Electrocardiosignal feature extraction and classification method and sensor system

Abstract

The invention discloses an electrocardiosignal feature extraction and classification method and a sensor system, and particularly relates to the technical field of electrocardiosignal processing and mode recognition. The core of the method is that an original electrocardiosignal is obtained through a sensor, after preprocessing, a self-adaptive strategy based on real-time signal quality evaluation is adopted to dynamically select and fuse time domain, frequency domain and nonlinear features, and then a machine learning model integrated with an incremental learning function is utilized to classify the features; the corresponding system comprises a sensor module, a preprocessing module, a feature extraction module, a classification module and a system control module which are connected with one another. According to the core structure, the dynamic optimization of the feature extraction strategy along with the signal quality is realized, and the classification model can be continuously self-updated according to new data, so that the adaptability of electrocardio analysis and the accuracy of long-term monitoring are improved.

Description

Technical Field

[0001] This invention relates to the field of electrocardiogram (ECG) signal processing and pattern recognition technology, and more specifically, to ECG signal feature extraction and classification methods and sensor systems. Background Technology

[0002] Electrocardiogram (ECG) signal analysis technology, as an important auxiliary means of diagnosing cardiovascular diseases, has evolved from analog circuits to digital systems, and from reliance on manual interpretation to automated computer analysis. Early ECG analysis equipment was mainly used in clinical settings, focusing on basic signal acquisition and display. With advancements in microelectronics and digital signal processing algorithms, portable and wearable ECG monitoring devices have become increasingly common, and their technological trend is towards miniaturization, intelligence, and long-term dynamic monitoring. In recent years, machine learning algorithms, especially deep learning models, have been increasingly applied to the automatic classification of ECG signals, aiming to improve the automation and accuracy of analysis.

[0003] However, existing automated ECG signal analysis technologies still have several shortcomings. In the feature extraction stage, many methods rely on preset fixed rules and thresholds to extract features from ECG waveforms. This static strategy struggles to adapt to individual differences in ECG signals and the signal morphology fluctuations of the same user under different physiological states, resulting in insufficient feature robustness. In the classification stage, most systems use classification models with fixed parameters after deployment, lacking continuous learning capabilities. When faced with slow drift in user ECG features over time or the appearance of abnormal patterns not fully covered in the training data, classification performance may decline. Furthermore, signal preprocessing, feature extraction, and classification modules are often designed as independent steps, lacking sufficient coordination and failing to dynamically optimize the overall analysis process based on real-time signal quality.

[0004] Therefore, this paper proposes a method and sensor system for ECG signal feature extraction and classification to address the aforementioned problems. The aim is to solve the key issues in the existing technologies, namely, how to improve the adaptability of the ECG feature extraction process to individual differences and signal quality variations, and how to enable the classification system to adapt to long-term changes in ECG patterns, thereby maintaining and improving the accuracy and reliability of the analysis in dynamic usage scenarios. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a method for extracting and classifying electrocardiogram signal features and a sensor system to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for extracting and classifying electrocardiogram (ECG) signal features, the method comprising the following steps:

[0007] S1. Acquire raw electrocardiogram (ECG) signals through a sensor system; preprocess the raw ECG signals, including noise filtering and baseline drift correction;

[0008] S2. Adaptively extract multi-dimensional features from the preprocessed electrocardiogram signal. The multi-dimensional features include time-domain features, frequency-domain features, and nonlinear features. The adaptive extraction is based on a dynamic adjustment of the feature selection strategy according to real-time signal quality assessment.

[0009] S3. Input the multi-dimensional features into the machine learning classification model to classify ECG abnormalities, wherein the machine learning classification model integrates incremental learning function and can automatically update the model parameters using newly input ECG signal data.

[0010] An electrocardiogram (ECG) signal feature extraction and classification sensor system includes: a sensor module for acquiring raw ECG signals; a preprocessing module connected to the sensor module for performing filtering and baseline correction on the raw ECG signals; a feature extraction module connected to the preprocessing module and configured to perform adaptive multi-dimensional feature extraction; a classification module connected to the feature extraction module, containing the machine learning classification model for outputting classification results; and a system control module connected to the aforementioned modules for coordinating data flow and control signals.

[0011] Preferably, the noise filtering in the preprocessing step employs a combined digital filter, which includes a bandpass filter with a cutoff frequency set in the range of 0.5Hz to 40Hz, and a notch filter for suppressing power frequency interference.

[0012] Preferably, the adaptive multi-dimensional feature extraction step specifically includes: first, calculating the short-term signal-to-noise ratio and waveform stability index of the signal as indicators for signal quality evaluation; then, selecting a feature subset from a preset feature pool based on the values ​​of the indicators, wherein the feature pool includes at least the RR interval, QRS group width, heart rate variability frequency domain index, and sample entropy; finally, normalizing the selected feature subset and concatenating them to form a feature vector.

[0013] Preferably, the machine learning classification model is a convolutional neural network or a support vector machine; when using a convolutional neural network, its input layer receives the feature vector, the hidden layer contains 1 to 3 fully connected layers, the number of neurons in each layer is configurable between 50 and 200, and the output layer uses the Softmax function to generate probability distributions corresponding to different ECG abnormality categories; the incremental learning function updates the model parameters through an online gradient descent algorithm or a least mean squares algorithm.

[0014] Preferably, the system further includes a data management and communication module; the data management and communication module includes a local storage unit for temporarily storing electrocardiogram signal data and feature data, with a storage capacity supporting continuous recording of data for at least 24 hours; and a wireless communication unit supporting Bluetooth or Wi-Fi protocols for transmitting the classification results and / or raw signal data to an external display device or a remote server.

[0015] Preferably, the system control module is configured to support two operating modes: a real-time mode, in which the electrocardiogram signal is processed continuously or frame by frame, with each frame being 2 to 10 seconds long; and a retrospective analysis mode, used to process and analyze batches of historical data stored in the local storage unit.

[0016] Preferably, it also includes a user-adaptive calibration mechanism; the user-adaptive calibration mechanism establishes a user-personalized baseline feature profile by collecting normal electrocardiogram signal samples of a specific user, and uses the profile to fine-tune the threshold in subsequent feature extraction or the decision boundary in the classification model.

[0017] Preferably, the adaptive multi-dimensional feature extraction step further includes a waveform component accurate identification sub-step; this sub-step uses a dynamic template matching algorithm to detect feature points of P-waves and T-waves, wherein the matching template is dynamically adjusted according to the morphology of the QRS group in the current signal period to improve positioning accuracy.

[0018] Preferably, the output of the classification module includes not only the predicted ECG abnormality category, but also a confidence score; when the confidence score is lower than a preset threshold, the system can trigger a secondary verification process, which includes re-analyzing the current signal using a different set of features or a simplified auxiliary classifier.

[0019] The technical effects and advantages of this invention are as follows:

[0020] Compared to existing technologies, this invention addresses the poor robustness of fixed feature extraction strategies by introducing an adaptive feature selection method based on real-time signal quality assessment. During operation, this method first dynamically calculates the signal-to-noise ratio and stability indices of the electrocardiogram (ECG) signal. Then, based on these quality indices, it intelligently selects the most suitable feature subset from a pre-defined feature pool containing time-domain, frequency-domain, and nonlinear features, and performs weighted fusion. This mechanism allows the system to automatically adapt to individual user differences and changes in the signal acquisition environment. Its advantage lies in effectively improving the representativeness and reliability of the extracted features, thus providing a higher-quality data foundation for subsequent classification.

[0021] Compared to existing technologies, this invention solves the problem that traditional static models cannot adapt to the long-term evolution of ECG patterns by integrating incremental learning functionality into the classification module. During system operation, this function continuously utilizes newly acquired, especially user-confirmed, valid ECG signal data to automatically fine-tune and update the internal parameters of the classification model through an online learning algorithm. The effect of this mechanism is that the classification model can continuously evolve along with the slow changes in the user's ECG characteristics. Its advantages include enhanced classification accuracy and personalized service levels in long-term dynamic monitoring scenarios, while avoiding performance degradation due to model aging. Attached Figure Description

[0022] Figure 1 This is a diagram of the overall system architecture of the present invention.

[0023] Figure 2 This is a flowchart of the main workflow for electrocardiogram signal processing and classification in this invention.

[0024] Figure 3 This is a detailed flowchart of the adaptive feature extraction of the present invention.

[0025] Figure 4 This is a flowchart of the incremental learning model update process of the present invention. Detailed Implementation

[0026] 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.

[0027] Example 1:

[0028] As attached Figures 1 to 4 The present invention illustrates a method and sensor system for ECG signal feature extraction and classification, providing a complete implementation process for such a method and sensor system. This system integrates adaptive feature selection and incremental learning functions to achieve intelligent analysis and long-term monitoring of ECG signals. The complete closed-loop process from signal acquisition to model update will be described in detail below.

[0029] 1. System initialization and parameter configuration

[0030] After the system is powered on, it first performs an initialization operation, which includes loading basic configuration parameters, starting various software modules, and completing hardware self-test.

[0031] Basic configuration parameters include signal sampling rate (typically set to 250Hz to 1000Hz), filter cutoff frequency (bandpass filter set to 0.5Hz to 40Hz, power frequency notch filter set to 50Hz or 60Hz), frame processing length (default 2 seconds to 10 seconds in real-time mode), and initial parameters for the classification model. The system control module checks the connection status of the sensor module to ensure good electrode contact. Simultaneously, the system attempts to load the stored user-customized configuration file. If this is the first time using the system or the user is not recognized, the system will run with the default parameter set and prompt the user to enter calibration mode. After initialization, the system enters standby mode, ready to begin ECG signal acquisition.

[0032] 2. ECG signal acquisition and buffering

[0033] The sensor module acquires raw analog electrocardiogram (ECG) signals via surface electrodes. The electrodes are typically configured with three or five leads to obtain signals of sufficient quality. The acquired analog signals are pre-amplified and then digitized by an analog-to-digital converter at a preset sampling rate to obtain the raw ECG signal S in discrete time series form. raw [n], where n represents the sampling point index.

[0034] The raw signal is sent to the buffer of the data management and communication module for temporary storage. The buffer adopts a first-in-first-out queue structure and its capacity is designed to hold at least 30 to 60 seconds of continuous data to ensure the continuity of the data flow in subsequent processing stages. The system control module monitors the amount of data in the buffer and triggers the subsequent preprocessing process when it reaches the length of a processing frame.

[0035] 3. Signal preprocessing

[0036] The preprocessing module reads a frame of raw signal S from the buffer. raw [n], the first step in preprocessing is noise filtering.

[0037] The combined digital filter processes the signal sequentially: First, a high-pass filter (cutoff frequency 0.5Hz) removes baseline drift and low-frequency motion artifacts; second, a low-pass filter (cutoff frequency 40Hz) suppresses high-frequency electromyographic noise; finally, a notch filter (center frequency 50Hz or 60Hz, bandwidth typically 2Hz) filters out power frequency interference. The filtered signal is denoted as S. filtered [n]. The second step is baseline correction, which uses a polynomial fitting method. The algorithm finds relatively stable segments in the signal (usually the TP segment) and fits the trend of this segment with a low-order polynomial (such as second- or third-order). Then, this fitted trend is subtracted from the entire signal frame to obtain the final preprocessed ECG signal S. processed [n], this signal will be used as input for subsequent feature extraction.

[0038] 4. Adaptive Feature Extraction

[0039] The feature extraction module receives the preprocessed signal S processed [n]. The core task of this module is to perform adaptive feature selection and fusion based on signal quality assessment.

[0040] 4.1 Signal Quality Assessment

[0041] First, the module calculates two key quality metrics for the current signal frame:

[0042] Local signal-to-noise ratio (SNR): In the time domain, it is estimated by calculating the logarithm of the ratio of the QRS group amplitude (usually the R-wave peak value) to the standard deviation of the signal in a preset quiet segment (such as the TP segment).

[0043] Waveform stability index: It measures the morphological consistency of the waveform within the current frame by dynamic time warping or by calculating the correlation coefficient between waveforms of consecutive heartbeat cycles.

[0044] 4.2 Feature Pool Calculation and Weight Allocation

[0045] The module computes a predefined feature pool in parallel, which contains three types of features:

[0046] Temporal characteristics: such as RR interval (time interval between adjacent R waves), QRS complex duration, ST segment elevation, etc.

[0047] Frequency domain characteristics: The low-frequency power (LF), high-frequency power (HF), and their ratio (LF / HF) of heart rate variability are calculated by fast Fourier transform.

[0048] Nonlinear characteristics: such as sample entropy (a measure of signal complexity) and Lyapunov exponent (a measure of signal chaotic characteristics).

[0049] For each i-th feature in the feature pool, the system calculates its fusion weight W_i according to formula (1):

[0050] Formula (1): Adaptive calculation of feature weights

[0051] W i =α·(SNR) i / SNR max )+β·(1-|ρ i -ρ target |)+γ·(Stability i )

[0052] In formula (1):

[0053] W i This represents the fusion weight of the i-th feature;

[0054] α, β, and γ are preset adjustment coefficients used to balance the influence of different evaluation indicators, and satisfy α + β + γ = 1;

[0055] SNR i This represents the local signal-to-noise ratio corresponding to the i-th feature in the current signal segment;

[0056] SNR max This represents the system's preset upper limit reference value for signal-to-noise ratio (SNR), used to measure SNR. i Perform normalization;

[0057] ρ i This represents the calculated value of the i-th feature within the current signal segment;

[0058] ρ target This indicates the reference value of the feature stored in the user's personalization baseline;

[0059] Stability i The stability index of the i-th feature over the most recent signal periods is obtained by calculating its coefficient of variation.

[0060] 4.3 Feature Selection and Vector Construction

[0061] The system calculates all weights W i Set a weight threshold W threshold Features with weights below this threshold are considered unreliable under the current signal quality and will be excluded. The remaining features form a feature subset. Each feature value in this subset is normalized (e.g., Z-score normalization) and then multiplied by its corresponding weight W. i Finally, all weighted eigenvalues ​​are concatenated into a one-dimensional eigenvector F. vector This eigenvector F vector That is, it serves as the input for the classification module.

[0062] 5. ECG signal classification and result output

[0063] The classification module receives the feature vector F vector The module comes pre-installed with machine learning models, such as convolutional neural networks or support vector machines.

[0064] Forward propagation / decision: F vector Input Model. For neural networks, data passes through the input layer and hidden layers (which may contain 1-3 fully connected layers, with the number of neurons in each layer configurable between 50-200). Finally, the output layer uses the Softmax function to obtain the probability distribution P = [p1, p2, ..., p_k] corresponding to different ECG abnormality categories (such as normal sinus rhythm, atrial fibrillation, premature ventricular contractions, etc.), where k is the number of categories.

[0065] Result determination: Select the category with the highest probability value as the initial classification result, Class_initial. At the same time, calculate the confidence score of this result, which is usually max(P).

[0066] Output and Triggering: The system control module outputs the classification result, Class_initial, and its confidence score, Score_confidence. Output methods include displaying it on the local screen, sending it to a mobile app or cloud server via wireless communication. If the confidence score, Score_confidence, falls below a preset confidence threshold (e.g., 0.8), the system triggers a secondary verification process. Secondary verification may involve reclassifying the data using a simpler, more robust set of features, prompting the user to remain still to re-collect a signal, or even, in some applications, prompting for manual review.

[0067] 6. Incremental learning and model update

[0068] The incremental update process is not performed for every signal frame, but is triggered when specific conditions are met.

[0069] Triggering conditions typically include: 1) The user manually confirms or corrects the system's automatic classification results; 2) The system makes a high-confidence, consistent classification of high-quality signals over multiple consecutive periods (considered as reliable automatic labeling); 3) The preset periodic update time is reached.

[0070] Data preparation: When the triggering conditions are met, the system will extract the recently confirmed, high-quality (i.e., excellent signal quality assessment indicators) ECG signal data and their corresponding labels (whether provided by the user or inferred by the system with high confidence) from the data storage module to form one or more mini-batch data B.

[0071] Parameter Update: For each mini-batch B, the system performs an incremental update of the model parameters once. The update process follows formula (2):

[0072] Formula (2): Incremental update of model parameters

[0073] θ new =θ old -η·▽J(θ old B)

[0074] In formula (2):

[0075] θ new This represents the updated model parameter vector;

[0076] θ oldThis represents the model parameter vector before the update;

[0077] η represents the learning rate, which controls the step size for each parameter update;

[0078] ▽J(θ old B) represents the loss function J with respect to the model parameters θ. old Gradient on mini-batch data B;

[0079] B represents a mini-batch sample randomly selected from newly acquired and verified valid electrocardiogram signal data.

[0080] Through this process of small, frequent updates, the model parameters θ are slowly adjusted, gradually adapting to the new ECG pattern without catastrophic forgetting of previously learned knowledge. The updated model parameters θ new It will take effect immediately, be used for subsequent signal classification, and be synchronously saved to the memory.

[0081] 7. User-adaptive calibration mechanism

[0082] The system also integrates a user adaptation mechanism, which complements incremental learning but focuses on establishing personalized feature baselines.

[0083] When a user uses the system for the first time or initiates calibration, the system guides the user to acquire a high-quality ECG signal for a duration (e.g., 5 minutes) while at rest. The system then performs routine feature extraction on this signal (but without adaptive weighting), calculates the statistical values ​​(mean, variance) of each feature (e.g., mean RR interval, QRS morphology parameters), establishes the user's personalized baseline feature profile (Profile_user), and stores it.

[0084] In subsequent use, when the system detects that the user is in a resting state and the signal quality is extremely high, it will occasionally fine-tune Profile_user with new resting signal data, such as updating the feature mean using the moving average method.

[0085] The Profile_user that is created is mainly used for ρ in formula (1) target This allows adaptive feature weight calculation to take the user's normal state as a reference, thereby more accurately identifying abnormal situations that deviate from the normal state.

[0086] 8. System Operating Modes and Cycles

[0087] Under the scheduling of the system control module, the entire system mainly supports two operating modes:

[0088] Real-time mode: The system continuously loops from step 2 to step 5. That is, it continuously acquires signals, preprocesses them, performs adaptive feature extraction, classifies them, and outputs the results. Incremental learning (step 6) is a background task that is executed asynchronously when conditions are met, without affecting the smoothness of real-time analysis.

[0089] Retrospective Analysis Mode: Users can choose to perform batch analysis on historical long-term data (such as 24-hour Holter ECG data) stored in the data storage module. In this mode, the system reads the data in chronological order, executes steps 3 (preprocessing), 4 (feature extraction), and 5 (classification) sequentially, and finally generates a comprehensive report. In this mode, model updates are usually not performed.

[0090] In summary, the implementation process of this invention is a dynamic closed-loop system integrating signal processing, intelligent decision-making, and continuous learning. Adaptive feature extraction ensures the quality of input information, incremental learning guarantees the model's viability, and user-adaptive calibration further significantly improves the accuracy, robustness, and personalization of ECG signal analysis. The entire process has a clear data flow, from analog signals to digital sequences, then to feature vectors and probability distributions, and finally back to model parameters, forming a continuously optimizing intelligent analysis system.

[0091] Finally, the following points should be noted: First, in the description of this application, it should be noted that, unless otherwise specified and limited, the terms "installation", "connection", and "linkage" should be interpreted broadly, and can be mechanical or electrical connections, or internal connections between two components, or direct connections. "Up", "down", "left", "right", etc. are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may change.

[0092] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other.

[0093] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for extracting and classifying electrocardiogram (ECG) signal features, characterized in that, The method includes the following steps: S1. Acquire raw electrocardiogram (ECG) signals through a sensor system; preprocess the raw ECG signals, including noise filtering and baseline drift correction; S2. Adaptively extract multi-dimensional features from the preprocessed ECG signal. The multi-dimensional features include time-domain features, frequency-domain features, and nonlinear features. The adaptive extraction is based on a dynamic adjustment of the feature selection strategy according to real-time signal quality assessment. S3. Input the multi-dimensional features into the machine learning classification model to classify ECG abnormalities, wherein the machine learning classification model integrates incremental learning function and can automatically update the model parameters using newly input ECG signal data.

2. A sensor system for extracting and classifying electrocardiogram (ECG) signal features, used to implement the method as described in claim 1, characterized in that, The system includes: a sensor module for acquiring raw electrocardiogram (ECG) signals; a preprocessing module connected to the sensor module for performing filtering and baseline correction on the raw ECG signals; a feature extraction module connected to the preprocessing module and configured to perform adaptive multi-dimensional feature extraction; a classification module connected to the feature extraction module, containing the machine learning classification model for outputting classification results; and a system control module connected to the aforementioned modules for coordinating data flow and control signals.

3. The method for extracting and classifying electrocardiogram signal features as described in claim 1, characterized in that, The noise filtering in the preprocessing step employs a combined digital filter, which includes a bandpass filter with a cutoff frequency set in the range of 0.5Hz to 40Hz, and a notch filter for suppressing power frequency interference.

4. The method for extracting and classifying electrocardiogram signal features as described in claim 1, characterized in that, The adaptive multi-dimensional feature extraction step specifically includes: first, calculating the short-term signal-to-noise ratio and waveform stability index of the signal as indicators for signal quality evaluation; then, selecting a feature subset from a preset feature pool based on the values ​​of the indicators, wherein the feature pool includes at least the RR interval, QRS group width, heart rate variability frequency domain index, and sample entropy; finally, normalizing the selected feature subset and concatenating them to form a feature vector.

5. The method for extracting and classifying electrocardiogram signal features as described in claim 1, characterized in that, The machine learning classification model is a convolutional neural network or a support vector machine. When using a convolutional neural network, its input layer receives the feature vector, the hidden layer contains 1 to 3 fully connected layers, the number of neurons in each layer is configurable between 50 and 200, and the output layer uses the Softmax function to generate probability distributions corresponding to different ECG abnormality categories. The incremental learning function updates the model parameters through an online gradient descent algorithm or a least mean squares algorithm.

6. The ECG signal feature extraction and classification sensor system as described in claim 2, characterized in that, The system further includes a data management and communication module; the data management and communication module includes a local storage unit for temporarily storing electrocardiogram signal data and feature data, with a storage capacity supporting continuous recording of at least 24 hours of data; It also includes a wireless communication unit that supports Bluetooth or Wi-Fi protocols for transmitting the classification results and / or raw signal data to an external display device or a remote server.

7. The ECG signal feature extraction and classification sensor system as described in claim 2, characterized in that, The system control module is configured to support two operating modes: real-time mode, in which the ECG signal is processed continuously or frame by frame, with each frame being 2 to 10 seconds in length; It also includes a retrospective analysis mode for processing and analyzing batches of historical data stored in the local storage unit.

8. The method and system for electrocardiogram signal feature extraction and classification as described in claim 1 or 2, characterized in that, It also includes a user-adaptive calibration mechanism; the user-adaptive calibration mechanism establishes a user-personalized baseline feature profile by collecting normal electrocardiogram signal samples of specific users, and uses the profile to fine-tune the threshold in subsequent feature extraction or the decision boundary in the classification model.

9. The method for extracting and classifying electrocardiogram signal features as described in claim 4, characterized in that, The adaptive multi-dimensional feature extraction step also includes a waveform component accurate identification sub-step; this sub-step uses a dynamic template matching algorithm to detect feature points of P-waves and T-waves, wherein the matching template is dynamically adjusted according to the morphology of the QRS group in the current signal period to improve positioning accuracy.

10. The method for extracting and classifying electrocardiogram signal features as described in claim 5, characterized in that, The output of the classification module includes not only the predicted ECG abnormality category, but also a confidence score. When the confidence score is lower than a preset threshold, the system can trigger a secondary verification process, which includes re-analyzing the current signal using a different set of features or a simplified auxiliary classifier.

Images