An electrocardio data processing method, device, equipment, medium and product

Abstract

This invention provides a method, apparatus, device, medium, and product for processing electrocardiogram (ECG) data. The method includes: acquiring target ECG data, which includes: a reference signal acquired by at least one reference electrode based on a zero-potential reference point and ECG signals acquired by spatial electrodes in at least three different body surface spatial sampling areas; inputting the target ECG data, the positions of the reference electrodes, and the spatial electrodes as input data into an ECG data processing model; constructing a three-dimensional ECG vector field based on the input data; and outputting corresponding risk warning results based on the activation-recovery interval parameters and stereo dispersion rate indices corresponding to each myocardial node. By configuring a zero-potential reference electrode and multi-directional spatial electrodes for collaborative acquisition, multi-channel signal calibration and alignment are effectively achieved; and stereo dispersion rate temporal trend analysis and dynamic risk scoring are implemented to improve the timeliness and stability of ECG risk warning output.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of medical diagnostic and treatment technology, and in particular to a method, apparatus, device, medium and product for processing electrocardiogram data. Background Technology

[0002] Electrocardiography (ECG) is an important non-invasive diagnostic method for cardiovascular disease screening and early warning of sudden cardiac death risk. Traditional ECG records the heart's electrical activity as a two-dimensional projection through surface leads, making it difficult to intuitively and quantitatively represent the conduction and diffusion processes in three-dimensional space. Increased heterogeneity in myocardial cell repolarization—i.e., "cardiac diffusion"—is a crucial predictor of malignant arrhythmias and sudden cardiac death. Traditional parameters such as QT dispersion (QTd) and T peak-T end interval (Tp-e) are based on limited two-dimensional leads and lack information in three-dimensional space.

[0003] Furthermore, traditional Holter monitoring systems can only record long-term, one-dimensional ECG signals and cannot provide three-dimensional spatial information on ECG diffusion, making it difficult to accurately and dynamically quantify the risk of sudden cardiac death. Summary of the Invention

[0004] This invention provides a method, apparatus, device, medium, and product for processing electrocardiogram (ECG) data to address the existing problem of difficulty in accurately and dynamically quantifying the risk of sudden cardiac death.

[0005] To solve the above-mentioned technical problems, the present invention is implemented as follows:

[0006] In a first aspect, embodiments of the present invention provide an electrocardiogram (ECG) data processing method, comprising:

[0007] Acquire target electrocardiogram (ECG) data, which includes: a reference signal acquired by at least one reference electrode based on a zero potential reference point and ECG signals acquired by spatial electrodes in at least three different body surface spatial sampling areas;

[0008] The target ECG data, the location of the reference electrode, and the location of the spatial electrode are input into the ECG data processing model. The ECG data processing model encodes the input data and uses a spatiotemporal joint attention module to complete channel alignment. At least one analysis window is divided for multi-level feature compression to obtain window-level latent features. A three-dimensional ECG vector field is constructed based on the latent features to obtain the activation-recovery interval parameters and the stereo diffusion rate index corresponding to each myocardial node. Based on the activation-recovery interval parameters and the stereo diffusion rate index corresponding to each myocardial node, the corresponding risk warning result is output.

[0009] Optionally, the electrocardiogram signals acquired by spatial electrodes in at least three different body surface spatial sampling areas include the following at least three types of body surface spatial sampling areas: precordial region, left chest wall, left posterolateral region, and left posterodorsolateral region.

[0010] Optionally, before inputting the target ECG data, the position of the reference electrode, and the position of the spatial electrode as input data into the ECG data processing model, the method further includes:

[0011] The target electrocardiogram data is preprocessed, including at least one of the following: high-pass filtering, wavelet denoising, notch filtering, amplitude normalization, artifact segment removal, severe dropout segment removal, low-quality window marking or quality downweighting, and R-peak alignment.

[0012] Optionally, the process of using a spatiotemporal joint attention module to perform channel alignment includes:

[0013] The time offset of each spatial channel is estimated based on the R-peak position of the reference signal, and a normalized scaling factor is generated based on the amplitude statistics of the reference signal. The attention weight is calculated and the channel alignment is completed using the channel features of the reference signal as the query and the channel features of the ECG signal as the key / value.

[0014] Optionally, the step of dividing at least one analysis window for multi-level feature compression to obtain window-level latent features includes:

[0015] The time-series segmentation module divides the data into at least one analysis window with a fixed window length and a sliding step size. The input data in each analysis window is encoded into multi-channel ECG signals to obtain window-level multi-channel features. The window-level multi-channel features are then compressed in at least two levels of time dimension and mapped in feature dimension to obtain window-level latent features.

[0016] Optionally, constructing a three-dimensional electrocardiogram vector field based on the latent features includes:

[0017] The global temporal features are extracted by mining long-term temporal dependencies between adjacent windows through a temporal modeling network. The temporal modeling network includes at least one of a long short-term memory network, a bidirectional long short-term memory network, a temporal Transformer, or a TCN. The global temporal features are then fused using a gating fusion mechanism to obtain fused features. The fused features are then reconstructed layer by layer and decoded and reconstructed at multiple scales using a bidirectional reconstruction decoder to build a three-dimensional electrocardiogram vector field.

[0018] Optionally, the bidirectional reconstruction decoder includes a forward 3D vector field reconstruction branch and a reverse virtual lead reconstruction branch. The forward 3D vector field reconstruction branch uses the fusion features as a low-scale latent representation, and sequentially performs upsampling, convolution / attention decoding, and jumper fusion to gradually restore the target temporal resolution, ultimately outputting a 3D ECG vector field. The reverse virtual lead reconstruction branch outputs virtual surface lead signals based on the fusion features, and generates consistency constraints based on the differences between the virtual surface lead signals and the labeled signals or the surface potentials obtained by forward modeling from the 3D ECG reconstruction results.

[0019] Optionally, the step of outputting corresponding risk warning results based on the activation-recovery interval parameters corresponding to each myocardial node and the stereo diffusion rate index includes:

[0020] The risk warning results include at least one of the following: warning of the absolute value threshold of the three-dimensional diffusion rate index, warning of the rate of change of the three-dimensional diffusion rate index, warning of the volatility of the three-dimensional diffusion rate index, warning of the sudden event of the three-dimensional diffusion rate index, and warning of the comprehensive dynamic risk score.

[0021] The three-dimensional diffusion rate index absolute value threshold early warning includes: a three-dimensional diffusion rate index less than a first threshold is normal and routine monitoring is performed; a three-dimensional diffusion rate index not less than the first threshold and less than a second threshold is low risk, requiring attention and increasing monitoring frequency; a three-dimensional diffusion rate index not less than the second threshold and less than a third threshold is medium risk, issuing an alert and outputting a review prompt message; a three-dimensional diffusion rate index not less than the third threshold is high risk, issuing an emergency early warning and outputting an alarm control signal or review prompt message; the first threshold is less than the second threshold, and the second threshold is less than the third threshold.

[0022] The warning for the rate of change of the three-dimensional dispersion index is triggered when the three-dimensional dispersion index is greater than the fourth threshold.

[0023] The three-dimensional dispersion rate index fluctuation warning is raised when the coefficient of variation is greater than the fifth threshold and continues to exceed the preset time threshold.

[0024] The three-dimensional dispersion rate index sudden change event warning is triggered when the three-dimensional dispersion rate index changes by a step greater than the sixth threshold, and the event is immediately warned and recorded.

[0025] The comprehensive dynamic risk score early warning system outputs an alarm control signal or a review prompt when the comprehensive dynamic risk score exceeds the seventh threshold.

[0026] Optionally, the formula for calculating the stereo dispersion index is:

[0027] SDR = ;

[0028] Where ARI_i is the activation-recovery time interval of the i-th myocardial node; ARI_mean is the mean of ARI across all nodes; and N is the total number of myocardial nodes.

[0029] Optionally, before inputting the target ECG data, the position of the reference electrode, and the position of the spatial electrode as input data into the ECG data processing model, the method further includes:

[0030] Training the ECG data processing model specifically includes:

[0031] Acquire training data, which includes publicly available electrocardiogram (ECG) data, high-density surface ECG mapping data, ECG waveform data, and / or three-dimensional electrical activity annotation data generated based on the human transmission model; wherein, the publicly available ECG data is used for pre-training and is not required to have three-dimensional electrical activity field supervision labels;

[0032] In the first training phase, the publicly available ECG data is input into the ECG data processing model to be trained, and pre-training tasks such as waveform reconstruction, mask prediction, rhythm classification, and / or R-peak localization are performed to obtain a pre-trained ECG data processing model. The ECG data processing model to be trained learns the nonlinear correlation between signals acquired by finite electrodes and the three-dimensional electrical activity characteristics of the heart. It encodes the ECG data samples and uses a spatiotemporal joint attention module to complete channel alignment. At least one analysis window is divided for multi-level feature compression to obtain window-level latent features. A three-dimensional ECG vector field is constructed based on the latent features to complete the initial construction of the model's basic feature extraction, spatiotemporal correlation fitting, and three-dimensional electrical activity reconstruction capabilities, obtaining the predicted activation-recovery interval parameters and stereo diffusion rate index corresponding to each myocardial node. Based on the predicted activation-recovery interval parameters and stereo diffusion rate index corresponding to each myocardial node, the supervision label of the three-dimensional electrical activity field corresponding to the first ECG data sample, and the loss function, the parameters of the ECG data processing model to be trained are optimized to obtain the pre-trained ECG data processing model.

[0033] In the second training phase, the pre-trained ECG data processing model is subjected to 3D reconstruction supervised training using the high-density surface ECG mapping data, ECG waveform data, and / or 3D electrical activity annotation data generated based on the human transmission model. In the third training phase, the risk warning output is calibrated using follow-up outcomes or expert review labels to obtain the trained ECG data processing model. The parameters of the pre-trained ECG data processing model are then further optimized to obtain the trained ECG data processing model.

[0034] Optionally, the ECG data processing model further aligns and averages the QRS complex based on the reference signal, and extracts high-frequency components in the 150-250Hz frequency band to construct high-frequency QRS auxiliary features. The high-frequency QRS auxiliary features include a high-frequency morphology index and / or a reduced amplitude region, and are jointly output with the excitation-recovery interval parameter and the stereo diffusion rate index to provide risk warning results.

[0035] In a second aspect, embodiments of the present invention provide an electrocardiogram (ECG) data processing device, comprising:

[0036] The acquisition module is used to acquire target electrocardiogram (ECG) data, which includes: a reference signal acquired by at least one reference electrode based on a zero potential reference point and ECG signals acquired by spatial electrodes in at least three different body surface spatial sampling areas.

[0037] The first processing module is used to input the target ECG data, the position of the reference electrode, and the position of the spatial electrode as input data into the ECG data processing model. The ECG data processing model encodes the input data and uses a spatiotemporal joint attention module to complete channel alignment; divides at least one analysis window for multi-level feature compression to obtain window-level latent features; and constructs a three-dimensional ECG vector field based on the latent features to obtain the activation-recovery interval parameters and the stereo diffusion rate index corresponding to each myocardial node. Based on the activation-recovery interval parameters and the stereo diffusion rate index corresponding to each myocardial node, the module outputs the corresponding risk warning result.

[0038] Optionally, the electrocardiogram signals acquired by spatial electrodes in at least three different body surface spatial sampling areas include the following at least three types of body surface spatial sampling areas: precordial region, left chest wall, left posterolateral region, and left posterodorsolateral region.

[0039] Optional, also includes:

[0040] The preprocessing module is used to perform data preprocessing on the target electrocardiogram data. The data preprocessing includes at least one of the following: high-pass filtering, wavelet denoising, notch filtering, amplitude normalization, artifact segment removal, severe dropout segment removal, low-quality window marking or quality downweighting, and R-peak alignment.

[0041] Optionally, the first processing module includes:

[0042] The first processing submodule is used to estimate the time offset of each spatial channel based on the R-peak position of the reference signal, and generate a normalized scaling factor based on the amplitude statistics of the reference signal; using the channel features of the reference signal as the query and the channel features of the electrocardiogram signal as the key / value, it calculates the attention weight and completes the channel alignment.

[0043] Optionally, the first processing module includes:

[0044] The second processing submodule is used to divide at least one analysis window according to a fixed window length and sliding step size through the windowing time-series segmentation module; to encode the input data in each analysis window into multi-channel ECG signals to obtain window-level multi-channel features; and to perform at least two levels of time dimension compression and feature dimension mapping on the window-level multi-channel features to obtain window-level latent features.

[0045] Optionally, the first processing module includes:

[0046] The third processing submodule is used to mine long-term temporal dependencies between adjacent windows through a temporal modeling network and extract global temporal features. The temporal modeling network includes at least one of a long short-term memory network, a bidirectional long short-term memory network, a temporal Transformer, or a TCN. The global temporal features are fused using a gating fusion mechanism to obtain fused features. The fused features are then restored layer by layer and decoded and reconstructed at multiple scales using a bidirectional reconstruction decoder to construct a three-dimensional electrocardiogram vector field.

[0047] Optionally, the bidirectional reconstruction decoder includes a forward 3D vector field reconstruction branch and a reverse virtual lead reconstruction branch. The forward 3D vector field reconstruction branch uses the fusion features as a low-scale latent representation, and sequentially performs upsampling, convolution / attention decoding, and jumper fusion to gradually restore the target temporal resolution, ultimately outputting a 3D ECG vector field. The reverse virtual lead reconstruction branch outputs virtual surface lead signals based on the fusion features, and generates consistency constraints based on the differences between the virtual surface lead signals and the labeled signals or the surface potentials obtained by forward modeling from the 3D ECG reconstruction results.

[0048] Optionally, the first processing module includes:

[0049] The fourth processing submodule is used to provide risk warning results that include at least one of the following: absolute value threshold warning for the three-dimensional diffusion rate index, rate of change warning for the three-dimensional diffusion rate index, volatility warning for the three-dimensional diffusion rate index, sudden event warning for the three-dimensional diffusion rate index, and comprehensive dynamic risk score warning.

[0050] The three-dimensional diffusion rate index absolute value threshold early warning includes: a three-dimensional diffusion rate index less than a first threshold is normal and routine monitoring is performed; a three-dimensional diffusion rate index not less than the first threshold and less than a second threshold is low risk, requiring attention and increasing monitoring frequency; a three-dimensional diffusion rate index not less than the second threshold and less than a third threshold is medium risk, issuing an alert and outputting a review prompt message; a three-dimensional diffusion rate index not less than the third threshold is high risk, issuing an emergency early warning and outputting an alarm control signal or review prompt message; the first threshold is less than the second threshold, and the second threshold is less than the third threshold.

[0051] The warning for the rate of change of the three-dimensional dispersion index is triggered when the three-dimensional dispersion index is greater than the fourth threshold.

[0052] The three-dimensional dispersion rate index fluctuation warning is raised when the coefficient of variation is greater than the fifth threshold and continues to exceed the preset time threshold.

[0053] The three-dimensional dispersion rate index sudden change event warning is triggered when the three-dimensional dispersion rate index changes by a step greater than the sixth threshold, and the event is immediately warned and recorded.

[0054] The comprehensive dynamic risk score early warning system outputs an alarm control signal or a review prompt when the comprehensive dynamic risk score exceeds the seventh threshold.

[0055] Optionally, the formula for calculating the stereo dispersion index is:

[0056] SDR = ;

[0057] Where ARI_i is the activation-recovery time interval of the i-th myocardial node; ARI_mean is the mean of ARI across all nodes; and N is the total number of myocardial nodes.

[0058] Optional, also includes:

[0059] The model training module is used to train the ECG data processing model, and specifically includes:

[0060] Acquire training data, which includes publicly available electrocardiogram (ECG) data, high-density surface ECG mapping data, ECG waveform data, and / or three-dimensional electrical activity annotation data generated based on the human transmission model; wherein, the publicly available ECG data is used for pre-training and is not required to have three-dimensional electrical activity field supervision labels;

[0061] In the first training phase, the publicly available ECG data is input into the ECG data processing model to be trained, and pre-training tasks such as waveform reconstruction, mask prediction, rhythm classification, and / or R-peak localization are performed to obtain a pre-trained ECG data processing model. The ECG data processing model to be trained learns the nonlinear correlation between signals acquired by finite electrodes and the three-dimensional electrical activity characteristics of the heart. It encodes the ECG data samples and uses a spatiotemporal joint attention module to complete channel alignment. At least one analysis window is divided for multi-level feature compression to obtain window-level latent features. A three-dimensional ECG vector field is constructed based on the latent features to complete the initial construction of the model's basic feature extraction, spatiotemporal correlation fitting, and three-dimensional electrical activity reconstruction capabilities, obtaining the predicted activation-recovery interval parameters and stereo diffusion rate index corresponding to each myocardial node. Based on the predicted activation-recovery interval parameters and stereo diffusion rate index corresponding to each myocardial node, the supervision label of the three-dimensional electrical activity field corresponding to the first ECG data sample, and the loss function, the parameters of the ECG data processing model to be trained are optimized to obtain the pre-trained ECG data processing model.

[0062] In the second training phase, the pre-trained ECG data processing model is subjected to 3D reconstruction supervised training using the high-density surface ECG mapping data, ECG waveform data, and / or 3D electrical activity annotation data generated based on the human transmission model. In the third training phase, the risk warning output is calibrated using follow-up outcomes or expert review labels to obtain the trained ECG data processing model. The parameters of the pre-trained ECG data processing model are then further optimized to obtain the trained ECG data processing model.

[0063] Thirdly, embodiments of the present invention provide an electronic device, including a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of the electrocardiogram data processing method as described in any one of the first aspects.

[0064] Fourthly, embodiments of the present invention provide a readable storage medium storing a program or instructions that, when executed by a processor, implement the steps of the electrocardiogram data processing method as described in any one of the first aspects.

[0065] Fifthly, embodiments of the present invention provide a computer program product including computer instructions that, when executed by a processor, implement the steps of the electrocardiogram data processing method as described in any one of the first aspects.

[0066] In this invention, a multi-channel ECG acquisition system adapted to three-dimensional ECG reconstruction is constructed by configuring a zero-potential reference electrode and a multi-directional spatial electrode for collaborative acquisition, effectively completing multi-channel signal calibration and alignment. Furthermore, at the algorithm level, windowed local feature extraction is employed, balancing the accuracy of local ECG detail representation with the global trend of long-term dynamic evolution. In terms of the evaluation system, a novel multi-dimensional ECG repolarization evaluation system is constructed by combining the three-dimensional diffusion rate quantification index with the activation-recovery interval parameters corresponding to each myocardial node. This enables temporal trend analysis of the three-dimensional diffusion rate and dynamic risk scoring, accurately capturing occult, regional, and progressive myocardial electrical activity disturbances, significantly improving the timeliness and stability of ECG risk warning output, and solving the existing problem of difficulty in accurately and dynamically quantifying the risk of sudden cardiac death. Attached Figure Description

[0067] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0068] Figure 1 This is a flowchart of an electrocardiogram (ECG) data processing method provided in an embodiment of the present invention;

[0069] Figure 2 This is a schematic diagram of the electrode distribution of an electrocardiogram data processing method provided in an embodiment of the present invention;

[0070] Figure 3 This is a schematic diagram of the electrode distribution for another electrocardiogram data processing method provided in an embodiment of the present invention;

[0071] Figure 4 This is a flowchart of the attention-guided module of an electrocardiogram data processing method provided in an embodiment of the present invention;

[0072] Figure 5 This is a static flowchart of an electrocardiogram data processing method provided in an embodiment of the present invention;

[0073] Figure 6 This is a dynamic mode flowchart of an electrocardiogram data processing method provided in an embodiment of the present invention;

[0074] Figure 7 This is a flowchart of a bidirectional decoding method for electrocardiogram data processing provided in an embodiment of the present invention;

[0075] Figure 8 This is a schematic diagram of the structure of an electrocardiogram data processing device provided in an embodiment of the present invention;

[0076] Figure 9This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0077] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0078] Please refer to Figure 1 This invention provides a method for processing electrocardiogram (ECG) data, including:

[0079] Step 11: Acquire target ECG data, which includes: a reference signal acquired by at least one reference electrode based on a zero potential reference point and ECG signals acquired by spatial electrodes in at least three different body surface spatial sampling areas;

[0080] In this embodiment of the invention, a reference reference signal is acquired by setting a zero-potential reference electrode, and multiple spatial electrocardiogram (ECG) signals are simultaneously acquired by spatial electrodes in at least three different body surface spatial sampling areas. This constructs an ECG acquisition system with a unified potential reference and spatial orientation dimension. It not only relies on the zero-potential reference to ensure the stability and uniformity of ECG signal acquisition and avoids the problems of phase shift and baseline disorder caused by the lack of a unified reference for multi-channel signals, but also effectively captures the electrical activity change information of the heart in different spatial orientations. While reducing the complexity of hardware deployment and usage costs, it provides comprehensive, reliable, and spatially dimensional raw data support for subsequent multi-channel signal alignment, spatiotemporal feature fusion, three-dimensional ECG solution, and accurate reconstruction of the three-dimensional electrical activity field.

[0081] In this embodiment of the invention, optionally, the electrocardiogram signals obtained by spatial electrodes in at least three different body surface spatial sampling areas include the following at least three types of body surface spatial sampling areas: the precordial region, the left chest wall, the left posterolateral region, and the left posterodorsolateral region.

[0082] In this embodiment of the invention, spatial electrocardiogram (ECG) signals are acquired by selecting body surface areas covering the anterior, lateral, and posterior multidimensional anatomical positions of the heart, such as the precordial region, left chest wall, left posterolateral region, and left posterodorsolateral region. This comprehensively captures electrical activity conduction and repolarization changes at different anatomical locations of the heart, effectively compensating for the shortcomings of unilateral acquisition methods, such as missing spatial information and missed local ECG abnormalities. It also enriches the spatial dimensional features of ECG signals and accurately covers the main areas of electrical activity dispersion in the heart.

[0083] In some embodiments, R-XYZ is used to represent a spatial sampling label on the body surface, and is not limited to a strict Cartesian coordinate axis; where X corresponds to the precordial region sampling area, Y corresponds to the left chest wall sampling area, and Z corresponds to the left posterolateral or left posterodorsolateral sampling area; the upper and lower extension electrodes are used to provide supplementary sampling information in the vertical direction of the body surface. The X-axis corresponds to the sagittal axis (anteroposterior direction), the Y-axis corresponds to the coronal axis (lateral direction), the Z-axis corresponds to the vertical axis (head-to-foot direction), and the R-axis is the reference axis. In specific implementations, the number of spatial electrodes can be increased based on the R-XYZ coordinate system according to actual conditions; please refer to [reference needed]. Figure 2 and Figure 3 The reference electrode R is an electrical zero-potential reference point, employing bidirectional signal capture, and can be set near the 2nd intercostal space on the right midclavicular line, i.e., V1 (the 4th intercostal space beside the right sternum); the spatial electrode X captures electrical activity in the anterior-posterior direction and can be set at the V3 lead position on the chest (located between V2 and V4, with V2 located in the 4th intercostal space beside the left sternum and V4 located in the 5th intercostal space on the left midclavicular line); the spatial electrode Y captures electrical activity in the lateral direction and can be set at the V7 lead position on the posterior axillary line (located on the left posterior axillary line, at the same level as V4, i.e., the 5th intercostal space on the left posterior axillary line); the spatial electrode Z captures electrical activity in the vertical direction and can be set at the V9 lead position beside the spine on the back (located in the back region between the inferior angle of the left scapula and the spine, at the same level as V4, i.e., the 5th intercostal space beside the left scapula).

[0084] The R reference electrode and the XYZ space electrodes form three differential pairs. Six-directional signal capture is achieved through basic data acquisition and reverse signal derivation. Based on the differential measurement principle, the derivation formula for the reverse signal is as follows:

[0085] V_{-X}(t) = 2*V_R(t)- V_X(t);

[0086] V_{-Y}(t) = 2*V_ R(t)- V_ Y(t);

[0087] V_{-Z}(t) = 2*V_ R(t)- V_ Z(t);

[0088] By employing an acquisition architecture that includes a zero-potential reference electrode and multi-directional spatial electrodes, only a small number of electrodes are needed to obtain six-directional spatial ECG signals through signal differential mechanisms. This greatly simplifies the acquisition hardware structure and deployment process, reducing the complexity of wearing and using the equipment. Simultaneously, the introduction of an independent zero-potential reference electrode constructs a standardized potential benchmark system, enabling bidirectional precise capture and reverse inference correction of ECG signals, ensuring the accuracy and stability of the acquired signals. Furthermore, this acquisition architecture possesses excellent general scalability, flexibly adaptable to configurations ranging from a basic 4-electrode setup to an extended N-electrode setup. It can be combined with upper and lower electrode linkage matching schemes to further refine the spatial dimensions of cardiac ECG acquisition and improve the spatial resolution of ECG signal acquisition. The acquisition configuration can be adaptively adjusted according to different detection scenarios and accuracy requirements. While maintaining the advantages of lightweight and low-cost deployment, it provides flexible, dimension-rich, and highly reliable ECG data support for subsequent high-precision 3D ECG modeling, spatial electrical activity feature analysis, and risk assessment.

[0089] Step 12: Input the target ECG data, the position of the reference electrode, and the position of the spatial electrode into the ECG data processing model. The ECG data processing model encodes the input data and uses a spatiotemporal joint attention module to complete channel alignment. Divide at least one analysis window for multi-level feature compression to obtain window-level latent features. Construct a three-dimensional ECG vector field based on the latent features to obtain the activation-recovery interval parameters and the stereo diffusion rate index corresponding to each myocardial node. Output the corresponding risk warning results based on the activation-recovery interval parameters and the stereo diffusion rate index corresponding to each myocardial node.

[0090] In this embodiment of the invention, electrocardiogram (ECG) signal data and electrode spatial location information are fused into the input model for joint encoding. Attention is guided by a reference signal to achieve precise multi-channel alignment, effectively eliminating temporal offset and spatial feature misalignment issues in multi-channel ECG signals. This ensures the spatiotemporal consistency of multi-source input features. By windowing and multi-level feature compression of the temporal data, redundant noise information in the original ECG data is effectively removed, thereby extracting window-level potential features with superior representation capabilities. Based on these potential features, a three-dimensional ECG vector field reconstruction is completed, enabling a refined restoration of the spatial conduction and temporal evolution of cardiac electrical activity across the entire cardiac domain. This allows for precise calculation of the activation-recovery interval parameters and overall three-dimensional diffusion rate index for each myocardial node, achieving a three-dimensional, refined, and quantitative assessment of the local repolarization state of the myocardium and the degree of spatial electrical activity diffusion disorder. This effectively captures regional and occult myocardial electrical activity abnormalities that are difficult to identify using traditional ECG assessment methods, significantly improving the accuracy, comprehensiveness, and reliability of dynamic risk warning for sudden cardiac death.

[0091] In some embodiments, the input data can be represented as (1, 4, 5000), where 1 represents one record input at a time during inference, 4 represents the four basic channels R, X, Y, and Z, and 5000 represents the number of sampling points for 10 seconds × 500Hz. In practical applications, if the hardware uses upper and lower extended electrodes, the input can also include extended channels of the same electrode array, and enter spatial position encoding through upper and lower level labels. Encoding based on the input data can be composed of a multi-channel ECG signal encoder, a spatial position encoder, and a multi-level feature compression module. The multi-channel ECG signal encoder can use 1D convolution, normalization layers, channel attention, and residual connections to extract waveform features. The spatial position encoder maps the chest leads of the R, X, Y, and Z electrodes to coordinates. The circumferential chest location labels, hierarchical labels, and channel quality scores are converted into location / quality vectors and superimposed on the signal features. Subsequently, the time length is gradually reduced and the feature dimension is increased through multi-level compression blocks. For example, the original window features are compressed into multi-scale latent features at 1 / 2, 1 / 4, and 1 / 8 scales for subsequent time series modeling and reconstruction decoding. The reference signal R and the XYZ ECG signals can be encoded separately. Since the reference signal R is closer to the reference benchmark and calibration anchor, it can be used for quality assessment, amplitude normalization, and elimination of individual differences. The XYZ ECG signals carry the spatial sampling information of the precordial region, left chest wall, and left posterolateral / dorsal region. The physical meaning of each channel is preserved from the encoding stage, and features at different scales are preserved during the compression process for subsequent jump-connection reconstruction.

[0092] In this embodiment of the invention, a multi-channel ECG acquisition system adapted to three-dimensional ECG reconstruction is constructed by configuring a zero-potential reference electrode and a multi-directional spatial electrode for collaborative acquisition, effectively completing the calibration and alignment of multi-channel signals. Furthermore, at the algorithm level, windowed local feature extraction is employed, balancing the accuracy of local ECG detail representation with the global trend of long-term dynamic evolution. In terms of the evaluation system, a novel multi-dimensional ECG repolarization evaluation system is constructed by using a three-dimensional diffusion rate quantification index combined with the activation-recovery interval parameters corresponding to each myocardial node. This enables temporal trend analysis of the three-dimensional diffusion rate and dynamic risk scoring, accurately capturing occult, regional, and progressive myocardial electrical activity disturbances, and significantly improving the timeliness and stability of ECG risk warning output.

[0093] In this embodiment of the invention, optionally, before inputting the target ECG data, the position of the reference electrode, and the position of the spatial electrode as input data into the ECG data processing model, the method further includes:

[0094] The target electrocardiogram data is preprocessed, including at least one of the following: high-pass filtering, wavelet denoising, notch filtering, amplitude normalization, artifact segment removal, severe dropout segment removal, low-quality window marking or quality downweighting, and R-peak alignment.

[0095] In this embodiment of the invention, before model input, the target ECG data undergoes multi-dimensional refined preprocessing operations, including 0.5 Hz high-pass filtering, wavelet denoising, 50 / 60 Hz notch filtering, amplitude normalization, artifact segment removal, severe missing segment removal, low-quality window marking or quality downweighting, and R-peak alignment. This effectively removes various noises such as baseline drift, high-frequency noise, and power frequency interference from ECG acquisition, effectively eliminates invalid and abnormal data segments, unifies the amplitude scale and timing reference of ECG signals, improves the purity and effectiveness of ECG data, avoids modeling deviations caused by acquisition noise and data differences from the source, and improves the accuracy of ECG feature extraction and the stability of risk warning results.

[0096] In this embodiment of the invention, optionally, the step of using a spatiotemporal joint attention module to complete channel alignment includes:

[0097] The time offset of each spatial channel is estimated based on the R-peak position of the reference signal, and a normalized scaling factor is generated based on the amplitude statistics of the reference signal. The attention weight is calculated and the channel alignment is completed using the channel features of the reference signal as the query and the channel features of the ECG signal as the key / value.

[0098] In this embodiment of the invention, the encoded features are stacked in multiple layers using a spatiotemporal joint attention module, for example, 6 layers; each layer includes temporal self-attention, spatial cross-attention, and R-reference guided attention. Temporal self-attention focuses on the changes of the same channel or the same region over time, spatial cross-attention focuses on the correlation between different directions or different spatial locations at the same moment; R-reference guided attention uses the R channel features as the query benchmark or alignment anchor point to guide the XYZ channel features to perform amplitude normalization, temporal alignment, and attention weighting.

[0099] Due to differences in patient body size, application location, and skin impedance, the amplitude and phase of the XYZ signals may vary. The R channel provides a unified reference, allowing each spatial channel to undergo baseline calibration before entering multi-stage compression and timing learning. Please refer to [reference needed]. Figure 4 Using reference signal channel features as a unified query benchmark and time-series alignment anchor, amplitude normalization, precise time-series alignment, and adaptive attention-weighted optimization are simultaneously implemented for the features of each spatial ECG channel. By standardizing amplitude scales, calibrating time-series positional deviations, and dynamically strengthening the weight of effective ECG features and suppressing ineffective interference features based on attention mechanisms, the temporal synchronization and feature correlation of ECG channel features in different spatial orientations are ensured. This effectively improves the accuracy and effectiveness of multi-source ECG spatial feature fusion, avoids modeling errors caused by channel inconsistency, and provides a multi-channel feature foundation for subsequent three-dimensional ECG vector field reconstruction and myocardial electrical activity parameter calculation, further enhancing the reliability and stability of overall ECG analysis and risk warning results.

[0100] In this embodiment of the invention, optionally, the step of dividing at least one analysis window for multi-level feature compression to obtain window-level latent features includes:

[0101] The time-series segmentation module divides the data into at least one analysis window with a fixed window length and a sliding step size. The input data in each analysis window is encoded into multi-channel ECG signals to obtain window-level multi-channel features. The window-level multi-channel features are then compressed in at least two levels of time dimension and mapped in feature dimension to obtain window-level latent features.

[0102] In this embodiment of the invention, the ECG time series data is finely segmented by a windowing time series segmentation module according to a fixed window length and a sliding step size, wherein, for example... Figure 5 As shown, in static mode, only one window signal is collected, typically a window signal of about 1 minute (e.g., 10 seconds) (5000 sampling points @ 500Hz), suitable for rapid screening in emergency and outpatient departments; Figure 6 As shown, in dynamic mode, continuous acquisitions are typically performed for 24-72 hours or even longer. Therefore, a windowing mechanism is adopted to segment the long-term ECG signal according to a fixed window length and step size. Each window is independently encoded, and global temporal modeling is achieved through a cross-window feature aggregation mechanism. Specifically, the window length W can be configured to a default of 10 seconds (5000 sampling points @ 500Hz), corresponding to approximately 10-12 cardiac cycles; the sliding step size S can be configured to a default of 2 seconds (1000 sampling points), with an 80% overlap rate between adjacent windows; each window independently passes through R-XYZ The encoder and spatiotemporal joint attention module extract local features; then the feature sequences of all windows are input into subsequent modules for global temporal modeling, which is suitable for inpatient monitoring, home remote monitoring, and long-term follow-up of high-risk groups of sudden death. The static mode takes a single 10-second R-XYZ window as input, and after encoding, compression and layer-by-layer reconstruction, it obtains a one-time ARI / SDR and risk results. The dynamic mode divides the continuous ECG according to the window length W and step size S, first generates H_comp compressed feature sequences window by window, and then uses LSTM or temporal Transformer to learn long-term temporal trends, outputting SDR time series, dynamic risk score and warning results. In other words, the static mode can also be regarded as a single-window special case of the dynamic mode.

[0103] By combining a multi-channel ECG signal encoder with a multi-level feature compression method to extract stable window-level latent features, it can adapt to the ECG data analysis needs in both static and dynamic modes. For the static mode of rapid screening in emergency and outpatient departments, it can meet the efficiency requirements of short-term rapid detection and immediate analysis screening. For the dynamic mode of inpatient monitoring, home remote monitoring, and long-term follow-up of high-risk groups, it can rely on the sliding window mechanism to perform uninterrupted temporal segmentation and real-time feature compression of 24 to 72-hour long-term continuous ECG data, realizing streaming and continuous feature extraction of long-term ECG data. At the same time, the multi-level feature compression mechanism can perform hierarchical purification of the spatiotemporal features after multi-channel encoding, retaining the core morphological features and spatial correlation information of ECG, filtering out noise and invalid redundant data, and generating window-level latent features with stronger stability, anti-interference and representation ability. This provides a high-quality feature foundation for subsequent long-term dependent modeling, three-dimensional ECG vector field reconstruction, and myocardial electrical activity risk parameter calculation.

[0104] In this embodiment of the invention, optionally, constructing a three-dimensional electrocardiogram vector field based on the latent features includes:

[0105] The global temporal features are extracted by mining long-term temporal dependencies between adjacent windows through a temporal modeling network. The temporal modeling network includes at least one of a long short-term memory network, a bidirectional long short-term memory network, a temporal Transformer, or a TCN. The global temporal features are then fused using a gating fusion mechanism to obtain fused features. The fused features are then reconstructed layer by layer and decoded and reconstructed at multiple scales using a bidirectional reconstruction decoder to build a three-dimensional electrocardiogram vector field.

[0106] In this embodiment of the invention, a Long Short-Term Memory (LSTM) module is introduced to mine the long-term temporal dependencies between potential features of adjacent analysis windows, accurately extracting global temporal features that reflect the dynamic trend of cardiac electrical activity. Then, a gated fusion mechanism is used to adaptively fuse the potential features of the window representing the short-term local ECG spatial morphology with the global temporal features representing the long-term dynamic evolution law, intelligently balancing the weight ratio of short-term detailed features and long-term trend features. Furthermore, a bidirectional reconstruction decoder is used to perform layer-by-layer feature restoration and multi-scale decoding reconstruction of the fused features, combining multi-scale feature information to achieve refined spatiotemporal feature restoration, efficiently completing the modeling and construction of the three-dimensional ECG vector field. This enables the three-dimensional, continuous, and dynamic restoration of the spatial conduction distribution and temporal evolution process of cardiac electrical activity, providing reliable and spatiotemporally complete three-dimensional ECG field data support for subsequent high-precision calculation of myocardial node-level excitation recovery interval parameters, three-dimensional diffusion rate indicators, and dynamic assessment and early warning of sudden death risk, significantly improving the refinement of ECG abnormality identification and the overall accuracy of dynamic risk assessment.

[0107] Specifically, the Long Short-Term Memory (LSTM) network module is used to capture long-term feature dependencies of ECG signals at minute or even hourly scales. The core design of the LSTM module includes: inputting the window-level feature sequence output by the windowing module into a bidirectional LSTM layer in chronological order, with each time step corresponding to a feature vector of a window; employing a 2-4 layer bidirectional LSTM with the hidden dimension consistent with the Transformer feature dimension (d_model=256), simultaneously capturing forward and backward temporal dependencies; applying an attention pooling mechanism to the LSTM output sequence to adaptively focus on the time period with the most diagnostic value (such as the time period when abnormal ECG events occur); and integrating the global temporal features output by the LSTM with the local spatiotemporal features output by the Transformer through a gating fusion mechanism, taking into account both local accuracy and global trends.

[0108] The formula for the gating fusion mechanism is:

[0109] g =σ(W_g * [H_local; H_global] + b_g);

[0110] H_fused = g⊙H_global + (1 - g)⊙H_local;

[0111] Where H_local represents the local spatiotemporal features of the Transformer, H_global represents the global temporal features of the LSTM, g represents the gating vector, σ represents the Sigmoid function, and ⊙ represents element-wise multiplication. The gating fusion mechanism enables the model to adaptively balance local fine reconstruction and global trend capture.

[0112] In this embodiment of the invention, optionally, the bidirectional reconstruction decoder includes a forward three-dimensional vector field reconstruction branch and a reverse virtual lead reconstruction branch; the forward three-dimensional vector field reconstruction branch uses the fusion feature as a low-scale latent representation, and sequentially performs upsampling, convolution / attention decoding, and jumper fusion to gradually restore to the target temporal resolution, and finally outputs a three-dimensional ECG vector field; the reverse virtual lead reconstruction branch outputs virtual body surface lead signals according to the fusion feature, and generates consistency constraints based on the difference between the virtual body surface lead signals and the labeled signals or the body surface potentials obtained by forward modeling from the three-dimensional ECG reconstruction results.

[0113] In embodiments of the present invention, such as Figure 7As shown, the decoding end does not simply classify and output, but rather performs layer-by-layer reconstruction based on the fused feature H_fused and the multi-scale ECG feature copies H_skip retained during the encoding stage. The forward 3D vector field reconstruction branch recovers the 3D ECG vector field through multi-level upsampling and skip fusion, while the reverse virtual lead reconstruction branch reconstructs the virtual 64-lead body surface potential. During model training, Boundary Element Model (BEM) 3D labels, real or simulated dense lead labels, and consistency constraints are used for supervision, corresponding the front-end multi-level feature compression with the back-end layer-by-layer reconstruction, approximating the encoding-decoding approach of U-Net. By constructing a bidirectional reconstruction decoder containing both a forward 3D vector field reconstruction branch and a reverse virtual lead reconstruction branch, multi-scale, multi-objective collaborative decoding optimization based on fused features is achieved. This effectively solves the problems of ill-posed solutions to the ECG inverse problem and large model reconstruction errors. Simultaneously, through the collaborative optimization of the bidirectional branches, the model's learning ability of ECG spatiotemporal features is strengthened, improving the consistency between 3D reconstruction accuracy and body surface potential restoration, and enhancing the overall model's generalization ability and evaluation accuracy.

[0114] In this embodiment of the invention, optionally, the step of outputting a corresponding risk warning result based on the activation-recovery interval parameter corresponding to each myocardial node and the stereo diffusion rate index includes:

[0115] The risk warning results include at least one of the following: warning of the absolute value threshold of the three-dimensional diffusion rate index, warning of the rate of change of the three-dimensional diffusion rate index, warning of the volatility of the three-dimensional diffusion rate index, warning of the sudden event of the three-dimensional diffusion rate index, and warning of the comprehensive dynamic risk score.

[0116] The three-dimensional diffusion rate index absolute value threshold early warning includes: a three-dimensional diffusion rate index less than a first threshold is normal and routine monitoring is performed; a three-dimensional diffusion rate index not less than the first threshold and less than a second threshold is low risk, requiring attention and increasing monitoring frequency; a three-dimensional diffusion rate index not less than the second threshold and less than a third threshold is medium risk, issuing an alert and outputting a review prompt message; a three-dimensional diffusion rate index not less than the third threshold is high risk, issuing an emergency early warning and outputting an alarm control signal or review prompt message; the first threshold is less than the second threshold, and the second threshold is less than the third threshold.

[0117] The warning for the rate of change of the three-dimensional dispersion index is triggered when the three-dimensional dispersion index is greater than the fourth threshold.

[0118] The three-dimensional dispersion rate index fluctuation warning is raised when the coefficient of variation is greater than the fifth threshold and continues to exceed the preset time threshold.

[0119] The three-dimensional dispersion rate index sudden change event warning is triggered when the three-dimensional dispersion rate index changes by a step greater than the sixth threshold, and the event is immediately warned and recorded.

[0120] The comprehensive dynamic risk score early warning system outputs an alarm control signal or a review prompt when the comprehensive dynamic risk score exceeds the seventh threshold.

[0121] In this embodiment of the invention, optionally, the formula for calculating the stereo dispersion index is:

[0122] SDR = ;

[0123] Where ARI_i is the activation-recovery time interval of the i-th myocardial node; ARI_mean is the mean of ARI across all nodes; and N is the total number of myocardial nodes.

[0124] In this embodiment of the invention, a three-dimensional and quantitative assessment of cardiac electrical activity repolarization heterogeneity is achieved through a spatial dispersion rate (SDR) index constructed based on the activation-recovery interval (ARI) parameters of each myocardial node. By employing a multi-dimensional early warning mechanism encompassing SDR absolute value thresholds, rate of change, volatility, mutation events, and a comprehensive dynamic risk score, a comprehensive sudden death risk assessment system covering static levels, dynamic trends, short-term fluctuations, and sudden abnormalities is established. This system enables stratified early warning and differentiated intervention guidance for electrocardiographic states at different risk levels, achieving precise hierarchical management from routine monitoring and follow-up to emergency medical treatment. The SDR index, through myocardial node-level ARI... By statistically modeling the spatial distribution of parameters, the spatial distribution and dynamic evolution characteristics of repolarization dispersion across the entire cardiac domain can be accurately quantified. This effectively captures occult, regional, and progressive myocardial electrical activity disturbances, significantly improving the sensitivity of identifying early and occult risks of sudden death. Combined with trend and fluctuation analysis in long-term ECG monitoring scenarios, it can dynamically track the changes in the degree of repolarization dispersion and provide early warnings of high-risk trends. This overcomes the shortcomings of traditional static single-assessment, which is prone to missing dynamic deterioration processes, and significantly improves the comprehensiveness, accuracy, and clinical applicability of sudden death risk warning.

[0125] In this embodiment of the invention, optionally, before inputting the target ECG data, the position of the reference electrode, and the position of the spatial electrode as input data into the ECG data processing model, the method further includes:

[0126] Training the ECG data processing model specifically includes:

[0127] Acquire training data, which includes publicly available electrocardiogram (ECG) data, high-density surface ECG mapping data, ECG waveform data, and / or three-dimensional electrical activity annotation data generated based on the human transmission model; wherein, the publicly available ECG data is used for pre-training and is not required to have three-dimensional electrical activity field supervision labels;

[0128] In the first training phase, the publicly available ECG data is input into the ECG data processing model to be trained, and pre-training tasks such as waveform reconstruction, mask prediction, rhythm classification, and / or R-peak localization are performed to obtain a pre-trained ECG data processing model. The ECG data processing model to be trained learns the nonlinear correlation between signals acquired by finite electrodes and the three-dimensional electrical activity characteristics of the heart. It encodes the ECG data samples and uses a spatiotemporal joint attention module to complete channel alignment. At least one analysis window is divided for multi-level feature compression to obtain window-level latent features. A three-dimensional ECG vector field is constructed based on the latent features to complete the initial construction of the model's basic feature extraction, spatiotemporal correlation fitting, and three-dimensional electrical activity reconstruction capabilities, obtaining the predicted activation-recovery interval parameters and stereo diffusion rate index corresponding to each myocardial node. Based on the predicted activation-recovery interval parameters and stereo diffusion rate index corresponding to each myocardial node, the supervision label of the three-dimensional electrical activity field corresponding to the first ECG data sample, and the loss function, the parameters of the ECG data processing model to be trained are optimized to obtain the pre-trained ECG data processing model.

[0129] In the second training phase, the pre-trained ECG data processing model is subjected to 3D reconstruction supervised training using the high-density surface ECG mapping data, ECG waveform data, and / or 3D electrical activity annotation data generated based on the human transmission model. In the third training phase, the risk warning output is calibrated using follow-up outcomes or expert review labels to obtain the trained ECG data processing model. The parameters of the pre-trained ECG data processing model are then further optimized to obtain the trained ECG data processing model.

[0130] In some embodiments, the raw signals are acquired as training data, specifically including: the raw ECG signals of the R reference channel and the XYZ spatial channel, and the chest lead mapping position, chest ring position, upper and lower level information, sampling time and channel quality information of each channel are acquired simultaneously.

[0131] The original signal is subjected to sampling rate unification, channel order standardization, polarity correction, baseline drift suppression, power frequency interference suppression, electromyographic noise suppression, and R-peak localization in sequence. The XYZ spatial channels are time-aligned and amplitude-normalized using the R channel as a reference. Specifically: Unifying the sampling rate involves resampling the sampling rates of different devices or batches to a preset frequency, such as 500Hz; if the original sampling rate is retained, it is recorded in the data packet and converted to an equivalent time scale at the model input layer; unifying the channel order involves arranging the channels in a fixed order of R, X, Y, Z, and extended channels, with missing channels marked by a mask, and zero values ​​are not used to impersonate real signals; unifying the units and polarity involves converting millivolts, microvolts, etc., to a unified unit, and correcting channels with reversed connections or inconsistent signs according to the device polarity table; unifying the time reference involves dynamically recording data sorted by acquisition timestamps, correcting breakpoints, repetitive segments, and time drift; breakpoints exceeding the allowable interval are divided into independent segments.

[0132] For low-quality channels or low-quality time periods, a quality mask and quality score are generated; severely detached or saturated windows are removed, while mildly low-quality windows are retained but downweighted in the training loss; specifically, quality evaluation is performed simultaneously at the channel and window levels; the channel-level quality score Q_ch reflects the reliability of a single electrode or differential channel, while the window-level quality score Q_win reflects whether an analysis window is suitable for participation in 3D reconstruction, ARI / SDR calculation, or risk calibration; the score can be jointly determined by impedance, saturation ratio, baseline drift amplitude, power frequency interference intensity, electromyographic noise intensity, R-peak detection consistency, and inter-channel correlation.

[0133] It is important to note that the data preprocessing stage should not disrupt the ST segment, T-wave terminal morphology, or repolarization. Configurable high-pass filtering is preferred to remove slow drift, low-pass or wavelet denoising should be used to reduce EMG noise, and 50Hz or 60Hz notch filtering should be used to reduce power frequency interference. Filtering parameters are fixed in the version configuration based on the device sampling rate and application scenario; training, validation, testing, and deployment must use the same configuration. For R-reference alignment, R-peak detection, rhythm segment localization, and amplitude scale calibration are performed using the R channel as a reference. The system provides location confidence for each candidate R-peak and performs time alignment and amplitude normalization of the XYZ channels near the corresponding time points. For abnormal beats with risk significance, such as premature ventricular contractions and short runs of ventricular tachycardia, they should not be simply deleted; instead, abnormal beat labels or rhythm status labels should be retained to avoid filtering out risk-related information during the preprocessing stage.

[0134] For continuous recordings, overlapping windows are divided according to window length W and step size S. Specifically, in static mode, a 10-second window is preferred for a single input. When the sampling rate is 500Hz, the basic input can be represented as (number of channels, 5000). In dynamic mode, continuous ECG recordings of 24 to 72 hours are slidably divided according to window length W and step size S, preferably W=10 seconds and S=2 seconds. Adjacent windows are retained for overlapping times, enabling the model to learn the continuous changes in SDR and repolarization morphology in subsequent time-series modules.

[0135] Specific configurations include: recording the start and end times, window number, patient anonymity identifier, channel mask, quality score, and hardware configuration for each window; outputting single-window preprocessing tensors for static windows; outputting a window sequence arranged in chronological order for dynamic records, prohibiting disruption of the chronological order within the same dynamic record; binding 3D vector field labels V_gt, virtual dense lead labels Phi_64, and ARI / SDR labels to windows with dense surface mapping or BEM labels; and binding only risk calibration labels to dynamic records with only clinical events or follow-up outcomes, and avoiding the forced use of non-existent 3D supervision items through loss masks.

[0136] For 3D reconstruction labels, the preferred method is to obtain them from dense body surface potentials via the inverse BEM problem. Specifically, this involves: using the body surface potential Phi_body and the human tissue transfer matrix A as a basis, obtaining the cardiac surface potential or 3D electrical activity representation through regularized inverse solution, and then organizing it into time series V_gt for multiple cardiac regions in three spatial directions. The activation time T_act and recovery time T_rec for each region are extracted from the V_gt or dense lead reconstruction results, and the ARI_i = T_rec - T_act is calculated. The SDR is then obtained from the spatial discretization of the ARI.

[0137] Dataset partitioning must be done on a patient-by-patient basis, not on a window-by-window basis. Windows from the same patient, the same dynamic recording, or adjacent time periods must not be included in both the training and test sets simultaneously. The training set is used for parameter learning, the validation set is used for hyperparameter and threshold selection, and the test set is used only for final performance evaluation. If multi-center or different device batches of data exist, independent batches should be maintained in validation and testing to check the model's generalization ability to device and attachment differences.

[0138] The specific training process includes: the model to be trained takes the preprocessed R-XYZ window tensor, position encoding, and quality mask as input; the front end includes an R-channel encoder, an XYZ spatial channel encoder, a spatial position encoder, and an R-reference guided attention module; in the middle, window-level latent features H_comp are obtained through multi-level feature compression, and multi-scale skip features H_skip are retained; in dynamic mode, the H_comp sequence is input into an LSTM or a temporal Transformer to obtain long temporal features H_global; the back end is gated and fused to obtain H_fused, and then the decoder is reconstructed layer by layer by a U-Net-like method to output the three-dimensional ECG vector field, virtual dense leads, ARI / SDR, and risk results.

[0139] Specifically, the first ECG data sample is input into the ECG data processing model to be trained for the first stage of training. By training the R-XYZ encoder and the R-reference guided attention module, the R channel features are used as reference anchors to guide the XYZ channels to complete the alignment of time, amplitude, and position. Then, the second ECG data sample is input into the pre-trained ECG data processing model for the second stage of training, namely: training the forward 3D reconstruction branch based on BEM 3D vector field labels, and training the reverse virtual lead branch based on dense lead or 64-lead labels. Subsequently, the ARI / SDR calculation branch and the dynamic temporal branch are trained, so that the compressed latent features of the continuous window can obtain the long-term trend representation through LSTM or temporal Transformer. Finally, the risk score and warning threshold are calibrated using follow-up outcomes, ventricular arrhythmia segments, or manual review labels.

[0140] In some embodiments, when the model enables high-frequency QRS (HFQRS) assisted analysis, the system does not require additional electrodes. Instead, it utilizes the time-aligned R reference channel and XYZ spatial channel signals. First, the start and end range of each QRS complex is determined by the R peak position. Multiple qualified heartbeat segments are aligned and averaged. Then, high-frequency components in the 150-250Hz frequency band are extracted from the averaged QRS segments. The High-Frequency Morphology Index (HFMI) reflects the overall change in the morphology of the high-frequency QRS, and the Reduced Amplitude Zone (RAZ) reflects the region where the high-frequency amplitude of the QRS is locally reduced. Both serve as window-level auxiliary features and are input into the risk score or threshold calibration module along with the activation-recovery interval (ARI), stereo dispersion rate (SDR), and their dynamic trends. For windows with low R peak location reliability, excessive noise, or insufficient effective heartbeats, the system may not calculate this high-frequency QRS auxiliary feature or may reduce its weight during fusion.

[0141] The training loss includes at least one of the following: R-reference alignment error, 3D vector field reconstruction error, virtual lead reconstruction error, jumper reconstruction consistency error, ARI / SDR error, dynamic trend error, and risk calibration error, and is weighted according to channel quality score and label integrity.

[0142] A validation set is then set up, independently divided by patient, to select model hyperparameters, loss weights, quality reduction functions, window parameters, and SDR warning thresholds. Validation metrics include 3D vector field error, virtual lead correlation, ARI / SDR error, SDR classification consistency rate, dynamic warning lead time, hourly false alarm rate, and risk calibration metrics. Model candidate versions are only frozen and tested if reconstruction performance, repolarization metrics, and risk calibration all meet preset requirements.

[0143] Finally, a test set was set up, constructed independently at the patient level, and excluding patients, adjacent windows, or the same dynamic recording from the training or validation sets. Model testing was performed after freezing model weights, preprocessing parameters, location encoding tables, quality score mappings, and risk thresholds. During testing, each recording underwent the same preprocessing and windowing as during training, outputting the 3D ECG vector field, virtual leads, ARI, SDR, risk level, and dynamic warning status window by window, and performing stratified statistics according to hardware configuration, acquisition mode, signal quality, label source, and rhythm status. Test metrics included preprocessing effectiveness, R-peak localization error, 3D vector field error, virtual lead correlation, ARI / SDR error, SDR trend consistency, warning lead time, false alarm rate, inference latency, and performance degradation in low-quality scenarios.

[0144] Once the test results meet the preset requirements, the model weights, R-XYZ position encoding table, channel quality scoring function, filtering and windowing parameters, SDR risk classification threshold, dynamic alarm hold time, model version number, and training data batch summary are collectively fixed into a deployment package. Upon receiving a new R-XYZ data packet, the deployment end performs preprocessing, encoding, reconstruction, ARI / SDR calculation, and risk output according to the fixed parameters, thereby ensuring consistency in data processing rules between training, testing, and actual deployment.

[0145] Please refer to Figure 8 This invention provides an electrocardiogram (ECG) data processing device, comprising:

[0146] Acquisition module 81 is used to acquire target electrocardiogram data, the target electrocardiogram data including: a reference signal acquired by at least one reference electrode based on a zero potential reference point and electrocardiogram signals acquired by spatial electrodes in at least three different body surface spatial sampling areas;

[0147] The first processing module 82 is used to input the target ECG data, the position of the reference electrode, and the position of the spatial electrode as input data into the ECG data processing model. The ECG data processing model encodes the input data and uses a spatiotemporal joint attention module to complete channel alignment; divides at least one analysis window for multi-level feature compression to obtain window-level latent features; and constructs a three-dimensional ECG vector field based on the latent features to obtain the activation-recovery interval parameters and the stereo diffusion rate index corresponding to each myocardial node. Based on the activation-recovery interval parameters and the stereo diffusion rate index corresponding to each myocardial node, the module outputs the corresponding risk warning result.

[0148] In this embodiment of the invention, optionally, the electrocardiogram signals obtained by spatial electrodes in at least three different body surface spatial sampling areas include the following at least three types of body surface spatial sampling areas: the precordial region, the left chest wall, the left posterolateral region, and the left posterodorsolateral region.

[0149] In this embodiment of the invention, optionally, it also includes:

[0150] The preprocessing module is used to perform data preprocessing on the target electrocardiogram data. The data preprocessing includes at least one of the following: high-pass filtering, wavelet denoising, notch filtering, amplitude normalization, artifact segment removal, severe dropout segment removal, low-quality window marking or quality downweighting, and R-peak alignment.

[0151] In this embodiment of the invention, optionally, the first processing module includes:

[0152] The first processing submodule is used to estimate the time offset of each spatial channel based on the R-peak position of the reference signal, and generate a normalized scaling factor based on the amplitude statistics of the reference signal; using the channel features of the reference signal as the query and the channel features of the electrocardiogram signal as the key / value, it calculates the attention weight and completes the channel alignment.

[0153] In this embodiment of the invention, optionally, the first processing module includes:

[0154] The second processing submodule is used to divide at least one analysis window according to a fixed window length and sliding step size through the windowing time-series segmentation module; to encode the input data in each analysis window into multi-channel ECG signals to obtain window-level multi-channel features; and to perform at least two levels of time dimension compression and feature dimension mapping on the window-level multi-channel features to obtain window-level latent features.

[0155] In this embodiment of the invention, optionally, the first processing module includes:

[0156] The third processing submodule is used to mine long-term temporal dependencies between adjacent windows through a temporal modeling network and extract global temporal features. The temporal modeling network includes at least one of a long short-term memory network, a bidirectional long short-term memory network, a temporal Transformer, or a TCN. The global temporal features are fused using a gating fusion mechanism to obtain fused features. The fused features are then restored layer by layer and decoded and reconstructed at multiple scales using a bidirectional reconstruction decoder to construct a three-dimensional electrocardiogram vector field.

[0157] In this embodiment of the invention, optionally, the bidirectional reconstruction decoder includes a forward three-dimensional vector field reconstruction branch and a reverse virtual lead reconstruction branch; the forward three-dimensional vector field reconstruction branch uses the fusion feature as a low-scale latent representation, and sequentially performs upsampling, convolution / attention decoding, and jumper fusion to gradually restore to the target temporal resolution, and finally outputs a three-dimensional ECG vector field; the reverse virtual lead reconstruction branch outputs virtual body surface lead signals according to the fusion feature, and generates consistency constraints based on the difference between the virtual body surface lead signals and the labeled signals or the body surface potentials obtained by forward modeling from the three-dimensional ECG reconstruction results.

[0158] In this embodiment of the invention, optionally, the first processing module includes:

[0159] The fourth processing submodule is used to provide risk warning results that include at least one of the following: absolute value threshold warning for the three-dimensional diffusion rate index, rate of change warning for the three-dimensional diffusion rate index, volatility warning for the three-dimensional diffusion rate index, sudden event warning for the three-dimensional diffusion rate index, and comprehensive dynamic risk score warning.

[0160] The three-dimensional diffusion rate index absolute value threshold early warning includes: a three-dimensional diffusion rate index less than a first threshold is normal and routine monitoring is performed; a three-dimensional diffusion rate index not less than the first threshold and less than a second threshold is low risk, requiring attention and increasing monitoring frequency; a three-dimensional diffusion rate index not less than the second threshold and less than a third threshold is medium risk, issuing an alert and outputting a review prompt message; a three-dimensional diffusion rate index not less than the third threshold is high risk, issuing an emergency early warning and outputting an alarm control signal or review prompt message; the first threshold is less than the second threshold, and the second threshold is less than the third threshold.

[0161] The warning for the rate of change of the three-dimensional dispersion index is triggered when the three-dimensional dispersion index is greater than the fourth threshold.

[0162] The three-dimensional dispersion rate index fluctuation warning is raised when the coefficient of variation is greater than the fifth threshold and continues to exceed the preset time threshold.

[0163] The three-dimensional dispersion rate index sudden change event warning is triggered when the three-dimensional dispersion rate index changes by a step greater than the sixth threshold, and the event is immediately warned and recorded.

[0164] The comprehensive dynamic risk score early warning system outputs an alarm control signal or a review prompt when the comprehensive dynamic risk score exceeds the seventh threshold.

[0165] In this embodiment of the invention, optionally, the formula for calculating the stereo dispersion index is:

[0166] SDR = ;

[0167] Where ARI_i is the activation-recovery time interval of the i-th myocardial node; ARI_mean is the mean of ARI across all nodes; and N is the total number of myocardial nodes.

[0168] In this embodiment of the invention, optionally, it also includes:

[0169] The model training module is used to train the ECG data processing model, and specifically includes:

[0170] Acquire training data, which includes publicly available electrocardiogram (ECG) data, high-density surface ECG mapping data, ECG waveform data, and / or three-dimensional electrical activity annotation data generated based on the human transmission model; wherein, the publicly available ECG data is used for pre-training and is not required to have three-dimensional electrical activity field supervision labels;

[0171] In the first training phase, the publicly available ECG data is input into the ECG data processing model to be trained, and pre-training tasks such as waveform reconstruction, mask prediction, rhythm classification, and / or R-peak localization are performed to obtain a pre-trained ECG data processing model. The ECG data processing model to be trained learns the nonlinear correlation between signals acquired by finite electrodes and the three-dimensional electrical activity characteristics of the heart. It encodes the ECG data samples and uses a spatiotemporal joint attention module to complete channel alignment. At least one analysis window is divided for multi-level feature compression to obtain window-level latent features. A three-dimensional ECG vector field is constructed based on the latent features to complete the initial construction of the model's basic feature extraction, spatiotemporal correlation fitting, and three-dimensional electrical activity reconstruction capabilities, obtaining the predicted activation-recovery interval parameters and stereo diffusion rate index corresponding to each myocardial node. Based on the predicted activation-recovery interval parameters and stereo diffusion rate index corresponding to each myocardial node, the supervision label of the three-dimensional electrical activity field corresponding to the first ECG data sample, and the loss function, the parameters of the ECG data processing model to be trained are optimized to obtain the pre-trained ECG data processing model.

[0172] In the second training phase, the pre-trained ECG data processing model is subjected to 3D reconstruction supervised training using the high-density surface ECG mapping data, ECG waveform data, and / or 3D electrical activity annotation data generated based on the human transmission model. In the third training phase, the risk warning output is calibrated using follow-up outcomes or expert review labels to obtain the trained ECG data processing model. The pre-trained model is then subjected to bidirectional distillation training to further optimize its parameters, resulting in the trained ECG data processing model.

[0173] The electrocardiogram data processing device provided in this embodiment of the invention can achieve Figure 1 The various processes implemented in the method embodiments achieve the same technical effect, and will not be described again here to avoid repetition.

[0174] This invention provides an electronic device 90, see [link to relevant documentation]. Figure 9 As shown, Figure 9 This is a schematic diagram of the electronic device 90 according to an embodiment of the present invention, including a processor 91, a memory 92, and a program or instructions stored in the memory 92 and executable on the processor 91. When the program or instructions are executed by the processor, they implement the steps in any of the electrocardiogram data processing methods of the present invention.

[0175] This invention provides a readable storage medium on which a program or instruction is stored. When the program or instruction is executed by a processor, it implements the various processes of the ECG data processing method of any of the above embodiments and achieves the same technical effect. To avoid repetition, it will not be described again here.

[0176] This application also provides a computer program product, including computer instructions, which, when executed by a processor, implement the above-described... Figure 1 The various processes of the method embodiments shown can achieve the same technical effect, and will not be described again here to avoid repetition.

[0177] Computer-readable media include both permanent and non-permanent, removable and non-removable media, which can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0178] It should be noted that the collection, gathering, updating, analysis, processing, use, transmission, and storage of user personal information involved in this disclosed technical solution all comply with relevant laws and regulations, are used for legitimate purposes, and do not violate public order and good morals. Necessary measures are taken to prevent unauthorized access to user personal information data and to safeguard user personal information security and network security.

[0179] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0180] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0181] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a computing device (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0182] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for processing electrocardiogram (ECG) data, characterized in that, include: Acquire target electrocardiogram (ECG) data, which includes: a reference signal acquired by at least one reference electrode based on a zero potential reference point and ECG signals acquired by spatial electrodes in at least three different body surface spatial sampling areas; The target ECG data, the location of the reference electrode, and the location of the spatial electrode are input into the ECG data processing model. The ECG data processing model encodes the input data and uses a spatiotemporal joint attention module to complete channel alignment. At least one analysis window is divided for multi-level feature compression to obtain window-level latent features. A three-dimensional ECG vector field is constructed based on the latent features to obtain the activation-recovery interval parameters and the stereo diffusion rate index corresponding to each myocardial node. Based on the activation-recovery interval parameters and the stereo diffusion rate index corresponding to each myocardial node, the corresponding risk warning result is output.

2. The electrocardiogram data processing method according to claim 1, characterized in that, ECG signals acquired by spatial electrodes in at least three different body surface spatial sampling areas include the following at least three types of body surface spatial sampling areas: precordial region, left chest wall, left posterolateral region, and left posterodorsal region.

3. The electrocardiogram data processing method according to claim 1, characterized in that, Before inputting the target ECG data, the position of the reference electrode, and the position of the spatial electrode into the ECG data processing model as input data, the method further includes: The target electrocardiogram data is preprocessed, including at least one of the following: high-pass filtering, wavelet denoising, notch filtering, amplitude normalization, artifact segment removal, severe dropout segment removal, low-quality window marking or quality downweighting, and R-peak alignment.

4. The electrocardiogram data processing method according to claim 1, characterized in that, The method of using a spatiotemporal joint attention module to complete channel alignment includes: The time offset of each spatial channel is estimated based on the R-peak position of the reference signal, and a normalized scaling factor is generated based on the amplitude statistics of the reference signal. The attention weight is calculated and the channel alignment is completed using the channel features of the reference signal as the query and the channel features of the ECG signal as the key / value.

5. The electrocardiogram data processing method according to claim 1, characterized in that, The process of dividing at least one analysis window for multi-level feature compression yields window-level latent features, including: The time-series segmentation module divides the data into at least one analysis window with a fixed window length and a sliding step size. The input data in each analysis window is encoded into multi-channel ECG signals to obtain window-level multi-channel features. The window-level multi-channel features are then compressed in at least two levels of time dimension and mapped in feature dimension to obtain window-level latent features.

6. The electrocardiogram data processing method according to claim 1, characterized in that, The construction of a three-dimensional electrocardiogram vector field based on the latent features includes: The global temporal features are extracted by mining long-term temporal dependencies between adjacent windows through a temporal modeling network. The temporal modeling network includes at least one of a long short-term memory network, a bidirectional long short-term memory network, a temporal Transformer, or a TCN. The global temporal features are then fused using a gating fusion mechanism to obtain fused features. The fused features are then reconstructed layer by layer and decoded and reconstructed at multiple scales using a bidirectional reconstruction decoder to build a three-dimensional electrocardiogram vector field.

7. The electrocardiogram data processing method according to claim 6, characterized in that, The bidirectional reconstruction decoder includes a forward 3D vector field reconstruction branch and a reverse virtual lead reconstruction branch. The forward 3D vector field reconstruction branch uses the fusion features as a low-scale latent representation, and sequentially performs upsampling, convolution / attention decoding, and jumper fusion to gradually restore the target temporal resolution, ultimately outputting a 3D ECG vector field. The reverse virtual lead reconstruction branch outputs virtual surface lead signals based on the fusion features, and generates consistency constraints based on the differences between the virtual surface lead signals and the labeled signals or the surface potentials obtained by forward modeling from the 3D ECG reconstruction results.

8. The electrocardiogram data processing method according to claim 1, characterized in that, The risk warning result is output based on the activation-recovery interval parameters corresponding to each myocardial node and the stereo diffusion rate index, including: The risk warning results include at least one of the following: warning of the absolute value threshold of the three-dimensional diffusion rate index, warning of the rate of change of the three-dimensional diffusion rate index, warning of the volatility of the three-dimensional diffusion rate index, warning of the sudden event of the three-dimensional diffusion rate index, and warning of the comprehensive dynamic risk score. The three-dimensional diffusion rate index absolute value threshold early warning includes: a three-dimensional diffusion rate index less than a first threshold is normal and routine monitoring is performed; a three-dimensional diffusion rate index not less than the first threshold and less than a second threshold is low risk, requiring attention and increasing monitoring frequency; a three-dimensional diffusion rate index not less than the second threshold and less than a third threshold is medium risk, issuing an alert and outputting a review prompt message; a three-dimensional diffusion rate index not less than the third threshold is high risk, issuing an emergency early warning and outputting an alarm control signal or review prompt message; the first threshold is less than the second threshold, and the second threshold is less than the third threshold. The warning for the rate of change of the three-dimensional dispersion index is triggered when the three-dimensional dispersion index is greater than the fourth threshold. The three-dimensional dispersion rate index fluctuation warning is raised when the coefficient of variation is greater than the fifth threshold and continues to exceed the preset time threshold. The three-dimensional dispersion rate index sudden change event warning is triggered when the three-dimensional dispersion rate index changes by a step greater than the sixth threshold, and the event is immediately warned and recorded. The comprehensive dynamic risk score early warning system outputs an alarm control signal or a review prompt when the comprehensive dynamic risk score exceeds the seventh threshold.

9. The electrocardiogram data processing method according to claim 1, characterized in that, The formula for calculating the stereo dispersion index is as follows: SDR = ; Where ARI_i is the activation-recovery time interval of the i-th myocardial node; ARI_mean is the mean of ARI across all nodes; and N is the total number of myocardial nodes.

10. The electrocardiogram data processing method according to claim 1, characterized in that, Before inputting the target ECG data, the position of the reference electrode, and the position of the spatial electrode into the ECG data processing model as input data, the method further includes: Training the ECG data processing model specifically includes: Acquire training data, which includes publicly available electrocardiogram (ECG) data, high-density surface ECG mapping data, ECG waveform data, and / or three-dimensional electrical activity annotation data generated based on the human transmission model; wherein, the publicly available ECG data is used for pre-training and is not required to have three-dimensional electrical activity field supervision labels; In the first training phase, the publicly available ECG data is input into the ECG data processing model to be trained, and pre-training tasks such as waveform reconstruction, mask prediction, rhythm classification, and / or R-peak localization are performed to obtain a pre-trained ECG data processing model. The ECG data processing model to be trained learns the nonlinear correlation between signals acquired by finite electrodes and the three-dimensional electrical activity characteristics of the heart. It encodes the ECG data samples and uses a spatiotemporal joint attention module to complete channel alignment. At least one analysis window is divided for multi-level feature compression to obtain window-level latent features. A three-dimensional ECG vector field is constructed based on the latent features to complete the initial construction of the model's basic feature extraction, spatiotemporal correlation fitting, and three-dimensional electrical activity reconstruction capabilities, obtaining the predicted activation-recovery interval parameters and stereo diffusion rate index corresponding to each myocardial node. Based on the predicted activation-recovery interval parameters and stereo diffusion rate index corresponding to each myocardial node, the supervision label of the three-dimensional electrical activity field corresponding to the first ECG data sample, and the loss function, the parameters of the ECG data processing model to be trained are optimized to obtain the pre-trained ECG data processing model. In the second training phase, the pre-trained ECG data processing model is subjected to 3D reconstruction supervised training using the high-density surface ECG mapping data, ECG waveform data, and / or 3D electrical activity annotation data generated based on the human transmission model. In the third training phase, the risk warning output is calibrated using follow-up outcomes or expert review labels to obtain the trained ECG data processing model. The parameters of the pre-trained ECG data processing model are then further optimized to obtain the trained ECG data processing model.

11. The electrocardiogram data processing method according to claim 1, characterized in that, The ECG data processing model also aligns and averages the QRS complex based on the reference signal, and extracts high-frequency components in the 150-250Hz frequency band to construct high-frequency QRS auxiliary features. The high-frequency QRS auxiliary features include high-frequency morphology index and / or reduced amplitude region, and are jointly output with the excitation-recovery interval parameter and the stereo diffusion rate index to output risk warning results.

12. An electrocardiogram (ECG) data processing device, characterized in that, include: The acquisition module is used to acquire target electrocardiogram (ECG) data, which includes: a reference signal acquired by at least one reference electrode based on a zero potential reference point and ECG signals acquired by spatial electrodes in at least three different body surface spatial sampling areas. The first processing module is used to input the target ECG data, the position of the reference electrode, and the position of the spatial electrode as input data into the ECG data processing model. The ECG data processing model encodes the input data and uses a spatiotemporal joint attention module to complete channel alignment; divides at least one analysis window for multi-level feature compression to obtain window-level latent features; and constructs a three-dimensional ECG vector field based on the latent features to obtain the activation-recovery interval parameters and the stereo diffusion rate index corresponding to each myocardial node. Based on the activation-recovery interval parameters and the stereo diffusion rate index corresponding to each myocardial node, the module outputs the corresponding risk warning result.

13. An electronic device, characterized in that, It includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of the electrocardiogram data processing method as described in any one of claims 1 to 11.

14. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a processor, implement the steps of the electrocardiogram data processing method as described in any one of claims 1 to 11.

15. A computer program product, characterized in that, It includes computer instructions that, when executed by a processor, implement the steps in the electrocardiogram data processing method as described in any one of claims 1 to 11.

Images