A method and system for intelligent localization of ventricular ectopic rhythms based on electrocardiogram vector map fusion
By using an intelligent positioning method that integrates electrocardiogram vector maps, precise localization of ventricular ectopic rhythms was achieved, solving the problems of insufficient spatial information from a single electrocardiogram and reliance on manual interpretation, thus improving positioning accuracy and surgical efficiency.
Patent Information
- Application Number
- CN202611356959.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-03
- Publication Date
- 2026-11-10
AI Technical Summary
In existing technologies, a single 12-lead surface electrocardiogram lacks spatial three-dimensional information when locating ventricular ectopic arrhythmias. It relies on manual interpretation and has a large positioning error, which makes it impossible to accurately delineate the ablation target area before the operation. During the operation, a large-scale mapping is required, which increases the operation time and risk.
An intelligent localization method based on ECG vector map fusion is adopted. By synchronously acquiring and preprocessing ECG and ECG vector map signals, precise spatiotemporal registration and multimodal feature fusion are performed. A deep learning model is used for hierarchical prediction, and a data feedback iterative closed loop is constructed to generate a preoperative localization reference report.
It improves the resolution of complex overlapping ectopic waveforms, enhances the localization accuracy of special areas, shortens intraoperative mapping time, reduces radiation exposure and myocardial damage, and is suitable for multi-center promotion.
Smart Images

Figure CN123058746A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of intelligent diagnosis of electrocardiogram signals and medical-engineering interdisciplinary artificial intelligence technology, and in particular to an intelligent localization method and system for ventricular ectopic rhythms based on electrocardiogram vector map fusion. Background Technology
[0002] Ectopic ventricular arrhythmias (including frequent premature ventricular contractions and idiopathic ventricular tachycardia) have a high incidence rate in clinical practice. For patients who do not respond well to drug treatment, intracardiac electrophysiological radiofrequency ablation is necessary for radical treatment. Preoperative accurate prediction of the anatomical origin of the ectopic pacemaker is a key step in shortening the operation time and reducing the patient's radiation exposure and myocardial damage.
[0003] Currently, clinical practice mainly relies on a single 12-lead surface electrocardiogram for manual interpretation and localization, which has the following inherent drawbacks: First, a single electrocardiogram can only reflect the two-dimensional projection of cardiac electrical activity on the body surface, lacking spatial three-dimensional conduction information. It has low recognition accuracy for complex, overlapping, and ectopic waveforms originating from special anatomical areas such as the Summit area of the left ventricle, papillary muscles, and para-His bundle. Second, manual interpretation is highly dependent on the physician's professional experience and lacks quantitative and standardized objective assessment methods. The interpretation results vary greatly among physicians with different years of experience, resulting in significant localization errors. Third, it is impossible to accurately delineate the ablation target area before the operation, which requires the reliance on a three-dimensional electroanatomical mapping system for large-scale point-by-point mapping during the operation. This results in a long operation time and a high risk of X-ray exposure and myocardial catheter injury.
[0004] Vectorcardiography (VCG) records the vector loop trajectory of cardiac electrical activity in three-dimensional space, displaying the spatial conduction characteristics of cardiac depolarization and repolarization processes from three planes: the frontal plane, the transverse plane, and the lateral plane. Theoretically, this can compensate for the lack of spatial dimension information in a single electrocardiogram (ECG). However, current technologies only use ECG and VCG as two independent diagnostic tools, failing to achieve high-precision synchronous registration and feature fusion between the two in both temporal and spatial dimensions. Furthermore, current technologies lack standardized datasets using successfully ablated intraoperative electrophysiological targets as the gold standard label, making it difficult to train automated, quantitative, and intelligent predictive models with clinical practical value. Summary of the Invention
[0005] The purpose of this invention is to overcome the shortcomings of the existing technology. To achieve the above objective, an intelligent localization method and system for ventricular ectopic rhythm based on electrocardiogram vector map fusion is adopted to solve the problems mentioned in the background technology.
[0006] This invention provides a method for intelligent localization of ventricular ectopic rhythms based on electrocardiogram vector map fusion, comprising: Simultaneously acquire raw signals of 12-lead surface electrocardiogram and Frank system electrocardiogram vector graph from the same patient; preprocess the acquired raw signals for noise reduction and baseline correction, and automatically segment the QRS feature points of ventricular ectopic beats.
[0007] Based on this, the spatiotemporal precise registration step is performed: using the ectopic QRS wave of the ventricle as the anchor point, the electrocardiogram time sequence signal and the three-dimensional vector loop of the electrocardiogram are synchronously aligned point-to-point, and the registered electrical signal is mapped to the three-dimensional anatomical space coordinates of the heart in combination with the anterior chest anatomical positioning marks.
[0008] Perform the multimodal feature fusion step: extract the temporal waveform features and the three-dimensional spatial vector features of the ECG vector map of the 12 leads respectively, and use the mid-term feature fusion strategy to fuse the temporal waveform features and spatial vector features to generate a unified spatiotemporal fusion feature vector.
[0009] Perform the hierarchical AI prediction step: input the spatiotemporal fusion feature vector into the pre-trained deep learning model, and output the anatomical prediction location of the ectopic pacemaker in a hierarchical manner.
[0010] The visualization process involves simultaneously displaying the AI-predicted target location along with the corresponding key ECG leads and feature vector loops, generating a pre-ablation localization reference report.
[0011] Furthermore, the spatiotemporal precise registration step specifically includes: using the QRS wave origin of each ectopic ventricular beat as the temporal alignment anchor point, performing point-to-point matching of the QRS origin, peak, and end point on the electrocardiogram temporal waveform of the same cardiac cycle with the corresponding turning nodes on the three-dimensional vector loop of the electrocardiogram vector map, and combining the synchronously acquired anterior chest anatomical positioning markers to map the registered dual-modal electrical signals to the standard cardiac three-dimensional anatomical space coordinate system.
[0012] Furthermore, the multimodal feature fusion step specifically includes: extracting temporal morphological features, including QRS wave width, amplitude, and transition leads, from the 12 leads of the electrocardiogram; extracting spatial conduction features, including the direction of rotation of the loop body, the area of the loop body, and the maximum vector angle, from the frontal, transverse, and lateral three-plane vector loops of the electrocardiogram; using a convolutional splicing method to perform intermediate fusion of the temporal morphological features and spatial conduction features, and filtering redundant high-dimensional features to generate a unified spatiotemporal fusion feature vector strongly correlated with ectopic lesions.
[0013] Furthermore, the pre-trained deep learning model adopts a hybrid architecture of ResNet and Transformer, with a built-in four-level progressive classifier for prediction. The first level is used to distinguish the origin of the left and right ventricles, the second level is used to distinguish large anatomical regions, the third level is used to refine subregions, and the fourth level is used to accurately output the estimated location of the earliest pacing target.
[0014] Furthermore, the pre-trained deep learning model also incorporates an attention mechanism to simultaneously calculate the contribution weights of each ECG lead and each ECG vector map feature vector loop to the prediction results during the hierarchical prediction process; the visualization step further includes highlighting the key ECG leads and key feature vector loops with the highest contribution weights as the basis for model judgment.
[0015] Furthermore, the method also includes a data backflow iteration step: collecting successful intraoperative electrophysiological ablation targets as gold standard data, associating and annotating the gold standard data with the corresponding bimodal electrocardiogram data and backflowing them to a standardized database, and periodically retraining and fine-tuning the pre-trained deep learning model using the expanded database through cross-validation.
[0016] Furthermore, in the synchronous acquisition step, a unified hardware sampling clock is used to ensure that the timing error between the electrocardiogram signal and the electrocardiogram vector graph signal is less than or equal to 1 millisecond.
[0017] Furthermore, the preprocessing noise reduction includes automatically filtering out electromyographic noise and power frequency noise, the baseline correction includes correcting respiratory baseline deviation, and the automatic segmentation of QRS feature points of ventricular ectopic beats includes automatically identifying the start point, peak value, end point of the QRS wave, and the turning point of the ECG vector loop.
[0018] The present invention also provides a positioning system employing a ventricular ectopic rhythm intelligent positioning method based on electrocardiogram vector map fusion as described in any of the preceding claims, comprising: The system includes a synchronous acquisition module for simultaneously acquiring raw 12-lead surface electrocardiogram (ECG) signals and Frank system vectorcardiogram (VCG) signals from the same patient; a signal preprocessing module for preprocessing and noise reduction of the acquired raw signals, baseline correction, and automatic segmentation of QRS feature points for ectopic ventricular beats; a spatiotemporal registration module for point-to-point synchronous alignment of the ECG time-series signals with the 3D vectorcardiogram using ectopic QRS waves as anchor points, and mapping the registered electrical signals to the 3D anatomical coordinates of the heart using chest anatomical landmarks; and a multimodal feature fusion module. The system comprises four modules: a module for extracting temporal waveform features from 12 leads of an electrocardiogram (ECG) and three-dimensional spatial vector features from the ECG vector map; a mid-term feature fusion strategy for fusing the temporal waveform features and spatial vector features to generate a unified spatiotemporal fusion feature vector; a tiered AI prediction model module for inputting the spatiotemporal fusion feature vector into a pre-trained deep learning model and outputting the anatomical prediction location of the ectopic pacemaker lesion in tiers; and a visualization output module for simultaneously displaying the AI-predicted target area location and the corresponding key ECG leads and feature vector loops, generating a pre-radiofrequency ablation localization reference report.
[0019] Furthermore, the system also includes a model iteration and update module, which is used to collect successful electrophysiological ablation targets during the operation as gold standard data, associate and label the gold standard data with the corresponding bimodal electrocardiogram data and then feed them back to the standardized database. The pre-trained deep learning model is periodically retrained and its weights fine-tuned using the expanded database through cross-validation.
[0020] Compared with the prior art, the present invention has the following technical advantages: By adopting the above technical solution, and by simultaneously acquiring dual-modal signals of electrocardiogram and electrocardiogram vector graph, and using ventricular ectopic QRS waves as anchor points for point-to-point spatiotemporal precise registration, the electrical signals are mapped to the three-dimensional anatomical spatial coordinate system of the heart, making up for the shortcomings of insufficient spatial information of a single electrocardiogram and greatly improving the resolution of complex overlapping ectopic waveforms.
[0021] By extracting the temporal morphological features of electrocardiograms and the spatial conduction features of electrocardiogram vector maps, and using a mid-term feature fusion strategy for convolutional splicing and fusion to filter redundant high-dimensional features, a unified spatiotemporal fusion feature vector strongly correlated with ectopic lesions was generated, providing high-quality feature input for subsequent accurate localization.
[0022] A deep learning model with a hybrid ResNet and Transformer architecture, trained using intraoperative electrophysiological ablation targets as the gold standard label, is used to output the anatomical prediction location of ectopic pacemakers through a four-level progressive classifier. This eliminates the reliance on physicians' human experience and significantly improves the accuracy of localization in special high-risk areas such as the left ventricular Summit area and papillary muscles.
[0023] By incorporating an attention mechanism into the model, key ECG leads and key ECG vector map feature vector loops can be displayed simultaneously while outputting localization results, enhancing the interpretability and clinical credibility of AI prediction results.
[0024] By constructing a closed-loop data feedback and iteration system, the gold standard data during surgery is continuously fed back to a standardized database and the model is retrained regularly. This allows the model's generalization ability to continuously improve with the accumulation of clinical samples, making it suitable for large-scale promotion needs in different application scenarios such as multi-center and primary hospitals.
[0025] By accurately delineating the ablation target area before the procedure and generating a visual reference report, the time for three-dimensional mapping during the procedure can be effectively shortened, and ineffective ablation, X-ray radiation exposure and myocardial catheter injury can be reduced, which has clear clinical translational value. Attached Figure Description
[0026] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings: Figure 1 This is a schematic diagram illustrating the steps of the intelligent localization method for ventricular ectopic rhythms according to an embodiment of this application. Detailed Implementation
[0027] 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 embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0028] Please refer to Figure 1 In this embodiment of the invention, a method for intelligent localization of ventricular ectopic rhythms based on electrocardiogram vector map fusion includes the following steps: Step S1: Simultaneously acquire the raw signals of the 12-lead surface electrocardiogram and the raw signals of the Frank system electrocardiogram vector diagram of the same patient.
[0029] Specifically, a unified hardware sampling clock is used to ensure that the timing error between the electrocardiogram signal and the electrocardiogram vector graph signal is less than or equal to 1 millisecond, so as to achieve strict synchronization of the dual-mode signals.
[0030] Step S2: Preprocess the acquired raw signal to reduce noise and correct the baseline, and automatically segment the QRS feature points of ectopic ventricular beats.
[0031] Specifically, the preprocessing noise reduction includes automatically filtering out electromyographic noise and power frequency noise; the baseline correction includes correcting respiratory baseline deviation; and the automatic segmentation of QRS feature points of ectopic ventricular beats includes automatically identifying the start point, peak value, end point of the QRS wave, and the turning point of the ECG vector loop.
[0032] Step S3: Perform precise spatiotemporal registration: Using the ectopic QRS wave of the ventricle as the anchor point, the electrocardiogram timing signal and the three-dimensional vector loop of the electrocardiogram are synchronously aligned point-to-point, and the registered electrical signal is mapped to the three-dimensional anatomical space coordinates of the heart in combination with the anterior chest anatomical positioning marks.
[0033] This step specifically includes: using the QRS wave origin of each ectopic ventricular beat as the timing alignment anchor point, matching the QRS origin, peak, and end point on the electrocardiogram timing waveform of the same cardiac cycle with the corresponding turning nodes on the three-dimensional vector loop of the electrocardiogram, and combining the synchronously acquired anterior chest anatomical positioning markers to map the registered dual-modal electrical signals to the standard cardiac three-dimensional anatomical space coordinate system.
[0034] Step S4: Perform multimodal feature fusion: Extract the temporal waveform features of the 12 leads of the electrocardiogram and the three-dimensional spatial vector features of the electrocardiogram vector map respectively. Use the mid-term feature fusion strategy to fuse the temporal waveform features and spatial vector features to generate a unified spatiotemporal fusion feature vector.
[0035] This step specifically includes: extracting temporal morphological features, including QRS width, amplitude, and transition leads, from the 12 leads of the electrocardiogram; extracting spatial conduction features, including the direction of rotation of the loop, the area of the loop, and the maximum vector angle, from the frontal, transverse, and lateral three-plane vector loops of the electrocardiogram; and using convolutional splicing to perform intermediate-term fusion of the temporal morphological features and spatial conduction features, filtering out redundant high-dimensional features, and generating a unified spatiotemporal fusion feature vector strongly correlated with ectopic lesions.
[0036] Step S5: Perform hierarchical AI prediction: Input the spatiotemporal fusion feature vector into the pre-trained deep learning model and output the anatomical prediction location of the ectopic pacemaker in a hierarchical manner.
[0037] In this embodiment, the pre-trained deep learning model adopts a hybrid architecture of ResNet and Transformer, and incorporates a four-level progressive classifier for prediction. The first level distinguishes between the origins of the left and right ventricles, the second level distinguishes between large anatomical regions, the third level refines subregions, and the fourth level accurately outputs the estimated location of the earliest pacing target. Simultaneously, the pre-trained deep learning model also incorporates an attention mechanism to simultaneously calculate the contribution weights of each ECG lead and each ECG vector map feature vector loop to the prediction results during the hierarchical prediction process.
[0038] Step S6: Perform visualization display: Simultaneously display the AI-predicted target area location and corresponding key ECG leads and feature vector loops, and generate a pre-radiofrequency ablation localization reference report.
[0039] In this embodiment, the visualization step further includes highlighting the key ECG leads and key feature vector loops with the highest contribution weight as the basis for model judgment, so as to enhance the interpretability of the prediction results.
[0040] Step S7: Perform data feedback iteration: Collect successful electrophysiological ablation targets during the procedure as gold standard data. After associating and annotating the gold standard data with the corresponding bimodal ECG data, feed the data back to the standardized database. Use cross-validation to periodically retrain and fine-tune the weights of the pre-trained deep learning model using the expanded database to achieve continuous improvement in model accuracy.
[0041] Based on the same inventive concept, embodiments of the present invention also provide a positioning system employing the above-described intelligent positioning method for ventricular ectopic rhythms, comprising: The synchronous acquisition module is used to simultaneously acquire the raw signals of the 12-lead surface electrocardiogram and the raw signals of the Frank system electrocardiogram vector map of the same patient. The signal preprocessing module is used to preprocess and reduce noise, correct the baseline, and automatically segment the QRS feature points of ectopic ventricular beats in the acquired raw signals. The spatiotemporal registration module is used to synchronously align the electrocardiogram time sequence signal with the three-dimensional vector loop of the electrocardiogram vector map point-to-point, using the ectopic QRS wave of the ventricle as the anchor point, and to map the registered electrical signal to the three-dimensional anatomical space coordinates of the heart in combination with the anterior chest anatomical positioning marks. The multimodal feature fusion module is used to extract the temporal waveform features and the three-dimensional spatial vector features of the ECG vector map of the 12 leads respectively. The temporal waveform features and spatial vector features are fused together using an intermediate feature fusion strategy to generate a unified spatiotemporal fusion feature vector. The hierarchical AI prediction model module is used to input the spatiotemporal fusion feature vector into a pre-trained deep learning model and output the anatomical prediction location of the ectopic pacemaker in a hierarchical manner. The visualization output module is used to simultaneously display the AI-predicted target area location and the corresponding key ECG leads and feature vector loops, generating a pre-radiofrequency ablation localization reference report.
[0042] In this embodiment, the system further includes a model iteration and update module, which is used to collect successful electrophysiological ablation targets during the operation as gold standard data, associate and label the gold standard data with the corresponding bimodal electrocardiogram data and then feed them back to the standardized database, and periodically retrain and fine-tune the pre-trained deep learning model using the expanded database through cross-validation.
[0043] The beneficial effects of this invention are: 1. Dual-modal spatiotemporal fusion makes up for the shortcomings of single electrocardiogram spatial information, greatly improves the ability to distinguish complex overlapping and ectopic waveforms, and solves the problem of "waveforms not being clear"; 2. By training an AI model based on the gold standard dataset of intracavitary electrophysiology, the location of lesions can be quantitatively output, eliminating reliance on human experience. The accuracy of locating special high-risk areas has been improved to over 90%, solving the problem of "inaccurate origin". 3. Precise preoperative delineation of the ablation target area reduces the time for intraoperative three-dimensional mapping by more than 30%, thereby reducing ineffective ablation, X-ray radiation, and myocardial catheter damage, and solving the problem of "difficulty in ablation calibration"; 4. Establish a closed loop for data feedback and iteration, continuously improve the model's generalization ability as clinical samples accumulate, and adapt to large-scale promotion in multi-center and primary hospitals; 5. The medical-engineering interdisciplinary architecture integrates hardware and software, making it compatible with existing ECG examination equipment and valuable for clinical translation and the implementation of standardized group standards.
[0044] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention. The scope of the invention is defined by the appended claims and their equivalents, all of which should be included within the scope of protection of the invention.
Claims
1. A method for intelligent localization of ventricular ectopic rhythms based on electrocardiogram vector map fusion, characterized in that, include: Simultaneously acquire raw signals of 12-lead surface electrocardiogram and Frank system vectorcardiogram of the same patient; The acquired raw signals are preprocessed for noise reduction and baseline correction, and the QRS feature points of ectopic ventricular beats are automatically segmented. The steps for precise spatiotemporal registration are as follows: using the ectopic QRS wave of the ventricle as the anchor point, the electrocardiogram time sequence signal and the three-dimensional vector loop of the electrocardiogram are synchronously aligned point by point, and the registered electrical signal is mapped to the three-dimensional anatomical space coordinates of the heart in combination with the anatomical positioning marks on the chest. Multimodal feature fusion steps: Extract the temporal waveform features and the three-dimensional spatial vector features of the ECG vector map from the 12 leads of the ECG respectively, and use the mid-term feature fusion strategy to fuse the temporal waveform features and spatial vector features to generate a unified spatiotemporal fusion feature vector; Hierarchical AI prediction steps: Input the spatiotemporal fusion feature vector into the pre-trained deep learning model and output the anatomical prediction location of the ectopic pacemaker in a hierarchical manner. Visualization steps: Simultaneously display the AI-predicted target area location and corresponding key ECG leads and feature vector loops, and generate a pre-radiofrequency ablation localization reference report.
2. The intelligent localization method for ventricular ectopic rhythm based on electrocardiogram vector map fusion according to claim 1, characterized in that, The precise spatiotemporal registration steps specifically include: Using the QRS wave origin of each ectopic ventricular beat as the timing alignment anchor point, the QRS origin, peak, and end point on the electrocardiogram timing waveform of the same cardiac cycle are matched point-to-point with the corresponding turning nodes on the three-dimensional vector loop of the electrocardiogram vector map. Combined with the synchronously acquired anterior chest anatomical positioning markers, the registered dual-modal electrical signals are mapped to the standard cardiac three-dimensional anatomical space coordinate system.
3. The intelligent localization method for ventricular ectopic rhythm based on electrocardiogram vector map fusion according to claim 1, characterized in that, The multimodal feature fusion step specifically includes: Temporal morphological features, including QRS width, amplitude, and transition leads, were extracted from the 12 leads of the electrocardiogram. Spatial conduction features, including the direction of rotation of the loop, the area of the loop, and the maximum vector angle, are extracted from the three-plane vector loops of the frontal, transverse, and lateral planes of the electrocardiogram vector map. The temporal morphological features and spatial transmission features are fused in the middle stage by convolutional splicing, and redundant high-dimensional features are filtered out to generate a unified spatiotemporal fusion feature vector that is strongly correlated with ectopic lesions.
4. The intelligent localization method for ventricular ectopic rhythm based on electrocardiogram vector map fusion according to claim 1, characterized in that, The pre-trained deep learning model adopts a hybrid architecture of ResNet and Transformer, with a built-in four-level progressive classifier for prediction. The first level is used to distinguish the origin of the left and right ventricles, the second level is used to distinguish large anatomical regions, the third level is used to refine subregions, and the fourth level is used to accurately output the estimated location of the earliest pacing target.
5. The intelligent localization method for ventricular ectopic rhythm based on electrocardiogram vector map fusion according to claim 4, characterized in that, The pre-trained deep learning model also has a built-in attention mechanism, which is used to simultaneously calculate the contribution weights of each ECG lead and each ECG vector map feature vector loop to the prediction results during the hierarchical prediction process. The visualization step also includes highlighting the key ECG leads and key feature vector loops with the highest contribution weight as the basis for model judgment.
6. The intelligent localization method for ventricular ectopic rhythm based on electrocardiogram vector map fusion according to claim 1, characterized in that, It also includes a data backflow iteration step: Successful electrophysiological ablation targets during the procedure are collected as gold standard data. The gold standard data is then associated and labeled with the corresponding bimodal electrocardiogram data and fed back into a standardized database. Cross-validation is used to periodically retrain and fine-tune the weights of the pre-trained deep learning model using the expanded database.
7. The intelligent localization method for ventricular ectopic rhythm based on electrocardiogram vector map fusion according to claim 1, characterized in that, In the synchronous acquisition step, a unified hardware sampling clock is used to ensure that the timing error between the electrocardiogram signal and the electrocardiogram vector graph signal is less than or equal to 1 millisecond.
8. The intelligent localization method for ventricular ectopic rhythm based on electrocardiogram vector map fusion according to claim 1, characterized in that, The preprocessing noise reduction includes automatically filtering out electromyographic noise and power frequency noise; the baseline correction includes correcting respiratory baseline deviation; and the automatic segmentation of QRS feature points for ectopic ventricular beats includes automatically identifying the start point, peak value, and end point of the QRS wave, as well as the turning point of the ECG vector loop.
9. A positioning system employing the intelligent positioning method for ventricular ectopic rhythms based on electrocardiogram vector map fusion as described in any one of claims 1 to 8, characterized in that, include: The synchronous acquisition module is used to simultaneously acquire the raw signals of the 12-lead surface electrocardiogram and the raw signals of the Frank system electrocardiogram vector map of the same patient. The signal preprocessing module is used to preprocess and reduce noise, correct the baseline, and automatically segment the QRS feature points of ectopic ventricular beats in the acquired raw signals. The spatiotemporal registration module is used to synchronously align the electrocardiogram time sequence signal with the three-dimensional vector loop of the electrocardiogram vector map point-to-point, using the ectopic QRS wave of the ventricle as the anchor point, and to map the registered electrical signal to the three-dimensional anatomical space coordinates of the heart in combination with the anterior chest anatomical positioning marks. The multimodal feature fusion module is used to extract the temporal waveform features and the three-dimensional spatial vector features of the ECG vector map of the 12 leads respectively. The temporal waveform features and spatial vector features are fused together using an intermediate feature fusion strategy to generate a unified spatiotemporal fusion feature vector. The hierarchical AI prediction model module is used to input the spatiotemporal fusion feature vector into a pre-trained deep learning model and output the anatomical prediction location of the ectopic pacemaker in a hierarchical manner. The visualization output module is used to simultaneously display the AI-predicted target area location and the corresponding key ECG leads and feature vector loops, generating a pre-radiofrequency ablation localization reference report.
10. The intelligent localization system for ventricular ectopic rhythm based on electrocardiogram vector map fusion according to claim 9, characterized in that, It also includes a model iteration and update module, which is used to collect successful electrophysiological ablation targets during the operation as gold standard data, associate and label the gold standard data with the corresponding bimodal electrocardiogram data and then feed them back to the standardized database. The pre-trained deep learning model is retrained and its weights are fine-tuned periodically using the expanded database through cross-validation.
