Portable real-time recording equipment and real-time generation method for electrocardiosignal three-dimensional excitation conduction dynamic graph based on deep learning

Abstract

The invention discloses portable real-time recording equipment and a real-time generation method for an electrocardiosignal three-dimensional excitation conduction dynamic graph based on deep learning, and belongs to the technical field of electrocardio physiological signal processing and visualization. In order to solve the problems that a traditional electrocardiogram cannot visually display the cardiac electrical excitation three-dimensional dynamic conduction process, and an existing three-dimensional mapping system is large in size, cannot be portable and used and is difficult to continuously display a three-dimensional excitation track in real time, the system comprises a body surface potential mapping electrode array for collecting electrocardiosignals; the portable collection box is used for synchronously collecting the electrocardiosignals, a deep learning reasoning module is integrated in the portable collection box, the electrocardiosignals are processed based on a CNN-DRNN model, and potential and three-dimensional coordinate information of an electrocardio excitation point are output in real time; and the terminal display equipment is connected with the portable acquisition box in a wireless communication mode, receives the potential and the three-dimensional coordinate information, and performs three-dimensional dynamic display according to the potential and the three-dimensional coordinate information so as to present an electrocardio excitation conduction track. The method is used for electrocardiosignal analysis.

Description

Technical Field

[0001] This invention relates to a portable real-time recording device and a real-time generation method for a three-dimensional dynamic diagram of electrocardiogram signal excitation conduction, belonging to the field of electrocardiographic signal processing and visualization technology. Background Technology

[0002] Electrocardiogram (ECG) signal analysis is one of the most widely used and non-invasive methods for assessing cardiac function in clinical practice, especially in locating abnormal activation points in atrial and ventricular tachyarrhythmias (such as supraventricular tachycardia, atrial tachycardia, and ventricular tachycardia). Since William Einthofen invented the string-type ECG machine in 1903, traditional ECG technology has evolved into 12-lead and even 18-lead systems. However, it still cannot directly reflect the real-time conduction process of cardiac electrical excitation in three-dimensional space, and suffers from limitations such as separation of temporal and spatial information and a single analytical perspective.

[0003] In recent years, some technologies have attempted to achieve multidimensional display of electrocardiogram (ECG) signals. For example, the 1989 patent "Imaging Method and Imager for Stereoscopic Electrocardiogram" and the 2011 patent "Implementation Method and System of Four-Dimensional ECG Diagnostic Instrument" proposed the concept of multidimensional ECG, providing a spatiotemporal approach to ECG analysis. However, they still lack detailed cardiac electroanatomical mapping functions and cannot achieve three-dimensional dynamic visualization of excitation conduction. In addition, although the abnormal excitation point localization method based on CNN and LSTM proposed by Zhejiang University in 2019 (CN110555388A) introduces deep learning technology, it only performs abnormal point localization for 12-lead data and does not achieve a three-dimensional real-time portable application of multi-lead excitation conduction process.

[0004] Regarding existing products, common Holter monitors (such as Philips DigiTrak XT, GE SEER 12, Mindray EPM10, etc.) can achieve long-term ECG recording, but none of them have the function of three-dimensional dynamic display of cardiac excitation conduction. Medtronic's "CardioInsight Cardiac Mapping System" can provide non-invasive three-dimensional cardiac electrical signal anatomy maps, but it needs to be used in conjunction with CT images, and the device is bulky (1300mm high, 664mm deep, 615mm wide, and weighs 102kg), which cannot meet the clinical needs for portability and real-time monitoring, and is especially unsuitable for long-term preoperative monitoring and pre-location of abnormalities.

[0005] Therefore, there is still a lack of existing technologies for a device that can realize real-time, portable three-dimensional dynamic display of electrocardiogram signal excitation and conduction, especially suitable for preoperative localization and observation of conditions such as paroxysmal ventricular arrhythmias that are not suitable for intraoperative induction. Summary of the Invention

[0006] The purpose of this invention is to solve the problems that traditional electrocardiograms cannot intuitively display the three-dimensional dynamic conduction process of cardiac electrical excitation, and that existing three-dimensional mapping systems are bulky, not portable, and difficult to display three-dimensional excitation trajectories in real time and continuously. This invention provides a portable real-time recording device and a real-time generation method for a dynamic three-dimensional excitation conduction map of electrocardiogram signals based on deep learning.

[0007] The present invention discloses a portable real-time recording device for a three-dimensional dynamic graph of electrocardiogram signal excitation conduction based on deep learning, comprising:

[0008] Surface potential mapping electrode array, in the form of The equipment is set up for collecting electrocardiogram (ECG) signals.

[0009] The portable acquisition box is connected to the body surface potential mapping electrode array to synchronously acquire electrocardiogram (ECG) signals. It integrates a deep learning inference module, which processes the acquired ECG signals based on the CNN-DRNN model and outputs the potential and three-dimensional coordinate information of the ECG excitation point in real time.

[0010] The terminal display device is connected to the portable acquisition box via wireless communication, receives the potential and three-dimensional coordinate information, and performs three-dimensional dynamic display to present the electrocardiogram conduction trajectory.

[0011] Preferably, the body surface potential mapping electrode array has at least eight reference electrodes, which are located at the top center, bottom center and symmetrical positions on the back of the array, respectively.

[0012] Eight reference electrodes are used to acquire reference potentials, and the average value of the reference potentials of each reference electrode is used as the potential reference point, which is set as the origin of the three-dimensional coordinate system.

[0013] Preferably, the portable data collection box includes:

[0014] A multi-channel switch is used to select each electrode channel;

[0015] A multi-channel analog-to-digital converter for acquiring electrocardiogram signals with microsecond-level synchronization accuracy;

[0016] The microcontroller unit integrates a neural network processing unit for running CNN-DRNN models for real-time inference;

[0017] Memory, used to store the acquired raw data or model parameters;

[0018] A wireless communication module for transmitting data with a terminal display device;

[0019] The power module supplies power to all components.

[0020] Preferably, the multi-channel analog-to-digital converter uses at least two eight-channel high-precision ECG front-end ADC chips, in conjunction with a multiplexer to achieve [the desired function]. Synchronous sampling of the electrode array.

[0021] Preferably, the CNN-DRNN model includes:

[0022] The first feature extraction module consists of four parallel two-dimensional convolutional neural network layers A1~A4 followed by four max-pooling layers B1~B4 connected sequentially, used for processing the input... Primary spatial feature extraction and dimensionality reduction of ECG data;

[0023] The second feature extraction module consists of four two-dimensional convolutional neural network layers C1~C4 and four subsequent max pooling layers D1~D4 connected in sequence. It is used to extract deep spatial features from primary features and further reduce dimensionality.

[0024] A fully connected layer F1 is used to fuse and map the dimensionality-reduced deep features;

[0025] The temporal fusion module consists of four DRNN hidden layers G1~G4 connected sequentially, which is used to model the temporal dependencies of the features output by the fully connected layer.

[0026] The output layer H1 is used to predict the quadrupole information of the ECG activation point based on the output of the temporal fusion module.

[0027] The output of each of the two-dimensional convolutional neural network layers is sequentially processed by standardization and ReLU activation function.

[0028] Preferably, the structural parameters of the CNN-DRNN model are configured as follows:

[0029] The two-dimensional convolutional neural network layers A1~A4 in the first feature extraction module all use two-dimensional convolutional kernels of size 4×24 with a stride of 1; The output dimension of the input data is ;

[0030] The max-pooling layers B1-B4 reduce the length and width of the input feature map by half, and their output dimension is... ;

[0031] In the second feature extraction module, the two-dimensional convolutional neural network layers C1-C4 use the same kernel size and stride as layers A1-A4. The subsequent max-pooling layers D1-D4 perform the same dimensionality reduction operation as layers B1-B4. The final output dimension of this module is... ;

[0032] The number of neurons in the fully connected layer F1 is ;

[0033] The number of neurons in each DRNN hidden layer G1~G4 in the temporal fusion module is: ;

[0034] The output layer H1 has 10 neurons, and each neuron outputs a quadruple of information about an electrocardiographic activation point. The quadruple includes a potential value and a three-dimensional spatial coordinate value.

[0035] The present invention discloses a real-time generation method for a three-dimensional dynamic graph of electrocardiogram signal excitation conduction based on deep learning, which includes the following steps:

[0036] Data collected via a body surface potential mapping electrode array ECG signal;

[0037] The acquired signals are preprocessed, including background removal, drift removal, zero-phase filtering, and normalization.

[0038] The preprocessed data is input into a pre-trained CNN-DRNN model, which infers the potential and three-dimensional coordinates of the electrocardiogram excitation point in real time.

[0039] The reasoning results are sent to the terminal device in the form of a 3D video stream for dynamic display.

[0040] Preferably, the CNN-DRNN model is trained in the following manner:

[0041] Simultaneously collect surface potential mapping data and intracardiac electrical signal mapping data as training samples;

[0042] The body surface data was used as input, and the excitation point quadruples in the intracardiac data were used as labels.

[0043] Convolutional neural networks are used to extract spatial features, DRNNs are used to fuse temporal features, and fully connected layers are used for coordinate prediction.

[0044] Preferably, during training, a grouped data collection method is used, with each group of data lasting for 2 seconds, and the body surface data dimension is [missing information]. The dimensions of the tag data are .

[0045] Preferably, the terminal display device is a smartphone or tablet computer, which has a built-in processor for 3D image reconstruction and display, and has backup inference capabilities.

[0046] The advantages of this invention are as follows: A portable real-time recording device and real-time generation method for a three-dimensional dynamic graph of cardiac electrical impulse conduction based on deep learning creatively realizes three-dimensional, dynamic, and real-time visualization of the cardiac electrical impulse conduction process, and for the first time completes the entire process integration from high-precision signal acquisition and intelligent real-time analysis to wireless visualization output in a truly portable system. Specifically, this is reflected in the following three aspects:

[0047] 1. This invention represents a leap in diagnostic capabilities from "static waveforms" to "dynamic 3D imaging," significantly improving the intuitiveness and efficiency of abnormal point localization. Traditional 12 / 18-lead electrocardiograms only provide two-dimensional waveforms of cardiac electrical activity, while this invention directly outputs the three-dimensional coordinates and potentials of the excitation point through a deep learning model and reconstructs them into a dynamic image. This allows doctors to intuitively observe the origin and conduction path of the excitation, much like watching a video, providing an unprecedented intuitive visualization method, especially for preoperative localization of abnormal excitation points that are difficult to induce, such as paroxysmal ventricular arrhythmias.

[0048] 2. This invention achieves the analytical performance of large-scale fixed systems on a portable device through precise microsecond-level synchronous sampling and embedded real-time inference. It is not simply a data forwarding process; its core technological advantage lies in completing the core computation within the portable acquisition box. The device achieves microsecond-level synchronous sampling of M×N electrodes (specifically a 16x12 array) via a high-speed multi-channel ADC and utilizes an integrated NPU (providing 6 TOPS of computing power) to run a CNN-DRNN model for online real-time inference, outputting excitation point quadruples with extremely low latency (the model can output 10 quadruples every 2 seconds of data). This ensures the "real-time" performance of the 3D dynamic display, enabling the device to combine the analytical depth of large-scale systems with the portability and flexibility of Holter devices.

[0049] 3. It expands the application scenarios of 3D ECG mapping, providing a portable new tool for preoperative diagnosis and long-term monitoring of arrhythmias. Due to its miniaturization, low power consumption, and wireless connectivity, its application is no longer limited to the catheterization lab. Patients can wear the device for long-term daily monitoring, capturing paroxysmal arrhythmia events and recording their three-dimensional excitation origins, providing crucial preoperative planning information for ablation surgery. This effectively solves the clinical challenge of accurately locating ventricular arrhythmias preoperatively due to their high induced risk, extending advanced 3D mapping technology to a wider range of screening, diagnostic, and preoperative assessment scenarios. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the internal structure of the portable data collection box described in this invention;

[0051] Figure 2 This is a schematic diagram of the deep learning framework based on CNN-DRNN described in this invention;

[0052] Figure 3 This is a schematic diagram of the CNN-DRNN network structure described in this invention;

[0053] Figure 4 This is a schematic diagram of the overall composition of the portable recording device described in this invention. Detailed Implementation

[0054] 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.

[0055] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0056] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the scope of the invention.

[0057] Example 1:

[0058] The following is combined Figures 1-4 This embodiment describes a portable real-time recording device for a three-dimensional dynamic graph of electrocardiogram signals based on deep learning, comprising:

[0059] Surface potential mapping electrode array, in the form of The equipment is set up for collecting electrocardiogram (ECG) signals.

[0060] The portable acquisition box is connected to the body surface potential mapping electrode array to synchronously acquire electrocardiogram (ECG) signals. It integrates a deep learning inference module, which processes the acquired ECG signals based on the CNN-DRNN model and outputs the potential and three-dimensional coordinate information of the ECG excitation point in real time.

[0061] The terminal display device is connected to the portable acquisition box via wireless communication, receives the potential and three-dimensional coordinate information, and performs three-dimensional dynamic display to present the electrocardiogram conduction trajectory.

[0062] Furthermore, the surface potential mapping electrode array is provided with at least eight reference electrodes, which are located at the top center, bottom center and symmetrical positions on the back of the array front, respectively.

[0063] Eight reference electrodes are used to acquire reference potentials, and the average value of the reference potentials of each reference electrode is used as the potential reference point, which is set as the origin of the three-dimensional coordinate system.

[0064] Furthermore, the portable data collection box includes:

[0065] A multi-channel switch is used to select each electrode channel;

[0066] A multi-channel analog-to-digital converter for acquiring electrocardiogram signals with microsecond-level synchronization accuracy;

[0067] The microcontroller unit integrates a neural network processing unit for running CNN-DRNN models for real-time inference;

[0068] Memory, used to store the acquired raw data or model parameters;

[0069] A wireless communication module for transmitting data with a terminal display device;

[0070] The power module supplies power to all components.

[0071] Furthermore, the multi-channel analog-to-digital converter employs at least two eight-channel high-precision ECG front-end ADC chips, in conjunction with a multiplexer to achieve... Synchronous sampling of the electrode array.

[0072] Furthermore, the CNN-DRNN model includes:

[0073] The first feature extraction module consists of four parallel two-dimensional convolutional neural network layers A1~A4 followed by four max-pooling layers B1~B4 connected sequentially, used for processing the input... Primary spatial feature extraction and dimensionality reduction of ECG data;

[0074] The second feature extraction module consists of four two-dimensional convolutional neural network layers C1~C4 and four subsequent max pooling layers D1~D4 connected in sequence. It is used to extract deep spatial features from primary features and further reduce dimensionality.

[0075] A fully connected layer F1 is used to fuse and map the dimensionality-reduced deep features;

[0076] The temporal fusion module consists of four DRNN hidden layers G1~G4 connected sequentially, which is used to model the temporal dependencies of the features output by the fully connected layer.

[0077] The output layer H1 is used to predict the quadrupole information of the ECG activation point based on the output of the temporal fusion module.

[0078] The output of each of the two-dimensional convolutional neural network layers is sequentially processed by standardization and ReLU activation function.

[0079] Furthermore, the structural parameters of the CNN-DRNN model are configured as follows:

[0080] The two-dimensional convolutional neural network layers A1~A4 in the first feature extraction module all use two-dimensional convolutional kernels of size 4×24 with a stride of 1; The output dimension of the input data is ;

[0081] The max-pooling layers B1-B4 reduce the length and width of the input feature map by half, and their output dimension is... ;

[0082] In the second feature extraction module, the two-dimensional convolutional neural network layers C1-C4 use the same kernel size and stride as layers A1-A4. The subsequent max-pooling layers D1-D4 perform the same dimensionality reduction operation as layers B1-B4. The final output dimension of this module is... ;

[0083] The number of neurons in the fully connected layer F1 is ;

[0084] The number of neurons in each DRNN hidden layer G1~G4 in the temporal fusion module is: ;

[0085] The output layer H1 has 10 neurons, and each neuron outputs a quadruple of information about an electrocardiographic activation point. The quadruple includes a potential value and a three-dimensional spatial coordinate value.

[0086] Example 2:

[0087] The real-time generation method for a three-dimensional dynamic graph of electrocardiogram signal based on deep learning described in this embodiment includes the following steps:

[0088] Data collected via a body surface potential mapping electrode array ECG signal;

[0089] The acquired signals are preprocessed, including background removal, drift removal, zero-phase filtering, and normalization.

[0090] The preprocessed data is input into a pre-trained CNN-DRNN model, which infers the potential and three-dimensional coordinates of the electrocardiogram excitation point in real time.

[0091] The reasoning results are sent to the terminal device in the form of a 3D video stream for dynamic display.

[0092] Furthermore, the CNN-DRNN model is trained in the following manner:

[0093] Simultaneously collect surface potential mapping data and intracardiac electrical signal mapping data as training samples;

[0094] The body surface data was used as input, and the excitation point quadruples in the intracardiac data were used as labels.

[0095] Convolutional neural networks are used to extract spatial features, DRNNs are used to fuse temporal features, and fully connected layers are used for coordinate prediction.

[0096] Furthermore, during training, a grouped data collection method was used, with each group of data lasting for 2 seconds, and the body surface data dimension being [missing information]. The dimensions of the tag data are .

[0097] Furthermore, the terminal display device is a smartphone or tablet computer, with a built-in processor for 3D image reconstruction and display, and has backup inference capabilities.

[0098] This invention discloses a portable real-time recording device and method for generating a three-dimensional dynamic graph of electrocardiogram (ECG) signal activation and conduction based on deep learning. For example... Figure 4 As shown, the device consists of three main parts: a flexible electrode array for body surface potential mapping (BSPM), a portable acquisition box, and a smart mobile device (such as a smartphone or tablet) serving as a display terminal. The electrode array and the portable acquisition box are connected via flexible circuitry, and the portable acquisition box and the display terminal communicate wirelessly via Bluetooth Low Energy (BLE).

[0099] Hardware Composition and Data Acquisition: The BSPM electrode array is made of flexible material and arranged in an M-row × N-column matrix for attaching to the patient's body surface to collect electrocardiogram (ECG) signals. In a specific configuration of this embodiment, a 16-row × 12-column (M=16, N=12) electrode layout is used. Eight green-marked electrodes are placed as reference electrodes at the top center, bottom center, and symmetrically positioned on the back of the array. The average potential value of the signals acquired by these eight reference electrodes is set as the measurement reference point, i.e., the origin of the three-dimensional coordinate system, and a cardiac three-dimensional coordinate system for localization is constructed based on this.

[0100] like Figure 1 As shown, the portable acquisition box is the core hardware unit for completing signal acquisition and real-time inference, and its internal components include:

[0101] Multi-channel switch: This embodiment uses eight 16-channel multiplexer chips TMUX9616 to sequentially select 16×12, a total of 192 electrode channels.

[0102] Multi-channel ADC: Two Texas Instruments (TI) ADS1292 eight-channel, 24-bit high-precision ECG front-end ADC chips are used. These chips have a sampling rate of 32Ksps, enabling synchronous acquisition of ECG signals from all channels with microsecond-level deviations, ensuring spatial synchronization of the signals. In this configuration, the system achieves 1Ksps real-time data sampling from 192 electrode channels.

[0103] Microcontroller Unit (MCU): The Rockchip RK3567 chip is used as the main controller. This MCU integrates eight high-performance ARM Cortex-A72 cores (2.2GHz) and a Neural Processing Unit (NPU), which can provide up to 6 TOPS of computing power and is specifically designed to support online inference operations for CNN-DRNN deep learning models.

[0104] Data storage: Onboard Flash memory is used to temporarily store the raw ECG data acquired by the ADC, which can be played back or read externally.

[0105] Communication and Power: An integrated BLE interface is used for wireless data transmission with the display terminal, and a dedicated power management module is provided to power the entire system.

[0106] NN-DRNN Deep Learning Model Construction and Training:

[0107] This invention employs a specific CNN-DRNN hybrid deep learning model to achieve the mapping from body surface potential to the three-dimensional coordinates of intracardiac excitation points. Figure 2 This demonstrates the overall training and inference framework of the model. Figure 3 Its network structure is described in detail.

[0108] Data preparation and preprocessing: For training the model, it is necessary to simultaneously collect paired surface potential data and intracardiac excitation point coordinate label data.

[0109] In practice, while using this device to collect surface BSPM data, a 10-electrode catheter from a STARTREK 3D mapping system, such as that from Jianhu Medical Technology, is used to simultaneously collect mapping data of intracardiac electrical signals, namely the quadruple information of activation points (3D spatial coordinates X, Y, Z and potential value V). Both types of data need to be accurately time-stamped for subsequent alignment in the time dimension.

[0110] Data was collected in 2-second increments. Within each data set, the surface BSPM data had a dimension of 16×12×2000 (electrode rows × columns × sampling points), while the intracardiac 3D mapping data, used as labels, had a dimension of 10×4×1000 (number of activation points × quadruplet dimension × sampling points). A total of 100 data sets were collected, divided in a 6:2:2 ratio into a training set (60 sets), a test set (20 sets), and a validation set (20 sets).

[0111] Before being input into the model, the BSPM data needs to be preprocessed, including removing background noise, removing baseline drift, performing zero-phase filtering to eliminate phase distortion, and finally normalizing to improve the stability and convergence speed of model training.

[0112] Model network structure, such as Figure 3 As shown, the specific structure and parameters of the CNN-DRNN model are as follows:

[0113] The first-level feature extraction consists of four parallel two-dimensional convolutional neural network layers (A1~A4). Each convolutional layer uses a 4×24 kernel with a stride of 1. For input M×N (16×12) data, the output feature map dimension is (M-4)×(N-24), which is the adjusted effective dimension of (16-4)×(12-24). Subsequently, four max-pooling layers (B1~B4) are connected. Each pooling layer halves the length and width of the feature map, and the output dimension becomes (M / 2 - 2)×(N / 2 - 12), which is the adjusted dimension of (8-2)×(6-12).

[0114] The second-level feature extraction consists of four 2D convolutional neural network layers (C1~C4) followed by four max-pooling layers (D1~D4). The convolutional layers (C1~C4) use the same kernel parameters as the first level (4×24, stride 1), and the pooling layers (D1~D4) perform the same halving operation as (B1~B4). After this level of processing, the dimension of the output feature map is (M / 4 - 1)×(N / 4 - 6), which is the adjusted dimension of (4-1)×(3-6).

[0115] Each convolutional layer is followed by batch normalization and ReLU activation.

[0116] Feature fusion and mapping: The feature map output by the second-level feature extraction module is flattened and fed into a fully connected layer (F1). The number of neurons in this fully connected layer is (M-4)×(N-24), that is, 12×(-12) adjusted actual number, which is responsible for high-level feature fusion and spatial mapping.

[0117] Temporal modeling: The output of the fully connected layer is then fed into the temporal fusion module, which consists of four DRNN (Deep Recurrent Neural Network) hidden layers (G1~G4) connected sequentially. Each DRNN hidden layer has (M-4) neurons and is used to capture the dynamic dependence of ECG excitation signals on the time series.

[0118] Output Layer: Finally, there is an output layer (H1) with a fixed number of 10 neurons. Each neuron outputs a complete four-tuple of information (V, X, Y, Z) for a specific electrocardiographic activation point. Therefore, the model can simultaneously output the state of 10 activation points at the current time step during each inference.

[0119] Model training and deployment:

[0120] Using an AI workstation equipped with an Intel i9-14900K processor, an NVIDIA RTX 5090 (32GB) graphics card, and 64GB of DDR5 memory, the aforementioned CNN-DRNN model was trained using a training set to minimize the loss between predicted coordinates and true labels. During training, a validation set was used to monitor model performance and prevent overfitting. After training, the finalized model parameters were optimized and lightweighted, then deployed to the RK3567 MCU in a portable acquisition box, where its integrated NPU handled real-time forward inference computation.

[0121] During the real-time application phase, the device operates according to the following process:

[0122] Signal acquisition and preprocessing: The electrode array continuously acquires surface electrocardiogram signals, which are converted into digital signals by a multiplexer and a high-speed ADC. The signals are then preprocessed in real time within the MCU (background removal, drift removal, zero-phase filtering, and normalization).

[0123] Real-time inference: The preprocessed M×N dimensional data stream is fed into the deployed CNN-DRNN model. The NPU performs model computation and outputs the quadruples (potentials and three-dimensional coordinates) of the 10 excitation points at the current time (or within a short time window) in real time.

[0124] Data transmission and visualization: The quadruple data obtained from inference is sent to the intelligent display terminal via the BLE interface. The display terminal uses its own GPU or CPU computing power to connect and render the discrete, time-series-arriving excitation point 3D coordinates in real time into a continuous 3D dynamic conduction trajectory map, and displays it on the screen in the form of a video stream.

[0125] The NPU of the display terminal can also serve as a backup computing resource, assisting or taking over some inference tasks when necessary.

[0126] Through the above implementation methods, this device achieves real-time, three-dimensional dynamic display of the cardiac electrical excitation conduction process. Clinicians can intuitively observe the origin and diffusion path of the excitation wave on the display terminal, which is particularly helpful in identifying and locating abnormal excitation points in paroxysmal arrhythmias preoperatively. The device is portable, wirelessly operable, and supports long-term monitoring, solving the clinical pain points of existing large-scale three-dimensional mapping systems being unusable due to their lack of portability and traditional electrocardiograms lacking spatial dimensional information. It provides an innovative tool for the accurate diagnosis and treatment planning of arrhythmias.

[0127] While the invention has been described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the invention. Therefore, it should be understood that many modifications can be made to the exemplary embodiments, and other arrangements can be designed without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that different dependent claims and features described herein can be combined in ways different from those described in the original claims. It is also understood that features described in conjunction with individual embodiments can be used in other described embodiments.

Claims

1. A portable real-time recording device for a three-dimensional dynamic graph of electrocardiogram signal excitation conduction based on deep learning, characterized in that, It includes: Surface potential mapping electrode array, in the form of The equipment is set up for collecting electrocardiogram (ECG) signals. The portable acquisition box is connected to the body surface potential mapping electrode array to synchronously acquire electrocardiogram (ECG) signals. It integrates a deep learning inference module, which processes the acquired ECG signals based on the CNN-DRNN model and outputs the potential and three-dimensional coordinate information of the ECG excitation point in real time. The terminal display device is connected to the portable acquisition box via wireless communication, receives the potential and three-dimensional coordinate information, and performs three-dimensional dynamic display to present the electrocardiogram conduction trajectory.

2. The portable real-time recording device for a three-dimensional dynamic graph of electrocardiogram signal excitation conduction based on deep learning according to claim 1, characterized in that, The body surface potential mapping electrode array has at least eight reference electrodes, which are located at the top center, bottom center and symmetrical positions on the back of the array front, respectively. Eight reference electrodes are used to acquire reference potentials, and the average value of the reference potentials of each reference electrode is used as the potential reference point, which is set as the origin of the three-dimensional coordinate system.

3. The portable real-time recording device for a three-dimensional dynamic graph of electrocardiogram signal excitation conduction based on deep learning according to claim 1, characterized in that, The portable data collection box includes: A multi-channel switch is used to select each electrode channel; A multi-channel analog-to-digital converter for acquiring electrocardiogram signals with microsecond-level synchronization accuracy; The microcontroller unit integrates a neural network processing unit for running CNN-DRNN models for real-time inference; Memory, used to store the acquired raw data or model parameters; A wireless communication module for transmitting data with a terminal display device; The power module supplies power to all components.

4. The portable real-time recording device for a three-dimensional dynamic graph of electrocardiogram signal excitation conduction based on deep learning according to claim 3, characterized in that, The multi-channel analog-to-digital converter uses at least two eight-channel high-precision ECG front-end ADC chips, along with a multiplexer to achieve [the desired output]. Synchronous sampling of the electrode array.

5. A portable real-time recording device for a three-dimensional dynamic graph of electrocardiogram signal excitation conduction based on deep learning, as described in claim 1, characterized in that, The CNN-DRNN model includes: The first feature extraction module consists of four parallel two-dimensional convolutional neural network layers A1~A4 followed by four max-pooling layers B1~B4 connected sequentially, used for processing the input... Primary spatial feature extraction and dimensionality reduction of ECG data; The second feature extraction module consists of four two-dimensional convolutional neural network layers C1~C4 and four subsequent max pooling layers D1~D4 connected in sequence. It is used to extract deep spatial features from primary features and further reduce dimensionality. A fully connected layer F1 is used to fuse and map the dimensionality-reduced deep features; The temporal fusion module consists of four DRNN hidden layers G1~G4 connected sequentially, which is used to model the temporal dependencies of the features output by the fully connected layer. The output layer H1 is used to predict the quadrupole information of the ECG activation point based on the output of the temporal fusion module. The output of each of the two-dimensional convolutional neural network layers is sequentially processed by standardization and ReLU activation function.

6. A portable real-time recording device for a three-dimensional dynamic graph of electrocardiogram signal excitation conduction based on deep learning, as described in claim 5, is characterized in that, The structural parameters of the CNN-DRNN model are configured as follows: The two-dimensional convolutional neural network layers A1~A4 in the first feature extraction module all use two-dimensional convolutional kernels of size 4×24 with a stride of 1; The output dimension of the input data is ; The max-pooling layers B1-B4 reduce the length and width of the input feature map by half, and their output dimension is... ; In the second feature extraction module, the two-dimensional convolutional neural network layers C1-C4 use the same kernel size and stride as layers A1-A4. The subsequent max-pooling layers D1-D4 perform the same dimensionality reduction operation as layers B1-B4. The final output dimension of this module is... ; The number of neurons in the fully connected layer F1 is ; The number of neurons in each DRNN hidden layer G1~G4 in the temporal fusion module is: ; The output layer H1 has 10 neurons, and each neuron outputs a quadruple of information about an electrocardiographic activation point. The quadruple includes a potential value and a three-dimensional spatial coordinate value.

7. A method for real-time generation of three-dimensional excitation conduction dynamic diagrams of electrocardiogram signals based on deep learning, characterized in that, It includes the following steps: Data collected via a body surface potential mapping electrode array ECG signal; The acquired signals are preprocessed, including background removal, drift removal, zero-phase filtering, and normalization. The preprocessed data is input into a pre-trained CNN-DRNN model, which infers the potential and three-dimensional coordinates of the electrocardiogram excitation point in real time. The reasoning results are sent to the terminal device in the form of a 3D video stream for dynamic display.

8. The real-time generation method for three-dimensional excitation conduction dynamic diagrams of electrocardiogram signals based on deep learning according to claim 7, characterized in that, The CNN-DRNN model is trained in the following way: Simultaneously collect surface potential mapping data and intracardiac electrical signal mapping data as training samples; The body surface data was used as input, and the excitation point quadruples in the intracardiac data were used as labels. Convolutional neural networks are used to extract spatial features, DRNNs are used to fuse temporal features, and fully connected layers are used for coordinate prediction.

9. The real-time generation method for a three-dimensional dynamic diagram of electrocardiogram signal excitation conduction based on deep learning according to claim 7, characterized in that, During training, a group data collection method was used, with each group of data lasting for 2 seconds. The body surface data dimension was [missing information]. The dimensions of the tag data are .

10. The real-time generation method for a three-dimensional dynamic diagram of electrocardiogram signal excitation conduction based on deep learning according to claim 7, characterized in that, The terminal display device is a smartphone or tablet computer, with a built-in processor for 3D image reconstruction and display, and has backup inference capabilities.

Images