurement accuracy of both the ECG electrode coordinates and the MCG sensor coordinates is better than 1 mm, and the repeatability error is <0.5 mm.

[0128] S32: Determine the target registration model from among multiple registration models based on the pre-set chest wall deformation degree.

[0129] The spatial coordinate mapping model includes multiple registration models, specifically rigid registration models and non-rigid registration models.

[0130] In one embodiment, the degree of chest wall deformation can be preset based on experience or selected manually.

[0131] In one embodiment, when the chest wall deformation is small, a rigid registration model is selected, which is suitable for scenarios where the body position is stable and there is no significant deformation; when the chest wall deformation is large, a non-rigid registration model is selected, which is suitable for situations where there is significant nonlinear deformation of the chest wall, such as obesity, scoliosis, and respiratory movements. In these cases, the rigid assumption fails and non-rigid spatial mapping needs to be used.

[0132] S33: Based on the target registration model, transform the MCG sensor coordinates and the ECG electrode coordinates to the same coordinate system, and perform spatial alignment processing on the MCG sensor coordinates and the ECG electrode coordinates.

[0133] In one embodiment, since the MCG sensor is usually located on a chest fixation bracket while the ECG electrode is directly attached to the skin, their initial coordinate systems are inconsistent. A spatial registration algorithm is needed to establish a mapping and spatially align the coordinates of the MCG sensor and the ECG electrode.

[0134] In one embodiment, the rigid registration model process is as follows:

[0135] Step 1: Generate the MCG surface projection point set.

[0136] Since the MCG sensor is not on the skin, it needs to be "projected" onto the body surface to align with the coordinates of the ECG electrodes.

[0137] Fitting the surface plane where the ECG electrode is located:

[0138]

[0139] In this formula, Π represents the surface plane where the ECG electrode is located; n represents the normal vector of plane Π (a vector perpendicular to the plane), obtained through principal component analysis (PCA) of leads V1-V6 (the direction of the third principal component is the direction of the normal vector); d represents the intercept parameter of plane Π, which, together with the normal vector, defines the position of the plane in space; X represents the three-dimensional coordinate point in space (usually represented as (x, y, z)). (), used to describe the location of MCG sensors or points on the body surface.

[0140] For the coordinates p of each MCG sensor j MCG Projected along the normal n onto plane Π:

[0141] n

[0142] In the formula, Represents the original three-dimensional coordinates of the j-th MCG sensor (j=1,2,...,M, where M is the number of MCG sensors, or the number of channels); The coordinates of the equivalent surface point (i.e., the "projection point") after the j-th MCG sensor is projected onto the ECG surface plane Π are represented; n represents the normal vector of plane Π (a vector perpendicular to the plane).

[0143] Obtain the equivalent surface projection coordinates of all MCG sensor projections. }

[0144] Step 2: Perform rigid ICP registration.

[0145] Objective: To find a rigid transformation (R) ICP ,t ICP )make and Alignment }, This represents the set of coordinate points of the ECG electrode on the body surface plane Π.

[0146] ICP optimization problem: finding the optimal rotation R ICP Peaceful movement ICP Make the MCG projection coordinate point set Q MCG With ECG electrode coordinate point set P ECG Get as close as possible (minimize the sum of squared distances between corresponding points).

[0147]

[0148] In this formula, NN(·) represents... Find the nearest neighbor in R; ICP Let denote the rotation matrix, which belongs to the special orthogonal group SO(3); The coordinates of the equivalent surface point after the j-th MCG sensor is projected onto the ECG surface plane Π are: M represents the number of MCG sensors; t represents the number of MCG sensors. ICP Represents a translation vector, belonging to , that is, three-dimensional Euclidean space; ‖·‖² represents the square of the Euclidean distance, used to measure the difference between two points.

[0149] Solution method (Umeyama algorithm):

[0150] Calculate the centroids of the two sets of points:

[0151]

[0152] In this formula, M represents the number of MCG sensors; This represents the equivalent surface point coordinates of the j-th MCG sensor after it is projected onto the ECG surface plane Π. This represents the centroid of the equivalent body surface point after the MCG sensor is projected onto the ECG body surface plane Π. This indicates the centroid of the ECG electrode within the plane Π on the body surface. Indicates the first l ECG electrode coordinates.

[0153] Rotation matrix R ICP Specific calculation steps:

[0154] Construct a covariance matrix H to describe the correlation between two coordinate point sets (MCG projection coordinate point set and ECG electrode coordinate point set): capture the spatial distribution correlation between the two by multiplying and summing the "centroid-decentrifuged" MCG point set and ECG point set.

[0155]

[0156] In the formula, This represents the equivalent surface point coordinates of the j-th MCG sensor after it is projected onto the ECG surface plane Π. This represents the centroid of the equivalent body surface point after the MCG sensor is projected onto the ECG body surface plane Π. Represents the coordinates of the j-th MCG projection point The coordinates of the matched ECG electrode point (obtained through "nearest neighbor matching"), where c(j) represents the index of the matched ECG electrode. This indicates the centroid of the ECG electrode within the plane Π on the body surface.

[0157] SVD decomposition: U represents the left singular matrix, and V represents the right singular matrix. The covariance matrix H is decomposed into orthogonal and diagonal matrices, from which the rotation matrix R is extracted. ICP This achieves directional alignment between two sets of coordinate points.

[0158] Optimal rotation matrix R ICP Calculation:

[0159]

[0160] In this formula, det(·) represents the determinant of the matrix; diag(1,1,det(·)) represents the diagonal correction matrix; U represents the left singular matrix; and V represents the right singular matrix.

[0161] Translation vector t ICPCalculation:

[0162]

[0163] In this formula, This indicates the centroid of the ECG electrode within the plane Π on the body surface. This represents the centroid of the equivalent body surface point after the MCG sensor is projected onto the ECG body surface plane Π. It is a rotation matrix.

[0164] Iterative optimization: Repeat "find nearest neighbor matching coordinates → solve R" ICP / t ICP "Until convergence (usually 3-5 iterations)."

[0165] Step 3: Apply the transformation.

[0166] Use the coordinates of the MCG sensor (non-projection point) Transform to the coordinate system of the ECG electrode to obtain the transformed coordinates. :

[0167]

[0168] In this formula, t is the rotation matrix; ICP is the translation vector; j is the j-th MCG sensor.

[0169] Output: All MCG sensors and ECG electrodes share the same coordinate system.

[0170] In one embodiment, the process of the non-rigid registration model is as follows:

[0171] Method selection: Thin Plate Spline (TPS). TPS is an interpolation-type non-rigid transformation that can maintain local smoothness and is suitable for modeling surface deformation.

[0172] Step 1: Define control point pairs.

[0173] Source: Set of MCG surface projection coordinate points or a subset thereof, a subset of the MCG surface projection coordinate point set refers to the set of MCG surface projection coordinate points. A representative subset of control points is selected from the data. The selection criteria for this subset include, but are not limited to: downsampling based on the spatial grid of the MCG sensor array (e.g., selecting the center point or key nodes of a tetrahedral grid), or based on the point set. The spatial density is uniformly sampled, or points located near anatomical landmarks are selected. By using this subset instead of the entire set, the computational efficiency and robustness of the model can be improved while ensuring registration accuracy.

[0174] Target point: The set of coordinate points corresponding to the ECG electrode. .

[0175] Virtual control points (such as the midpoint of the sternum or the intersection of the anterior axillary line) can be added to improve coverage.

[0176] Step 2: Solve for the TPS transformation function .

[0177] TPS maps any MCG surface projection coordinate point x´ to:

[0178]

[0179] In this formula, For the affine part; Non-linear weights; For radial basis functions, r= I represents the total number of control points; Represents the coordinates of the MCG surface projection point {q j MCG The i-th point in}.

[0180] The parameters are solved by minimizing the following energy function:

[0181]

[0182] In this formula, λ is the smoothing regularization term (typically 10). -3 ~10 -1 ); For the set of ECG electrode coordinate points {p l ECG The i-th point in}; Transformation function The Laplace operator measures the magnitude of the curvature of the transformation.

[0183] Step 3: Apply nonlinear mapping.

[0184] For each MCG sensor coordinate First, project the data onto the ECG surface plane Π to obtain the equivalent surface point coordinates. Then apply the TPS transformation to obtain the transformed coordinates. : .

[0185] Note: After non-rigid registration, the MCG sensor no longer maintains rigid body geometry, but it is more consistent with real anatomy.

[0186] In one embodiment, after completing rigid or non-rigid registration, the following list of MCG sensor coordinates in a unified coordinate system is output: ECG electrode coordinates list: .

[0187] In one embodiment, the rigid registration quality metric is the average nearest point distance (preferably <5mm). The non-rigid registration quality metrics are the control point fitting residuals and regularization energy.

[0188] This application yields a list of MCG sensor coordinates in the same coordinate system: ECG electrode coordinates list: This will serve as the spatial input for subsequent multimodal fusion inversion modules, used to construct the spatial embedding of feedforward models or deep learning networks.

[0189] S4: Based on a pre-set cardiac geometry and conductance model, and according to the mapping relationship between ECG time-series feature sequences, MCG spatiotemporal feature sets, MCG sensor coordinates and ECG electrode coordinates, the abnormal pulsation source is determined through a Bayesian framework model, a deep learning model and a sparse inversion model.

[0190] After completing the time alignment of the MCG and ECG signals in step S2 and unifying them to the same coordinate system in step S3, this application enters the core stage: jointly inverting the spatial location of the abnormal cardiac pulsation source. This application supports the following three complementary or combinable fusion strategies.

[0191] In one embodiment, such as Figure 6 As shown, step S4, based on a pre-set cardiac geometry and conductance model, determines the abnormal pulsation source according to the mapping relationship between the ECG time-series feature sequence, the MCG spatiotemporal feature set, the MCG sensor coordinates, and the ECG electrode coordinates, using a Bayesian framework model, a deep learning model, and a sparse inversion model, including:

[0192] S41: Based on a pre-set cardiac geometry and conductance model, and according to the mapping relationship between the ECG time-series feature sequence, the MCG spatiotemporal feature set, the coordinates of the MCG sensor, and the coordinates of the ECG electrode, the first abnormal pulsation source is obtained through the Bayesian framework model.

[0193] In one embodiment, the Bayesian estimation framework is described below.

[0194] 1) Core idea.

[0195] The high temporal resolution prior information (such as QRS initiation time and conduction velocity) composed of ECG temporal feature sequences is used as the prior distribution, and the high spatial sensitivity magnetic field distribution composed of MCG spatiotemporal feature sets is used as the likelihood observation. The location of the abnormal pulsation source with the highest posterior probability is inferred through Bayesian inference.

[0196] 2) Mathematical modeling.

[0197] Let the intracardiac candidate source space be a discrete grid. (Usually generated from the cardiac surface or myocardial volume in a cardiac geometry and conduction model reconstructed from MRI, where L is the number of grid points in the cardiac geometry and conduction model, approximately 500-2000).

[0198] Objective: Estimate the activation source location r∈Ω at a certain time t in the heartbeat cycle.

[0199] Bayes' theorem:

[0200]

[0201] In this formula: ∈R M M-channel MCG observations; ∈R N ECG observations in lead N; Likelihood function (forward model based on MCG spatiotemporal feature set); : Prior distribution (driven by ECG time series feature sequence); R is the real number domain; t is time.

[0202] 3) Key components.

[0203] (1) Forward Model.

[0204] The transfer matrix from each candidate source r to the body surface (unified to the same coordinate system for MCG sensor and ECG electrode coordinates) was calculated using an individualized FEM (finite element method) model:

[0205]

[0206] In this formula, s∈R L Let L be the source current density vector, and L represent the number of grid points in the cardiac geometry and conductance model. ∈R M×L For the MCG forward operator, M represents the number of sensors or channels, and L represents the number of grid points in the cardiac geometry and conductance model; ∈R N×L For the ECG forward operator, N is the number of leads.

[0207] (2) Prior distribution driven by ECG time-series feature sequences The introduction is as follows.

[0208] Based on ECG time series characteristics (especially QRS initiation time) And pre-defined normal cardiac electrical conduction patterns (such as the His-Purkinje system) for each candidate source location on the cardiac geometry and conduction model. Construct a theoretical benchmark excitation time :

[0209]

[0210] In this formula, For the first heart model One candidate source location (three-dimensional spatial coordinates); These are the approximate anatomical coordinates of the atrio-ventricular node. Preset electrical conduction velocity for myocardial tissue (typical value for ventricular myocardium: 0.5-1.0 m / s); This is a reference time offset (usually set to 0, or based on the QRS initiation time of a normal sinus rhythm). Euclidean distance represents the distance between two points in space.

[0211] Prior distribution modeling:

[0212] If the observed abnormal electrical activity occurs at the time of (For example, the onset time of premature beats), then the location Prior probability It can be modeled as a theoretical baseline excitation time Gaussian distribution centered at:

[0213]

[0214] In this formula, The starting point of the abnormal electrical activity actually observed from the ECG signal; The time uncertainty variance is used to control the "width" or ambiguity of the prior distribution along the time dimension. This formula indicates that if at a certain location... Theoretical excitation time With actual observation time The closer the location is, the higher the probability (prior probability) that the location is the source of abnormal pulsation.

[0215] Special case handling:

[0216] If the onset time of abnormal electrical activity The QRS initiation point earlier than that of normal sinus rhythm relative to t R Time offset Therefore, the prior probability quality should be concentrated more in atypical origin regions (such as the right ventricular outflow tract, pulmonary vein orifice, etc.), which can be achieved by adjusting the theoretical baseline activation time. Computational models (e.g., using different conduction pathways or preset electrical conduction velocities in myocardial tissue) This can be achieved through [the following].

[0217] (3) Likelihood function based on MCG spatiotemporal feature set The introduction is as follows.

[0218] Assuming the observed noise is Gaussian white noise, given the location of the cardiac source... At that time, the observed MCG signal The conditional probability distribution (likelihood function) can be modeled as:

[0219]

[0220] In this formula, For observation Channel MCG signal vector (at a specific point in time or during a period of time); For the first heart model One candidate source location; The vector derived from the forward model represents the position when... When a current source of unit intensity exists at a location, in all The signal predicted on each MCG sensor (i.e., the forward matrix) The List); For at any time The source current intensity scalar, which can be estimated from the synchronously acquired ECG signal, for example, by proportional estimation based on the R-wave amplitude or QRS integral area; Let be the covariance matrix of the MCG observation noise, an intensity scalar under the assumption of "white Gaussian noise," which is typically a diagonal matrix. ,in For noise variance, It is the identity matrix; To represent the square of the weighted Euclidean norm, it is defined as follows: , = .

[0221] The physical meaning of the likelihood function: This formula measures the actual observed MCG signal. With candidate sources Source current intensity scalar MCG signal predicted by physical forward model The smaller the difference (i.e., the closer the prediction matches the observation), the better the candidate source location. Likelihood probability The higher the value, the better.

[0222] Source current intensity scalar It reflects the overall intensity of cardiac electrical activity and is often considered a constant or slowly varying parameter within a single time point or short time window. Source current intensity scalar. The estimation method uses synchronously recorded ECG signals for calculation:

[0223]

[0224] or

[0225]

[0226] In this formula, R-wave amplitude; The proportionality coefficient is obtained through experimental calibration or model fitting; This is the integration time window (typically covering the QRS group).

[0227] (4) Solve and output.

[0228] Based on Bayes' theorem, for all candidate source locations in the cardiac geometry and conductance model Its posterior probability The following is obtained by combining prior and likelihood function calculations:

[0229]

[0230] In this formula, It is the likelihood function; This represents the prior probability.

[0231] ① Solution process:

[0232] a. Time-series scanning: For each time point at which abnormal pulsation might occur. (in The onset time of abnormal electrical activity. For the corresponding R-wave peak timestamp (i.e., time anchor point), calculate all candidate sources respectively. The posterior probability at that moment.

[0233] Maximum a posteriori (MAP) estimation: at each time step Select the source position r that maximizes the posterior probability. As the most likely source at that moment:

[0234] r

[0235] In this formula, For all candidate sources at time t The corresponding posterior probability.

[0236] b. Determine the origin point: Compare r at different times The earliest significant electrical activity will appear (usually 100%). The location corresponding to the time was determined as the origin of the abnormal pulsation. : .

[0237] ② Output results:

[0238] Spatial location of the first abnormal pulsation source: (3D coordinates, unit: mm); Anatomical label of the first abnormal pulsation source: automatically mapped based on a pre-set cardiac anatomy atlas (e.g., "anterior septum of the right ventricular outflow tract"); Activation time of the first abnormal pulsation source: (Relative to reference time, unit: ms); Confidence score of the first abnormal pulsation source is calculated using the maximum posterior probability value. The confidence level of this positioning result measure:

[0239]

[0240] Normalized to Range or convert to percentage form.

[0241] Physical meaning: This output represents the location and reliability of the single cardiac electrical activity source that is most likely to cause abnormal MCG and ECG signals within the Bayesian probabilistic framework, providing a primary target estimate for clinical ablation therapy.

[0242] S42: Based on a pre-set cardiac geometry and conductance model, and according to the mapping relationship between the ECG time-series feature sequence, the MCG spatiotemporal feature set, the coordinates of the MCG sensor, and the coordinates of the ECG electrode, the second abnormal pulsation source is obtained through the deep learning model.

[0243] In one embodiment, a deep learning model (GNN / Transformer) is described below.

[0244] 1) Network architecture selection: Spatial GNN (Spatial Graph Neural Network) + Temporal Transformer (Temporal Transformer). Specifically, a deep learning architecture of "spatial graph encoding - cross-modal attention fusion - temporal Transformer modeling - grid decoding (upsampling decoding to the heart grid)" is adopted, and the specific process is as follows:

[0245] 2) Input data construction.

[0246] MCG spatiotemporal feature set input: MCG spatiotemporal feature tensor: ,in For time steps, For the number of MCG sensors, MCG feature dimension;

[0247] ECG time series feature sequence input: ECG time series feature tensor: ,in For ECG lead number, For ECG feature dimensions;

[0248] Spatial embedding: unifying the MCG sensor coordinates and ECG electrode coordinates to the same coordinate system. i Encoded as positional encoding.

[0249] Graph Structure: Construct K-nearest neighbor graphs or graphs based on anatomical adjacency relationships for the MCG sensor array and ECG electrode array, respectively. and For example, a K-nearest neighbor graph (K=4) is established among MCG sensors, with each MCG sensor connected only to its four nearest MCG sensors; ECG electrodes are mapped according to anatomical adjacency relationships, and connection relationships are defined based on human physiological structure.

[0250] 3) Deep learning model network process

[0251] (1) Spatial map coding

[0252] Encoding process: Two independent graph neural networks are used to spatially encode the features of the two modalities respectively, capturing the topological relationship between the sensor / electrode.

[0253]

[0254]

[0255] In this formula, For the MCG spatiotemporal feature tensor; For ECG time series feature tensor; It is an MCG graph structure; The structure is an ECG graph; The spatial encoding of the output is MCG features; The output spatially encoded ECG features.

[0256] (2) Cross-modal attention fusion

[0257] By using an attention mechanism, the spatiotemporal features of MCG are used as query vectors Q to retrieve and fuse highly relevant information key vectors K and value vectors V- (Key,Value) from the temporal features of ECG, thereby achieving modal complementarity.

[0258] Cross-modal attention computation is expressed by the following formula:

[0259]

[0260] In this formula, From MCG features (query vector); Derived from ECG features (key vectors); Derived from ECG features (value vectors); The projective weight matrix is ​​a learnable matrix. This is a scaling factor used to stabilize the gradient.

[0261] Fusion output :

[0262]

[0263] (3) Temporal Transformer Modeling

[0264] The fused features Input a Transformer encoder to capture long-term dependencies across time steps.

[0265]

[0266] In this formula, It is a spatiotemporal feature representation rich in spatiotemporal context information.

[0267] (4) Decoding and Output

[0268] Use an upsampling decoder (such as a transposed convolution or graph deconvolution network) to extract spatiotemporal features. Mapping to contain A 3D mesh of the heart at each vertex. Above, generate the activation probability Pˊ for each vertex (candidate source position).

[0269]

[0270] The output is a probability graph composed of the activation probabilities P' of all vertices. .

[0271] (5) Loss function

[0272] a. Supervised scenarios (e.g., with internal standard gold testing): using a weighted cross-entropy loss function. Higher weights are assigned to key anatomical regions (such as the area near the His bundle) or positive samples.

[0273]

[0274] In this formula, This is a real label; For sample weights; The mesh representing the cardiac geometry and conductance model. The probability that a vertex is an abnormal pulsation source; L is the number of grid points in the cardiac geometry and conductance model; l The first in the geometric and conductance model of the heart l Each grid point. Furthermore, the ground truth label refers to the accurate location information of the abnormal beat source determined by the invasive "gold standard" method—intracardiac electrophysiological mapping. The ground truth label is a binary label, taking only 0 or 1 values. 1 indicates that the location of the vertex is (or very close to) the abnormal beat source found in the real world through intracardiac mapping (such as the origin of premature beats or atrial fibrillation trigger points), while 0 indicates that the location of the vertex is not an abnormal beat source.

[0275] b. Unsupervised / weakly supervised scenarios: Combine the residuals of the physical forward model as regularization terms, leverage domain knowledge to guide learning, and construct a loss function.

[0276]

[0277] In this formula, For signal reconstruction loss (e.g., MSE); To compensate for the loss of forward physical consistency, the network output is forced to conform to biophysical laws. These are MCG observations. Let P be the MCG forward operator, and P be the probability graph. These are ECG observations. For ECG forward operator; This is the regularization coefficient.

[0278] Output: Probability graph The vertex position with the highest probability is the second abnormal pulsation source predicted by the deep learning model. It can also output the top-K candidate source locations for clinical reference.

[0279] S43: Based on a pre-set cardiac geometry and conductance model, and according to the mapping relationship between the ECG time-series feature sequence, the MCG spatiotemporal feature set, the coordinates of the MCG sensor, and the coordinates of the ECG electrode, the third abnormal pulsation source is obtained through the sparse inversion model.

[0280] In one embodiment, the sparse source inversion algorithm is described below:

[0281] 1) Problem Modeling

[0282] Assume that the abnormal pulsation is triggered by a few (or even a single) intracardiac point sources, and that the source distribution is sparse.

[0283] Constructing a joint observation model: ;

[0284] A more compact form of fusion is: ;

[0285] In the two formulas above: : Channel MCG observations; : ECG readings from leads; The source current intensity scalar to be solved. This represents the number of grid points in the cardiac geometry and conductance model. , : These correspond to the MCG forward operator and the ECG forward operator (transfer matrix), respectively. : The fused forward matrix.

[0286] Objective: To solve for sparse vectors This minimizes the reconstruction error while satisfying the sparsity constraint.

[0287] 2) Optimization methods

[0288] (1) L1 Regularized Least Squares (LASSO)

[0289]

[0290] In this formula, Data fidelity measures the difference between the predicted signal and the observed signal. Regularization parameter, which controls the sparsity intensity (usually selected through L-curve criterion or cross-validation); L1 regularization terms promote sparsity of solutions; The source current intensity scalar to be solved. This represents the number of grid points in the cardiac geometry and conductance model. l The first in the geometric and conductance model of the heart l 1 grid point; : The fused forward matrix.

[0291] Solvers: ADMM (Alternating Direction Multiplier Method), FISTA (Fast Iterative Shrinking Threshold Algorithm), SPGL1 (Spectral Projection Gradient L1 Algorithm), etc.

[0292] (2) FOCUSS (FOCalUnderdeterminedSystemSolver) is a sparse solution algorithm specifically designed for underdetermined systems (the number of unknowns > the number of observation data), and it is used here to locate the source of abnormal heartbeats.

[0293] Algorithm flow:

[0294] ① Initialization: Set the source current intensity scalar (Zero vector) or a non-zero initial estimate.

[0295] ② Iterative update (for the first) (nth iteration)

[0296] a. Calculate the weight matrix : In this formula, The decay exponent is usually taken as... ; This represents the number of grid points in the cardiac geometry and conductance model. : The source current intensity scalar to be solved.

[0297] b. Solving the weighted ridge regression problem:

[0298]

[0299] This problem has a closed-ended solution: In this formula It is a small regularization parameter; This is the fused forward matrix; This is the weight matrix; This is the fused observation data vector, which is composed of observation data from MCG and ECG.

[0300] ③ Convergence judgment: When Stop iteration when (preset tolerance) is reached.

[0301] ④ Output: Final iteration result .

[0302] Locating the source of abnormal pulsation: from convergent solutions Extract the position corresponding to the component with the largest amplitude. :

[0303]

[0304] As the estimated location of the third abnormal pulsation source.

[0305] S44: Based on a pre-set voting mechanism, determine the target abnormal pulsation source according to the first abnormal pulsation source, the second abnormal pulsation source, and the third abnormal pulsation source.

[0306] In one embodiment, a multi-model voting mechanism is used to calculate the Euclidean distance between each pair of the first abnormal pulsation source, the second abnormal pulsation source, and the third abnormal pulsation source. If the Euclidean distance is less than 1 cm, the location point, i.e. the target abnormal pulsation source, is marked on the heart model, and the confidence level is indicated; otherwise, uncertainty is indicated.

[0307] In one embodiment, such as Figure 7 As shown, the method for obtaining the cardiac geometry and conductance model includes:

[0308] S01: Acquire cardiac imaging data.

[0309] In one embodiment, the data type and requirements for cardiac imaging are as follows:

[0310] Preferred images: high-resolution ECG-gated cardiac MRI (e.g., steady-state free precession sequence SSFP) images; slice thickness ≤ 2mm, slice spacing 0, covering the entire heart; including the left ventricle, right ventricle, atrium, and pericardium;

[0311] Secondary option: Enhanced CT (iodine contrast agent required to distinguish cardiac chambers from myocardium) imaging data; slice thickness ≤1mm, ECG acquired simultaneously; Auxiliary: If only general chest CT imaging data is available, deep learning can be used to complete the cardiac structure.

[0312] In one embodiment, acquiring cardiac imaging data further includes preprocessing the cardiac imaging data to obtain a high-quality individualized anatomical model.

[0313] Preprocessing steps: 1) Noise reduction: Non-local mean filtering or the BM4D algorithm is used to reduce image noise. 2) Field correction: For MRI data, the N4ITK algorithm is used to correct intensity inhomogeneities. 3) Resampling: All image data are resampled to a uniform isotropic resolution (e.g., ...). 4) Registration and Alignment: If there are multiple phase data (such as end-diastolic (ED) and end-systolic (ES), register them to the same three-dimensional coordinate system. Output: After the above preprocessing, high-quality, isotropic three-dimensional volume data is obtained. : In this formula, This represents the preprocessed three-dimensional cardiac image volume data. The variable is a three-dimensional array, with each voxel storing an intensity value (such as the Henle unit in CT or the signal intensity in MRI). This represents the dimensional space of the three-dimensional array. These represent the number of voxels in the three spatial dimensions of the volume data. The number of voxels in the height direction (usually corresponding to the head-to-foot direction of the human body or the row direction of the image coordinate system). The number of voxels in the width direction (usually corresponding to the left-right direction of the human body or the column direction of the image coordinate system). The number of voxels in the depth / layer direction (usually corresponding to the front-back direction of the human body or the slice direction of the image coordinate system). Representing coordinates in three-dimensional space The coordinate index, where , , All are integer indices. Or to write more explicitly This represents the voxel intensity value at this coordinate.

[0314] S02: Based on a pre-set deep learning segmentation algorithm, multiple tissue structures are determined according to the cardiac imaging data.

[0315] In one embodiment, multi-tissue automatic segmentation is performed using a deep learning segmentation algorithm.

[0316] The key organizational structures were segmented using deep learning segmentation algorithms (such as nnU-Net and SwinUNETR) as shown in Table 1. The segmentation process was pixel-level segmentation.

[0317] Table 1 shows the segmented organizational structure.

[0318]

[0319] Segmentation output: Multi-label voxel mask Seg(x,y,z)∈{0,1,…,C s}, where C s Number of organization categories (usually C) s =6~8), to obtain the geometric model of the heart.

[0320] S03: Based on a pre-set cardiac geometry mesh generation algorithm, generate a cardiac geometry model according to multiple tissue structures.

[0321] In one embodiment, a volumetric mesh is generated on a 3D cardiac model based on multiple segmented tissue structures (for FEM analysis in step S4). Specifically, tetrahedral meshes are generated using TetGen or Gmsh; the mesh is refined (cell size ≤ 2 mm) in the myocardial region and sparsed in the gaseous tissue structures; tissue labels are retained for each cell. Output: Tetrahedral meshes with tissue labels.

[0322] S04: Based on pre-set conductivity parameters, generate a cardiac conductance model according to multiple tissue structures.

[0323] In one embodiment, the anisotropic conductivity (unit: S / m) of different tissue structures directly affects electromagnetic field propagation. The anisotropic conductivity of different tissue structures is shown in Table 2, supporting personalized adjustments. For example, if the patient has emphysema, the pulmonary conductivity is set to the lower limit (0.005 S / m); if the myocardial fiber direction is known (DT-MRI), a tensor conductivity σ(x) is constructed. The isotropic approximation (longitudinal value) is used by default to reduce computational complexity. Based on the conductivity parameters, a cardiac geometric model and a pre-defined myocardial path are assigned, resulting in a cardiac conductance model.

[0324] Table 2 Anisotropic electrical conductivity of tissue structure

[0325]

[0326] In one embodiment, such as Figure 7 As shown, the method after step S04 further includes:

[0327] S05: Construct an individualized forward model based on the cardiac geometry and conductance model, and calculate the transfer matrix using the finite element method (FEM). The finite element method (FEM) is applicable to complex geometries / anisotropic conditions, directly on tetrahedral meshes. Discretizes the governing equations; supports tensor conductivity and nonlinear materials; provides more accurate calculations, but takes longer (suitable for offline reconstruction).

[0328] In one embodiment, such as Figure 7 As shown, the method after step S04 further includes:

[0329] S06: The individualized forward model and the transfer matrix obtained from the cardiac geometry and conductance model are embedded into the Bayesian framework model, the deep learning model, and the sparse inversion model. They are used as physical constraint priors in the fusion inversion, as shown in Table 3.

[0330] Table 3 Application of cardiac geometry and conductance models in fusion inversion

[0331]

[0332] Results: Experiments show that individualized cardiac geometry and conductance models can reduce the localization error from 15-20 mm in general models to 5-8 mm (close to the accuracy of intracardiac mapping).

[0333] Output: Heart / thoracic cavity surface mesh; forward matrix L ECG ,L MCG Conductivity configuration for each region.

[0334] Integrated Interface: The output heart / thoracic cavity surface mesh and forward operator matrix L are displayed via the API interface. ECG / L MCG The conductivity configuration of each region is provided to the fusion inversion module to determine the source of abnormal pulsation, and it can support real-time query of the body surface response corresponding to a source point.

[0335] Step S4 outputs the localization results: after determining the abnormal pulsation source, it outputs the three-dimensional spatial coordinates of the abnormal pulsation source, the localization reliability score, and the excitation propagation sequence that changes with the heartbeat cycle; it supports visualization and clinical decision support.

[0336] Specifically, the three-dimensional spatial coordinates of the abnormal pulsation source are defined as the location of the earliest onset point of abnormal electrical activity within the heart (i.e., the "source focus") in a personalized anatomical coordinate system. The three-dimensional spatial coordinates are represented using Cartesian coordinates: (Unit: mm), with the sternal angle or cardiac centroid as the origin. Localization results may also include anatomical labels (optional), which are automatically mapped to standard anatomical regions, such as: "Right Ventricular Outflow Tract (RVOT) Anterior Septal", "Left Ventricular Summit", "Pulmonary Vein Orifice (LSPV)". The AHA17 segmental model or PACES ablation zoning standard is used. Accuracy requirement: Spatial error ≤ 8 mm (clinically validated threshold). Specifically, the "tissue structure label" description: Tissue labels in steps S02 / S03: These are basic, coarse-grained classifications, such as myocardium, left ventricular blood pool, and right ventricular blood pool, used to distinguish tissue types. Anatomical labels in the output stage: These are fine-grained, clinically significant anatomical location descriptions, such as the RVOT anterior septum, left atrial appendage orifice, and mitral valve annulus P2 region, used to predict the specific location of abnormal pulsation sources in the heart. Anatomical labels are built upon tissue structure labels by pre-importing or pre-defining a set of standard clinical anatomical atlases (such as the AHA17 segment model or PACES ablation atlas) and mapping them onto an individual cardiac model. For example, within the "myocardium," clinical regions such as the "basal segment of the interventricular septum," the "mid-anterior wall," and the "apex" are further subdivided. Specifically, anatomical labels are obtained by matching the coordinates output by S4 with an individualized cardiac geometry model that has been associated with the standard clinical anatomical atlas. This requires pre-constructing a standard anatomical atlas and registering and mapping it with the patient's image segmentation results.

[0337] Specifically, the Localization Confidence Score (LCS) for abnormal pulsation sources is used to quantify the reliability of localization results and avoid blindly relying on algorithm output. The scoring mechanism (0-100 points) is shown in Table 4, taking into account the following factors:

[0338] Table 4. Scoring Mechanism of Location Reliability Score for Abnormal Pulsations

[0339]

[0340] The process of the myocardial electrical excitation wavefront spreading in time and space from the point of origin of abnormal pulsation is called the excitation propagation sequence.

[0341] The method for generating excitation propagation sequences is as follows:

[0342] (1) Extract the temporal source distribution.

[0343] Extract each discrete time point within the anomalous excitation time window from the results of the Bayesian or sparse inversion model. The source current density distribution vector on is denoted as , The discrete time point at the start of abnormal electrical activity within the abnormal excitation time window; For discrete-time variables; is the discrete time point of the time anchor within the abnormal excitation time window; L is the number of grid points in the cardiac geometry and conductance model.

[0344] (2) Calculate the activation time map.

[0345] For the first in the heart model grid vertices (positions) ), its activation time Defined as the time when the source current intensity at this location first exceeds a specific threshold:

[0346]

[0347] In this formula: It is a vector The Each component represents a position. At any moment The source strength; This is a relative threshold parameter, typically set to 0.1 to 0.3 (i.e., 10%-30% of the maximum intensity); It is the maximum absolute source intensity at that location at all points in time.

[0348] (3) Construct the excitation propagation sequence (spatiotemporal trajectory).

[0349] All activated mesh vertices are sorted by their activation time. Sort the sequences from morning to night to obtain the excitation propagation sequence. :

[0350]

[0351] In the formula The excitation propagation sequence It describes the spatiotemporal path of electrical excitation propagating in the myocardium from the earliest activation point (i.e., the abnormal pulsatile source).

[0352] (4) Output.

[0353] Dynamic activation timeline: a form of visualization that typically displays... The values ​​are rendered on a 3D heart model using color mapping to dynamically display the propagation process of excitation waves.

[0354] Excitation propagation sequence : As quantifiable data output.

[0355] Specifically, the method further includes: generating visualization information and clinical decision support information based on the three-dimensional spatial coordinates of the abnormal pulsation source of the target abnormal pulsation source, the location reliability score, the excitation propagation sequence that changes with the heartbeat cycle, and the cardiac geometry and conductance model.

[0356] Specifically, the visualized information is described as follows: This application provides a multi-level interactive visualization interface to support rapid clinical interpretation.

[0357] 1) Three-dimensional anatomical fusion view

[0358] The located abnormal cardiac pulsation source is superimposed on the individualized three-dimensional cardiac model (from the MRI / CT reconstruction in step S03).

[0359] Supports multi-view rotation and transparency adjustment (can see through the heart chambers);

[0360] Color coding: Red = high probability origin region, blue = late activation region.

[0361] 2) Dynamic Excitement Communication Animation

[0362] Based on the excitation propagation sequence that varies with the heartbeat cycle, play the process of electrical excitation wavefront diffusion within the heartbeat cycle (frame rate = 50fps).

[0363] It can simultaneously display the MCG signal magnetic field map and ECG signal waveform, realizing the "signal-anatomy-timing" three-linkage.

[0364] 3) Two-dimensional projection diagram (compatible with traditional drawing reading habits)

[0365] Generate surface potential / magnetic field contour maps corresponding to 64-lead ECG signals;

[0366] The equivalent dipole direction and depth obtained by superposition inversion.

[0367] 4) Reports are generated automatically.

[0368] PDF format clinical report, including:

[0369] Coordinates and anatomical description of the source of abnormal cardiac pulsation, for example, premature ventricular contractions (PVCs): Location conclusion: The abnormal pulsation source is located in the anterior septum of the right ventricular outflow tract (RVOT), approximately 8 mm from the lower edge of the pulmonary valve annulus. Clinical correspondence: Consistent with the typical morphology of PVCs in the outflow tract with left bundle branch block (LBBB) and downward axis deviation.

[0370] The confidence score and its basis are obtained from Table 4.

[0371] A schematic diagram of the excitation propagation path was obtained based on the excitation propagation sequence that varies with the heartbeat cycle;

[0372] Recommended ablation target areas (e.g., "Recommended mapping on the anterior wall of the RVOT").

[0373] Specifically, the clinical decision support information in this application is described as follows:

[0374] 1) Interfacing with electrophysiological navigation systems

[0375] Output in a standard format (such as CARTO®.txt or EnSite™.xml);

[0376] It can be directly imported into a 3D mapping system as a preliminary ablation target, reducing surgical exploration time.

[0377] 2) Uncertainty Warning

[0378] If LCS < 70, this application outputs an uncertainty highlight message: "Location is low confidence, please confirm with intracardiac mapping."

[0379] Provides the top-3 candidate regions and their respective probabilities.

[0380] 3) Historical comparison

[0381] If a patient has undergone multiple examinations, the system automatically compares the drift of abnormal cardiac pulsation sources (such as atrial fibrillation trigger foci migration) across the multiple examinations.

[0382] 4) AI-assisted diagnostic suggestions

[0383] Based on the location of the abnormal cardiac pulsation source and ECG characteristics, the system pushes predicted diagnostic results, such as "RVOT originating from premature ventricular contractions, typical LBBB morphology"; "high probability of left posterior fascicular reentrant ventricular tachycardia".

[0384] This application achieves precise spatial localization and temporal reconstruction of abnormal pulsation sources by simultaneously acquiring MCG and ECG signals, and through preprocessing, time alignment, spatiotemporal alignment and multimodal fusion inversion, combined with individualized cardiac geometry and conductance models.

[0385] As can be seen, in the above scheme, this application integrates MCG and ECG signals to locate the source of abnormal cardiac pulsation, thereby solving the problems of low positioning accuracy, poor anti-interference ability, and neglect of individual anatomical differences in the prior art, and achieving high spatiotemporal resolution, high stability, and individualized non-invasive cardiac electrical activity source localization. It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this invention.

[0386] In one embodiment, a device for locating the source of abnormal cardiac pulsation is provided, which corresponds one-to-one with the method for locating the source of abnormal cardiac pulsation described in the above embodiments. For example... Figure 8 As shown, the cardiac abnormality pulsation source localization device includes an acquisition module 101, a feature extraction module 102, a spatiotemporal alignment module 103, and a fusion and inversion module 104. Detailed descriptions of each functional module are as follows:

[0387] Acquisition module 101 is used to simultaneously acquire MCG and ECG signals and perform preprocessing.

[0388] The feature extraction module 102 is used to extract features from the preprocessed MCG signal and the ECG signal respectively based on a pre-set feature extraction model to obtain the ECG time series feature sequence and the MCG spatiotemporal feature set.

[0389] The spatiotemporal alignment module 103 is used to acquire the coordinates of the MCG sensor and the coordinates of the ECG electrode, and perform spatial alignment processing on the coordinates of the MCG sensor and the coordinates of the ECG electrode based on a pre-set spatial coordinate mapping model.

[0390] The fusion inversion module 104 is used to determine the abnormal pulsation source based on a pre-set cardiac geometry and conductance model, according to the mapping relationship between the ECG time-series feature sequence, the MCG spatiotemporal feature set, the coordinates of the MCG sensor, and the coordinates of the ECG electrode, through a Bayesian framework model, a deep learning model, and a sparse inversion model.

[0391] Specifically, the acquisition module 101 is also used to perform bandpass filtering, power frequency notch filtering, spatial filtering, and wavelet denoising on the MCG signal; to perform bandpass filtering, power frequency notch filtering, and electromyography interference suppression on the ECG signal; to correct the low-frequency baseline drift of the MCG signal and the ECG signal; to detect the QRS complex in the MCG signal and the ECG signal and locate the R-wave peak timestamp, and generate an anchor time series.

[0392] Specifically, the feature extraction module 102 is further configured to extract the spatiotemporal magnetic field distribution features in the preprocessed MCG signal based on a pre-set feature extraction model. The spatial magnetic field distribution features include at least one of the following: maximum magnetic field gradient point, zero contour line position, magnetic field polarity reversal time, and dominant component time series. Based on the pre-set feature extraction model, it also extracts the time series features in the preprocessed ECG signal, which include at least one of the following: QRS start point, R-wave peak value, QRS end point, and T-wave end point. The module acquires the anchor point time series obtained after preprocessing, obtains the ECG time series feature series based on the anchor point time series and the time series features, and obtains the MCG spatiotemporal feature set based on the anchor point time series and the spatial magnetic field distribution features.

[0393] Specifically, the feature extraction module 102 is further configured to divide the MCG signal and the ECG signal into multiple heartbeat cycles; determine the time anchor point of each heartbeat cycle according to the anchor point time series; take the time anchor point as the origin, and obtain the ECG time series feature subsequence and the MCG spatiotemporal feature subset according to the time series features and the spatial magnetic field distribution features within each heartbeat cycle; obtain the ECG time series feature sequence according to the multiple ECG time series feature subsequences, and obtain the MCG spatiotemporal feature set according to the multiple MCG spatiotemporal feature subsets.

[0394] Specifically, the spatiotemporal alignment module 103 is also used to obtain the coordinates of the MCG sensor and the coordinates of the ECG electrode based on a pre-set three-dimensional anatomical coordinate system; determine a target registration model among multiple registration models according to a pre-set chest wall deformation degree; transform the coordinates of the MCG sensor and the coordinates of the ECG electrode to the same coordinate system according to the target registration model; and perform spatial alignment processing on the coordinates of the MCG sensor and the coordinates of the ECG electrode.

[0395] Specifically, the fusion inversion module 104 is further configured to, based on a pre-set cardiac geometry and conductance model, obtain a first abnormal pulsation source through the Bayesian framework model according to the mapping relationship between the ECG time-series feature sequence, the MCG spatiotemporal feature set, the MCG sensor coordinates, and the ECG electrode coordinates; based on the pre-set cardiac geometry and conductance model, obtain a second abnormal pulsation source through the deep learning model according to the mapping relationship between the ECG time-series feature sequence, the MCG spatiotemporal feature set, the MCG sensor coordinates, and the ECG electrode coordinates; based on the pre-set cardiac geometry and conductance model, obtain a third abnormal pulsation source through the sparse inversion model according to the mapping relationship between the ECG time-series feature sequence, the MCG spatiotemporal feature set, the MCG sensor coordinates, and the ECG electrode coordinates; and based on a pre-set voting mechanism, determine a target abnormal pulsation source according to the first abnormal pulsation source, the second abnormal pulsation source, and the third abnormal pulsation source.

[0396] Specifically, such as Figure 8 As shown, the cardiac abnormal pulsation source localization device also includes an individualized modeling module 105, which is used to acquire cardiac image data; determine multiple tissue structures based on the cardiac image data according to a pre-set deep learning segmentation algorithm; generate a cardiac geometric model based on the multiple tissue structures according to a pre-set cardiac geometric mesh generation algorithm; and generate a cardiac conductance model based on the multiple tissue structures according to a pre-set conductivity parameter.

[0397] Specifically, such as Figure 8 As shown, the cardiac abnormal pulsation source localization device also includes a real-time display module 106. The real-time display module 106 is used to generate visual display information and clinical decision support information based on the three-dimensional spatial coordinates of the abnormal pulsation source of the target abnormal pulsation source, the location reliability score, the excitation propagation sequence that changes with the heartbeat cycle, and the cardiac geometry and conductance model.

[0398] This invention provides a device for locating the source of abnormal cardiac pulsation. This application integrates MCG and ECG signals to locate the source of abnormal cardiac pulsation, addressing the problems of low positioning accuracy, poor anti-interference ability, and neglect of individual anatomical differences in existing technologies. It achieves high spatiotemporal resolution, high stability, and individualized non-invasive cardiac electrical activity source localization. Specific limitations of the device for locating the source of abnormal cardiac pulsation can be found in the limitations of the method for locating the source of abnormal cardiac pulsation described above, and will not be repeated here. Each module in the above-mentioned device for locating the source of abnormal cardiac pulsation can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in hardware or independently of the processor in a computer device, or stored in software in the memory of a computer device, so that the processor can call and execute the operations corresponding to each module.

[0399] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 9 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used for communication with external devices via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a server-side method for locating abnormal cardiac pulsations.

[0400] In one embodiment, a computer device is provided, which may be a device terminal, and its internal structure diagram may be as follows: Figure 10 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a method for locating the source of abnormal cardiac pulsation on the device side.

[0401] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps:

[0402] Simultaneously acquire and preprocess MCG and ECG signals;

[0403] Based on a pre-set feature extraction model, features are extracted from the pre-processed MCG signal and ECG signal respectively to obtain ECG time-series feature sequences and MCG spatiotemporal feature sets.

[0404] Obtain the coordinates of the MCG sensor and the coordinates of the ECG electrode, and perform spatial alignment processing on the MCG sensor coordinates and the ECG electrode coordinates based on a pre-set spatial coordinate mapping model;

[0405] Based on a pre-set cardiac geometry and conductance model, and according to the mapping relationship between ECG time-series feature sequences, MCG spatiotemporal feature sets, MCG sensor coordinates, and ECG electrode coordinates, the abnormal pulsation source is determined through a Bayesian framework model, a deep learning model, and a sparse inversion model.

[0406] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0407] Simultaneously acquire and preprocess MCG and ECG signals;

[0408] Based on a pre-set feature extraction model, features are extracted from the pre-processed MCG signal and ECG signal respectively to obtain ECG time-series feature sequences and MCG spatiotemporal feature sets.

[0409] Obtain the coordinates of the MCG sensor and the coordinates of the ECG electrode, and perform spatial alignment processing on the MCG sensor coordinates and the ECG electrode coordinates based on a pre-set spatial coordinate mapping model;

[0410] Based on a pre-set cardiac geometry and conductance model, and according to the mapping relationship between ECG time-series feature sequences, MCG spatiotemporal feature sets, MCG sensor coordinates, and ECG electrode coordinates, the abnormal pulsation source is determined through a Bayesian framework model, a deep learning model, and a sparse inversion model.

[0411] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and device side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0412] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0413] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0414] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method of locating a source of abnormal heartbeats, characterized by, The method comprises the following steps: synchronously collecting and preprocessing MCG signals and ECG signals; extracting features from the preprocessed MCG signals and ECG signals based on a pre-set feature extraction model to obtain an ECG time sequence feature sequence and an MCG space-time feature set; obtaining MCG sensor coordinates and ECG electrode coordinates, and performing spatial alignment processing on the MCG sensor coordinates and the ECG electrode coordinates based on a pre-set spatial coordinate mapping model; determining an abnormal beat source based on a pre-set cardiac geometry and conductivity model, a mapping relationship between the ECG time sequence feature sequence, the MCG space-time feature set, the MCG sensor coordinates and the ECG electrode coordinates, and a Bayesian framework model, a deep learning model and a sparse inversion model; determining an abnormal beat source based on a pre-set cardiac geometry and conductivity model, a mapping relationship between the ECG time sequence feature sequence, the MCG space-time feature set, the MCG sensor coordinates and the ECG electrode coordinates, and a Bayesian framework model, a deep learning model and a sparse inversion model, comprising: determining a first abnormal beat source based on a pre-set cardiac geometry and conductivity model, a mapping relationship between the ECG time sequence feature sequence, the MCG space-time feature set, the MCG sensor coordinates and the ECG electrode coordinates, and the Bayesian framework model; determining a second abnormal beat source based on a pre-set cardiac geometry and conductivity model, a mapping relationship between the ECG time sequence feature sequence, the MCG space-time feature set, the MCG sensor coordinates and the ECG electrode coordinates, and the deep learning model; determining a third abnormal beat source based on a pre-set cardiac geometry and conductivity model, a mapping relationship between the ECG time sequence feature sequence, the MCG space-time feature set, the MCG sensor coordinates and the ECG electrode coordinates, and the sparse inversion model; determining a target abnormal beat source based on the first abnormal beat source, the second abnormal beat source and the third abnormal beat source based on a pre-set voting mechanism.

2. The method according to claim 1, wherein the synchronous collection and preprocessing of the MCG signals and the ECG signals comprises: band-pass filtering, power frequency notch filtering, spatial filtering and wavelet denoising processing of the MCG signals; band-pass filtering, power frequency notch filtering and electromyographic interference suppression processing of the ECG signals; correction of low-frequency baseline drift of the MCG signals and the ECG signals; detection and positioning of R-wave peak timestamps of QRS complexes in the MCG signals and the ECG signals to generate anchor point time sequences.

3. The method according to claim 1, wherein the feature extraction from the preprocessed MCG signals and ECG signals based on a pre-set feature extraction model to obtain an ECG time sequence feature sequence and an MCG space-time feature set comprises: ​ ​ Based on a pre-set feature extraction model, the spatiotemporal magnetic field distribution features in the preprocessed MCG signal are extracted. The spatial magnetic field distribution features include at least one of the following: the maximum magnetic field gradient point, the zero contour line position, the magnetic field polarity reversal time, and the dominant component time series. Based on a pre-set feature extraction model, the time-series features in the pre-processed ECG signal are extracted. The time-series features include at least one of the following: QRS start point, R-wave peak value, QRS end point, and T-wave end point. The anchor time series obtained after preprocessing is obtained, the ECG time series feature series is obtained based on the anchor time series and the time series features, and the MCG spatiotemporal feature set is obtained based on the anchor time series and the spatial magnetic field distribution features.

4. The method for locating the source of abnormal cardiac pulsation according to claim 3, characterized in that, The process involves obtaining the preprocessed anchor time series, deriving the ECG time series feature series based on the anchor time series and the time series features, and obtaining the MCG spatiotemporal feature set based on the anchor time series and the spatial magnetic field distribution features, including: The MCG signal and the ECG signal are each divided into multiple heartbeat cycles; The time anchor point for each heartbeat cycle is determined based on the anchor point time series. Using the time anchor point as the origin, based on the temporal characteristics and spatial magnetic field distribution characteristics within each heartbeat cycle, ECG temporal feature subsequence and MCG spatiotemporal feature subset are obtained; The ECG time-series feature sequence is obtained based on multiple ECG time-series feature subsequences, and the MCG spatiotemporal feature set is obtained based on multiple MCG spatiotemporal feature subsets.

5. The method for locating the source of abnormal cardiac pulsation according to claim 1, characterized in that, Obtain the MCG sensor coordinates and ECG electrode coordinates, and perform spatial alignment processing on the MCG sensor coordinates and ECG electrode coordinates based on a pre-set spatial coordinate mapping model, including: Based on a pre-set three-dimensional anatomical coordinate system, the coordinates of the MCG sensor and the coordinates of the ECG electrode are obtained; Based on the pre-set chest wall deformation degree, the target registration model is determined from multiple registration models; According to the target registration model, the MCG sensor coordinates and the ECG electrode coordinates are transformed into the same coordinate system, and the MCG sensor coordinates and the ECG electrode coordinates are spatially aligned.

6. The method for locating the source of abnormal cardiac pulsation according to claim 1, characterized in that, The method for obtaining the cardiac geometry and conductance model includes: Acquire cardiac imaging data; Based on a pre-set deep learning segmentation algorithm, multiple tissue structures are determined according to the cardiac imaging data; Based on a pre-set cardiac geometry mesh generation algorithm, a cardiac geometry model is generated according to multiple tissue structures. Based on pre-set conductivity parameters, a cardiac conductance model is generated according to multiple tissue structures.

7. A device for locating a source of abnormal heartbeats, characterized in that, include: The acquisition module is used to simultaneously acquire MCG and ECG signals and perform preprocessing. The feature extraction module is used to extract features from the preprocessed MCG signal and the ECG signal based on a pre-set feature extraction model, so as to obtain the ECG time series feature sequence and the MCG spatiotemporal feature set. The spatiotemporal alignment module is used to obtain the coordinates of the MCG sensor and the coordinates of the ECG electrode, and to perform spatial alignment processing on the coordinates of the MCG sensor and the coordinates of the ECG electrode based on a pre-set spatial coordinate mapping model. The fusion inversion module is used to determine the abnormal pulsation source based on a pre-set cardiac geometry and conductance model, according to the mapping relationship between the ECG time-series feature sequence, the MCG spatiotemporal feature set, the coordinates of the MCG sensor and the coordinates of the ECG electrode, through a Bayesian framework model, a deep learning model and a sparse inversion model. The fusion inversion module is further configured to, based on a pre-set cardiac geometry and conductance model, obtain a first abnormal pulsation source using the Bayesian framework model according to the mapping relationship between the ECG time-series feature sequence, the MCG spatiotemporal feature set, the MCG sensor coordinates, and the ECG electrode coordinates; based on the pre-set cardiac geometry and conductance model, obtain a second abnormal pulsation source using the deep learning model according to the mapping relationship between the ECG time-series feature sequence, the MCG spatiotemporal feature set, the MCG sensor coordinates, and the ECG electrode coordinates; based on the pre-set cardiac geometry and conductance model, obtain a third abnormal pulsation source using the sparse inversion model according to the mapping relationship between the ECG time-series feature sequence, the MCG spatiotemporal feature set, the MCG sensor coordinates, and the ECG electrode coordinates; and based on a pre-set voting mechanism, determine a target abnormal pulsation source according to the first, second, and third abnormal pulsation sources.

8. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method for locating the source of abnormal cardiac pulsation as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 8. When the computer program is executed by the processor, it implements the steps of the method for locating the source of abnormal cardiac pulsation as described in any one of claims 1 to 6.

Images