Electrocardiogram data anomaly recognition method, device and equipment based on deep learning and storage medium

Abstract

The invention provides an electrocardiogram data anomaly recognition method, device and equipment based on deep learning and a storage medium. An original electrocardiogram signal of a patient is collected, and a target electrocardiogram signal is obtained through multi-stage preprocessing. And identifying an abnormal signal in the target electrocardiogram signal by using a deep fusion learning model, and generating a preliminary identification result. And then, the preliminary recognition result is combined with a pre-trained machine learning model, mixed model processing is performed through a stack generalization fusion method, a Bayesian network is constructed based on the detailed information of the patient, a reasoning result with context sensing is obtained, and a target recognition result is finally generated. And generating an abnormal report according to the target identification result, and storing and displaying the abnormal report to the user in a classified manner according to a predefined standard. According to the method, high precision and personalized diagnosis of anomaly detection are ensured, and the reliability and efficiency of electrocardiogram analysis are improved.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of data processing technology, and in particular to a method, device, equipment and storage medium for identifying abnormal electrocardiogram data based on deep learning. Background Art

[0002] With the digital transformation of the healthcare sector, real-time monitoring and analysis of ECG data has become increasingly important. Especially in the areas of telemedicine and personalized health management, being able to quickly and accurately identify abnormalities in the ECG is important for preventing the occurrence and development of heart disease. Therefore, an efficient and accurate method is needed to automatically detect abnormal signals in the ECG.

[0003] At present, traditional anomaly detection methods mainly include threshold-based methods, template matching, and simple machine learning algorithms.

[0004] However, these methods usually rely on expert knowledge to define thresholds or templates, and these thresholds or templates are often difficult to adapt to the differences of all individuals, resulting in poor efficiency in ECG data processing. Summary of the invention

[0005] The embodiments of the present application provide a method, device, equipment and storage medium for identifying abnormal ECG data based on deep learning, so as to solve the problem of poor efficiency of ECG data processing in the prior art.

[0006] In a first aspect, an embodiment of the present application provides a method for identifying abnormalities in electrocardiogram data based on deep learning, comprising: Collecting the original electrocardiogram signal of the patient and performing a multi-stage preprocessing process to obtain a target electrocardiogram signal, the multi-stage preprocessing process at least includes: using a bandpass filter to remove high-frequency noise and power frequency interference in the original electrocardiogram signal; applying wavelet transform or adaptive filtering technology to suppress myoelectric artifacts and low-frequency noise in the original electrocardiogram signal; using a differential operational amplifier to enhance the original electrocardiogram signal; A deep fusion learning model is trained using a target electrocardiogram signal, wherein the deep fusion learning model combines a convolutional neural network and a recurrent neural network, wherein the convolutional neural network is used to extract local features of the target electrocardiogram signal, and the recurrent neural network is used to capture time series features of the target electrocardiogram signal, and an attention mechanism is introduced to enable the deep fusion learning model to focus on abnormal signals in the target electrocardiogram signal; The abnormal signal in the target electrocardiogram signal is identified by the deep fusion learning model to obtain a preliminary recognition result; the preliminary recognition result is combined with the pre-trained machine learning model, and the preliminary recognition result is subjected to hybrid model fusion processing by a stacked generalization fusion method, and the inference result of the preliminary recognition result is obtained by a Bayesian network constructed by the detailed information of the patient to generate a target recognition result; Based on the target recognition result, a rule engine is used to generate an exception report, which includes the abnormal type, location, and severity of the target electrocardiogram signal. The exception report is classified and stored according to predefined standards, and displayed through a customized visual interface.

[0007] Optionally, identifying an abnormal signal in the target electrocardiogram signal by using the deep fusion learning model to obtain a preliminary recognition result includes: Performing time window segmentation processing on the target electrocardiogram signal to obtain a plurality of time windows of fixed length, wherein each time window represents an electrocardiogram data segment within a time period, and the electrocardiogram data segments corresponding to all the time windows can constitute the target electrocardiogram signal; The electrocardiogram data segments in each time window are fed into the convolutional neural network in the deep fusion learning model, so as to extract the local morphological features of the electrocardiogram data in each time window through the convolutional neural network, and obtain a local feature representation characterizing the local characteristics of the cardiac activity, wherein the local feature representation includes at least the morphological changes of the QRS complex, the P wave and the T wave, and the heart rate variability; Sending the ECG data segments in each time window to the recurrent neural network in the deep fusion learning model, so as to capture the time series features between multiple time windows through the recurrent neural network, and determining the change pattern and long-term dependency of the heart rhythm based on the time series features to obtain a time series feature representation; Introducing an attention mechanism into the deep fusion learning model, so that when the deep fusion learning model processes the electrocardiogram data fragments in each time window, it identifies and highlights the key features that are most likely to indicate an abnormality from the electrocardiogram data fragments, and generates a key feature representation with a weight, wherein the weight is used to reflect the importance of each key feature for abnormal signal detection; Integrate the local feature representation, the time series feature representation, and the key feature representation into a comprehensive feature representation, and use the comprehensive feature representation as input to train the deep fusion learning model, so that the deep fusion learning model can predict the anomaly type and confidence; Applying a multi-task learning framework in the trained deep fusion learning model to predict the abnormality type and confidence of the electrocardiogram data segment in each time window, wherein the abnormality type includes atrial fibrillation, ventricular premature beat, and conduction block, and based on the abnormality type and confidence of the electrocardiogram data segment in each time window, identifying the abnormal signal whose confidence is higher than a preset threshold, and marking the abnormality type, location, and severity of the abnormal signal; Evaluate the patient's cardiac health status score based on the abnormality types and confidence levels of the electrocardiogram data segments within all time windows; A preliminary recognition result is generated based on the corresponding abnormal signals in all time windows and the patient's heart health status score.

[0008] Optionally, applying a multi-task learning framework in the trained deep fusion learning model to predict the abnormal type and confidence of the electrocardiogram data segment in each time window, and based on the abnormal type and confidence of the electrocardiogram data segment in each time window, identifying abnormal signals with a confidence higher than a preset threshold, including: Using the multi-task learning framework in the trained deep fusion learning model, identify the abnormality type contained in the electrocardiogram data segment in each time window, and obtain the probability distribution corresponding to each abnormality type, wherein the probability distribution corresponding to each abnormality type is used to indicate the possibility of the abnormality type existing in each time window; Evaluate the confidence of each anomaly type and generate a confidence score, wherein the confidence score refers to the reliability score of the prediction result of the deep fusion learning model for each anomaly type; Filter out time windows with confidence scores higher than a preset threshold, identify abnormal signals within the time windows, mark the time windows identified as abnormal, and record the abnormal type, location and severity of the abnormal signals, where the location refers to the time window corresponding to the abnormal signal.

[0009] Optionally, the preliminary recognition result is combined with a pre-trained machine learning model, the preliminary recognition result is subjected to hybrid model fusion processing by a stacked generalization fusion method, and the inference result of the preliminary recognition result is obtained by a Bayesian network constructed by the detailed information of the patient to generate a target recognition result, including: Training a machine learning model based on the comprehensive feature representation and the enhanced feature representation additionally extracted from the target electrocardiogram signal, so that the machine learning model can predict whether each time window contains an abnormal signal and the abnormal type and confidence level corresponding to the abnormal signal; the enhanced feature representation includes statistical features, frequency domain features, time domain features and nonlinear features additionally extracted from the target electrocardiogram signal; Using stacked generalization technology, the preliminary recognition result output by the deep fusion learning model is used as the first layer feature input, and the prediction result of the machine learning model is used as the second layer feature input to generate an intermediate recognition result; Based on the acquired detailed information of the patient, a Bayesian network is constructed, and a context-aware reasoning result is generated for the preliminary identification through the Bayesian network, so as to provide additional context information for the detection of abnormal signals through the reasoning result, wherein the detailed information includes medical history data, historical electrocardiogram records and clinical symptoms; Based on the inference result, combined with the acquired individual differences and historical heart data of the patient, the intermediate recognition result is inferred to adjust the probability distribution corresponding to the intermediate recognition result to generate a target recognition result.

[0010] Optionally, the inference processing of the intermediate recognition result based on the inference result and in combination with the acquired individual differences and historical heart data of the patient to adjust the probability distribution corresponding to the intermediate recognition result and generate the target recognition result includes: According to the acquired individual differences and historical heart data of the patient, the probability distribution of abnormal signals in each time window in the intermediate recognition result is adjusted to obtain an adjusted probability distribution; Based on the inference result, the possibly omitted or uncertain parts in the intermediate recognition result are supplemented to enhance the reliability of the intermediate recognition result, thereby obtaining an optimized intermediate recognition result; Integrating the adjusted concept distribution and the optimized intermediate recognition result to generate a target recognition result; The method further comprises: The time series consistency of the target identification results is optimized based on the long-term trends and short-term fluctuations in the patient's historical cardiac data.

[0011] Optionally, the evaluating the patient's heart health status score based on the abnormality types and confidence levels of the electrocardiogram data segments within all time windows includes: Based on the abnormal type of the electrocardiogram data segment in each time window and its corresponding confidence score, and in combination with the local feature representation, time series feature representation and key feature representation in the time window, the patient's heart health status score is evaluated; wherein the heart health status score is calculated by the following formula: ; in, is the heart health status score, is the number of time windows, is the index of the time window; is the number of exception types, is the index of the exception type; It is In the time window A confidence score for each anomaly type, wherein the confidence score refers to a reliability score of the prediction result of the deep fusion learning model for each anomaly type; is the exponential factor of the confidence score, which is used to adjust the influence of the confidence score. ,in is the maximum value among all confidence scores; It is The weights of the abnormality types reflect the importance of different abnormality types to the heart health status; It is The influence factor of the local feature representation within a time window is calculated by the standardized score of the local feature representation; It is The impact factor of the time series feature representation within a time window is calculated by the standardized score of the time series feature representation; It is The impact factor of the key feature representation within a time window is calculated by the standardized score of the key feature representation; They are the weight coefficients of local feature representation, time series feature representation and key feature representation, which are used to adjust the influence of each feature representation and are determined according to the importance and contribution of each feature; They are the exponential factors of local feature representation, time series feature representation, and key feature representation, which are used to adjust the influence of each feature representation and are determined according to the sensitivity and importance of each feature; is a Gaussian kernel function used to smooth the effect of confidence scores, where is the average of all confidence scores, is the standard deviation of the confidence scores; It is The comprehensive impact factor of a time window is defined as: ; in, is the number of additional features, which include at least the patient's age, gender, and medical history; It is The time window and relevance scores of additional features; It is The weight of the additional feature is used to adjust the influence of the additional feature.

[0012] Optionally, adjusting the probability distribution of abnormal signals in each time window in the intermediate recognition result according to the acquired individual differences and historical heart data of the patient to obtain an adjusted probability distribution includes: The adjusted probability distribution is calculated by the following formula: ; in, It is In the time window The adjusted probability distribution of anomaly types, It is In the time window The initial probability distribution of anomaly types; It is The influencing factors of individual differences among patients within a time window are calculated based on the patient's age, gender, and medical history; is the average of the individual difference influencing factors in all time windows; is the standard deviation of the individual difference influencing factors in all time windows; is the number of historical specific features, which include specific indicators in the patient's historical cardiac data, including historical ECG records, treatment responses, It is a historically specific index; It is The time window and relevance scores for specific historical features; It is The weight of the historical specific feature is used to adjust the influence of the historical specific feature; The method of supplementing the possibly omitted or uncertain parts of the intermediate recognition result based on the inference result, enhancing the reliability of the intermediate recognition result, and obtaining an optimized intermediate recognition result includes: The optimized intermediate recognition result is calculated by the following formula: ; in, It is The intermediate recognition result after time window optimization, is the number of exception types; It is In the time window Adjusted probability distribution of anomaly types; It is the context-aware reasoning result provided by the Bayesian network. In the time window Confidence scores for anomaly types; The step of integrating the adjusted probability distribution and the optimized intermediate recognition result to generate a target recognition result includes: The target recognition result is calculated by the following formula: ; in, It is The target recognition result of time window is It is The result corresponding to the maximum adjusted probability distribution within a time window; It is The optimized intermediate recognition result of the time window; It is the weight coefficient between the adjusted probability distribution and the optimized intermediate recognition result, which is used to balance the influence of the two.

[0013] In a second aspect, an embodiment of the present application provides a device for identifying abnormal ECG data based on deep learning, comprising: An acquisition and processing module is used to acquire the original electrocardiogram signal of the patient and perform a multi-stage preprocessing process to obtain a target electrocardiogram signal, wherein the multi-stage preprocessing process at least includes: using a bandpass filter to remove high-frequency noise and power frequency interference in the original electrocardiogram signal; using wavelet transform or adaptive filtering technology to suppress myoelectric artifacts and low-frequency noise in the original electrocardiogram signal; and using a differential operational amplifier to enhance the original electrocardiogram signal; A training module, for training a deep fusion learning model using a target electrocardiogram signal, wherein the deep fusion learning model combines a convolutional neural network and a recurrent neural network, wherein the convolutional neural network is used to extract local features of the target electrocardiogram signal, and the recurrent neural network is used to capture time series features of the target electrocardiogram signal, and at the same time, an attention mechanism is introduced to enable the deep fusion learning model to focus on key parts indicating abnormal signals, so as to improve the sensitivity of abnormal information recognition; An identification output module is used to identify abnormal signals in the target electrocardiogram signal through the deep fusion learning model to obtain a preliminary identification result; combine the preliminary identification result with the pre-trained machine learning model, perform hybrid model fusion processing on the preliminary identification result through a stacked generalization fusion method, and obtain an inference result of the preliminary identification result through a Bayesian network constructed by the detailed information of the patient to generate a target identification result; A generation and display module is used to generate an exception report based on the target recognition result using a rule engine. The exception report includes the abnormal type, location, and severity of the target electrocardiogram signal, classifies and stores the exception report according to predefined standards, and displays the exception report through a customized visual interface.

[0014] In a third aspect, an embodiment of the present application provides a computing device, comprising a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement a method for identifying anomalies in electrocardiogram data based on deep learning as described in any one of the first aspects.

[0015] In a fourth aspect, an embodiment of the present application provides a computer storage medium storing a computer program. When the computer program is executed by a computer, it implements a method for identifying anomalies in electrocardiogram data based on deep learning as described in any one of the first aspects.

[0016] In an embodiment of the present application, the original electrocardiogram signal of the patient is collected, and a multi-stage preprocessing process is performed to obtain a target electrocardiogram signal, and the multi-stage preprocessing process at least includes: using a bandpass filter to remove high-frequency noise and power frequency interference in the original electrocardiogram signal; applying wavelet transform or adaptive filtering technology to suppress myoelectric artifacts and low-frequency noise in the original electrocardiogram signal; using a differential operational amplifier to enhance the original electrocardiogram signal; using the target electrocardiogram signal to train a deep fusion learning model, the deep fusion learning model combines a convolutional neural network and a recurrent neural network, wherein the convolutional neural network is used to extract local features of the target electrocardiogram signal, and the recurrent neural network is used to capture the time series features of the target electrocardiogram signal, and at the same time, by introducing an attention mechanism, the deep fusion learning model can focus on abnormal signals in the target electrocardiogram signal; The deep fusion learning model is used to identify abnormal signals in the target electrocardiogram signal to obtain a preliminary recognition result; the preliminary recognition result is combined with a pre-trained machine learning model, and the preliminary recognition result is subjected to hybrid model fusion processing through a stacked generalization fusion method, and the inference result of the preliminary recognition result is obtained through a Bayesian network constructed based on the patient's detailed information to generate a target recognition result; based on the target recognition result, an abnormality report is generated using a rule engine, and the abnormality report includes the abnormal type, location, and severity of the target electrocardiogram signal, and the abnormality report is classified and stored according to predefined standards, and the abnormality report is displayed through a customized visual interface.

[0017] The embodiments of the present application have the following beneficial effects: This application uses a bandpass filter to remove high-frequency noise and power frequency interference, applies wavelet transform or adaptive filtering technology to suppress electromyographic artifacts and low-frequency noise, and uses a differential operational amplifier to enhance the original electrocardiogram signal, ensuring the signal quality input to the subsequent model and improving the accuracy of abnormality detection. Combining convolutional neural network (CNN) and recurrent neural network (RNN), while introducing an attention mechanism, the model can not only extract local features, but also capture time series features and focus on potential abnormal signals, thereby significantly improving the accuracy of abnormality detection. The preliminary recognition results are combined with the pre-trained machine learning model, processed by a stacked generalization fusion method, and the Bayesian network is constructed using the patient's detailed information to generate context-aware reasoning results, making the final recognition results more personalized and reflecting the patient's specific situation and historical data. According to the target recognition results, the rule engine is used to generate an abnormality report containing the abnormality type, location, and severity, and the abnormality is stored according to predefined standards, ensuring the comprehensiveness and interpretability of the diagnosis results. The abnormality report is displayed through a customized visual interface, which facilitates doctors to quickly understand the diagnosis results, supports clinical decision-making, and improves work efficiency. By integrating multi-source information (such as patients' medical records, historical ECG records, and clinical symptoms), the model's robustness and adaptability are enhanced, enabling it to maintain high detection performance under different conditions. The context-aware reasoning results provided by the Bayesian network can supplement the parts that may be omitted or uncertain in the intermediate recognition results, further enhancing the reliability of the recognition results. Through real-time monitoring and anomaly detection of ECG signals, potential heart problems can be discovered in a timely manner, achieving early warning, and helping to improve treatment outcomes and reduce medical costs.

[0018] To sum up, through the above-mentioned technical means, this application realizes efficient and accurate identification of abnormal signals in electrocardiogram signals, provides personalized diagnostic results, supports efficient clinical applications, and has good robustness and adaptability, thereby significantly improving the utilization efficiency of medical resources and the quality of medical services.

[0019] These and other aspects of the present application will become more clearly understood in the description of the following embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0021] Figure 1A flowchart of a method for identifying abnormal ECG data based on deep learning provided in an embodiment of the present application; Figure 2 A schematic diagram of the structure of an abnormal ECG data device based on deep learning provided in an embodiment of the present application; Figure 3 A schematic diagram of the structure of a computing device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0022] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0023] In some of the processes described in the specification and claims of this application and the above-mentioned figures, multiple operations that appear in a specific order are included, but it should be clearly understood that these operations may not be executed in the order in which they appear in this article or executed in parallel. The serial numbers of the operations, such as 101, 102, etc., are only used to distinguish between different operations, and the serial numbers themselves do not represent any execution order. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions of "first", "second", etc. in this article are used to distinguish different messages, devices, modules, etc., do not represent the order of precedence, and do not limit the "first" and "second" to be different types.

[0024] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of this application.

[0025] Figure 1 A flowchart of a method for identifying abnormal ECG data based on deep learning is provided for an embodiment of the present application. Figure 1 As shown, the method includes: 101. Collect the original electrocardiogram signal of the patient and perform a multi-stage preprocessing process to obtain a target electrocardiogram signal.

[0026] In this step, after collecting the patient's original electrocardiogram (ECG) signal, a multi-stage preprocessing process is performed to obtain a high-quality target ECG signal. The multi-stage preprocessing includes using a bandpass filter to remove high-frequency noise and power frequency interference, applying wavelet transform or adaptive filtering technology to suppress myoelectric artifacts and low-frequency noise, and using a differential operational amplifier to enhance the original ECG signal. These steps ensure that the signal input into the subsequent analysis model has a higher signal-to-noise ratio and reduces the impact of external factors on signal quality.

[0027] In the embodiment of the present application, firstly, the appropriate frequency range is set by the bandpass filter to remove high-frequency noise beyond the ECG signal frequency band and common 50Hz or 60Hz power frequency interference. Then, the wavelet transform or adaptive filtering technology is used to effectively suppress the myoelectric artifacts caused by muscle activity and the low-frequency noise from physiological activities such as breathing. Finally, the differential operational amplifier is used to enhance the signal amplitude so that the weak ECG signal can be properly amplified, thereby improving the accuracy of subsequent feature extraction and abnormality detection.

[0028] In one specific embodiment, the patient wears a portable ECG monitoring device to continuously record ECG data for 24 hours. The built-in preprocessing module of the device first limits the signal to between 0.5-150Hz through a bandpass filter, effectively removing the high-frequency interference in the environment and the noise caused by the power line frequency. Then, the data is processed segment by segment using wavelet transform technology to eliminate electromyographic artifacts and low-frequency drift. The preprocessed data is transmitted to the cloud server for further analysis. The application of differential operational amplifiers ensures that good resolution can be maintained even in the case of weak signals, providing high-quality input for subsequent deep learning model training.

[0029] 102. Use the target electrocardiogram signal to train a deep fusion learning model, wherein the deep fusion learning model combines a convolutional neural network and a recurrent neural network.

[0030] Among them, the convolutional neural network is used to extract the local features of the target electrocardiogram signal, and the recurrent neural network is used to capture the time series features of the target electrocardiogram signal. At the same time, the attention mechanism is introduced to enable the deep fusion learning model to focus on abnormal signals in the target electrocardiogram signal.

[0031] In this step, a deep fusion learning model combining convolutional neural network (CNN) and recurrent neural network (RNN) is trained using the target ECG signal. The convolutional neural network is used to extract local morphological features, such as QRS complex, changes in P wave and T wave, and heart rate variability; the recurrent neural network captures the time series features between multiple time windows and determines the change pattern and long-term dependency of the heart rhythm. At the same time, the attention mechanism is introduced to enable the model to focus on the key features that are most likely to indicate abnormalities, thereby improving the sensitivity and specificity of abnormal signal detection.

[0032] In the embodiment of the present application, during the training process, the convolutional layer automatically learns and extracts the local morphological features of the ECG signal, while the recurrent layer is responsible for capturing the dynamic changes between different time points. After the introduction of the attention mechanism, the model can identify and highlight key features when processing each time window, giving them higher weights, so as to better locate abnormal signals. Through supervised learning with a large amount of annotated ECG data, the model gradually optimizes the parameters and finally has high-precision anomaly detection capabilities.

[0033] Based on the above preprocessed 24-hour ECG data, the researchers used a pre-designed deep fusion learning model architecture that integrates multiple convolutional layers and recurrent layers composed of long short-term memory (LSTM) units. In collaboration with clinical experts, thousands of ECG samples with detailed annotations were collected for model training. During training, the attention mechanism helps the model focus on those parts of the waveform that best reflect potential abnormalities. As the training progressed, the model learned to identify a variety of common types of cardiac abnormalities, such as atrial fibrillation, ventricular premature beats, and conduction blocks, laying a solid foundation for subsequent hybrid model fusion processing.

[0034] 103. Identify abnormal signals in the target electrocardiogram signal through the deep fusion learning model to obtain a preliminary recognition result; combine the preliminary recognition result with the pre-trained machine learning model, perform hybrid model fusion processing on the preliminary recognition result through a stacked generalization fusion method, and obtain the inference result of the preliminary recognition result through a Bayesian network constructed based on the patient's detailed information to generate a target recognition result.

[0035] In this step, the trained deep fusion learning model is used to identify abnormal signals in the target ECG signal to obtain a preliminary recognition result. This preliminary result is then combined with the pre-trained machine learning model, and a hybrid model fusion process is performed using the stacked generalization fusion method. The Bayesian network is constructed using the patient's detailed information to generate context-aware reasoning results, and finally the target recognition result is generated. This method not only improves the accuracy of recognition, but also enhances the personalization and reliability of the results.

[0036] In the embodiment of the present application, the preliminary recognition result output by the deep fusion learning model is used as the first-layer feature input, and the prediction result of the machine learning model optimized based on the enhanced feature representation is used as the second-layer feature input, and the intermediate recognition result is generated through the stacked generalization technology. Subsequently, a Bayesian network is constructed based on the patient's medical records, historical electrocardiogram records, and clinical symptoms to generate context-aware reasoning results for preliminary recognition. Finally, the probability distribution of the intermediate recognition results is adjusted in combination with the individual differences of the patients and historical heart data to generate more accurate target recognition results.

[0037] After completing the model training, the system begins to process the newly collected ECG data in real time. The deep fusion learning model quickly identifies possible abnormal signals and their confidence levels, and passes these preliminary results to the hybrid model fusion module. This module comprehensively considers multiple feature representations, such as statistical features, frequency domain features, time domain features, and nonlinear features to optimize the prediction results. Next, the system builds a Bayesian network based on the patient's specific circumstances (such as age, gender, medical history, etc.) to generate personalized reasoning results. For example, for a patient with a history of heart disease, the system will pay special attention to abnormal patterns that are similar to historical records, thereby improving the accuracy and credibility of the diagnosis.

[0038] 104. Generate an exception report based on the target recognition result using a rule engine, wherein the exception report includes the abnormal type, location, and severity of the target electrocardiogram signal, classifies and stores the exception report according to predefined standards, and displays the exception report through a customized visual interface.

[0039] In this step, based on the generated target recognition results, the rule engine is used to generate an abnormality report containing the abnormality type, location, and severity, and the abnormality report is classified and stored according to predefined standards. In addition, the abnormality report is displayed through a customized visual interface, allowing doctors to intuitively view and understand the diagnosis results and support clinical decision-making. This approach not only improves work efficiency, but also ensures the transparency and interpretability of the diagnosis results.

[0040] In an embodiment of the present application, once the target recognition result is generated, the rule engine automatically writes a detailed abnormality report based on preset logical rules. The report content covers the specific type of abnormal signal, the location of occurrence, and indicators for assessing its severity. Subsequently, the system classifies and archives the reports according to predefined criteria to facilitate subsequent retrieval and analysis. For the convenience of doctors, the system also provides a customized visual interface that allows users to interactively browse ECG data, abnormal markers and their related annotations, thereby accelerating the diagnostic process.

[0041] In actual application scenarios, when the system detects an abnormal signal, it will immediately trigger the rule engine to generate a detailed abnormality report. This report not only includes the type, location and severity of the abnormal signal, but also includes relevant references and recommended follow-up examination items. All generated reports are classified and stored according to the hospital's internal coding system to ensure the standardization and consistency of data management. At the same time, doctors can view the latest ECG data in real time through a customized visual interface, and the marked abnormal areas are clear at a glance, which significantly shortens the time cycle from data collection to clinical decision-making and improves the quality and efficiency of medical services.

[0042] In the above steps 101 to 104 of this application, the quality of the input signal is ensured through multi-level preprocessing. The deep fusion learning model combines the convolutional neural network and the recurrent neural network to achieve efficient and accurate anomaly detection. The hybrid model fusion processing and Bayesian network reasoning enhance the personalization and reliability of the recognition results. Finally, the rule engine generates anomaly reports and provides a customized visualization interface, which significantly improves the utilization efficiency of medical resources and the quality of service. The entire solution not only improves the accuracy and robustness of anomaly detection, but also supports efficient clinical applications by integrating multi-source information and providing detailed diagnostic reports, and realizes an integrated solution from data collection to presentation of diagnostic results.

[0043] In order to solve the problem of insufficient capture of detail features and insufficient processing of time series dependencies in abnormal signal detection, in some embodiments, the step of identifying the abnormal signal in the target electrocardiogram signal by the deep fusion learning model in step 103 to obtain a preliminary recognition result includes: The target electrocardiogram signal is segmented into time windows to obtain multiple time windows of fixed length, and each time window represents an electrocardiogram data segment within a time period, and the electrocardiogram data segments corresponding to all time windows can constitute the target electrocardiogram signal; the electrocardiogram data segments in each time window are sent to the convolutional neural network in the deep fusion learning model, so as to extract the local morphological features of the electrocardiogram data in each time window through the convolutional neural network, and obtain a local feature representation characterizing the local characteristics of cardiac activity, and the local feature representation at least includes the morphological changes of the QRS complex, P wave and T wave, and heart rate variability; the electrocardiogram data segments in each time window are sent to the recurrent neural network in the deep fusion learning model, so as to capture the time series features between multiple time windows through the recurrent neural network, and determine the change pattern and long-term dependency of the cardiac rhythm based on the time series features to obtain a time series feature representation; an attention mechanism is introduced in the deep fusion learning model so that the deep fusion learning model can extract the local morphological features of the electrocardiogram data segments from the electrocardiogram data segments when processing the electrocardiogram data segments in each time window. The method comprises the steps of: identifying and highlighting the key features most likely to indicate an abnormality, and generating a key feature representation with a weight, wherein the weight is used to reflect the importance of each key feature for abnormal signal detection; integrating the local feature representation, the time series feature representation and the key feature representation into a comprehensive feature representation, and using the comprehensive feature representation as an input to train the deep fusion learning model so that the deep fusion learning model can predict the abnormality type and confidence; applying a multi-task learning framework to the trained deep fusion learning model to predict the abnormality type and confidence of the electrocardiogram data segment in each time window, wherein the abnormality type includes atrial fibrillation, ventricular premature beat, and conduction block, and based on the abnormality type and confidence of the electrocardiogram data segment in each time window, identifying the abnormal signal with a confidence higher than a preset threshold, and marking the abnormality type, location, and severity of the abnormal signal; evaluating the patient's heart health status score based on the abnormality type and confidence of the electrocardiogram data segment in all time windows; and generating a preliminary recognition result based on the corresponding abnormal signals in all time windows and the patient's heart health status score.

[0044] In this solution, time window segmentation refers to dividing the continuous ECG data into multiple time segments of fixed length, and each time window represents the ECG data within a time period. This not only helps to reduce the computational burden, but also ensures that the model can focus on the local features within each time period.

[0045] ECG data segments refer to data segments corresponding to a specific time window extracted from the original ECG signal. These segments contain the heart activity information within the time period and are the basis for subsequent feature extraction.

[0046] Convolutional neural network (CNN) is a deep learning architecture that is mainly used for feature extraction of image and sequence data. In ECG analysis, CNN is used to extract local morphological features of ECG data within each time window, such as QRS complex, P wave and T wave morphological changes, and heart rate variability.

[0047] Local morphological features refer to the characteristic manifestations of cardiac activity in a short period of time, such as the morphological changes of QRS complexes, P waves and T waves, and heart rate variability. These features reflect the activity pattern of the heart in a specific time period and are crucial for identifying abnormal signals.

[0048] Recurrent Neural Network (RNN) is a neural network suitable for processing sequence data, especially for capturing time series features. In electrocardiogram analysis, RNN is used to capture the heart rhythm change pattern and long-term dependencies between multiple time windows.

[0049] Time series features refer to the dynamic change characteristics of cardiac activity over a period of time, including the changing patterns of cardiac rhythm and long-term dependencies. These features help understand the overall trend of cardiac activity and improve the accuracy of anomaly detection.

[0050] The attention mechanism is a technique used to enhance the model's ability to identify key features, enabling the model to identify and highlight key features that are most likely to indicate anomalies and give them higher weights when processing data in each time window.

[0051] Key feature representations are those ECG features that are identified by the attention mechanism as most likely to indicate abnormalities. These features are weighted to reflect their importance for abnormal signal detection.

[0052] Comprehensive feature representation refers to a feature set that is integrated from local feature representation, time series feature representation, and key feature representation. This comprehensive feature representation is used as input to train the deep fusion learning model, enabling it to predict the type of anomaly and its confidence.

[0053] The multi-task learning framework is a machine learning method that allows the model to complete multiple related tasks simultaneously. In ECG anomaly detection, the multi-task learning framework is used to predict the anomaly type and confidence of the ECG data segment within each time window.

[0054] The Heart Health Status Score is a quantitative indicator that assesses a patient's overall heart health based on the type and confidence of abnormalities in ECG data segments within all time windows. The score combines the severity and distribution of abnormal signals to provide a comprehensive assessment of heart health.

[0055] In the embodiment of the present application, the target ECG signal is first segmented into time windows, and then the ECG data segments in each time window are respectively sent to the convolutional neural network and the recurrent neural network to extract local morphological features and capture time series features, respectively. At the same time, the attention mechanism is introduced to enhance the model's ability to recognize key features. Finally, the three feature representations are integrated, the model is trained to predict the type of abnormality and its confidence, and a multi-task learning framework is applied to identify high-confidence abnormal signals, evaluate the patient's heart health status score, and generate preliminary recognition results.

[0056] Here is a specific example: In a practical application scenario, suppose we have a 24-hour ECG record. The system first divides this long ECG data into multiple fixed-length time windows, for example, each window is 5 seconds long. The data segments in each time window are processed separately to ensure that the model can focus on the detailed features in each short time period.

[0057] Next, the data fragments of each time window are fed into a convolutional neural network (CNN), which automatically learns and extracts important local morphological features, such as the morphological changes of the QRS complex, P wave and T wave, and heart rate variability. These features reflect the activity pattern of the heart in each time period and are the basis for identifying abnormal signals.

[0058] At the same time, data fragments of the same time window are also fed into a recurrent neural network (RNN), which captures the dynamic changes between multiple time windows and determines the changing patterns and long-term dependencies of the heart rhythm. This step helps the model understand the overall trend of heart activity and improves the accuracy of anomaly detection.

[0059] To further improve the performance of the model, the system introduces an attention mechanism that enables the model to identify and highlight the key features that are most likely to indicate anomalies when processing the data in each time window. For example, if significant heart rate variability or abnormal QRS wave morphology appears in a certain time window, the attention mechanism will give these features a higher weight, so that the model pays more attention to these possible anomalies.

[0060] Finally, the system integrates local feature representation, time series feature representation, and key feature representation into a comprehensive feature representation, and uses these comprehensive features to train the deep fusion learning model. After training, the model is able to predict the type of anomaly and its confidence level in each time window. By applying a multi-task learning framework, the model identifies abnormal signals with a confidence level higher than a preset threshold, and marks their abnormality type, location, and severity. In addition, the system also evaluates the patient's heart health status score based on the abnormality type and confidence level in all time windows, and generates preliminary recognition results including abnormality type, location, and severity.

[0061] Through this method, doctors can obtain a detailed abnormality report to support them in making more accurate diagnostic decisions. This process not only improves the accuracy of abnormality detection, but also provides personalized diagnostic information, significantly improving the quality and efficiency of medical services.

[0062] In order to solve the multi-task processing and confidence assessment problems in abnormal signal detection, the above optional solution of "applying a multi-task learning framework in the trained deep fusion learning model to predict the abnormal type and confidence of the electrocardiogram data segment in each time window, and based on the abnormal type and confidence of the electrocardiogram data segment in each time window, identifying the abnormal signal with a confidence higher than a preset threshold" specifically includes: Utilizing the multi-task learning framework in the trained deep fusion learning model, the abnormality types contained in the electrocardiogram data segments within each time window are identified, and the probability distribution corresponding to each abnormality type is obtained, where the probability distribution corresponding to each abnormality type is used to indicate the possibility of whether the abnormality type exists within each time window; the confidence of each abnormality type is evaluated, and a confidence score is generated, where the confidence score refers to the reliability score of the prediction result of each abnormality type by the deep fusion learning model; the time windows with confidence scores higher than a preset threshold are screened out, the abnormal signals within the time windows are identified, the time windows identified as abnormal are marked, and the abnormality type, location and severity of the abnormal signals are recorded, where the location refers to the time window corresponding to the abnormal signal.

[0063] In this scheme, the multi-task learning framework: a machine learning method that allows the model to complete multiple related tasks simultaneously. In ECG anomaly detection, the multi-task learning framework is used to predict multiple anomaly types and their corresponding probability distributions for ECG data segments within each time window. This approach improves the learning efficiency and generalization ability of the model.

[0064] Abnormality types refer to different types of abnormal conditions that may appear in the ECG signal, such as atrial fibrillation, ventricular premature beats, conduction block, etc. Each abnormality type corresponds to a specific heart activity pattern and is an important basis for diagnosing heart disease.

[0065] The probability distribution indicates the possibility of a certain type of anomaly occurring in each time window. The probability distribution predicted by the model reflects the probability of the occurrence of the anomaly type and is the basis for subsequent confidence evaluation.

[0066] The confidence score is the reliability score of the deep fusion learning model's prediction results for each anomaly type. The higher the confidence score, the more reliable the model's prediction of the anomaly type. Confidence is usually quantified using indicators such as the cross entropy loss function.

[0067] The preset threshold is a numerical standard set to filter out time windows with higher confidence scores. Only when the confidence score of an anomaly type in a time window exceeds this threshold will it be considered a reliable anomaly signal.

[0068] The position refers to the time window corresponding to the abnormal signal. This includes not only the specific position of the time window in the overall ECG record (such as the start time and end time), but also the detailed time point information within the window.

[0069] In the embodiment of the present application, the multi-task learning framework in the trained deep fusion learning model is used to first identify the abnormal types contained in the ECG data fragments in each time window, and obtain the probability distribution corresponding to each abnormal type. Then, the confidence of each abnormal type is evaluated to generate a confidence score. Next, the system screens out time windows with confidence scores higher than a preset threshold, identifies abnormal signals in these time windows, marks the time windows identified as abnormal, and records the abnormal type, location, and severity of the abnormal signal.

[0070] The process of identifying abnormal signals in these time windows includes: for each time window, if the confidence score of a certain abnormal type exceeds a preset threshold, it is considered that the corresponding abnormal signal exists in the time window.

[0071] Here is a specific example: In a practical application scenario, suppose we have a trained deep fusion learning model that can identify multiple types of cardiac abnormalities, such as atrial fibrillation, ventricular premature beats, conduction block, etc. In order to improve the detection accuracy, the system adopts a multi-task learning framework, which enables the model to predict the probability distribution of multiple abnormality types at the same time.

[0072] For each time window (e.g., a 5-second segment of ECG data), the model outputs a vector where each element represents the probability of the presence of an abnormality type. For example, if three abnormality types are considered, the output might be [0.2, 0.7, 0.1], representing the probabilities of atrial fibrillation, ventricular premature beats, and conduction block, respectively.

[0073] Next, the system calculates a confidence score for each abnormality type based on these probability distributions. The confidence score reflects the model's prediction reliability for each abnormality type. For example, for the above output vector, the system may generate a confidence score of [0.6, 0.9, 0.5], indicating that the model is most confident in predicting premature ventricular contractions.

[0074] To ensure that only high-confidence results are reported, the system sets a preset threshold (such as 0.8). Only when the confidence score of a certain abnormality type exceeds this threshold will the system consider it a reliable abnormal signal. In this example, only the confidence score of ventricular premature beats exceeds the threshold, so the system marks the time window as abnormal and records its abnormality type (ventricular premature beats), location (specific time period) and severity (indirectly reflected by the confidence score).

[0075] Finally, all time windows marked as abnormal will be summarized into a detailed abnormality report. Doctors can view the specific information of these abnormal signals through a customized visual interface to make more accurate diagnostic decisions. This method not only improves the accuracy of abnormality detection, but also provides personalized diagnostic information, significantly improving the quality and efficiency of medical services.

[0076] In order to solve the limitation that the preliminary recognition result may have omissions or uncertainties, and further improve the accuracy and personalization level of abnormal signal detection, optionally, the preliminary recognition result in step 103 is combined with a pre-trained machine learning model, and the preliminary recognition result is subjected to hybrid model fusion processing by a stacked generalization fusion method, and the inference result of the preliminary recognition result is obtained through a Bayesian network constructed by the detailed information of the patient to generate a target recognition result, including: A machine learning model is trained based on a comprehensive feature representation and an enhanced feature representation additionally extracted from the target electrocardiogram signal, so that the machine learning model can predict whether each time window contains an abnormal signal and the abnormal type and confidence corresponding to the abnormal signal; the enhanced feature representation includes statistical features, frequency domain features, time domain features and nonlinear features additionally extracted from the target electrocardiogram signal; using stacked generalization technology, the preliminary recognition result output by the deep fusion learning model is used as the first layer feature input, and the prediction result of the machine learning model is used as the second layer feature input to generate an intermediate recognition result; based on the detailed information of the patient obtained, a Bayesian network is constructed, and a context-aware reasoning result for the preliminary recognition is generated through the Bayesian network, so as to provide additional context information for the detection of abnormal signals through the reasoning result, and the detailed information includes medical records, historical electrocardiogram records and clinical symptoms; based on the reasoning result, the intermediate recognition result is reasoned in combination with the individual differences and historical heart data of the patient obtained to adjust the probability distribution corresponding to the intermediate recognition result to generate a target recognition result.

[0077] In this scheme, the comprehensive feature representation is a feature set that is integrated from the local feature representation, time series feature representation, and key feature representation. These features reflect the multifaceted information in the electrocardiogram data and are the basis for training machine learning models.

[0078] Enhanced feature representation refers to the statistical features, frequency domain features, time domain features and nonlinear features extracted from the target ECG signal. These features can provide richer information and help the machine learning model understand the ECG data more comprehensively.

[0079] The machine learning model is an algorithm based on statistical principles, which is used to predict whether each time window contains an abnormal signal and its corresponding abnormal type and confidence. In this embodiment, the model is trained using comprehensive feature representation and enhanced feature representation to enhance its ability to capture complex patterns.

[0080] Stacking generalization is an integrated learning method that improves overall performance by combining the outputs of different models. In this embodiment, the initial recognition results of the deep fusion learning model are used as the first-layer feature input, and the prediction results of the machine learning model are used as the second-layer feature input, thereby generating a more reliable intermediate recognition result.

[0081] A Bayesian network is a probabilistic graphical model used to represent conditional dependencies between variables. In this scenario, the Bayesian network is constructed based on detailed patient information (such as medical records, historical electrocardiogram records, and clinical symptoms), providing context-aware reasoning results for the initial recognition results, increasing the personalization of the diagnosis.

[0082] The inference result refers to the context-aware information generated by the Bayesian network for preliminary recognition. This inference result not only takes into account the current ECG data, but also combines the individual differences of patients and historical heart data to improve the accuracy of abnormal signal detection.

[0083] The target recognition result refers to the final abnormal signal detection result, which is analyzed and optimized at multiple levels to ensure the high accuracy and reliability of the result.

[0084] In an embodiment of the present application, a machine learning model is first trained based on a comprehensive feature representation and an enhanced feature representation additionally extracted from the target electrocardiogram signal, so that the model can predict whether each time window contains an abnormal signal and its corresponding abnormal type and confidence. Then, using the stacked generalization technique, the preliminary recognition results output by the deep fusion learning model are used as the first-layer feature input, and the prediction results of the machine learning model are used as the second-layer feature input to generate intermediate recognition results. Next, a Bayesian network is constructed based on the patient's detailed information, and a context-aware reasoning result is generated through the network to provide additional contextual information for the detection of abnormal signals. Finally, based on the reasoning results, combined with the individual differences of the patient and historical heart data, the intermediate recognition results are reasoned, their probability distribution is adjusted, and the final target recognition results are generated.

[0085] Here is a specific example: In a practical application scenario, suppose we have a deep fusion learning model that has been trained to identify multiple types of cardiac abnormalities. To further improve detection accuracy, the system introduces an additional machine learning model that is trained using comprehensive feature representations (such as local morphological features, time series features, and key features) and enhanced feature representations (such as statistical features, frequency domain features, time domain features, and nonlinear features). These enhanced features provide more information dimensions, which help the model better understand and classify complex ECG data.

[0086] For each time window (e.g., a 5-second ECG data segment), the deep fusion learning model first outputs a preliminary recognition result, including the possible abnormality type and its confidence score. This preliminary result is then used as the first-layer feature input for the stacked generalization technique, while the machine learning model predicts whether each time window contains an abnormal signal and its corresponding abnormality type and confidence based on the comprehensive feature representation and enhanced feature representation. These predictions are used as the second-layer feature input. Through this two-layer input structure, the system generates more accurate intermediate recognition results.

[0087] The system then builds a Bayesian network based on the patient's detailed information, such as medical records, historical ECG records, and clinical symptoms. The network not only considers the current ECG data, but also combines the patient's historical health records and personal characteristics to generate context-aware reasoning results. For example, if the patient has a history of heart disease, the Bayesian network will pay special attention to abnormal patterns that are similar to historical records, thereby improving the accuracy and personalization of the diagnosis.

[0088] Finally, the system further infers the intermediate recognition results based on the inference results generated by the Bayesian network, combined with the individual differences of the patients and historical heart data. This step adjusts the probability distribution of the intermediate recognition results, making the final target recognition results more reliable. For example, for an abnormal signal within a certain time window, if the Bayesian network infers that the signal is highly correlated with the patient's medical history, the system will increase its confidence score accordingly and mark the time window as abnormal.

[0089] In this way, doctors can obtain a detailed abnormality report to support them in making more accurate diagnostic decisions. This process not only improves the accuracy of abnormality detection, but also provides personalized diagnostic information, significantly improving the quality and efficiency of medical services. In addition, the system can automatically classify and store all generated reports according to the hospital's internal coding system, ensuring the standardization and consistency of data management.

[0090] In order to solve the limitation that the preliminary recognition results may have omissions or uncertainties, and further improve the accuracy and personalization level of abnormal signal detection, the above optional solution of "based on the inference result, combined with the acquired individual differences and historical heart data of the patient, the intermediate recognition result is inferred to adjust the probability distribution corresponding to the intermediate recognition result to generate the target recognition result" specifically includes: According to the individual differences and historical heart data of the patients obtained, the probability distribution of abnormal signals in each time window in the intermediate recognition result is adjusted to obtain an adjusted probability distribution; based on the reasoning result, the possible omitted or uncertain parts in the intermediate recognition result are supplemented to enhance the reliability of the intermediate recognition result to obtain an optimized intermediate recognition result; the adjusted concept distribution and the optimized intermediate recognition result are integrated to generate a target recognition result.

[0091] In this scheme, individual differences refer to differences in physiological characteristics between different patients, such as age, gender, medical history, etc. These differences affect the expression of ECG signals and therefore require special consideration in abnormality detection.

[0092] Historical cardiac data includes a patient’s previous ECG records, treatment responses, and other relevant medical information. This data can provide additional contextual information to help more accurately evaluate the current ECG signal.

[0093] The probability distribution indicates the possibility of a certain type of anomaly occurring in each time window. The probability distribution predicted by the model reflects the probability of the occurrence of the anomaly type and is the basis for subsequent confidence evaluation.

[0094] The adjusted probability distribution refers to the probability distribution adjusted for individual patient differences and historical cardiac data. This adjustment makes abnormality detection more personalized and improves the accuracy of diagnosis.

[0095] The inference results are generated by the Bayesian network with context-aware information based on the patient's detailed information. This inference result not only takes into account the current ECG data, but also combines the individual differences of the patient and historical heart data, providing a more comprehensive basis for diagnosis.

[0096] The optimized intermediate recognition result refers to the intermediate recognition result whose reliability is enhanced after the possible missing or uncertain parts are supplemented by the reasoning result. This step ensures the accuracy and completeness of the final recognition result.

[0097] In the embodiment of the present application, firstly, the probability distribution of abnormal signals in each time window in the intermediate recognition result is adjusted according to the individual differences and historical heart data of the patients obtained, and the adjusted probability distribution is obtained. Then, based on the reasoning results, the possible missing or uncertain parts in the intermediate recognition result are supplemented to enhance its reliability and obtain the optimized intermediate recognition result. Finally, the adjusted probability distribution and the optimized intermediate recognition result are integrated to generate the final target recognition result.

[0098] Here is a specific example: In a practical application scenario, suppose we have a trained system that can recognize multiple types of cardiac abnormalities and generate preliminary intermediate recognition results. In order to further improve the detection accuracy, the system introduces personalized adjustment and optimization steps.

[0099] The system first obtains the individual differences of the patient (such as age, gender, medical history) and historical heart data (such as previous electrocardiogram records, treatment response). This information is used to adjust the probability distribution of abnormal signals in each time window in the intermediate recognition results.

[0100] For example, if a patient has a history of atrial fibrillation, the system will increase the probability score of that abnormality type in all time windows. For a young patient with no history of heart disease, the system may reduce the probability scores of some common abnormality types, thereby focusing more on less common abnormal patterns.

[0101] The adjusted probability distribution is more consistent with the actual situation of the patient and improves the personalization level of abnormality detection.

[0102] Next, the system uses the inference results generated by the Bayesian network to supplement the parts that may be missing or uncertain in the intermediate recognition results. The Bayesian network is built based on the patient's detailed information (such as medical records, historical electrocardiogram records, and clinical symptoms) to generate context-aware inference results.

[0103] For example, if the confidence score of an abnormal signal within a time window is low, but the Bayesian network infers that the signal is highly correlated with the patient's medical history, the system will increase its confidence score accordingly and mark the time window as an anomaly.

[0104] This supplement enhances the reliability and completeness of intermediate identification results and reduces the possibility of misdiagnosis and missed diagnosis.

[0105] Finally, the system integrates the adjusted probability distribution and the optimized intermediate recognition results to generate the final target recognition result. This result not only includes the type, location and severity of the abnormality in each time window, but also integrates the individual differences of the patient and historical heart data, providing more comprehensive and reliable diagnostic information. Doctors can view these detailed abnormality reports through a customized visual interface, supporting them to make more accurate diagnostic decisions.

[0106] In this way, doctors can obtain a detailed abnormality report to support them in making more accurate diagnostic decisions. This process not only improves the accuracy of abnormality detection, but also provides personalized diagnostic information, significantly improving the quality and efficiency of medical services. In addition, the system can automatically classify and store all generated reports according to the hospital's internal coding system, ensuring the standardization and consistency of data management.

[0107] In order to solve the problem of insufficient time series consistency in the preliminary recognition results and further improve the accuracy and consistency of abnormal signal detection, in one or more of the above embodiments, the method in step 103 also includes: optimizing the time series consistency of the target recognition results based on the long-term trends and short-term fluctuations in the historical cardiac data of the patient.

[0108] In this scenario, long-term trends refer to changes in cardiac activity patterns over long periods of time in a patient’s historical cardiac data. These trends reflect the evolution of a patient’s cardiac health status and are important for understanding the context of the current ECG signal.

[0109] Short-term fluctuations are changes in heart activity over a shorter period of time (e.g., hours to days). These fluctuations may be caused by specific events (e.g., exercise, mood changes), or they may indicate an acute heart problem.

[0110] Time series consistency refers to the coherence and logic of target recognition results in the time dimension. Optimizing time series consistency means ensuring that the detected abnormal signals not only conform to the current time window characteristics, but also match the long-term trends and short-term fluctuations in historical data, thereby providing more reliable diagnostic information.

[0111] In the embodiment of the present application, the system not only adjusts the probability distribution of the intermediate recognition results based on the individual differences of the patients and the historical heart data, but also optimizes the time series consistency of the target recognition results according to the long-term trends and short-term fluctuations in the historical heart data of the patients. This step ensures that the detected abnormal signal not only conforms to the current time window characteristics, but also matches the long-term trends and short-term fluctuations in the historical data, providing more reliable and coherent diagnostic information.

[0112] Here is a specific example: In a practical application scenario, suppose we have a trained system that can recognize multiple types of cardiac abnormalities and generate preliminary target recognition results. In order to further improve the detection accuracy and consistency, the system introduces a time series consistency optimization step based on historical cardiac data.

[0113] The system first analyzes the patient's historical cardiac data for long-term trends. For example, if the patient has a long history of atrial fibrillation, the system checks whether the current ECG data shows a similar abnormal pattern.

[0114] Long-term trend analysis helps the system understand the evolution of the patient's heart health status, ensuring that abnormal signals detected are not isolated events but are consistent with historical records.

[0115] Next, the system evaluates short-term fluctuations in the patient’s historical cardiac data. For example, if the patient has had an acute cardiac event (such as an acute myocardial infarction) within the past few days, the system will pay special attention to abnormal signals during those time periods.

[0116] Short-term volatility assessment helps the system catch potential acute problems and ensures these signals are properly flagged and reported.

[0117] Finally, the system optimizes the time series consistency of the target recognition results based on long-term trends and short-term fluctuations. This step ensures that the detected abnormal signal not only conforms to the current time window characteristics, but also matches the long-term trends and short-term fluctuations in the historical data. For example, if the confidence of the abnormal signal in a certain time window is low, but the Bayesian network infers that the signal is highly correlated with the patient's medical history, the system will increase its confidence score accordingly and mark the time window as abnormal. At the same time, the system will also check whether the abnormal signal is consistent with the long-term trends and short-term fluctuations in the historical data to ensure its time series consistency.

[0118] After optimization, the system generates a final diagnostic report, which not only includes the type, location and severity of the abnormality in each time window, but also integrates the individual differences of patients, historical cardiac data and the results of time series consistency optimization.

[0119] Doctors can view these detailed abnormality reports through a customized visual interface, which supports them in making more accurate diagnostic decisions. In addition, the system can automatically classify and store all generated reports according to the hospital's internal coding system, ensuring the standardization and consistency of data management.

[0120] This application takes into account that in order to more comprehensively and accurately evaluate the patient's heart health status, traditional methods only rely on the detection results and confidence scores of abnormal types, ignoring the influence of local features, time series features and key features. In addition, existing methods fail to fully consider individual differences in patients and long-term trends and short-term fluctuations in historical heart data. Therefore, this application proposes a new heart health status scoring formula, which is not only based on the abnormal type of electrocardiogram data fragments in each time window and their corresponding confidence scores, but also combines local feature representation, time series feature representation and key feature representation, while taking into account individual differences in patients and the influence of additional features. In this way, the system can generate a more comprehensive and reliable heart health status score to support doctors in making more accurate diagnostic decisions.

[0121] The program includes: Based on the abnormal type of the electrocardiogram data segment in each time window and its corresponding confidence score, and in combination with the local feature representation, time series feature representation and key feature representation in the time window, the patient's heart health status score is evaluated; wherein the heart health status score is calculated by the following formula: ; in, is the heart health status score, is the number of time windows, is the index of the time window; is the number of exception types, is the index of the exception type; It is In the time window A confidence score for each anomaly type, wherein the confidence score refers to a reliability score of the prediction result of the deep fusion learning model for each anomaly type; is the exponential factor of the confidence score, which is used to adjust the influence of the confidence score. ,in is the maximum value among all confidence scores; It is The weights of the abnormality types reflect the importance of different abnormality types to the heart health status; It is The influence factor of the local feature representation within a time window is calculated by the standardized score of the local feature representation; It is The impact factor of the time series feature representation within a time window is calculated by the standardized score of the time series feature representation; It is The impact factor of the key feature representation within a time window is calculated by the standardized score of the key feature representation; They are the weight coefficients of local feature representation, time series feature representation and key feature representation, which are used to adjust the influence of each feature representation and are determined according to the importance and contribution of each feature; They are the exponential factors of local feature representation, time series feature representation, and key feature representation, which are used to adjust the influence of each feature representation and are determined according to the sensitivity and importance of each feature; is a Gaussian kernel function used to smooth the effect of confidence scores, where is the average of all confidence scores, is the standard deviation of the confidence scores; It is The comprehensive impact factor of a time window is defined as: ; in, is the number of additional features, which include at least the patient's age, gender, and medical history; It is The time window and relevance scores of additional features; It is The weight of the additional feature is used to adjust the influence of the additional feature.

[0122] The following are detailed explanations of the parameters involved in the above formula: : Heart health status score. This score comprehensively reflects the overall status of the patient's heart health and is the weighted sum of abnormal types and their feature representations in all time windows.

[0123] : The number of time windows. Each time window represents a fixed-length segment of ECG data and is used to analyze ECG signals segment by segment.

[0124] : The index of the time window. Used to identify a specific time window.

[0125] : The number of abnormality types. Each abnormality type corresponds to a specific cardiac activity pattern, such as atrial fibrillation, premature beats, etc.

[0126] : Index of exception type. Used to identify a specific exception type.

[0127] : No. In the time window The confidence score of each anomaly type is obtained through deep fusion learning model prediction, which reflects the model's prediction reliability for each anomaly type.

[0128] : Exponential factor for confidence score. Set to , used to adjust the influence of confidence scores and increase the importance of high confidence scores. Maximum confidence score is the maximum value among all confidence scores.

[0129] : No. The weight of each abnormality type is determined according to its importance to the heart health status, usually based on clinical experience and statistical data.

[0130] : No. The influencing factors of the local features in a time window are extracted and standardized by convolutional neural network, reflecting the influence of the local morphological features in the time window (such as changes in QRS complex, P wave and lower wave) on the heart health status.

[0131] : No. The influencing factors of the time series features within a time window are captured and standardized by the recurrent neural network, reflecting the impact of the heart rhythm change pattern and long-term dependency between multiple time windows on the heart health status.

[0132] : No. The influence factor of the key features in the time window is identified and the score calculation is standardized through the attention mechanism, reflecting the influence of the key features that are most likely to indicate abnormalities on the heart health status.

[0133] : are the weight coefficients of local feature representation, time series feature representation and key feature representation. They are determined according to the importance and contribution of each feature and are used to balance the impact of different feature representations.

[0134] : are the exponential factors of local feature representation, time series feature representation and key feature representation respectively. They are determined according to the sensitivity and importance of each feature and are used to further adjust the influence of each feature representation.

[0135] : Gaussian kernel function, used to smooth the influence of confidence score. is the average of all confidence scores, is the standard deviation of the confidence score. This function ensures that the confidence score does not have a large impact difference when it fluctuates within a certain range.

[0136] : No. The comprehensive impact factor of a time window. It is defined as: ; in is the number of additional characteristics, including at least the patient’s age, sex, and medical history; It is The time window and The correlation score of the additional feature is obtained by statistical analysis or machine learning methods, which reflects the strength of the association between the data in the time window and the specific additional feature. For example, the association can be quantified by correlation coefficient, mutual information or other appropriate statistical measures; It is The weight of the additional feature is taken into account by the patient's individual differences and the impact of historical cardiac data.

[0137] The following is a brief explanation of the reasons for the design of each item in the above formula: : This sub-item combines the weight of the anomaly type and the influencing factors of the three feature representations. Anomaly type weight It reflects the importance of different abnormality types, while the impact factor of feature representation evaluates the characteristics of cardiac activity from different perspectives. and exponential factor , ensuring the reasonable distribution of different feature representations in the scores.

[0138] : This sub-item uses a Gaussian kernel function to smooth the effect of confidence scores. This helps reduce the undue influence of extreme confidence scores and ensures stability when scores fluctuate within a certain range.

[0139] :This sub-item is calculated by exponential factor Adjusts the confidence score's influence. Exponential Factor The design enhances the importance of high confidence scores, thus better highlighting those more reliable anomaly detection results.

[0140] :This sub-item comprehensively considers the impact of individual differences and historical cardiac data of patients. By introducing additional features (such as age, gender, and medical history), the score is made more personalized and comprehensive.

[0141] The above formula will ensure that the abnormal signal in each time window is comprehensively evaluated by multiplying the confidence score, the influencing factor of the feature representation, the Gaussian kernel function and the smoothing factor. The multiplication operation can focus on the specific impact of each factor on the heart health status score, while avoiding a single factor dominating the score result. By adding the scores in all time windows, the comprehensive information of the entire ECG record can be summarized. The addition operation ensures the overall coherence and logic of the score, reflecting the overall picture of the heart health status.

[0142] The overall design of the formula aims to provide a comprehensive, accurate and personalized assessment method for evaluating the patient's heart health status. It not only takes into account the test results and confidence scores of abnormal types, but also combines the influence of multiple feature representations and individual differences of patients. In this way, the system can generate a more reliable heart health status score to support doctors in making more accurate diagnostic decisions. In addition, the design of the formula also takes into account the consistency and coherence of the time series, reduces the possibility of misdiagnosis and missed diagnosis, and significantly improves the quality and efficiency of medical services. .

[0143] Here is a specific example: Suppose we have a specific application scenario involving a 50-year-old male patient with a history of hypertension. The system collects 24 hours of ECG data and divides it into multiple fixed-length time windows (for example, each window is 5 seconds long). We will use the above formula to calculate the patient's heart health status score .

[0144] First set the parameters: Number of time windows (12 windows per minute, 24 hours in total).

[0145] Number of exception types (atrial fibrillation, ventricular premature beats, conduction block).

[0146] Confidence score : Assume that the confidence score output by the model is Corresponding to three exception types.

[0147] Weight coefficient : respectively , reflecting the importance of different anomaly types.

[0148] Influencing factors of local feature representation , Influencing factors of time series feature representation and the impact factor of key features : Assume the standardized score is .

[0149] Weight coefficient Set as .

[0150] Exponential Factor Set as .

[0151] Gaussian kernel function parameters: Assume .

[0152] Comprehensive impact factor Additional feature relevance score and weight : Assumption They are They are , 0.2.

[0153] Second, calculate the exponential factor for the confidence score : ; Assuming the maximum confidence score is 0.8, then .

[0154] The influencing factors of local feature representation, time series feature representation and key feature representation are further calculated: ; Further apply the Gaussian kernel function: ; Further calculation of the comprehensive impact factor : ; Finally, bring in the above values ​​to calculate the score: ; Assume that after all windows and abnormality types are combined, the final heart health status score is The score is 85 out of 100. This score indicates that the patient's heart health is relatively good, but there is a certain risk of abnormality.

[0155] Specifically: A higher score indicates that abnormal signals in most time windows are reliably detected and the severity of these abnormal signals is low. The comprehensive influencing factors take into account additional characteristics such as the patient's age, gender, and medical history, and the score further reflects the impact of individual differences in patients on their heart health status. The score also takes into account long-term trends and short-term fluctuations in historical heart data to ensure the time series consistency of the diagnostic results. Doctors can develop more personalized treatment plans or preventive measures based on this detailed scoring report, combined with other clinical information. This not only improves the accuracy of diagnosis, but also provides a comprehensive assessment of heart health.

[0156] This application takes into account that in the detection of ECG abnormalities, there may be omissions or uncertainties in the initial recognition results, especially when faced with different individual differences and complex historical cardiac data. Traditional single models often find it difficult to fully capture the impact of these factors, resulting in inaccurate and in personalized diagnostic results. In order to overcome these problems, a method that can comprehensively consider individual differences in patients, historical cardiac data, and context-aware reasoning is needed to improve the accuracy and reliability of abnormal signal detection.

[0157] Therefore, this application proposes a new formula scheme, which aims to optimize the intermediate recognition results by adjusting the probability distribution of abnormal signals in each time window and combining the inference results provided by the Bayesian network, and finally generate more accurate target recognition results. This scheme not only improves the accuracy of detection, but also provides personalized diagnostic information, significantly improving the quality and efficiency of medical services. The scheme includes: The step of adjusting the probability distribution of abnormal signals in each time window in the intermediate recognition result according to the acquired individual differences and historical heart data of the patient to obtain an adjusted probability distribution includes: The adjusted probability distribution is calculated by the following formula: ; in, It is In the time window The adjusted probability distribution of anomaly types, It is In the time window The initial probability distribution of anomaly types; It is The influencing factors of individual differences among patients within a time window are calculated based on the patient's age, gender, and medical history; is the average of the individual difference influencing factors in all time windows; is the standard deviation of the individual difference influencing factors in all time windows; is the number of historical specific features, which include specific indicators in the patient's historical cardiac data, including historical ECG records, treatment responses, It is a historically specific index; It is The time window and relevance scores for specific historical features; It is The weight of a historically specific feature is used to adjust the influence of the historically specific feature.

[0158] The following is a detailed explanation of the parameters in the above formula: It is In the time window The adjusted probability distribution of the abnormality type. This value reflects the model's latest assessment of the presence of a specific abnormality type after taking into account the impact of individual patient differences and historical cardiac data. It is calculated by the initial probability distribution It is calculated based on a series of adjustment factors to ensure that the test results are more personalized and accurate.

[0159] It is In the time window The initial probability distribution of abnormal types. This value is directly predicted by the deep fusion learning model and represents the model's initial assessment of abnormal signals without considering individual differences in patients and historical cardiac data. It is obtained by processing the ECG data fragments of each time window through the trained deep learning model and outputting the corresponding abnormal type and its confidence score.

[0160] It is The influencing factor of individual differences of patients in a time window. This factor comprehensively considers the patient's age, gender, medical history and other information to quantify the impact of these factors on the ECG data in the current time window. The acquisition method is to extract relevant information from the patient's basic medical records (such as electronic medical records) and use predefined formulas or algorithms to calculate the individual difference influencing factors of each time window.

[0161] is the average of the individual difference influencing factors in all time windows. This value is used to standardize the individual difference influencing factors in each time window, making the comparison between different time windows more reasonable. The way to obtain it is to calculate , thereby obtaining a benchmark value representing the overall individual difference level.

[0162] is the standard deviation of the individual difference influencing factors in all time windows. The standard deviation reflects the degree of dispersion of the individual difference influencing factors in all time windows and is used to measure the volatility of the data. The way to obtain it is to calculate , so as to facilitate the subsequent Gaussian kernel function smoothing and ensure that the adjusted probability distribution is more stable.

[0163] is the number of historical specific features, which include specific indicators in the patient's historical cardiac data, such as historical ECG records, treatment response, etc. Each historical specific feature provides additional contextual information that helps to more accurately evaluate the current ECG data. The acquisition path is to extract relevant historical cardiac data from the patient's electronic medical record and determine which indicators are most important for the current analysis.

[0164] It is The time window and The score reflects the correlation strength between the ECG data in the current time window and a specific historical feature. The score is obtained by calculating the similarity or correlation between the data in each time window and the historical specific feature through statistical analysis or machine learning methods, thereby providing a basis for adjusting the probability distribution.

[0165] It is The weight of a historical specific feature is used to adjust the influence of the historical specific feature. This weight reflects the importance of different historical specific features for the current abnormality detection, and is usually determined based on clinical experience and statistical data. The way to obtain it is to assign appropriate weights to each historical specific feature through expert evaluation or data analysis to ensure that it plays an appropriate role in adjusting the probability distribution.

[0166] Through the combined effect of the above parameters, by adjusting the probability distribution of abnormal signals in each time window, the final recognition result is not only based on the current ECG data, but also fully considers the individual differences of patients and historical heart data, thereby improving the accuracy and personalization of diagnosis. This method ensures that abnormal signal detection is more reliable and supports doctors to make more accurate clinical decisions.

[0167] The method of supplementing the possibly omitted or uncertain parts of the intermediate recognition result based on the inference result, enhancing the reliability of the intermediate recognition result, and obtaining an optimized intermediate recognition result includes: The optimized intermediate recognition result is calculated by the following formula: ; in, It is The intermediate recognition result after time window optimization, is the number of exception types; It is In the time window Adjusted probability distribution of anomaly types; It is the context-aware reasoning result provided by the Bayesian network. In the time window Confidence scores for each anomaly type.

[0168] The following is a detailed explanation of the parameters in the above formula: It is The intermediate recognition result after optimization of each time window. This value comprehensively considers the adjusted probability distribution and the context-aware reasoning results provided by the Bayesian network, aiming to enhance the reliability and accuracy of the intermediate recognition results. The results calculated by this formula can more comprehensively reflect the existence and severity of abnormal signals in each time window, providing a solid foundation for the final target recognition results.

[0169] is the number of abnormality types, which refers to the total number of different cardiac abnormality types that the system can identify. For example, common abnormality types may include atrial fibrillation, ventricular premature beats, conduction block, etc. The acquisition path is determined based on clinical needs and the types of abnormalities covered during model training. This parameter ensures that the formula can handle all possible abnormalities, thereby improving the comprehensiveness of detection.

[0170] It is In the time window Adjusted probability distribution of abnormality types. This value reflects the model's latest assessment of the presence of a specific abnormality type after considering the impact of individual patient differences and historical cardiac data. It is obtained by calculating the formula introduced in the previous step, that is, adjusting the initial probability distribution using the patient's individual difference influencing factors and historical specific characteristics. , ensuring that the test results are more personalized and accurate.

[0171] It is the context-aware reasoning result provided by the Bayesian network. In the time window The confidence score of each abnormality type is generated by the Bayesian network based on the patient's detailed information (such as medical records, historical ECG records, and clinical symptoms). It provides context-aware information and supplements the parts that may be missed or uncertain in the initial recognition results. The way to obtain it is to build and train the Bayesian network, input the patient's detailed information, and output the confidence score of the abnormality type in each time window, thereby enhancing the reliability of the recognition results.

[0172] Indicates The maximum value in the adjusted probability distribution of all anomaly types in a time window. This value is used to select the most likely anomaly type and serves as a basis to enhance the reliability of intermediate recognition results. It is obtained by comparing the adjusted probability distribution of all anomaly types in each time window. , and select the maximum value. This step ensures that even if the probability of some anomaly types is low, the most likely cases are given priority.

[0173] The above formula The purpose is to optimize the The intermediate recognition results of the time window are more reliable and accurate. Specifically: Highest probability option: Pass Select the most likely anomaly type to ensure that important anomaly signals are not missed.

[0174] Bayesian network supplement: through Adding the context-aware reasoning results provided by the Bayesian network supplements the parts that may be missed or uncertain in the initial recognition results, enhancing the reliability of the recognition results.

[0175] This method not only improves the accuracy of anomaly detection, but also provides personalized diagnostic information, significantly improving the quality and efficiency of medical services. By combining the adjusted probability distribution and the inference results of the Bayesian network, doctors can obtain more comprehensive and reliable abnormality reports, supporting them to make more accurate diagnostic decisions.

[0176] The step of integrating the adjusted probability distribution and the optimized intermediate recognition result to generate a target recognition result includes: The target recognition result is calculated by the following formula: ; in, It is The target recognition result of time window is It is The result corresponding to the maximum adjusted probability distribution within a time window; It is The optimized intermediate recognition result of the time window; It is the weight coefficient between the adjusted probability distribution and the optimized intermediate recognition result, which is used to balance the influence of the two.

[0177] The following is a detailed explanation of the parameters in the above formula: It is The target recognition result of each time window. This value comprehensively considers the adjusted probability distribution and the optimized intermediate recognition results to generate the final abnormal signal detection result. The results calculated by this formula can more comprehensively reflect the existence and severity of abnormal signals in each time window, providing doctors with the most accurate and reliable diagnostic information.

[0178] It is The result corresponding to the maximum adjusted probability distribution in the time window. This value reflects the most likely situation in the model's latest assessment of the presence of a specific abnormality type after considering the influence of individual patient differences and historical cardiac data. It is obtained by comparing the first Adjusted probability distribution of all anomaly types in a time window , select the maximum value among them This step ensures that even if some anomaly types have a lower probability, the most likely cases are prioritized.

[0179] It is The optimized intermediate recognition result of the time window. This value comprehensively considers the adjusted probability distribution and the context-aware reasoning results provided by the Bayesian network, enhancing the reliability and accuracy of the intermediate recognition results. The way to obtain it is through the formula in the previous step The calculation ensures that the recognition result is not only based on the current ECG data, but also fully considers the individual differences and historical heart data of the patient, thereby improving the accuracy and personalization level of diagnosis.

[0180] It is the weight coefficient between the adjusted probability distribution and the optimized intermediate recognition result, which is used to balance the influence of the two. This coefficient determines the adjusted probability distribution. And the optimized intermediate recognition result In the final target recognition result The relative importance of the weights can be determined by expert evaluation or experimental verification. The value range is The specific value can be adjusted according to the needs of different application scenarios. For example, if you want to pay more attention to the initial prediction of the model, you can choose a larger On the contrary, if you attach more importance to the context-aware reasoning results provided by the Bayesian network, you can choose a smaller value.

[0181] The above formula The purpose is to integrate The adjusted probability distribution of the time window and the optimized intermediate recognition results generate the final target recognition result. Specifically: Adjusted probability distribution: The introduction of the adjusted probability distribution ensures that the model's initial assessment of abnormal signals still accounts for a certain proportion.

[0182] Optimized intermediate recognition results: The introduction of optimized intermediate recognition results supplements the context-aware reasoning results provided by the Bayesian network and enhances the reliability of the recognition results.

[0183] Weight coefficient : By adjusting The value of can flexibly control the balance between the adjusted probability distribution and the optimized intermediate recognition results, ensuring that the final result is both accurate and personalized.

[0184] This method not only improves the accuracy of anomaly detection, but also provides personalized diagnostic information, significantly improving the quality and efficiency of medical services. By combining the adjusted probability distribution and the optimized intermediate recognition results, doctors can obtain more comprehensive and reliable abnormality reports, supporting them to make more accurate diagnostic decisions. Final target recognition results It provides a solid foundation for clinical application and ensures the scientificity and reliability of diagnosis.

[0185] Here is a specific example for the complete solution above: Suppose we have a specific application scenario involving a 50-year-old male patient with a history of hypertension. The system collects 24 hours of ECG data and divides it into multiple fixed-length time windows (for example, each window is 5 seconds long). We will use the above formula to calculate the patient's heart health status score and explain the calculation results.

[0186] First set the parameters: Number of time windows (12 windows per minute, 24 hours in total); Number of exception types (atrial fibrillation, ventricular premature beats, conduction block); Initial probability distribution : Assume that the initial probability distribution of the model output is Corresponding to three types of exceptions. ; Individual difference influencing factors :Assuming that the patient's age, gender, and medical history are calculated They are .

[0187] The average value of the individual difference influencing factors ; Standard deviation of individual difference influencing factors ; Number of history-specific features (historical ECG records, treatment response, medication use); Relevance score :Assume that ; Weights of history-specific features : respectively ; Confidence scores provided by Bayesian networks : Assume that ; Weight coefficient ; Second, calculate the adjusted probability distribution: a. Gaussian kernel function: ; For example, for , the calculation result is: ; b. The influence of historically specific characteristics: ; c. Adjusted probability distribution: ; Assumptions ,but: ; Further calculate the optimized intermediate recognition results: a. Maximum adjusted probability distribution: ; b. Bayesian network supplement: ; c. Optimized intermediate recognition results: ; Further calculate the target recognition results: a. The value corresponding to the maximum adjusted probability distribution: ; b. Target recognition results: ; The final calculated target recognition result The score is 1.12. This score indicates that after considering the individual differences of patients and the influence of historical heart data, the system has a more accurate assessment of the presence and severity of abnormal signals in this time window. Specifically: In the ECG abnormality detection, the target recognition result The value itself is a comprehensive score, which reflects the The presence and severity of abnormal signals within a time window. In order to draw specific conclusions from this value, we need to interpret it in combination with clinical thresholds, historical data, and the professional judgment of doctors.

[0188] The following are the detailed explanation steps and conclusions: First, based on the set clinical threshold, the normal range is assumed to be , the warning range is , the exception range is .because Falling within the warning range, it indicates that potential heart problems may exist within that time window, but have not yet reached a serious level.

[0189] Secondly, if the patient has a history of atrial fibrillation and has had similar high scores many times in the recent period, special attention should be paid. If this is the first time that a high score has appeared, it is recommended to further monitor the data of other time periods to confirm whether there is a persistent problem.

[0190] Finally, the doctor can The patient's heart rate and blood pressure are measured by the blood pressure sensor, and the patient's heart rate and blood pressure are measured by the blood pressure sensor. The patient's heart rate and blood pressure are measured by the blood pressure sensor, and the patient's heart rate and blood pressure are measured by the blood pressure sensor. The patient's heart rate and blood pressure are measured by the blood pressure sensor, and the patient's heart rate and blood pressure sensor are measured by the blood pressure sensor. The patient's heart rate and blood pressure sensor ...

[0191] Based on the above analysis, the target recognition results This indicates that there may be potential heart problems within this time window, but they have not yet reached a serious level. Specifically: Mild to moderate abnormality: The score falls within the warning range, indicating a certain risk of abnormality.

[0192] Further observation: It is recommended to continue monitoring data in other time periods to confirm whether there is a persistent problem.

[0193] Personalized diagnosis: Combining the patient's individual differences and historical heart data, doctors can develop more personalized diagnosis and treatment plans.

[0194] By introducing adjusted probability distribution, optimized intermediate recognition results and final target recognition results, this solution effectively enhances the accuracy and personalization of ECG abnormality detection. This approach ensures more reliable abnormal signal detection and supports doctors to make more accurate clinical decisions. The final target recognition results provide a solid foundation for clinical applications and ensure the scientificity and reliability of diagnosis.

[0195] Figure 2 A schematic diagram of a structure of an abnormal ECG data device based on deep learning is provided for an embodiment of the present application. Figure 2 As shown, the device comprises: The acquisition and processing module 21 is used to acquire the original electrocardiogram signal of the patient and perform a multi-stage preprocessing process to obtain a target electrocardiogram signal, wherein the multi-stage preprocessing process at least includes: using a bandpass filter to remove high-frequency noise and power frequency interference in the original electrocardiogram signal; using wavelet transform or adaptive filtering technology to suppress myoelectric artifacts and low-frequency noise in the original electrocardiogram signal; and using a differential operational amplifier to enhance the original electrocardiogram signal. A training module 22 is used to train a deep fusion learning model using a target electrocardiogram signal, wherein the deep fusion learning model combines a convolutional neural network and a recurrent neural network, wherein the convolutional neural network is used to extract local features of the target electrocardiogram signal, and the recurrent neural network is used to capture time series features of the target electrocardiogram signal, and at the same time, an attention mechanism is introduced to enable the deep fusion learning model to focus on key parts indicating abnormal signals, so as to improve the sensitivity of abnormal information recognition; The recognition output module 23 is used to recognize abnormal signals in the target electrocardiogram signal through the deep fusion learning model to obtain a preliminary recognition result; combine the preliminary recognition result with the pre-trained machine learning model, perform hybrid model fusion processing on the preliminary recognition result through a stacked generalization fusion method, and obtain the inference result of the preliminary recognition result through a Bayesian network constructed by the detailed information of the patient to generate a target recognition result; The generation and display module 24 is used to generate an abnormality report based on the target recognition result using a rule engine. The abnormality report includes the abnormality type, location, and severity of the target electrocardiogram signal, classifies and stores the abnormality report according to predefined standards, and displays the abnormality report through a customized visual interface.

[0196] Figure 2 The device for identifying abnormal ECG data based on deep learning can be used to perform Figure 1The implementation principle and technical effect of the method for identifying abnormal ECG data based on deep learning described in the embodiment shown are not repeated here. The specific way in which each module and unit performs operations in the device for identifying abnormal ECG data based on deep learning in the above embodiment has been described in detail in the embodiment of the method, and will not be elaborated here.

[0197] In one possible design, Figure 2 The device for identifying abnormal ECG data based on deep learning in the embodiment shown can be implemented as a computing device, such as Figure 3 As shown, the computing device may include a storage component 31 and a processing component 32; The storage component 31 stores one or more computer instructions, wherein the one or more computer instructions are called and executed by the processing component 32 .

[0198] The processing component 32 is used to: collect the original electrocardiogram signal of the patient, and perform a multi-stage preprocessing process to obtain a target electrocardiogram signal, wherein the multi-stage preprocessing process at least includes: using a bandpass filter to remove high-frequency noise and power frequency interference in the original electrocardiogram signal; applying wavelet transform or adaptive filtering technology to suppress myoelectric artifacts and low-frequency noise in the original electrocardiogram signal; using a differential operational amplifier to enhance the original electrocardiogram signal; using the target electrocardiogram signal to train a deep fusion learning model, wherein the deep fusion learning model combines a convolutional neural network and a recurrent neural network, wherein the convolutional neural network is used to extract local features of the target electrocardiogram signal, and the recurrent neural network is used to capture the time series features of the target electrocardiogram signal, and at the same time, by introducing an attention mechanism, the deep fusion learning model can focus on abnormal signals in the target electrocardiogram signal; The deep fusion learning model is used to identify abnormal signals in the target electrocardiogram signal to obtain a preliminary recognition result; the preliminary recognition result is combined with a pre-trained machine learning model, and the preliminary recognition result is subjected to hybrid model fusion processing through a stacked generalization fusion method, and the inference result of the preliminary recognition result is obtained through a Bayesian network constructed based on the patient's detailed information to generate a target recognition result; based on the target recognition result, an abnormality report is generated using a rule engine, and the abnormality report includes the abnormal type, location, and severity of the target electrocardiogram signal, and the abnormality report is classified and stored according to predefined standards, and the abnormality report is displayed through a customized visual interface.

[0199] The processing component 32 may include one or more processors to execute computer instructions to complete all or part of the steps in the above method. Of course, the processing component may also be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors or other electronic components to perform the above method.

[0200] The storage component 31 is configured to store various types of data to support operations at the terminal. The storage component can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0201] Of course, the computing device may also include other components, such as input / output interfaces, display components, communication components, etc.

[0202] The input / output interface provides an interface between the processing component and the peripheral interface module, which may be an output device, an input device, etc.

[0203] The communication component is configured to facilitate, among other things, wired or wireless communications between the computing device and other devices.

[0204] Among them, the computing device can be a physical device or an elastic computing host provided by a cloud computing platform, etc. In this case, the computing device can refer to a cloud server, and the above-mentioned processing components, storage components, etc. can be basic server resources rented or purchased from the cloud computing platform.

[0205] The present application also provides a computer storage medium storing a computer program, wherein the computer program can achieve the above-mentioned Figure 1 The illustrated embodiment is a method for detecting abnormal ECG data based on deep learning.

[0206] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0207] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0208] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0209] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for identifying abnormal electrocardiogram data based on deep learning, characterized in that: include: Collecting the original electrocardiogram signal of the patient and performing a multi-stage preprocessing process to obtain a target electrocardiogram signal, the multi-stage preprocessing process at least includes: using a bandpass filter to remove high-frequency noise and power frequency interference in the original electrocardiogram signal; applying wavelet transform or adaptive filtering technology to suppress myoelectric artifacts and low-frequency noise in the original electrocardiogram signal; using a differential operational amplifier to enhance the original electrocardiogram signal; A deep fusion learning model is trained using a target electrocardiogram signal, wherein the deep fusion learning model combines a convolutional neural network and a recurrent neural network, wherein the convolutional neural network is used to extract local features of the target electrocardiogram signal, and the recurrent neural network is used to capture time series features of the target electrocardiogram signal, and an attention mechanism is introduced to enable the deep fusion learning model to focus on abnormal signals in the target electrocardiogram signal; The abnormal signal in the target electrocardiogram signal is identified by the deep fusion learning model to obtain a preliminary recognition result; the preliminary recognition result is combined with the pre-trained machine learning model, and the preliminary recognition result is subjected to hybrid model fusion processing by a stacked generalization fusion method, and the inference result of the preliminary recognition result is obtained by a Bayesian network constructed by the detailed information of the patient to generate a target recognition result; Based on the target recognition result, a rule engine is used to generate an exception report, which includes the abnormal type, location, and severity of the target electrocardiogram signal. The exception report is classified and stored according to predefined standards, and displayed through a customized visual interface.

2. The method according to claim 1, characterized in that The identifying abnormal signal in the target electrocardiogram signal by the deep fusion learning model to obtain a preliminary identification result includes: Performing time window segmentation processing on the target electrocardiogram signal to obtain a plurality of time windows of fixed length, wherein each time window represents an electrocardiogram data segment within a time period, and the electrocardiogram data segments corresponding to all the time windows can constitute the target electrocardiogram signal; The electrocardiogram data segments in each time window are fed into the convolutional neural network in the deep fusion learning model, so as to extract the local morphological features of the electrocardiogram data in each time window through the convolutional neural network, and obtain a local feature representation characterizing the local characteristics of the cardiac activity, wherein the local feature representation includes at least the morphological changes of the QRS complex, the P wave and the T wave, and the heart rate variability; Sending the ECG data segments in each time window to the recurrent neural network in the deep fusion learning model, so as to capture the time series features between multiple time windows through the recurrent neural network, and determining the change pattern and long-term dependency of the heart rhythm based on the time series features to obtain a time series feature representation; Introducing an attention mechanism into the deep fusion learning model, so that when the deep fusion learning model processes the electrocardiogram data fragments in each time window, it identifies and highlights the key features that are most likely to indicate an abnormality from the electrocardiogram data fragments, and generates a key feature representation with a weight, wherein the weight is used to reflect the importance of each key feature for abnormal signal detection; Integrate the local feature representation, the time series feature representation, and the key feature representation into a comprehensive feature representation, and use the comprehensive feature representation as input to train the deep fusion learning model, so that the deep fusion learning model can predict the anomaly type and confidence; Applying a multi-task learning framework in the trained deep fusion learning model to predict the abnormality type and confidence of the electrocardiogram data segment in each time window, wherein the abnormality type includes atrial fibrillation, ventricular premature beat, and conduction block, and based on the abnormality type and confidence of the electrocardiogram data segment in each time window, identifying the abnormal signal whose confidence is higher than a preset threshold, and marking the abnormality type, location, and severity of the abnormal signal; Evaluate the patient's cardiac health status score based on the abnormality types and confidence levels of the electrocardiogram data segments within all time windows; A preliminary recognition result is generated based on the corresponding abnormal signals in all time windows and the patient's heart health status score.

3. The method according to claim 2, characterized in that Applying a multi-task learning framework in the trained deep fusion learning model to predict the abnormal type and confidence of the electrocardiogram data segment in each time window, and identifying abnormal signals with a confidence higher than a preset threshold based on the abnormal type and confidence of the electrocardiogram data segment in each time window, including: Using the multi-task learning framework in the trained deep fusion learning model, identify the abnormality type contained in the electrocardiogram data segment in each time window, and obtain the probability distribution corresponding to each abnormality type, wherein the probability distribution corresponding to each abnormality type is used to indicate the possibility of the abnormality type existing in each time window; Evaluate the confidence of each anomaly type and generate a confidence score, wherein the confidence score refers to the reliability score of the prediction result of the deep fusion learning model for each anomaly type; Filter out time windows with confidence scores higher than a preset threshold, identify abnormal signals within the time windows, mark the time windows identified as abnormal, and record the abnormal type, location and severity of the abnormal signals, where the location refers to the time window corresponding to the abnormal signal.

4. The method according to claim 1, characterized in that The method combines the preliminary recognition result with the pre-trained machine learning model, performs hybrid model fusion processing on the preliminary recognition result through a stacked generalization fusion method, and obtains the inference result of the preliminary recognition result through a Bayesian network constructed by the detailed information of the patient to generate a target recognition result, including: Training a machine learning model based on the comprehensive feature representation and the enhanced feature representation additionally extracted from the target electrocardiogram signal, so that the machine learning model can predict whether each time window contains an abnormal signal and the abnormal type and confidence level corresponding to the abnormal signal; the enhanced feature representation includes statistical features, frequency domain features, time domain features and nonlinear features additionally extracted from the target electrocardiogram signal; Using stacked generalization technology, the preliminary recognition result output by the deep fusion learning model is used as the first layer feature input, and the prediction result of the machine learning model is used as the second layer feature input to generate an intermediate recognition result; Based on the acquired detailed information of the patient, a Bayesian network is constructed, and a context-aware reasoning result is generated for the preliminary identification through the Bayesian network, so as to provide additional context information for the detection of abnormal signals through the reasoning result, wherein the detailed information includes medical history data, historical electrocardiogram records and clinical symptoms; Based on the inference result, combined with the acquired individual differences and historical heart data of the patient, the intermediate recognition result is inferred to adjust the probability distribution corresponding to the intermediate recognition result to generate a target recognition result.

5. The method according to claim 4, characterized in that The inference processing of the intermediate recognition result based on the inference result and in combination with the acquired individual differences and historical heart data of the patient to adjust the probability distribution corresponding to the intermediate recognition result and generate a target recognition result includes: According to the acquired individual differences and historical heart data of the patient, the probability distribution of abnormal signals in each time window in the intermediate recognition result is adjusted to obtain an adjusted probability distribution; Based on the inference result, the possibly omitted or uncertain parts in the intermediate recognition result are supplemented to enhance the reliability of the intermediate recognition result, thereby obtaining an optimized intermediate recognition result; Integrating the adjusted concept distribution and the optimized intermediate recognition result to generate a target recognition result; The method further comprises: The time series consistency of the target identification results is optimized based on the long-term trends and short-term fluctuations in the patient's historical cardiac data.

6. The method according to claim 2, characterized in that The step of evaluating the patient's cardiac health status score based on the abnormality type and confidence of the electrocardiogram data segments in all time windows includes: Based on the abnormal type of the electrocardiogram data segment in each time window and its corresponding confidence score, and in combination with the local feature representation, time series feature representation and key feature representation in the time window, the patient's heart health status score is evaluated; wherein the heart health status score is calculated by the following formula: ； in, is the heart health status score, is the number of time windows, is the index of the time window; is the number of exception types, is the index of the exception type; It is In the time window A confidence score for each anomaly type, wherein the confidence score refers to a reliability score of the prediction result of the deep fusion learning model for each anomaly type; is the exponential factor of the confidence score, which is used to adjust the influence of the confidence score. ,in is the maximum value among all confidence scores; It is The weights of the abnormality types reflect the importance of different abnormality types to the heart health status; It is The influence factor of the local feature representation within a time window is calculated by the standardized score of the local feature representation; It is The impact factor of the time series feature representation within a time window is calculated by the standardized score of the time series feature representation; It is The impact factor of the key feature representation within a time window is calculated by the standardized score of the key feature representation; They are the weight coefficients of local feature representation, time series feature representation and key feature representation, which are used to adjust the influence of each feature representation and are determined according to the importance and contribution of each feature; They are the exponential factors of local feature representation, time series feature representation, and key feature representation, which are used to adjust the influence of each feature representation and are determined according to the sensitivity and importance of each feature; is a Gaussian kernel function used to smooth the effect of confidence scores, where is the average of all confidence scores, is the standard deviation of the confidence scores; It is The comprehensive impact factor of a time window is defined as: ； in, is the number of additional features, which include at least the patient's age, gender, and medical history; It is The time window and relevance scores for additional features; It is The weight of the additional feature is used to adjust the influence of the additional feature.

7. The method according to claim 5, characterized in that The step of adjusting the probability distribution of abnormal signals in each time window in the intermediate recognition result according to the acquired individual differences and historical heart data of the patient to obtain an adjusted probability distribution includes: The adjusted probability distribution is calculated by the following formula: ； in, It is In the time window The adjusted probability distribution of anomaly types, It is In the time window The initial probability distribution of anomaly types; It is The influencing factors of individual differences among patients within a time window are calculated based on the patient's age, gender, and medical history; is the average of the individual difference influencing factors in all time windows; is the standard deviation of the individual difference influencing factors in all time windows; is the number of historical specific features, which include specific indicators in the patient's historical cardiac data, including historical ECG records, treatment responses, It is a historically specific index; It is The time window and relevance scores for specific historical features; It is The weight of the historical specific feature is used to adjust the influence of the historical specific feature; The method of supplementing the possibly omitted or uncertain parts of the intermediate recognition result based on the inference result, enhancing the reliability of the intermediate recognition result, and obtaining an optimized intermediate recognition result includes: The optimized intermediate recognition result is calculated by the following formula: ； in, It is The intermediate recognition result after time window optimization, is the number of exception types; It is In the time window Adjusted probability distribution of anomaly types; It is the context-aware reasoning result provided by the Bayesian network. In the time window Confidence scores for each anomaly type; The step of integrating the adjusted probability distribution and the optimized intermediate recognition result to generate a target recognition result includes: The target recognition result is calculated by the following formula: ； in, It is The target recognition result of time window is It is The result corresponding to the maximum adjusted probability distribution within a time window; It is The optimized intermediate recognition result of the time window; It is the weight coefficient between the adjusted probability distribution and the optimized intermediate recognition result, which is used to balance the influence of the two.

8. A device for identifying abnormal electrocardiogram data based on deep learning, characterized in that: include: An acquisition and processing module is used to acquire the original electrocardiogram signal of the patient and perform a multi-stage preprocessing process to obtain a target electrocardiogram signal, wherein the multi-stage preprocessing process at least includes: using a bandpass filter to remove high-frequency noise and power frequency interference in the original electrocardiogram signal; using wavelet transform or adaptive filtering technology to suppress myoelectric artifacts and low-frequency noise in the original electrocardiogram signal; and using a differential operational amplifier to enhance the original electrocardiogram signal; A training module, for training a deep fusion learning model using a target electrocardiogram signal, wherein the deep fusion learning model combines a convolutional neural network and a recurrent neural network, wherein the convolutional neural network is used to extract local features of the target electrocardiogram signal, and the recurrent neural network is used to capture time series features of the target electrocardiogram signal, and at the same time, an attention mechanism is introduced to enable the deep fusion learning model to focus on key parts indicating abnormal signals, so as to improve the sensitivity of abnormal information recognition; An identification output module is used to identify abnormal signals in the target electrocardiogram signal through the deep fusion learning model to obtain a preliminary identification result; combine the preliminary identification result with the pre-trained machine learning model, perform hybrid model fusion processing on the preliminary identification result through a stacked generalization fusion method, and obtain an inference result of the preliminary identification result through a Bayesian network constructed by the detailed information of the patient to generate a target identification result; A generation and display module is used to generate an exception report based on the target recognition result using a rule engine. The exception report includes the abnormal type, location, and severity of the target electrocardiogram signal, classifies and stores the exception report according to predefined standards, and displays the exception report through a customized visual interface.

9. A computing device, characterized in that It comprises a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement a method for identifying abnormal ECG data based on deep learning as described in any one of claims 1 to 7.

10. A computer storage medium, characterized in that: A computer program is stored, and when the computer program is executed by a computer, the method for identifying abnormal electrocardiogram data based on deep learning as described in any one of claims 1 to 7 is implemented.

Images