A method and system for electrocardiogram signal analysis

Abstract

The application relates to the technical field of data analysis, and discloses an electrocardiosignal analysis method and system, which comprises the following steps: collecting original electrocardiosignal and physiological index data of a target user; performing adaptive filtering processing on the original electrocardiosignal based on the motion sensing state of an electrocardiograph unit to obtain target electrocardiosignal; extracting electrocardiosignal waveform features of the target electrocardiosignal by using a preset multi-scale morphological algorithm, performing data fusion on the electrocardiosignal waveform features and the physiological index data to obtain a feature fusion vector; generating an electrocardiosignal analysis model according to the embedded memory resources of a medical device and the network structure of a preset lightweight neural network model; training the electrocardiosignal analysis model according to pre-acquired medical health training data, analyzing the feature fusion vector by using the trained electrocardiosignal analysis model, and obtaining an electrocardiosignal graph of the target user. The application can improve the accuracy of electrocardiosignal analysis.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, and in particular to a method and system for electrocardiogram (ECG) signal analysis. Background Technology

[0002] In the current medical context, early management of cardiovascular diseases and chronic diseases such as diabetes and gout has become a crucial need in the healthcare field. Based on this, integrated medical monitoring devices are gradually becoming key tools in home and primary healthcare settings. These devices can collect real-time electrocardiogram signals and multiple physiological indicators, providing doctors with diagnostic information and assisting patients in self-management of their health.

[0003] While some existing medical devices can collect electrocardiogram (ECG) signals and physiological indicators, they suffer from significant technical deficiencies. For example, during signal acquisition, ECG signals and physiological indicator data are often acquired independently, lacking a strict time synchronization mechanism. This weakens the temporal correlation between the two, affecting the accuracy of subsequent data analysis. In signal processing, fixed-parameter filters are often used to process the raw ECG signals, resulting in numerous artifacts even under motion interference, reducing signal quality. Consequently, existing devices struggle to provide accurate and efficient analysis results in ECG signal analysis, leading to low accuracy. Summary of the Invention

[0004] This invention provides a method and system for electrocardiogram (ECG) signal analysis, the main purpose of which is to solve the problem of low accuracy in ECG signal analysis.

[0005] To achieve the above objectives, the present invention provides an electrocardiogram (ECG) signal analysis method, comprising:

[0006] The system collects the target user's raw electrocardiogram signal through the built-in electrocardiograph unit of the medical device and simultaneously collects the target user's physiological indicator data through the built-in physiological indicator monitoring unit of the medical device.

[0007] Based on the motion sensing state of the electrocardiograph unit, the original electrocardiogram signal is adaptively filtered to obtain the target electrocardiogram signal.

[0008] The electrocardiogram (ECG) waveform features of the target ECG signal are extracted using a preset multi-scale morphological algorithm. The ECG waveform features are then fused with the physiological index data to obtain a feature fusion vector.

[0009] An electrocardiogram (ECG) signal analysis model is generated based on the embedded memory resources of the medical device and the network structure of a preset lightweight neural network model.

[0010] The ECG signal analysis model is trained based on pre-acquired medical and health training data. The feature fusion vector is then analyzed using the trained ECG signal analysis model to obtain the ECG signal map of the target user.

[0011] Optionally, the synchronous collection of the target user's physiological indicator data through the physiological indicator monitoring unit built into the medical device includes:

[0012] A synchronous trigger signal is generated based on the acquisition timing of the electrocardiograph unit;

[0013] The synchronization trigger signal is sent to the physiological indicator monitoring unit, and the physiological indicator monitoring unit is controlled to perform data acquisition.

[0014] The system receives the timestamp-aligned physiological indicator feedback data from the physiological indicator monitoring unit based on the data acquisition action performed after execution, and verifies the temporal consistency between the physiological indicator feedback data and the original electrocardiogram signal.

[0015] When the verification result is consistent with the timing, the physiological indicator feedback data is used as the physiological indicator data, and the physiological indicator data is cached in the embedded memory of the medical device.

[0016] Optionally, the step of adaptively filtering the original electrocardiogram (ECG) signal based on the motion sensing state of the ECG machine unit to obtain the target ECG signal includes:

[0017] Acquire the triaxial acceleration data output by the built-in accelerometer of the electrocardiograph unit;

[0018] The motion intensity index and motion status signal of the electrocardiograph unit are determined based on the vector amplitude of the triaxial acceleration data.

[0019] Determine whether the target user is in a state of motion interference based on the motion state signal;

[0020] When the target user is in a state of motion interference, obtain the mapping relationship between the preset motion intensity threshold range and the filter type, and match the motion intensity index with the mapping relationship;

[0021] Extract the filter parameters corresponding to the matched filter type, reconstruct the filter parameters into a preset adaptive filter, and use the reconstructed adaptive filter to suppress motion artifacts in the original ECG signal to obtain the target ECG signal.

[0022] Optionally, the step of extracting the electrocardiogram waveform features of the target electrocardiogram signal using a preset multi-scale morphological algorithm includes:

[0023] Identify flat structural elements and morphological gradient operators at different scales in a pre-defined multi-scale morphological algorithm;

[0024] Morphological analysis of the target electrocardiogram signal was performed using the flat structural elements of different scales to obtain multi-scale electrocardiogram signals;

[0025] The waveform boundary features of the multi-scale electrocardiogram signal are enhanced according to the morphological gradient algorithm.

[0026] Extract waveform amplitude, waveform slope, and waveform duration features from the enhanced multiscale electrocardiogram signal;

[0027] The waveform amplitude feature, the waveform slope feature, and the waveform duration feature are determined as the ECG waveform features of the target ECG signal.

[0028] Optionally, the step of fusing the electrocardiogram waveform features with the physiological index data to obtain a feature fusion vector includes:

[0029] Extract the trend characteristics of physiological indicators from the physiological indicator data;

[0030] The electrocardiogram waveform features and the physiological index trend features are standardized respectively.

[0031] Based on pre-defined clinical prior knowledge, the first and second weighting coefficients of the standardized electrocardiogram waveform features and physiological index trend features are generated respectively.

[0032] Based on the first weighting coefficient and the second weighting coefficient, the standardized ECG waveform features and physiological indicator trend features are fused at the feature level to obtain a feature fusion vector.

[0033] Optionally, generating the electrocardiogram signal analysis model based on the embedded memory resources of the medical device and the network structure of a preset lightweight neural network model includes:

[0034] The available storage capacity and resource computing space are determined based on the embedded memory resources of the medical device.

[0035] The model parameter quantity of the preset lightweight neural network model is determined based on the available storage capacity and the resource computing space;

[0036] Identify the number of neural network layers and convolutional kernels in the network structure of a preset lightweight neural network model;

[0037] The number of neural network layers and the number of convolutional kernels in the network structure are adjusted according to the model parameters, and the adjusted network structure is compiled into an electrocardiogram signal analysis model that can be embedded into medical devices.

[0038] Optionally, adjusting the number of neural network layers and the number of convolutional kernels in the network structure according to the model parameters includes:

[0039] Using a preset network architecture search algorithm, different combinations of configuration parameters for the number of neural network layers and the number of convolutional kernels in the network structure are searched based on the model parameter quantity;

[0040] Analyze the model accuracy and inference speed of lightweight neural network models under different combinations of configuration parameters;

[0041] The configuration parameter combination that maximizes both model accuracy and inference speed is selected as the target configuration combination.

[0042] The number of neural network layers and the number of convolutional kernels in the network structure are adjusted based on the target configuration combination.

[0043] Optionally, training the electrocardiogram signal analysis model based on pre-acquired medical and health training data includes:

[0044] Obtain the sample combination feature vector corresponding to historical electrocardiogram signals and physiological index data, and obtain the electrocardiogram abnormality type corresponding to the sample combination feature vector;

[0045] Based on the complexity of the feature vector of the sample combination and the rarity of the ECG abnormality type, the pre-acquired medical and health training data is sorted from easy to difficult to generate a target training data sequence.

[0046] The probability distribution of ECG abnormality types corresponding to the target training data sequence is analyzed using the ECG signal analysis model.

[0047] The loss value of the probability distribution is calculated using a preset transentropy loss function. When the loss value is greater than or equal to a preset loss threshold, the learning rate and batch parameters in the ECG signal analysis model training process are dynamically adjusted according to the target training data sequence until the loss value is less than the preset loss threshold.

[0048] When the loss value is less than the preset loss threshold, the model parameters of the ECG signal analysis model are adjusted based on the adjusted learning rate and batch parameters, and the trained ECG signal model is determined according to the model parameters.

[0049] Optionally, the step of analyzing the feature fusion vector using a trained electrocardiogram signal analysis model to obtain the electrocardiogram of the target user includes:

[0050] The feature fusion vector is input into the input layer of the electrocardiogram signal analysis model;

[0051] The deep implicit correlation features between the ECG waveform features and physiological index features in the feature fusion vector are extracted by using the multi-branch deep separable convolutional layer inside the ECG signal analysis model.

[0052] The deep implicit correlation features are fused to obtain fused features, and the feature contribution weights of the electrocardiogram waveform features and the physiological index features are calculated by a preset adaptive attention weighting unit.

[0053] The fusion features are weighted by the feature contribution weights to obtain weighted fusion features. The target probability distribution of the ECG abnormality type corresponding to the weighted fusion features is calculated by the fully connected layer in the trained ECG signal analysis model.

[0054] A structured electrocardiogram (ECG) signal analysis report is generated based on the target probability distribution, and the ECG signal analysis report is visualized to obtain the ECG signal graph of the target user.

[0055] To address the above problems, the present invention also provides an electrocardiogram (ECG) signal analysis system, the system comprising:

[0056] The data acquisition module is used to acquire the target user's raw electrocardiogram signal through the electrocardiograph unit built into the medical device, and to simultaneously acquire the target user's physiological indicator data through the physiological indicator monitoring unit built into the medical device.

[0057] An adaptive filtering processing module is used to perform adaptive filtering processing on the original electrocardiogram signal based on the motion sensing state of the electrocardiograph unit to obtain the target electrocardiogram signal.

[0058] The data fusion module is used to extract the electrocardiogram waveform features of the target electrocardiogram signal using a preset multi-scale morphological algorithm, and to fuse the electrocardiogram waveform features with the physiological index data to obtain a feature fusion vector.

[0059] The electrocardiogram (ECG) signal analysis model generation module is used to generate an ECG signal analysis model based on the embedded memory resources of the medical device and the network structure of a preset lightweight neural network model.

[0060] The electrocardiogram (ECG) signal generation module is used to train the ECG signal analysis model based on pre-acquired medical and health training data, and to analyze the feature fusion vector through the trained ECG signal analysis model to obtain the ECG signal of the target user.

[0061] This invention, through simultaneous acquisition of raw electrocardiogram (ECG) signals and physiological indicator data, ensures the temporal correlation between the two, providing a reliable data foundation for subsequent analysis. Adaptive filtering of the raw ECG signals during motion perception dynamically adjusts the filtering strategy based on the user's motion state, effectively removing motion artifacts and improving ECG signal quality. Utilizing a multi-scale morphological algorithm to extract ECG waveform features comprehensively captures the complex characteristics of ECG signals. Feature fusion, combined with prior clinical knowledge, results in a fused feature vector containing richer information, enhancing feature representativeness. An ECG signal analysis model is generated based on the embedded memory resources of the medical device, ensuring efficient model operation and improving its practicality and efficiency. Analysis of the fused feature vector yields a more accurate ECG signal graph. Therefore, the ECG signal analysis method and system proposed in this invention can solve the problem of low accuracy in ECG signal analysis. Attached Figure Description

[0062] Figure 1 This is a flowchart illustrating an electrocardiogram (ECG) signal analysis method according to an embodiment of the present invention.

[0063] Figure 2 This is a functional block diagram of an electrocardiogram signal analysis system provided in an embodiment of the present invention.

[0064] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0065] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0066] This application provides a method for electrocardiogram (ECG) signal analysis. The execution entity of the ECG signal analysis method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the ECG signal analysis method can be executed by software or hardware installed on a terminal device or a server device, and the software can be a blockchain platform. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0067] Reference Figure 1The diagram shown is a flowchart illustrating an electrocardiogram (ECG) signal analysis method according to an embodiment of the present invention. In this embodiment, the ECG signal analysis method includes:

[0068] S1. Collect the target user's original electrocardiogram signal through the electrocardiograph unit built into the medical device, and simultaneously collect the target user's physiological indicator data through the physiological indicator monitoring unit built into the medical device.

[0069] In this embodiment of the invention, the pre-set medical device is a device that integrates functional modules such as an electrocardiograph unit and a physiological indicator monitoring unit, enabling the acquisition, processing, and analysis of human electrocardiogram signals and physiological indicators. The raw electrocardiogram signal refers to the unprocessed electrical activity signal emitted from the target user's heart and directly acquired by the electrocardiograph unit.

[0070] In detail, the electrocardiogram (ECG) unit is activated based on a user-triggered physiological monitoring command. This command can be triggered by pressing a corresponding button on the medical device or clicking a relevant option on the device's touchscreen. Upon triggering the command, the medical device's control system receives the signal and sends a start command to the ECG unit, switching it from standby to operational status, ready for data acquisition. For example, when a user feels unwell and wants to monitor their heart condition and related physiological indicators, pressing the start monitoring button on the device will activate the ECG unit.

[0071] Specifically, the raw electrocardiogram (ECG) signal is acquired by an electrocardiograph (ECG) unit at a set sampling frequency. The set sampling frequency is a pre-defined frequency for acquiring signals based on the frequency characteristics of the ECG signal and clinical diagnostic needs. The ECG unit, according to the set sampling frequency, senses the weak electrical signals generated by heart activity through electrodes in contact with the skin and converts them into storable and processable electrical signal data, i.e., the raw ECG signal. For example, a set sampling frequency of 500Hz means that the ECG unit acquires 500 ECG signal data points per second, which can capture the details of the ECG signal relatively completely.

[0072] In this embodiment of the invention, the physiological indicator monitoring unit is a functional module in a medical device used to collect physiological indicator data such as blood glucose, uric acid, and blood ketones from the target user. Physiological indicator data are specific numerical values ​​of various indicators reflecting the target user's physiological state, such as blood glucose levels, uric acid levels, and blood ketone levels.

[0073] In this embodiment of the invention, the synchronous collection of physiological indicator data of the target user through the physiological indicator monitoring unit built into the medical device includes:

[0074] A synchronous trigger signal is generated based on the acquisition timing of the electrocardiograph unit;

[0075] The synchronization trigger signal is sent to the physiological indicator monitoring unit, and the physiological indicator monitoring unit is controlled to perform data acquisition.

[0076] The system receives the timestamp-aligned physiological indicator feedback data from the physiological indicator monitoring unit based on the data acquisition action performed after execution, and verifies the temporal consistency between the physiological indicator feedback data and the original electrocardiogram signal.

[0077] When the verification result is consistent with the timing, the physiological indicator feedback data is used as the physiological indicator data, and the physiological indicator data is cached in the embedded memory of the medical device.

[0078] In detail, the acquisition sequence of the electrocardiogram (ECG) unit refers to the time order and time interval of its raw ECG signal acquisition. The control module of the medical device tracks the acquisition sequence of the ECG unit in real time and generates a synchronization trigger signal based on the start time, interval, and other information of this sequence. This signal instructs the physiological indicator monitoring unit when to start data acquisition to ensure that the data acquisition time of both is synchronized. For example, if the ECG unit acquires data at 1 second, 2 seconds, 3 seconds, etc., the synchronization trigger signal will be generated and sent to the physiological indicator monitoring unit at these time points. The synchronization trigger signal is sent to the physiological indicator monitoring unit through the internal communication line of the medical device. Upon receiving the signal, the physiological indicator monitoring unit immediately activates its internal sensing module and performs data acquisition actions according to the same time rhythm as the ECG unit. For example, when the synchronization trigger signal arrives, the physiological indicator monitoring unit activates the blood glucose sensor to collect the target user's blood glucose data.

[0079] Specifically, after collecting physiological indicator data, the physiological indicator monitoring unit adds a corresponding timestamp to each data point. This timestamp is consistent with the timestamp of the original ECG signal collected by the electrocardiograph unit, resulting in timestamp-aligned physiological indicator feedback data. The medical device's processing module receives this data and compares the timestamps of the physiological indicator data with those of the original ECG signal to verify temporal consistency. If the timestamps of the physiological indicator data and the original ECG signal match perfectly, the timing is consistent; otherwise, it is inconsistent. Once the timing of the physiological indicator feedback data is verified to be consistent with the original ECG signal, the physiological indicator feedback data is determined to be valid physiological indicator data. The medical device temporarily stores this data in its embedded memory. Embedded memory is a storage device within the medical device used for temporary data storage, featuring fast read / write capabilities for rapid data retrieval and processing in subsequent steps. For example, time-consistent blood glucose, uric acid, and blood ketone data are cached in the embedded memory to prepare for subsequent feature extraction and fusion.

[0080] Furthermore, after the raw ECG signal is acquired by the ECG machine unit, it is transmitted to the embedded memory of the medical device for caching. Simultaneously, verified timing-consistent physiological indicator data is also cached in this embedded memory. In this way, the raw ECG signal and physiological indicator data are stored in the same memory, facilitating collaborative processing and analysis in subsequent steps. For example, the raw ECG signal can be stored as a data sequence in a specific area of ​​the embedded memory, while the physiological indicator data can be stored in an adjacent area, allowing for rapid retrieval and correlation operations.

[0081] Furthermore, by synchronously acquiring raw electrocardiogram (ECG) signals and physiological index data, a reliable data foundation is provided for subsequent accurate data analysis. Only by acquiring synchronous and high-quality raw ECG signals can effective filtering be performed to obtain a purer target ECG signal.

[0082] S2. Based on the motion sensing state of the electrocardiograph unit, the original electrocardiogram signal is adaptively filtered to obtain the target electrocardiogram signal.

[0083] In this embodiment of the invention, motion sensing status refers to the motion status of the target user sensed by the electrocardiograph unit through its built-in sensors, including whether or not the user is moving, and the intensity of the movement. Adaptive filtering is a processing method that can automatically adjust filter parameters according to the characteristics of the input signal to achieve the best filtering effect. The target electrocardiogram signal is a clean electrocardiogram signal that has been purified by adaptive filtering to remove interference such as motion artifacts.

[0084] In this embodiment of the invention, the adaptive filtering of the original electrocardiogram (ECG) signal based on the motion sensing state of the ECG machine unit to obtain the target ECG signal includes:

[0085] Acquire the triaxial acceleration data output by the built-in accelerometer of the electrocardiograph unit;

[0086] The motion intensity index and motion status signal of the electrocardiograph unit are determined based on the vector amplitude of the triaxial acceleration data.

[0087] Determine whether the target user is in a state of motion interference based on the motion state signal;

[0088] When the target user is in a state of motion interference, obtain the mapping relationship between the preset motion intensity threshold range and the filter type, and match the motion intensity index with the mapping relationship;

[0089] Extract the filter parameters corresponding to the matched filter type, reconstruct the filter parameters into a preset adaptive filter, and use the reconstructed adaptive filter to suppress motion artifacts in the original ECG signal to obtain the target ECG signal.

[0090] In detail, the accelerometer built into the electrocardiograph unit can sense acceleration changes in three mutually perpendicular directions (usually the X, Y, and Z axes), and its output triaxial acceleration data consists of the acceleration values ​​in each of the three directions. For example, when the target user walks during monitoring, the accelerometer will detect acceleration changes in the X, Y, and Z axes and output corresponding data, such as 0.5 m / s² for the X-axis, 0.3 m / s² for the Y-axis, and 0.2 m / s² for the Z-axis. The vector amplitude is the comprehensive acceleration value calculated based on the triaxial acceleration data, which can be expressed by the formula: Calculate the vector magnitude, where The magnitude of the vector. for Acceleration on the axis, for Acceleration on the axis, for Acceleration on the axis.

[0091] Specifically, the exercise intensity index is a quantitative indicator reflecting the intensity of exercise, determined based on vector amplitude, while the exercise state signal indicates whether the target user is in an exercise state. For example, if the calculated vector amplitude is 0.65 m / s², and this value is greater than a preset exercise threshold, then the exercise state signal is exercise, and the exercise intensity index is moderate. The exercise threshold is determined based on statistical analysis of the target user's triaxial acceleration data in a static state. A large amount of triaxial acceleration data output by the accelerometer built into the electrocardiograph unit is collected from the target user in a static state. The vector amplitude of this data is calculated, and the calculated vector amplitude is statistically processed to determine its distribution range. Usually, the maximum value or 95th percentile of the vector amplitude in the static state is set as the initial exercise threshold, or it can be adjusted in conjunction with clinical trial data. Users of different ages and physical conditions are selected for testing. Acceleration data is collected when the user is in a state of slight exercise and at rest, and the difference in vector amplitude between the two states is analyzed to ensure that the exercise threshold can accurately distinguish between static and exercise states. For example, statistical analysis shows that the vector amplitude is mostly below 0.3 m / s² when stationary, and mostly above 0.3 m / s² during slight movement. Therefore, the motion threshold is set to 0.3 m / s². When the calculated vector amplitude is greater than this value, the target user is judged to be in a motion interference state. When the motion status signal is in motion, it indicates that the target user's movement may interfere with the original ECG signal, i.e., in a motion interference state; when the motion status signal is stationary, the target user is not in a motion interference state. For example, if the motion status signal is in motion, the target user is judged to be in a motion interference state, and the original ECG signal may contain motion artifacts.

[0092] Furthermore, the mapping relationship between the preset exercise intensity threshold range and the filter type is pre-defined. Different exercise intensity threshold ranges correspond to different types of filters. For example, low-intensity exercise corresponds to a first-order Butterworth filter, medium-intensity exercise corresponds to a second-order Butterworth filter, and high-intensity exercise corresponds to an adaptive Notch filter. The calculated exercise intensity index is compared with these threshold ranges to find the corresponding filter type. For example, a medium exercise intensity index corresponds to a threshold range of 0.5-1.0 m / s², matching a second-order Butterworth filter. The filter parameters corresponding to the matched filter type are extracted, including cutoff frequency, order, etc., and different types of filters have different parameters. After extracting the parameters corresponding to the matched filter type, these parameters are set into the preset adaptive filter so that the filter can adapt to the current motion interference. The reconstructed adaptive filter processes the original ECG signal to remove motion artifacts, obtaining a clean target ECG signal. For example, setting the cutoff frequency of the second-order Butterworth filter to 35Hz filters the original ECG signal containing motion artifacts, obtaining the artifact-free target ECG signal.

[0093] Furthermore, by adaptively adjusting the filtering parameters based on the motion state, motion artifacts in the original ECG signal are effectively removed, solving the problem of poor processing effect with fixed filtering parameters and improving the quality of the ECG signal.

[0094] S3. Extract the ECG waveform features of the target ECG signal using a preset multi-scale morphological algorithm, and fuse the ECG waveform features with the physiological index data to obtain a feature fusion vector.

[0095] In this embodiment of the invention, ECG waveform features refer to various features extracted from the target ECG signal that can reflect the characteristics of the ECG waveform, such as amplitude, slope, and duration.

[0096] In this embodiment of the invention, the step of extracting the electrocardiogram waveform features of the target electrocardiogram signal using a preset multi-scale morphological algorithm includes:

[0097] Identify flat structural elements and morphological gradient operators at different scales in a pre-defined multi-scale morphological algorithm;

[0098] Morphological analysis of the target electrocardiogram signal was performed using the flat structural elements of different scales to obtain multi-scale electrocardiogram signals;

[0099] The waveform boundary features of the multi-scale electrocardiogram signal are enhanced according to the morphological gradient algorithm.

[0100] Extract waveform amplitude, waveform slope, and waveform duration features from the enhanced multiscale electrocardiogram signal;

[0101] The waveform amplitude feature, the waveform slope feature, and the waveform duration feature are determined as the ECG waveform features of the target ECG signal.

[0102] In detail, the pre-defined multi-scale morphological algorithm is an algorithm that performs morphological analysis and processing on signals based on structuring elements of different scales. Flat structuring elements of different scales refer to templates with flat structures of different lengths or sizes used in morphological analysis to perform different levels of signal analysis. Morphological gradient operators are used to calculate the gradient changes at the boundaries of signal waveforms, enhancing the boundary features of the waveform. For example, a pre-defined multi-scale morphological algorithm may contain flat structuring elements with lengths of 5, 10, and 15 sampling points, and corresponding morphological gradient operators. Morphological analysis includes operations such as erosion, dilation, opening, and closing. By performing these operations on flat structuring elements of different scales with the target ECG signal, signal representations at different scales can be obtained, i.e., multi-scale ECG signals.

[0103] For example, three flat structure elements of different scales are selected. These three scales are chosen based on the frequency range and waveform characteristics of the target ECG signal, respectively adapting to the analysis needs of different frequency components and waveform details in the signal. For instance, flat structure elements with lengths of 3, 7, and 11 sampling points are selected, corresponding to small, medium, and large scales, respectively. Noise suppression is performed using the smallest scale structure element. The smallest scale flat structure element can effectively identify and remove high-frequency noise in the target ECG signal, as high-frequency noise typically manifests as small-scale fluctuations. For example, using a flat structure element with a length of 3 sampling points to perform an opening operation on the target ECG signal can filter out spike noise in the signal. Waveform enhancement is performed using intermediate scale structure elements. Intermediate scale flat structure elements can process the main waveform components in the ECG signal, enhancing the waveform characteristics and making the waveform clearer and more discernible. For example, using a flat structure element with a length of 7 sampling points to perform a combination of dilation and erosion operations on the noise-suppressed signal enhances the characteristics of the QRS complex. Baseline extraction is performed using the largest scale structure element. The largest-scale flat structuring element can ignore local fluctuations in the signal and extract the baseline trend of the ECG signal, that is, the overall DC component variation of the signal. For example, using a flat structuring element with a length of 11 sampling points to perform a closing operation on the signal yields a smooth baseline signal.

[0104] Specifically, morphological gradient operators are applied to multi-scale ECG signals. By calculating the gradient changes of the signal at different locations, the rising and falling edges of the waveform are made clearer, enhancing the boundary features of the waveform and facilitating subsequent feature extraction. For example, after processing with morphological gradient operators, the boundaries of the QRS complex in the ECG signal are more distinct. Then, the waveform amplitude, waveform slope, and waveform duration features of the enhanced multi-scale ECG signal are extracted. Waveform amplitude features refer to the magnitude of the peak and trough values ​​of the waveform; waveform slope features refer to the degree of inclination during the rise or fall of the waveform; waveform duration features refer to the time elapsed from the start to the end of the waveform. For example, the peak amplitude of the QRS complex is extracted to be 0.8 mV, the rising slope is 0.5 mV / ms, and the duration is 80 ms. Finally, the waveform amplitude, waveform slope, and waveform duration features are determined as the ECG waveform features of the target ECG signal, thus comprehensively reflecting the waveform characteristics of the target ECG signal and constituting the ECG waveform features, providing a foundation for subsequent fusion with physiological index data.

[0105] In this embodiment of the invention, the feature fusion vector is a comprehensive feature vector formed by fusing electrocardiogram waveform features with physiological index data, containing key information from both.

[0106] In this embodiment of the invention, the step of fusing the electrocardiogram waveform features with the physiological index data to obtain a feature fusion vector includes:

[0107] Extract the trend characteristics of physiological indicators from the physiological indicator data;

[0108] The electrocardiogram waveform features and the physiological index trend features are standardized respectively.

[0109] Based on pre-defined clinical prior knowledge, the first and second weighting coefficients of the standardized electrocardiogram waveform features and physiological index trend features are generated respectively.

[0110] Based on the first weighting coefficient and the second weighting coefficient, the standardized ECG waveform features and physiological indicator trend features are fused at the feature level to obtain a feature fusion vector.

[0111] In detail, physiological indicator trend characteristics refer to the changing trend of physiological indicator data over a period of time, such as the rising or falling trend of blood glucose levels, and the fluctuation range of uric acid levels. For example, extracting the trend of a gradual increase in blood glucose over 2 hours from continuously collected blood glucose data. Standardization processing is performed on both electrocardiogram waveform characteristics and physiological indicator trend characteristics. Standardization is the process of converting features of different magnitudes and units into a uniform magnitude, usually by converting feature values ​​to the range of 0-1 or 1-1 to eliminate the influence of magnitude differences on subsequent fusion. For example, the amplitude values ​​in electrocardiogram waveform characteristics are converted from mV to standardized values ​​between 0 and 1, and the rate of change of blood glucose in physiological indicator trend characteristics is converted to standardized values ​​between 0 and 1.

[0112] Specifically, prior clinical knowledge refers to knowledge about the importance of electrocardiogram (ECG) characteristics and physiological indicators in disease diagnosis, summarized based on medical research and clinical experience. For example, if prior clinical knowledge indicates that the rhythmic characteristics of ECG waveforms contribute approximately 70% to clinical outcomes, while physiological indicators such as blood glucose trends contribute approximately 30%, then the first weighting coefficient is directly determined to be 0.7, and the second weighting coefficient is determined to be 0.3. For cases where specific proportions are not explicitly given in the literature, weight values ​​are obtained through statistical analysis of clinical case data. A certain number of case data containing both characteristics and corresponding diagnostic results are collected, and regression analysis and other methods are used to calculate the correlation strength between the two characteristics and the diagnostic results. For example, analyzing 1000 cases, the regression coefficient for ECG waveform characteristics is 0.6, and the regression coefficient for physiological indicator trends is 0.4, obtained through a logistic regression model. After standardization, the first weighting coefficient is set to 0.6, and the second weighting coefficient is set to 0.4 to reflect the relative importance of the two in the diagnosis of the disease. The first weighting coefficient reflects the importance of ECG waveform characteristics, and the second weighting coefficient reflects the importance of physiological indicator trends.

[0113] Furthermore, the standardized ECG waveform features and physiological indicator trend features are vectorized. This can be achieved through vector transformation using the BERT model. The transformed vectors are then fused at the feature level. Feature-level fusion involves weighting the two types of features according to their respective weight coefficients to form a comprehensive feature vector containing information from both. For example, the standardized ECG features are multiplied by a first weight coefficient, and the standardized physiological indicator trend features are multiplied by a second weight coefficient. These two are then added together to obtain the fused feature vector.

[0114] Furthermore, comprehensive ECG waveform features are extracted using a multi-scale morphological algorithm, and then combined with prior clinical knowledge to fuse ECG waveform features and physiological index data, resulting in a feature fusion vector containing richer information.

[0115] S4. Generate an electrocardiogram signal analysis model based on the embedded memory resources of the medical device and the network structure of the preset lightweight neural network model.

[0116] In this embodiment of the invention, embedded memory resources refer to the available memory resources within the medical device, including available storage capacity and computational space. The preset lightweight neural network model is a neural network model with fewer parameters and lower computational load, suitable for running on resource-constrained devices. The electrocardiogram (ECG) signal analysis model is specifically designed to analyze ECG-related features to achieve ECG signal analysis functionality.

[0117] In this embodiment of the invention, generating an electrocardiogram signal analysis model based on the embedded memory resources of the medical device and the network structure of a preset lightweight neural network model includes:

[0118] The available storage capacity and resource computing space are determined based on the embedded memory resources of the medical device.

[0119] The model parameter quantity of the preset lightweight neural network model is determined based on the available storage capacity and the resource computing space;

[0120] Identify the number of neural network layers and the number of convolutional kernels in the network structure of a preset lightweight neural network model;

[0121] The number of neural network layers and the number of convolutional kernels in the network structure are adjusted according to the model parameters, and the adjusted network structure is compiled into an electrocardiogram signal analysis model that can be embedded into medical devices.

[0122] In detail, available storage capacity refers to the unused storage space in the embedded memory of a medical device, usually measured in bytes; resource computing space refers to the resources that the processor of the medical device can use for model calculations, such as the number of computing units and the processing speed. For example, if the available storage capacity of a medical device is 512MB, the corresponding processor computing power is 100MIPS. The lightweight neural network classifier is a deep separable convolutional neural network suitable for embedded microcontrollers, and its model size is constrained to less than 128KB (directly determined by the memory resources (SRAM / Flash) of the specific embedded microcontroller (such as the STM32 series or Nordic series chips) selected by the enterprise) to adapt to the limited memory resources of the medical testing device. The number of model parameters refers to the total number of parameters such as weights and biases included in the neural network model, and its size is directly related to the storage capacity occupied by the model. Based on the available storage capacity, the upper limit of the number of model parameters that can be accommodated is calculated to ensure that the model will not fail to run due to insufficient storage capacity. For example, if the available storage capacity is 512MB and each parameter occupies 4 bytes, then the upper limit of the number of model parameters is approximately 130 million.

[0123] Specifically, the lightweight neural network model's network structure consists of an input layer, three depthwise separable convolutional layers, a global average pooling layer, and an output layer connected sequentially. The number of filters in the three depthwise separable convolutional layers increases progressively. Each of the three depthwise separable convolutional layers is followed by a batch normalization layer and a ReLU activation function layer. The number of neural network layers refers to the number of hidden layers in the model, and the number of convolutional kernels refers to the number of convolutional kernels in each convolutional layer. These two parameters directly affect the model's parameter count and computational cost. For example, a preset lightweight neural network model is identified as having a 5-layer neural network with 32 convolutional kernels. Based on a determined upper limit for the model's parameter count, the number of neural network layers and the number of convolutional kernels are adjusted to ensure that the adjusted model's parameter count does not exceed the upper limit.

[0124] In this embodiment of the invention, adjusting the number of neural network layers and the number of convolutional kernels in the network structure according to the model parameters includes:

[0125] Using a preset network architecture search algorithm, different combinations of configuration parameters for the number of neural network layers and the number of convolutional kernels in the network structure are searched based on the model parameter quantity;

[0126] Analyze the model accuracy and inference speed of lightweight neural network models under different combinations of configuration parameters;

[0127] The configuration parameter combination that maximizes both model accuracy and inference speed is selected as the target configuration combination.

[0128] The number of neural network layers and the number of convolutional kernels in the network structure are adjusted based on the target configuration combination.

[0129] In detail, the preset network architecture search algorithm is an algorithm that can automatically search for neural network structure parameters suitable for specific conditions. Based on the limitation of the number of model parameters, this algorithm generates multiple combinations of different neural network layers and convolutional kernels, calculates the number of model parameters corresponding to each combination, and filters out configuration parameter combinations with parameter numbers within the allowable range. For example, it generates different combinations of neural network layers with 4, 5, and 6 layers, and convolutional kernel numbers with 16, 24, and 32, filtering out combinations with parameter numbers that meet the requirements. Model accuracy refers to the model's prediction accuracy on test data; inference speed refers to the speed at which the model analyzes input data and outputs results. For each qualified configuration parameter combination, the corresponding model is trained, and its accuracy and inference speed on the validation set are tested. For example, analysis shows that the model accuracy of one configuration combination is 90%, and the inference speed is 0.5 seconds / inference; another combination has a model accuracy of 88% and an inference speed of 0.3 seconds / inference.

[0130] Specifically, a trade-off is made between model accuracy and inference speed, and the optimal combination of configuration parameters is selected as the target configuration combination. For example, if a certain combination has a model accuracy of 89% and an inference speed of 0.4 seconds / inference, which is better than other combinations, then this combination is selected as the target configuration combination. The original network structure is modified according to the number of neural network layers and convolutional kernels determined in the target configuration combination to make the network structure conform to the target configuration. For example, if the target configuration combination has 5 neural network layers and 24 convolutional kernels, then the original network structure is adjusted to 5 layers, and the number of convolutional kernels is set to 24. After adjustment, the network structure is converted into machine code that the medical device can recognize and run, i.e., compiled into an ECG signal analysis model, and embedded into the medical device. For example, if the upper limit of the model parameters is 100 million, and the original model has 120 million parameters, then one neural network layer can be reduced, and the number of convolutional kernels can be reduced to 24, reducing the number of parameters to 90 million, before compiling into an embedded model. Before being deployed on the medical testing equipment, the lightweight neural network classifier underwent weight quantization, converting the 32-bit floating-point weights into 8-bit integer weights to improve inference speed on the embedded processor. Quantization can significantly reduce memory usage and processor computational load, protecting the model's final form on the terminal device (8-bit integer), rather than just its form during training (32-bit floating-point).

[0131] Furthermore, by adjusting the structure of the neural network model according to the embedded memory resources of the medical device, an electrocardiogram signal analysis model suitable for device operation is generated, which solves the problem of incompatibility between the model and device resources in the existing technology, resulting in poor operation.

[0132] S5. Train the ECG signal analysis model based on the pre-acquired medical and health training data, and analyze the feature fusion vector through the trained ECG signal analysis model to obtain the ECG signal map of the target user.

[0133] In this embodiment of the invention, medical and health training data refers to a data set containing historical electrocardiogram signals, physiological index data, and corresponding diagnostic results used to train the electrocardiogram signal analysis model.

[0134] In this embodiment of the invention, training the electrocardiogram signal analysis model based on pre-acquired medical and health training data includes:

[0135] Obtain the sample combination feature vector corresponding to historical electrocardiogram signals and physiological index data, and obtain the electrocardiogram abnormality type corresponding to the sample combination feature vector;

[0136] Based on the complexity of the feature vector of the sample combination and the rarity of the ECG abnormality type, the pre-acquired medical and health training data is sorted from easy to difficult to generate a target training data sequence.

[0137] The probability distribution of ECG abnormality types corresponding to the target training data sequence is analyzed using the ECG signal analysis model.

[0138] The loss value of the probability distribution is calculated using a preset transentropy loss function. When the loss value is greater than or equal to a preset loss threshold, the learning rate and batch parameters in the ECG signal analysis model training process are dynamically adjusted according to the target training data sequence until the loss value is less than the preset loss threshold.

[0139] When the loss value is less than the preset loss threshold, the model parameters of the ECG signal analysis model are adjusted based on the adjusted learning rate and batch parameters, and the trained ECG signal model is determined according to the model parameters.

[0140] In detail, historical ECG signals and physiological index data are data collected in the past for model training. These are fused using the same method as in step S3 to obtain sample combination feature vectors. ECG abnormality types refer to the types of heart diseases corresponding to the sample combination feature vectors, such as myocardial infarction and arrhythmia. For example, if 1000 sets of sample combination feature vectors are obtained, each set corresponding to a type of ECG abnormality, then 200 sets might correspond to arrhythmia. The complexity of the sample combination feature vector refers to the complexity of the information contained in the feature vector. More complex feature vectors correspond to more difficult-to-diagnose cases. The more features, the more complex the correlations, and the greater the fluctuation range, the higher the complexity of the sample combination feature vector. The number of features is thus defined as complexity. The rareness of an ECG abnormality type refers to the probability of that type of disease occurring in the population. Rare disease types correspond to more difficult-to-diagnose cases, and the rareness is determined based on the proportion of a certain type of disease. The medical and health training data are sorted in order from simple to complex and from common to rare to obtain the target training data sequence. For example, samples with simple features corresponding to common arrhythmias are sorted first, followed by samples with complex features corresponding to rare myocardial infarction.

[0141] Specifically, the feature vectors of sample combinations from the target training data sequence are input into the ECG signal analysis model. The model outputs the probability that each sample belongs to various ECG abnormality types, forming a probability distribution. For example, after a sample is input into the model, the output probability of it being an arrhythmia is 80%, and the probability of it being normal is 20%. The preset transentropy loss function is a function used to measure the difference between the model's predicted probability distribution and the actual label distribution. The smaller the loss value, the more accurate the model's prediction. The preset loss threshold is a pre-set threshold for judging whether the model has been trained properly. When the loss value is greater than or equal to the preset loss threshold, the training process is optimized by adjusting the learning rate (controlling the magnitude of model parameter updates) and batch parameters (controlling the number of samples input to the model for each training session). For example, increasing the learning rate to speed up convergence and adjusting the batch size to improve training stability, until the loss value is less than the preset loss threshold. When the loss value reaches the preset requirement, the model is trained again using the adjusted learning rate and batch parameters, updating the model's weights, biases, and other parameters until the model converges. The model corresponding to the parameters at this point is the trained ECG signal model.

[0142] In this embodiment of the invention, the trained electrocardiogram (ECG) signal analysis model refers to a model that, after being trained with medical and health training data, can accurately analyze and predict input features. The target user's ECG signal graph is a visual chart containing the target user's ECG signal analysis results, which can assist doctors in diagnosis.

[0143] In this embodiment of the invention, the step of analyzing the feature fusion vector using a trained electrocardiogram signal analysis model to obtain the electrocardiogram of the target user includes:

[0144] The feature fusion vector is input into the input layer of the electrocardiogram signal analysis model;

[0145] The deep implicit correlation features between the ECG waveform features and physiological index features in the feature fusion vector are extracted by using the multi-branch deep separable convolutional layer inside the ECG signal analysis model.

[0146] The deep implicit correlation features are fused to obtain fused features, and the feature contribution weights of the electrocardiogram waveform features and the physiological index features are calculated by a preset adaptive attention weighting unit.

[0147] The fusion features are weighted by the feature contribution weights to obtain weighted fusion features. The target probability distribution of the ECG abnormality type corresponding to the weighted fusion features is calculated by the fully connected layer in the trained ECG signal analysis model.

[0148] A structured electrocardiogram (ECG) signal analysis report is generated based on the target probability distribution, and the ECG signal analysis report is visualized to obtain the ECG signal graph of the target user.

[0149] In detail, the input layer is the first layer in the ECG signal analysis model that receives input data. The feature fusion vector enters the model through the input layer for processing. For example, the feature fusion vector, containing 100 feature values, is input into the model's input layer. A multi-branch depthwise separable convolutional layer is a special type of convolutional layer structure in the model that can process different types of features (ECG waveform features and physiological indicator features) in the feature fusion vector separately, extracting deep, implicit correlations between them. That is, different branches of the multi-branch depthwise separable convolutional layer focus on different parts of the feature fusion vector. One branch mainly processes information related to ECG waveform features, while another branch mainly processes information related to physiological indicator features. For the branch processing ECG waveform features, the depthwise convolution part performs individual convolution operations on each channel of the ECG waveform feature, capturing local features within that channel, such as the variation patterns of specific waveform segments. The pointwise convolution part then combines these local features to extract the correlation information within the ECG waveform features. Simultaneously, this branch interacts with relevant information from physiological indicators to uncover deep correlations between ECG waveform features and physiological indicators, such as the hidden link between heart rate changes and blood glucose fluctuations. For the branch processing physiological indicator features, it also captures local trends within each physiological indicator feature channel through depthwise convolution, such as the rise and fall of blood glucose over different time periods. These local trends are then combined through pointwise convolution to extract the internal correlation information of the physiological indicator features. Furthermore, this branch combines information from ECG waveform features to uncover deep correlations between physiological indicators and cardiac activity, such as the potential link between abnormal uric acid levels and cardiac rhythm disorders. The two branches continuously interact during the extraction process, passing information through inter-layer connections to ensure a comprehensive capture of various deep, implicit correlations between the two types of features. Finally, a multi-branch depthwise separable convolutional layer integrates the correlation information extracted from the two branches to obtain deep, implicit correlation features that reflect the complex intrinsic relationship between ECG waveform features and physiological indicator features.

[0150] Specifically, different correlated features can be combined sequentially into a higher-dimensional feature vector; or information from different correlated features can be fused through element-wise addition, multiplication, and other operations, so that the fused feature retains the unique information of each deep implicit correlated feature while also reflecting their synergistic relationship. For example, features reflecting the correlation between heart rate changes and blood glucose fluctuations and features reflecting the correlation between electrocardiogram rhythm and abnormal uric acid levels can be concatenated to form a fused feature that includes both types of correlated information.

[0151] Furthermore, the pre-defined adaptive attention weighting unit is a unit in the model used to automatically assign weights based on feature importance. It calculates the magnitude of the contribution of ECG waveform features and physiological indicator features to the fused features, i.e., the feature contribution weight. By learning the association patterns between the two features and ECG abnormality types in the training data, it automatically adjusts the degree of attention given to different feature components. The adaptive attention weighting unit assigns an attention score to each element in the fused features. These scores are normalized using a softmax function or similar method to convert them into the corresponding feature contribution weight. For example, the calculated contribution weight of the ECG waveform feature is 0.6, and the contribution weight of the physiological indicator feature is 0.4. By using the feature contribution weights to weight the fused features, important features occupy a larger proportion in the fused features, resulting in a weighted fused feature. The fully connected layer is the layer in the model responsible for mapping the weighted fused features to the output results. Through the calculation of this layer, the probability of the target user belonging to various ECG abnormality types, i.e., the target probability distribution, is output. For example, the probability of the target user being normal is 70%, the probability of having mild arrhythmia is 25%, and the probability of having other abnormalities is 5%.

[0152] Furthermore, a structured electrocardiogram (ECG) signal analysis report is a report organized according to a certain format, containing information such as the target probability distribution and major abnormality types. This report is then visualized, presenting the data in the form of charts, curves, etc., forming an ECG signal graph of the target user, allowing doctors to intuitively understand the user's cardiac condition. For example, the ECG signal graph includes ECG waveform curves, physiological indicator change curves, and abnormality probability annotations.

[0153] Furthermore, a high-precision model is obtained through reasonable training methods, and the model is used to analyze the feature fusion vector to obtain the electrocardiogram signal, which solves the problem that the model analysis effect is poor in the existing technology and cannot provide doctors with effective diagnostic basis.

[0154] like Figure 2 The diagram shown is a functional block diagram of an electrocardiogram signal analysis system provided in an embodiment of the present invention.

[0155] The electrocardiogram (ECG) signal analysis system 100 of this invention can be installed in an electronic device. Depending on the functions implemented, the ECG signal analysis system 100 may include a data acquisition module 101, an adaptive filtering processing module 102, a data fusion module 103, an ECG signal analysis model generation module 104, and an ECG signal graph generation module 105. The module described in this invention can also be referred to as a unit, which refers to a series of computer program segments that can be executed by the processor of an electronic device and can perform a fixed function, and are stored in the memory of the electronic device.

[0156] In this embodiment, the functions of each module / unit are as follows:

[0157] The data acquisition module 101 is used to acquire the target user's original electrocardiogram signal through the electrocardiograph unit built into the medical device, and to simultaneously acquire the target user's physiological indicator data through the physiological indicator monitoring unit built into the medical device.

[0158] The adaptive filtering processing module 102 is used to perform adaptive filtering processing on the original electrocardiogram signal based on the motion sensing state of the electrocardiograph unit to obtain the target electrocardiogram signal.

[0159] The data fusion module 103 is used to extract the electrocardiogram waveform features of the target electrocardiogram signal using a preset multi-scale morphological algorithm, and to fuse the electrocardiogram waveform features with the physiological index data to obtain a feature fusion vector.

[0160] The electrocardiogram signal analysis model generation module 104 is used to generate an electrocardiogram signal analysis model based on the embedded memory resources of the medical device and the network structure of a preset lightweight neural network model.

[0161] The electrocardiogram (ECG) signal generation module 105 is used to train the ECG signal analysis model based on pre-acquired medical and health training data, and to analyze the feature fusion vector through the trained ECG signal analysis model to obtain the ECG signal of the target user.

[0162] In detail, the modules in the electrocardiogram signal analysis system 100 described in this embodiment of the invention employ the same methods as described above. Figures 1 to 2 The method used is the same as the electrocardiogram signal analysis method described above, and it can produce the same technical effect, so it will not be repeated here.

[0163] In the several embodiments provided by this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.

[0164] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0165] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0166] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0167] Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects. The scope of the invention is not limited to the foregoing description, and all variations within the meaning and scope of equivalents falling within the protection scope are intended to be included in the invention.

[0168] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence (AI) refers to the theories, methods, technologies, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0169] Furthermore, it is clear that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units or systems stated in a system claim may also be implemented by a single unit or system through software or hardware. The terms "first," "second," etc., are used to indicate names and do not indicate any specific order.

[0170] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for analyzing electrocardiogram (ECG) signals, characterized in that, The method includes: The system collects the target user's raw electrocardiogram signal through the built-in electrocardiograph unit of the medical device and simultaneously collects the target user's physiological indicator data through the built-in physiological indicator monitoring unit of the medical device. Based on the motion sensing state of the electrocardiograph unit, the original electrocardiogram signal is adaptively filtered to obtain the target electrocardiogram signal. The method involves extracting ECG waveform features from the target ECG signal using a pre-defined multi-scale morphological algorithm, including: identifying flattened structural elements and morphological gradient operators at different scales in the algorithm; performing morphological analysis on the target ECG signal using the flattened structural elements at different scales to obtain a multi-scale ECG signal; enhancing the waveform boundary features of the multi-scale ECG signal according to the morphological gradient algorithm; extracting waveform amplitude, slope, and duration features from the enhanced multi-scale ECG signal; determining the waveform amplitude, slope, and duration features as the ECG waveform features of the target ECG signal; and fusing the ECG waveform features with physiological index data to obtain a feature fusion vector, including: extracting physiological index trend features from the physiological index data; standardizing the ECG waveform features and physiological index trend features respectively; generating a first weighting coefficient and a second weighting coefficient for the standardized ECG waveform features and physiological index trend features based on pre-defined clinical prior knowledge; and performing feature-level fusion of the standardized ECG waveform features and physiological index trend features according to the first weighting coefficient and the second weighting coefficient to obtain a feature fusion vector. An electrocardiogram (ECG) signal analysis model is generated based on the embedded memory resources of the medical device and the network structure of a preset lightweight neural network model. The ECG signal analysis model is trained based on pre-acquired medical and health training data. The feature fusion vector is then analyzed using the trained ECG signal analysis model to obtain the ECG signal map of the target user.

2. The electrocardiogram signal analysis method as described in claim 1, characterized in that, The process of synchronously collecting physiological indicator data of the target user through the physiological indicator monitoring unit built into the medical device includes: A synchronous trigger signal is generated based on the acquisition timing of the electrocardiograph unit; The synchronization trigger signal is sent to the physiological indicator monitoring unit, and the physiological indicator monitoring unit is controlled to perform data acquisition. The system receives the timestamp-aligned physiological indicator feedback data from the physiological indicator monitoring unit based on the data acquisition action performed after execution, and verifies the temporal consistency between the physiological indicator feedback data and the original electrocardiogram signal. When the verification result is consistent with the timing, the physiological indicator feedback data is used as the physiological indicator data, and the physiological indicator data is cached in the embedded memory of the medical device.

3. The electrocardiogram signal analysis method as described in claim 1, characterized in that, The adaptive filtering of the original electrocardiogram (ECG) signal based on the motion sensing state of the ECG machine unit to obtain the target ECG signal includes: Acquire the triaxial acceleration data output by the built-in accelerometer of the electrocardiograph unit; The motion intensity index and motion status signal of the electrocardiograph unit are determined based on the vector amplitude of the triaxial acceleration data. Determine whether the target user is in a state of motion interference based on the motion state signal; When the target user is in a state of motion interference, obtain the mapping relationship between the preset motion intensity threshold range and the filter type, and match the motion intensity index with the mapping relationship; Extract the filter parameters corresponding to the matched filter type, reconstruct the filter parameters into a preset adaptive filter, and use the reconstructed adaptive filter to suppress motion artifacts in the original ECG signal to obtain the target ECG signal.

4. The electrocardiogram signal analysis method as described in claim 1, characterized in that, The step of generating an electrocardiogram signal analysis model based on the embedded memory resources of the medical device and the network structure of a preset lightweight neural network model includes: The available storage capacity and resource computing space are determined based on the embedded memory resources of the medical device. The model parameter quantity of the preset lightweight neural network model is determined based on the available storage capacity and the resource computing space; Identify the number of neural network layers and the number of convolutional kernels in the network structure of a preset lightweight neural network model; The number of neural network layers and the number of convolutional kernels in the network structure are adjusted according to the model parameters, and the adjusted network structure is compiled into an electrocardiogram signal analysis model that can be embedded into medical devices.

5. The electrocardiogram signal analysis method as described in claim 4, characterized in that, The adjustment of the number of neural network layers and the number of convolutional kernels in the network structure based on the model parameters includes: Using a preset network architecture search algorithm, different combinations of configuration parameters for the number of neural network layers and the number of convolutional kernels in the network structure are searched based on the model parameter quantity; Analyze the model accuracy and inference speed of lightweight neural network models under different combinations of configuration parameters; The configuration parameter combination that maximizes both model accuracy and inference speed is selected as the target configuration combination. The number of neural network layers and the number of convolutional kernels in the network structure are adjusted based on the target configuration combination.

6. The electrocardiogram signal analysis method as described in claim 1, characterized in that, The step of training the electrocardiogram signal analysis model based on pre-acquired medical and health training data includes: Obtain the sample combination feature vector corresponding to historical electrocardiogram signals and physiological index data, and obtain the electrocardiogram abnormality type corresponding to the sample combination feature vector; Based on the complexity of the feature vector of the sample combination and the rarity of the ECG abnormality type, the pre-acquired medical and health training data is sorted from easy to difficult to generate a target training data sequence. The probability distribution of ECG abnormality types corresponding to the target training data sequence is analyzed using the ECG signal analysis model. The loss value of the probability distribution is calculated using a preset transentropy loss function. When the loss value is greater than or equal to a preset loss threshold, the learning rate and batch parameters in the ECG signal analysis model training process are dynamically adjusted according to the target training data sequence until the loss value is less than the preset loss threshold. When the loss value is less than the preset loss threshold, the model parameters of the ECG signal analysis model are adjusted based on the adjusted learning rate and batch parameters, and the trained ECG signal model is determined according to the model parameters.

7. The electrocardiogram signal analysis method as described in claim 1, characterized in that, The step of analyzing the feature fusion vector using a trained electrocardiogram signal analysis model to obtain the target user's electrocardiogram includes: The feature fusion vector is input into the input layer of the electrocardiogram signal analysis model; The deep implicit correlation features between the ECG waveform features and physiological index features in the feature fusion vector are extracted by using the multi-branch deep separable convolutional layer inside the ECG signal analysis model. The deep implicit correlation features are fused to obtain fused features, and the feature contribution weights of the electrocardiogram waveform features and the physiological index features are calculated by a preset adaptive attention weighting unit. The fusion features are weighted by the feature contribution weights to obtain weighted fusion features. The target probability distribution of the ECG abnormality type corresponding to the weighted fusion features is calculated by the fully connected layer in the trained ECG signal analysis model. A structured electrocardiogram (ECG) signal analysis report is generated based on the target probability distribution, and the ECG signal analysis report is visualized to obtain the ECG signal graph of the target user.

8. An electrocardiogram (ECG) signal analysis system, characterized in that, The system for performing the electrocardiogram signal analysis method as described in any one of claims 1-7 includes: The data acquisition module is used to acquire the target user's raw electrocardiogram signal through the electrocardiograph unit built into the medical device, and to simultaneously acquire the target user's physiological indicator data through the physiological indicator monitoring unit built into the medical device. An adaptive filtering processing module is used to perform adaptive filtering processing on the original electrocardiogram signal based on the motion sensing state of the electrocardiograph unit to obtain the target electrocardiogram signal. The data fusion module is used to extract the ECG waveform features of the target ECG signal using a preset multi-scale morphological algorithm, including: identifying flattened structural elements and morphological gradient operators at different scales in the preset multi-scale morphological algorithm; performing morphological analysis on the target ECG signal using the flattened structural elements at different scales to obtain a multi-scale ECG signal; enhancing the waveform boundary features of the multi-scale ECG signal according to the morphological gradient algorithm; extracting the waveform amplitude features, waveform slope features, and waveform duration features of the enhanced multi-scale ECG signal; and confirming the waveform amplitude features, waveform slope features, and waveform duration features. The ECG waveform features are defined as the target ECG signal. The ECG waveform features are then fused with the physiological indicator data to obtain a feature fusion vector. This includes: extracting physiological indicator trend features from the physiological indicator data; standardizing the ECG waveform features and the physiological indicator trend features respectively; generating first and second weighting coefficients for the standardized ECG waveform features and physiological indicator trend features based on preset clinical prior knowledge; and performing feature-level fusion of the standardized ECG waveform features and physiological indicator trend features according to the first and second weighting coefficients to obtain a feature fusion vector. The electrocardiogram (ECG) signal analysis model generation module is used to generate an ECG signal analysis model based on the embedded memory resources of the medical device and the network structure of a preset lightweight neural network model. The electrocardiogram (ECG) signal generation module is used to train the ECG signal analysis model based on pre-acquired medical and health training data, and to analyze the feature fusion vector through the trained ECG signal analysis model to obtain the ECG signal of the target user.

Images