A deep learning-based electrocardio data anomaly recognition method, device, equipment and storage medium

Abstract

The application provides an electrocardio data anomaly recognition method and device based on deep learning, equipment and storage medium. Wherein, the original electrocardiogram signal of the patient is collected, and the target electrocardiogram signal is obtained through multi-stage preprocessing. The abnormal signal in the target electrocardiogram signal is identified by using a deep fusion learning model to generate a preliminary recognition result. Then, the preliminary recognition result is combined with a pre-trained machine learning model, a hybrid model processing is performed through a stacked generalization fusion method, a Bayesian network is constructed based on the detailed information of the patient, a reasoning result with context awareness is obtained, and finally a target recognition result is generated. According to the target recognition result, an abnormal report is generated, which is classified and stored according to the pre-defined standard and displayed to the user. This method ensures high precision and personalized diagnosis of anomaly detection, and improves the reliability and efficiency of electrocardiogram analysis.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method, apparatus, device and storage medium for identifying abnormal electrocardiogram data based on deep learning. Background Technology

[0002] With the digital transformation of the healthcare sector, real-time monitoring and analysis of electrocardiogram (ECG) data are becoming increasingly important. Especially in areas such as telemedicine and personalized health management, the ability to quickly and accurately identify abnormalities in ECGs is crucial for preventing the occurrence and progression of heart disease. Therefore, an efficient and accurate method is needed to automatically detect abnormal signals in ECGs.

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

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

[0005] This application provides a method, apparatus, device, and storage medium for identifying abnormal electrocardiogram (ECG) data based on deep learning, in order to solve the problem of poor ECG data processing efficiency in the prior art.

[0006] In a first aspect, embodiments of this application provide a method for identifying abnormal electrocardiogram (ECG) data based on deep learning, including:

[0007] The patient's raw electrocardiogram (ECG) signal is acquired and subjected to a multi-level preprocessing procedure to obtain the target ECG signal. The multi-level preprocessing procedure includes at least: using a bandpass filter to remove high-frequency noise and power frequency interference from the raw ECG signal; applying wavelet transform or adaptive filtering techniques to suppress electromyographic artifacts and low-frequency noise from the raw ECG signal; and using a differential operational amplifier to enhance the raw ECG signal.

[0008] A deep fusion learning model is trained using a target electrocardiogram (ECG) signal. The deep fusion learning model combines a convolutional neural network (CNN) and a recurrent neural network (RNN). The CNN is used to extract local features of the target ECG signal, and the RNN is used to capture the time-series features of the target ECG signal. At the same time, an attention mechanism is introduced to enable the deep fusion learning model to focus on abnormal signals in the target ECG signal.

[0009] Abnormal signals in the target electrocardiogram signal are identified by the deep fusion learning model to obtain preliminary identification results; the preliminary identification results are combined with a pre-trained machine learning model, and the preliminary identification results are fused by a stacked generalization fusion method; and the inference results of the preliminary identification results are obtained by a Bayesian network constructed from the patient's detailed information to generate target identification results.

[0010] Based on the target identification results, an anomaly report is generated using a rule engine. The anomaly report includes the type, location, and severity of the anomaly in the target electrocardiogram signal. The anomaly report is classified and stored according to predefined standards and displayed through a customized visualization interface.

[0011] Optionally, the step of identifying abnormal signals in the target electrocardiogram signal through the deep fusion learning model to obtain preliminary identification results includes:

[0012] The target electrocardiogram signal is segmented into time windows to obtain multiple time windows of fixed length. Each time window represents an electrocardiogram data segment within a time period. All the electrocardiogram data segments corresponding to the time windows can form the target electrocardiogram signal.

[0013] The electrocardiogram (ECG) data segments within each time window are fed into the convolutional neural network in the deep fusion learning model to extract local morphological features of the ECG data within each time window through the convolutional neural network, thereby obtaining a local feature representation that characterizes the local properties of cardiac activity. The local feature representation includes at least the morphological changes of the QRS complex, P wave, and T wave, as well as heart rate variability.

[0014] Electrocardiogram data segments within each time window are fed into the recurrent neural network in the deep fusion learning model to capture time series features between multiple time windows through the recurrent neural network, and to determine the change pattern and long-term dependency of heart rhythm based on the time series features, thereby obtaining a time series feature representation.

[0015] An attention mechanism is introduced into the deep fusion learning model so that when processing ECG data segments in each time window, the deep fusion learning model can identify and highlight the key features most likely to indicate abnormalities from the ECG data segments and generate key feature representations with weights, the weights being used to reflect the importance of each key feature for abnormal signal detection.

[0016] The local feature representation, the time series feature representation, and the key feature representation are integrated into a comprehensive feature representation, and the comprehensive feature representation is used as input to train the deep fusion learning model so that the deep fusion learning model can predict anomaly types and confidence levels.

[0017] A multi-task learning framework is applied in the trained deep fusion learning model to predict the abnormality type and confidence level of ECG data segments within each time window. The abnormality types include atrial fibrillation, premature ventricular contractions, and conduction block. Based on the abnormality type and confidence level of ECG data segments within each time window, abnormal signals with confidence levels higher than a preset threshold are identified, and the abnormality type, location, and severity of the abnormal signals are marked.

[0018] The patient's cardiac health status score was assessed based on the abnormality type and confidence level of all electrocardiogram data segments within all time windows.

[0019] Preliminary identification results are generated based on the abnormal signals corresponding to all time windows and the patient's cardiac health status score.

[0020] Optionally, a multi-task learning framework is applied to the trained deep fusion learning model to predict the abnormality type and confidence level of ECG data segments within each time window, and based on the abnormality type and confidence level of the ECG data segments within each time window, abnormal signals with a confidence level higher than a preset threshold are identified, including:

[0021] Using 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. The probability distribution corresponding to each abnormality type is used to represent the possibility of the abnormality type existing within each time window.

[0022] The confidence level of each anomaly type is evaluated to generate a confidence score, which refers to the reliability score of the prediction results of the deep fusion learning model for each anomaly type.

[0023] The system filters out time windows with confidence scores higher than a preset threshold, identifies abnormal signals within the time windows, marks the time windows identified as abnormal, and records the abnormality type, location, and severity of the abnormal signals, where the location refers to the time window corresponding to the abnormal signal.

[0024] Optionally, the step of combining the preliminary identification results with a pre-trained machine learning model, performing hybrid model fusion processing on the preliminary identification results through a stacked generalization fusion method, and obtaining the inference results of the preliminary identification results through a Bayesian network constructed from the patient's detailed information to generate target identification results includes:

[0025] 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 abnormality 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;

[0026] 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 intermediate recognition results.

[0027] Based on the detailed information of the patient obtained, a Bayesian network is constructed, and a context-aware inference result is generated through the Bayesian network for the preliminary identification, so as to provide additional contextual information for the detection of abnormal signals through the inference result. The detailed information includes medical records, historical electrocardiogram records and clinical symptoms.

[0028] Based on the reasoning results, and combined with the acquired individual differences and historical cardiac data of the patients, the intermediate identification results are processed to adjust the probability distribution corresponding to the intermediate identification results and generate target identification results.

[0029] Optionally, based on the reasoning result, and in conjunction with the acquired individual differences and historical cardiac data of the patient, the intermediate identification result is subjected to reasoning processing to adjust the probability distribution corresponding to the intermediate identification result and generate the target identification result, including:

[0030] Based on the individual differences and historical cardiac data of the patients obtained, the probability distribution of abnormal signals in each time window of the intermediate identification results is adjusted to obtain the adjusted probability distribution;

[0031] Based on the reasoning results, supplement the intermediate identification results to supplement any parts that may be omitted or uncertain, thereby enhancing the reliability of the intermediate identification results and obtaining optimized intermediate identification results;

[0032] By integrating the adjusted concept distribution and the optimized intermediate recognition results, a target recognition result is generated.

[0033] The method further includes:

[0034] The temporal 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.

[0035] Optionally, the assessment of the patient's cardiac health status score based on the abnormality type and confidence level of electrocardiogram data segments within all time windows includes:

[0036] Based on the abnormality type and corresponding confidence score of the electrocardiogram data segment within each time window, and combined with the local feature representation, time series feature representation, and key feature representation within that time window, the patient's cardiac health status score is assessed; wherein, the cardiac health status score is calculated using the following formula:

[0037] ;

[0038] in, It is a heart health status score. It is the number of time windows. It is the index of the time window; It is the number of exception types. It is an index of an exception type; It is the first Within the first time window The confidence score for each anomaly type refers to the reliability score of the prediction result of the deep fusion learning model for each anomaly type. It is an index factor for the confidence score, used to adjust the degree of influence of the confidence score, and is set... ,in It is the maximum value among all confidence scores; It is the first The weights of different abnormality types reflect the importance of different abnormality types to cardiac health status; It is the first The influence factor of local feature representation within a time window is calculated using the standardized score of the local feature representation. It is the first The influence factors represented by time series features within a time window are calculated using the standardized scores of the time series feature representations. It is the first The influence factor of key feature representation within a time window is calculated by the standardized score of key feature representation; These are the weighting coefficients for local feature representation, time series feature representation, and key feature representation, respectively, used to adjust the influence of each feature representation, and determined according to the importance and contribution of each feature; These are the index factors for local feature representation, time series feature representation, and key feature representation, respectively, used to adjust the degree of influence of each feature representation, and determined according to the sensitivity and importance of each feature;

[0039] It is a Gaussian kernel function, used to smooth the influence of confidence scores, where It is the average of all confidence scores. It is the standard deviation of the confidence score;

[0040] It is the first The comprehensive impact factor for each time window is defined as:

[0041] ;

[0042] in, This refers to the number of additional features, which include at least the patient's age, gender, and medical history. It is the first The first time window and the first The relevance score of the additional features; It is the first The weights of additional features are used to adjust the degree of influence of these additional features.

[0043] Optionally, adjusting the probability distribution of abnormal signals within each time window of the intermediate identification results based on the acquired individual differences and historical cardiac data of the patient to obtain the adjusted probability distribution includes:

[0044] The adjusted probability distribution is calculated using the following formula:

[0045] ;

[0046] in, It is the first Within the first time window Adjusted probability distribution of the anomaly types It is the first Within the first time window Initial probability distribution of the anomaly types; It is the first The influencing factors of individual patient differences within a time window are calculated based on the patient's age, gender, and medical history. It is the average of the individual difference influencing factors across all time windows; It is the standard deviation of the factors influencing individual differences across all time windows; It refers to the quantity of historically specific characteristics, which include specific indicators in the patient's historical cardiac data, including historical electrocardiogram records and treatment response. It is a historically specific index; It is the first The first time window and the first The relevance score of a specific historical feature; It is the first The weights of specific historical features are used to adjust the degree of influence of those features.

[0047] The process of supplementing the intermediate identification results with potentially omitted or uncertain parts based on the reasoning results enhances the reliability of the intermediate identification results, resulting in optimized intermediate identification results, including:

[0048] The optimized intermediate recognition result is calculated using the following formula:

[0049] ;

[0050] in, It is the first Intermediate recognition results after time window optimization It represents the number of exception types; It is the first Within the first time window Adjusted probability distribution of various anomaly types; The first in the context-aware reasoning results provided by the Bayesian network Within the first time window Confidence scores for different anomaly types;

[0051] The process of integrating the adjusted probability distribution and the optimized intermediate recognition results to generate target recognition results includes:

[0052] The target recognition result is calculated using the following formula:

[0053] ;

[0054] in, It is the first Target recognition results within a time window It is the first The result corresponding to the maximum adjusted probability distribution within each time window; It is the first Optimized intermediate recognition results for each time window; It is the weighting coefficient between the adjusted probability distribution and the optimized intermediate recognition result, used to balance the influence of the two.

[0055] Secondly, embodiments of this application provide a deep learning-based electrocardiogram (ECG) data anomaly identification device, comprising:

[0056] The acquisition and processing module is used to acquire the patient's raw electrocardiogram (ECG) signal and perform a multi-level preprocessing procedure to obtain the target ECG signal. The multi-level preprocessing procedure includes at least: using a bandpass filter to remove high-frequency noise and power frequency interference from the raw ECG signal; applying wavelet transform or adaptive filtering techniques to suppress electromyographic artifacts and low-frequency noise from the raw ECG signal; and using a differential operational amplifier to enhance the raw ECG signal.

[0057] The training module is used to train a deep fusion learning model using the target electrocardiogram signal. The deep fusion learning model combines a convolutional neural network and a recurrent neural network. 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. At the same time, an attention mechanism is introduced to enable the deep fusion learning model to focus on the key parts indicating abnormal signals, thereby improving the sensitivity to abnormal information identification.

[0058] The 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 a pre-trained machine learning model, perform hybrid model fusion processing on the preliminary identification result through a stacked generalization fusion method, and obtain the inference result of the preliminary identification result through a Bayesian network constructed from the patient's detailed information to generate the target identification result;

[0059] The generation and display module is used to generate an anomaly report based on the target recognition result using a rule engine. The anomaly report includes the anomaly type, location, and severity of the target electrocardiogram signal. The anomaly report is classified and stored according to predefined standards and displayed through a customized visualization interface.

[0060] Thirdly, embodiments of this application provide a computing device, including a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are to be invoked and executed by the processing component to implement a deep learning-based electrocardiogram data anomaly identification method as described in any of the first aspects.

[0061] Fourthly, embodiments of this application provide a computer storage medium storing a computer program, which, when executed by a computer, implements a deep learning-based method for identifying abnormal electrocardiogram data as described in any of the first aspects.

[0062] In this embodiment, the patient's original electrocardiogram (ECG) signal is acquired and a multi-level preprocessing process is performed to obtain a target ECG signal. The multi-level preprocessing process includes at least: using a bandpass filter to remove high-frequency noise and power frequency interference from the original ECG signal; applying wavelet transform or adaptive filtering techniques to suppress electromyographic artifacts and low-frequency noise from the original ECG signal; using a differential operational amplifier to enhance the original ECG signal; and training a deep fusion learning model using the target ECG signal. The deep fusion learning model combines a convolutional neural network (CNN) and a recurrent neural network (RNN). The CNN is used to extract local features of the target ECG signal, and the RNN is used to capture the time-series features of the target ECG signal. Simultaneously, an attention mechanism is introduced to enable the deep fusion learning model to focus on abnormal signals in the target ECG signal. Abnormal signals in the target electrocardiogram signal are identified using the deep fusion learning model to obtain preliminary identification results. These preliminary identification results are then combined with a pre-trained machine learning model, and a hybrid model fusion process is performed using a stacked generalization fusion method. Furthermore, a Bayesian network constructed from the patient's detailed information is used to obtain inference results from the preliminary identification results, thereby generating a target identification result. Based on the target identification result, an anomaly report is generated using a rule engine. This report includes the type, location, and severity of the abnormality in the target electrocardiogram signal. The anomaly report is categorized and stored according to predefined standards and displayed through a customized visualization interface.

[0063] The embodiments of this application have the following beneficial effects:

[0064] This application removes high-frequency noise and power frequency interference using a bandpass filter, suppresses electromyography artifacts and low-frequency noise using wavelet transform or adaptive filtering techniques, and enhances the original electrocardiogram signal using a differential operational amplifier, ensuring the signal quality input to the subsequent model and improving the accuracy of anomaly detection. By combining convolutional neural networks (CNNs) and recurrent neural networks (RNNs) and introducing an attention mechanism, the model can not only extract local features but also capture time-series features and focus on potential abnormal signals, thus significantly improving the accuracy of anomaly detection. The preliminary identification results are combined with a pre-trained machine learning model, processed using a stacked generalization fusion method, and a Bayesian network is constructed using detailed patient information to generate context-aware inference results, making the final identification results more personalized and reflecting the patient's specific situation and historical data. Based on the target identification results, a rule engine is used to generate an anomaly report containing the anomaly type, location, and severity, and the reports are classified and stored according to predefined standards, ensuring the comprehensiveness and interpretability of the diagnostic results. A customized visual interface is used to display the anomaly reports, facilitating doctors' quick understanding of the diagnostic results, supporting clinical decision-making, and improving work efficiency. By integrating multi-source information (such as patient medical records, historical electrocardiogram records, and clinical symptoms), the robustness and adaptability of the model are enhanced, enabling it to maintain high detection performance under different conditions. The context-aware inference results provided by the Bayesian network can supplement potentially missed or uncertain parts of the intermediate identification results, further enhancing the reliability of the identification results. Real-time monitoring and anomaly detection of electrocardiogram signals can promptly identify potential cardiac problems, enabling early warning and helping to improve treatment outcomes and reduce medical costs.

[0065] In summary, this application, through the aforementioned technical means, achieves efficient and accurate identification of abnormal signals in electrocardiogram signals, provides personalized diagnostic results, supports efficient clinical applications, and possesses good robustness and adaptability, thereby significantly improving the utilization efficiency of medical resources and the quality of medical services.

[0066] These or other aspects of this application will become more apparent in the following description of the embodiments. Attached Figure Description

[0067] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0068] Figure 1A flowchart illustrating a deep learning-based method for identifying abnormal electrocardiogram (ECG) data, provided as an embodiment of this application;

[0069] Figure 2 A schematic diagram of a deep learning-based ECG data anomaly device is provided as an embodiment of this application;

[0070] Figure 3 This is a schematic diagram of the structure of a computing device provided in an embodiment of this application. Detailed Implementation

[0071] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0072] In some of the processes described in the specification, claims, and accompanying drawings of this application, multiple operations appearing in a specific order are included. However, it should be clearly understood that these operations may not be executed in the order they appear herein, or may be executed in parallel. The operation numbers, such as 101, 102, etc., are merely used to distinguish different operations and do not themselves represent any execution order. Furthermore, these processes may include more or fewer operations, and these operations may be executed sequentially or in parallel. It should be noted that the descriptions such as "first," "second," etc., in this document are used to distinguish different messages, devices, modules, etc., and do not represent a chronological order, nor do they limit "first" and "second" to different types.

[0073] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0074] Figure 1 A flowchart of a method for identifying abnormal electrocardiogram data based on deep learning is provided in this application embodiment, as follows: Figure 1 As shown, the method includes:

[0075] 101. Collect the patient's raw electrocardiogram signal and perform a multi-level preprocessing procedure to obtain the target electrocardiogram signal.

[0076] In this step, after acquiring the patient's raw electrocardiogram (ECG) signal, a multi-stage preprocessing procedure is performed to obtain a high-quality target ECG signal. This multi-stage preprocessing includes using a bandpass filter to remove high-frequency noise and power line interference, applying wavelet transform or adaptive filtering techniques to suppress electromyographic artifacts and low-frequency noise, and using a differential operational amplifier to enhance the raw ECG signal. These steps ensure that the signal input to the subsequent analysis model has a higher signal-to-noise ratio and reduce the impact of external factors on signal quality.

[0077] In this embodiment, firstly, a bandpass filter is used to set an appropriate frequency range to remove high-frequency noise exceeding the ECG signal frequency band and common 50Hz or 60Hz power frequency interference. Next, wavelet transform or adaptive filtering techniques are used to effectively suppress electromyographic artifacts caused by muscle activity and low-frequency noise from physiological activities such as respiration. Finally, a differential operational amplifier is used to enhance the signal amplitude, appropriately amplifying the weak ECG signal, thereby improving the accuracy of subsequent feature extraction and anomaly detection.

[0078] In one specific embodiment, the patient wears a portable ECG monitoring device to continuously record ECG data for 24 hours. The device's built-in preprocessing module first limits the signal to between 0.5-150Hz using a bandpass filter, effectively removing high-frequency interference from the environment and noise from power line frequencies. Then, wavelet transform technology is used to process the data segment by segment, eliminating electromyographic artifacts and low-frequency drift. The preprocessed data is then transmitted to a cloud server for further analysis. The application of a differential operational amplifier ensures good resolution even under weak signal conditions, providing high-quality input for subsequent deep learning model training.

[0079] 102. A deep fusion learning model is trained using the target electrocardiogram signal, wherein the deep fusion learning model combines convolutional neural networks and recurrent neural networks.

[0080] The convolutional neural network is used to extract local features of the target electrocardiogram signal, the recurrent neural network is used to capture the 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.

[0081] In this step, a deep fusion learning model combining convolutional neural networks (CNNs) and recurrent neural networks (RNNs) is trained using the target electrocardiogram (ECG) signal. The CNNs are used to extract local morphological features, such as changes in the QRS complex, P wave, and T wave, as well as heart rate variability; the RNNs capture time-series features across multiple time windows to determine patterns of cardiac rhythm changes and long-term dependencies. Simultaneously, an attention mechanism is introduced, enabling the model to focus on key features most likely indicating abnormalities, improving the sensitivity and specificity of abnormal signal detection.

[0082] In this embodiment, during training, the convolutional layers automatically learn and extract local morphological features of the electrocardiogram (ECG) signal, while the recurrent layers are responsible for capturing dynamic changes between different time points. By introducing an attention mechanism, the model can identify and highlight key features when processing each time window, assigning them higher weights to better locate abnormal signals. Through supervised learning using a large amount of labeled ECG data, the model gradually optimizes its parameters, ultimately achieving high-precision anomaly detection capabilities.

[0083] Based on the preprocessed 24-hour ECG data, researchers used a pre-designed deep fusion learning model architecture that integrated multiple convolutional layers and recurrent layers composed of Long Short-Term Memory (LSTM) units. In collaboration with clinical experts, thousands of meticulously labeled ECG samples were collected for model training. During training, an attention mechanism helped the model focus on waveform segments that best reflected potential abnormalities. As training progressed, the model learned to identify various common cardiac abnormalities, such as atrial fibrillation, premature ventricular contractions, and conduction blocks, laying a solid foundation for subsequent hybrid model fusion processing.

[0084] 103. Identify abnormal signals in the target electrocardiogram signal through the deep fusion learning model to obtain preliminary identification results; combine the preliminary identification results with a pre-trained machine learning model, perform hybrid model fusion processing on the preliminary identification results through a stacked generalization fusion method, and obtain the inference results of the preliminary identification results through a Bayesian network constructed from the patient's detailed information to generate target identification results.

[0085] In this step, a pre-trained deep fusion learning model is used to identify abnormal signals in the target electrocardiogram signal, obtaining a preliminary identification result. This preliminary result is then combined with a pre-trained machine learning model, undergoing hybrid model fusion processing using a stacked generalization fusion method. A Bayesian network is constructed using detailed patient information to generate context-aware inference results, ultimately producing the target identification result. This method not only improves the accuracy of identification but also enhances the personalization and reliability of the results.

[0086] In this embodiment, the preliminary identification result output by the deep fusion learning model serves as the first layer of feature input, while the prediction result of the machine learning model optimized based on enhanced feature representation serves as the second layer of feature input. Intermediate identification results are generated through stacked generalization techniques. Subsequently, a Bayesian network is constructed based on the patient's medical records, historical electrocardiogram records, and clinical symptoms to generate context-aware inference results for the preliminary identification. Finally, the probability distribution of the intermediate identification results is adjusted by incorporating the patient's individual differences and historical cardiac data to generate more accurate target identification results.

[0087] After model training is complete, the system begins processing newly acquired electrocardiogram (ECG) data in real time. The deep fusion learning model quickly identifies potential abnormal signals and their confidence levels, and passes these preliminary results to the hybrid model fusion module. This module comprehensively considers various feature representations, such as statistical features, frequency domain features, time domain features, and nonlinear features, to optimize the prediction results. Next, the system constructs a Bayesian network based on the patient's specific circumstances (such as age, gender, and medical history) to generate personalized inference results. For example, for a patient with a history of heart disease, the system pays special attention to abnormal patterns similar to historical records, thereby improving the accuracy and reliability of the diagnosis.

[0088] 104. Based on the target identification results, generate an anomaly report using a rule engine. The anomaly report includes the anomaly type, location, and severity of the target electrocardiogram signal. The anomaly report is classified and stored according to predefined standards and displayed through a customized visualization interface.

[0089] In this step, based on the generated target identification results, a rule engine is used to generate anomaly reports containing the anomaly type, location, and severity, and these reports are categorized and stored according to predefined criteria. Furthermore, a customized visual interface is used to display the anomaly reports, allowing doctors to intuitively view and understand the diagnostic results, supporting clinical decision-making. This approach not only improves work efficiency but also ensures the transparency and interpretability of diagnostic results.

[0090] In this embodiment, once the target identification result is generated, the rule engine automatically generates a detailed anomaly report based on preset logical rules. The report covers the specific type of the abnormal signal, its location, and indicators for assessing its severity. Subsequently, the system categorizes and archives the report according to predefined standards for easy retrieval and analysis. To facilitate use by doctors, the system also provides a customized visual interface, allowing users to interactively browse electrocardiogram data, abnormal markers, and related annotations, thereby accelerating the diagnostic process.

[0091] In practical applications, when the system detects an abnormal signal, it immediately triggers the rule engine to generate a detailed anomaly report. This report not only includes the type, location, and severity of the abnormal signal, but also includes relevant references and suggested follow-up examinations. All generated reports are categorized and stored according to the hospital's internal coding system, ensuring the standardization and consistency of data management. Simultaneously, doctors can view the latest ECG data in real time through a customized visualization interface, with clearly marked abnormal areas, significantly shortening the time cycle from data collection to clinical decision-making and improving the quality and efficiency of medical services.

[0092] In steps 101-104 of this application, multi-level preprocessing ensures the quality of the input signal. A deep fusion learning model combining convolutional neural networks and recurrent neural networks achieves efficient and accurate anomaly detection. Hybrid model fusion processing and Bayesian network inference enhance the personalization and reliability of the recognition results. Finally, a rule engine generates anomaly reports and provides a customized visualization interface, significantly improving the utilization efficiency and service quality of medical resources. 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, realizing an integrated solution from data acquisition to diagnostic result presentation.

[0093] To address the issues of insufficient capture of detailed features and inadequate handling of time-series dependencies in abnormal signal detection, in some embodiments, step 103, which involves identifying abnormal signals in the target electrocardiogram signal using the deep fusion learning model to obtain preliminary identification results, includes:

[0094] The target electrocardiogram (ECG) signal is segmented into multiple fixed-length time windows, each representing a segment of ECG data within a given time period. All ECG data segments corresponding to the time windows can form the target ECG signal. The ECG data segments within each time window are fed into a convolutional neural network (CNN) in the deep fusion learning model to extract local morphological features of the ECG data within each time window, obtaining a local feature representation characterizing local cardiac activity. This local feature representation includes at least the morphological changes of the QRS complex, P wave, and T wave, as well as heart rate variability. The ECG data segments within each time window are then fed into a recurrent neural network (RNN) in the deep fusion learning model to capture time-series features between multiple time windows. Based on these time-series features, the model determines the pattern of cardiac rhythm changes and long-term dependencies, obtaining a time-series feature representation. An attention mechanism is introduced into the deep fusion learning model to allow the model to process the ECG data segments within each time window. The system identifies and highlights key features most likely indicating abnormalities, generating weighted key feature representations that reflect the importance of each key feature for abnormal signal detection. The local feature representations, time-series feature representations, and key feature representations are integrated into a comprehensive feature representation, which is then used as input to train the deep fusion learning model, enabling it to predict abnormality types and confidence levels. A multi-task learning framework is applied to the trained deep fusion learning model to predict the abnormality type and confidence level of ECG data segments within each time window. Abnormality types include atrial fibrillation, premature ventricular contractions, and conduction block. Based on the abnormality type and confidence level of ECG data segments within each time window, abnormal signals with confidence levels exceeding a preset threshold are identified, and their abnormality type, location, and severity are labeled. Based on the abnormality type and confidence level of ECG data segments across all time windows, the patient's cardiac health status score is assessed. Preliminary identification results are generated based on the corresponding abnormal signals within all time windows and the patient's cardiac health status score.

[0095] In this scheme, time window segmentation refers to dividing continuous electrocardiogram (ECG) data into multiple fixed-length time segments, with each time window representing ECG data within a specific time period. This not only helps reduce computational burden but also ensures that the model can focus on local features within each time period.

[0096] Electrocardiogram (ECG) data segments refer to data segments extracted from raw ECG signals that correspond to a specific time window. These segments contain information about cardiac activity within that time period and form the basis for subsequent feature extraction.

[0097] Convolutional Neural Networks (CNNs) are a deep learning architecture primarily used for feature extraction from image and sequence data. In electrocardiogram (ECG) analysis, CNNs are used to extract local morphological features of ECG data within each time window, such as morphological changes in the QRS complex, P wave, and T wave, as well as heart rate variability.

[0098] Local morphological features refer to the characteristic manifestations of cardiac activity over a short period of time, such as morphological changes in the QRS complex, P wave, and T wave, as well as heart rate variability. These features reflect the heart's activity patterns over a specific time period and are crucial for identifying abnormal signals.

[0099] Recurrent Neural Networks (RNNs) are a type of neural network suitable for processing sequential data, particularly well-suited for capturing time-series features. In electrocardiogram (ECG) analysis, RNNs are used to capture patterns of cardiac rhythm changes and long-term dependencies across multiple time windows.

[0100] Time series characteristics refer to the dynamic changes in cardiac activity over a period of time, including patterns of change in heart rhythm and long-term dependencies. These characteristics help to understand the overall trends in cardiac activity and improve the accuracy of abnormality detection.

[0101] Attention mechanisms are a technique used to enhance a model’s ability to identify key features. They enable the model to identify and highlight the key features most likely to indicate anomalies and assign them higher weights when processing data in each time window.

[0102] Key feature representation refers to those electrocardiogram features that the attention mechanism identifies as most likely to indicate abnormalities. These features are weighted to reflect their importance in the detection of abnormal signals.

[0103] A comprehensive feature representation refers to a feature set that integrates local feature representations, time-series feature representations, and key feature representations. This comprehensive feature representation is used as input to train a deep fusion learning model, enabling it to predict anomaly types and their confidence levels.

[0104] Multi-task learning frameworks are machine learning methods that allow models to perform multiple related tasks simultaneously. In electrocardiogram (ECG) anomaly detection, multi-task learning frameworks are used to predict the anomaly type and confidence level of ECG data segments within each time window.

[0105] The cardiac health status score is a quantitative indicator that assesses a patient's overall cardiac health based on the types and confidence levels of abnormalities in electrocardiogram (ECG) data segments across all time windows. This score integrates the severity and distribution of abnormal signals, providing a comprehensive assessment of cardiac health.

[0106] In this embodiment, the target electrocardiogram (ECG) signal is first segmented into time windows. Then, the ECG data segments within each time window are fed into a convolutional neural network (CNN) and a recurrent neural network (RNN) to extract local morphological features and capture time-series features, respectively. Simultaneously, an attention mechanism is introduced to enhance the model's ability to identify key features. Finally, the three feature representations are integrated to train the model to predict abnormality types and their confidence levels. A multi-task learning framework is then applied to identify high-confidence abnormal signals, assess the patient's cardiac health status score, and generate preliminary identification results.

[0107] Here is a specific example:

[0108] In a real-world application scenario, suppose we have a 24-hour electrocardiogram (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 within each time window are processed separately to ensure that the model can focus on the detailed features within each short time interval.

[0109] Next, the data segments from each time window are fed into a convolutional neural network (CNN). The CNN automatically learns and extracts important local morphological features, such as the morphological changes of the QRS complex, P wave, and T wave, as well as heart rate variability. These features reflect the heart's activity patterns in each time period and are the basis for identifying abnormal signals.

[0110] Simultaneously, data segments within the same time window are also fed into a recurrent neural network (RNN). The RNN captures the dynamic changes across multiple time windows, determining patterns of heart rhythm changes and long-term dependencies. This step helps the model understand the overall trend of cardiac activity and improves the accuracy of anomaly detection.

[0111] To further improve the model's performance, an attention mechanism is introduced, enabling the model to identify and highlight key features most likely indicating anomalies when processing data within each time window. For example, if significant heart rate variability or abnormal QRS wave morphology occurs within a time window, the attention mechanism assigns higher weights to these features, thus making the model pay more attention to these potential anomalies.

[0112] Finally, the system integrates local feature representations, time-series feature representations, and key feature representations into a comprehensive feature representation, and uses these comprehensive features to train a deep fusion learning model. After training, the model can predict the type of abnormality and its confidence level within each time window. By applying a multi-task learning framework, the model identifies abnormal signals with confidence levels higher than a preset threshold and labels their abnormality type, location, and severity. In addition, based on the abnormality type and confidence level across all time windows, the system assesses the patient's cardiac health status score and generates preliminary identification results including abnormality type, location, and severity.

[0113] This method allows doctors to obtain a detailed report of abnormalities, supporting them in making more accurate diagnostic decisions. This process not only improves the accuracy of abnormality detection but also provides personalized diagnostic information, significantly enhancing the quality and efficiency of healthcare services.

[0114] To address the issues of multi-task processing and confidence assessment in abnormal signal detection, the aforementioned optional solution, specifically "applying a multi-task learning framework to the trained deep fusion learning model to predict the abnormality type and confidence level of ECG data segments within each time window, and identifying abnormal signals with a confidence level higher than a preset threshold based on the abnormality type and confidence level of the ECG data segments within each time window," includes:

[0115] Using 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. The probability distribution corresponding to each abnormality type is used to represent the possibility of the existence of the abnormality type within each time window. The confidence level of each abnormality type is evaluated, and a confidence score is generated. The confidence score refers to the reliability score of the prediction result of the deep fusion learning model for each abnormality type. Time windows with confidence scores higher than a preset threshold are selected, abnormal signals within the time windows are identified, and the time windows identified as abnormal are marked. 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.

[0116] In this scheme, a multi-task learning framework—a machine learning method—allows the model to perform multiple related tasks simultaneously. In electrocardiogram (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 model's learning efficiency and generalization ability.

[0117] Abnormality type refers to different types of abnormalities that may appear in electrocardiogram (ECG) signals, such as atrial fibrillation, premature ventricular contractions (PVCs), and conduction blocks. Each abnormality type corresponds to a specific pattern of cardiac activity and is an important basis for diagnosing heart disease.

[0118] A probability distribution represents the likelihood of a certain type of anomaly occurring within each time window. The probability distribution predicted by the model reflects the probability of that anomaly occurring and forms the basis for subsequent confidence assessments.

[0119] Confidence score refers to the reliability score of a deep fusion learning model's predictions for each anomaly type. A higher confidence score indicates that the model's predictions for that anomaly type are more reliable. Metrics such as cross-entropy loss function are typically used to quantify confidence.

[0120] The preset threshold is a set numerical standard used to filter out time windows with high confidence scores. Only when the confidence score of an anomaly type within a certain time window exceeds this threshold will it be considered a reliable anomaly signal.

[0121] Location refers to the time window corresponding to the abnormal signal. This includes not only the specific position of the time window in the overall electrocardiogram recording (such as the start and end times), but also detailed time point information within that window.

[0122] In this embodiment, the multi-task learning framework in the trained deep fusion learning model is used to first identify the abnormality types contained in the electrocardiogram data segments within each time window and obtain the probability distribution corresponding to each abnormality type. Then, the confidence level of each abnormality type is evaluated to generate a confidence score. Next, the system filters out time windows with confidence scores higher than a preset threshold, identifies abnormal signals within these time windows, marks the time windows identified as abnormal, and records the abnormality type, location, and severity of the abnormal signals.

[0123] The process of identifying anomalous signals within these time windows includes: for each time window, if the confidence score of a certain anomalous type exceeds a preset threshold, then the corresponding anomalous signal is considered to exist within that time window.

[0124] Here is a specific example:

[0125] In a practical application scenario, suppose we have a pre-trained deep fusion learning model that can identify various types of cardiac abnormalities, such as atrial fibrillation, premature ventricular contractions, and conduction blocks. To improve detection accuracy, the system employs a multi-task learning framework, enabling the model to simultaneously predict the probability distributions of multiple abnormality types.

[0126] For each time window (e.g., a 5-second ECG data segment), the model outputs a vector where each element represents the probability of a certain type of abnormality. For example, if three types of abnormalities are considered, the output might be [0.2, 0.7, 0.1], representing the probabilities of atrial fibrillation, premature ventricular contractions, and conduction block, respectively.

[0127] Next, the system calculates a confidence score for each anomaly type based on these probability distributions. The confidence score reflects the model's predictive reliability for each anomaly type. For example, for the output vector above, the system might generate a confidence score of [0.6, 0.9, 0.5], indicating that the model is most confident in predicting premature ventricular contractions (PVCs).

[0128] To ensure that only high-confidence results are reported, the system sets a preset threshold (e.g., 0.8). The system only considers an abnormal signal reliable when the confidence score for a particular type of abnormality exceeds this threshold. In this example, only the confidence score for premature ventricular contractions (PVCs) exceeds the threshold, so the system marks this time window as abnormal and records its abnormality type (PVCs), location (specific time period), and severity (indirectly reflected by the confidence score).

[0129] Finally, all time windows marked as abnormal are compiled into a detailed anomaly report. Doctors can view the specific information of these abnormal signals through a customized visualization interface, enabling them to make more accurate diagnostic decisions. This method not only improves the accuracy of anomaly detection but also provides personalized diagnostic information, significantly enhancing the quality and efficiency of healthcare services.

[0130] To address the limitations of potential omissions or uncertainties in preliminary identification results and further improve the accuracy and personalization of abnormal signal detection, optionally, step 103 involves combining the preliminary identification results with a pre-trained machine learning model. This is achieved by performing a hybrid model fusion process on the preliminary identification results using a stacked generalization fusion method, and obtaining the inference results from the preliminary identification results through a Bayesian network constructed from the patient's detailed information to generate target identification results. This includes:

[0131] A machine learning model is trained based on a comprehensive feature representation and an enhanced feature representation additionally extracted from the target electrocardiogram (ECG) signal. This enables the machine learning model to predict whether each time window contains an abnormal signal and the corresponding abnormality type and confidence level. The enhanced feature representation includes statistical features, frequency domain features, time domain features, and nonlinear features additionally extracted from the target ECG signal. Using a stacked generalization technique, the preliminary identification 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 intermediate identification results. Based on the acquired detailed patient information, a Bayesian network is constructed, and the Bayesian network generates context-aware inference results for the preliminary identification. These inference results provide additional contextual information for the detection of abnormal signals. The detailed information includes medical records, historical ECG records, and clinical symptoms. Based on the inference results, combined with the acquired individual differences and historical cardiac data of the patient, the intermediate identification results are processed to adjust the probability distribution corresponding to the intermediate identification results, thereby generating the target identification result.

[0132] In this scheme, the comprehensive feature representation is a feature set that integrates local feature representation, time-series feature representation, and key feature representation. These features reflect multifaceted information in the electrocardiogram data and form the basis for training machine learning models.

[0133] Enhanced feature representation refers to the additional statistical, frequency domain, time domain, and nonlinear features extracted from the target electrocardiogram (ECG) signal. These features provide richer information, helping machine learning models to understand ECG data more comprehensively.

[0134] Machine learning models are algorithms based on statistical principles used to predict whether each time window contains anomalous signals and their corresponding anomalous types and confidence levels. In this embodiment, the model is trained using comprehensive feature representations and enhanced feature representations, thereby improving its ability to capture complex patterns.

[0135] Stacking generalization is an ensemble learning method that improves overall performance by combining the outputs of different models. In this embodiment, the preliminary recognition results of the deep fusion learning model are used as the first layer of feature input, and the prediction results of the machine learning model are used as the second layer of feature input, thereby generating more reliable intermediate recognition results.

[0136] A Bayesian network is a probabilistic graphical model used to represent conditional dependencies between variables. In this scenario, the Bayesian network is built based on detailed patient information (such as medical records, historical electrocardiogram records, and clinical symptoms), providing context-aware inference results for the initial identification, thus increasing the personalization of the diagnosis.

[0137] The inference result refers to the context-aware information generated by a Bayesian network for the initial identification. This inference result not only considers the current electrocardiogram data but also incorporates individual patient differences and historical cardiac data, improving the accuracy of abnormal signal detection.

[0138] The target recognition result refers to the final abnormal signal detection result, which has undergone multi-level analysis and optimization to ensure high accuracy and reliability.

[0139] In this embodiment, 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. This model is able to predict whether each time window contains an abnormal signal and its corresponding abnormality type and confidence level. Then, using a stacked generalization technique, the preliminary identification 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 intermediate identification results. Next, a Bayesian network is constructed based on the patient's detailed information, and this network generates context-aware inference results, providing additional contextual information for the detection of abnormal signals. Finally, based on the inference results, combined with the patient's individual differences and historical cardiac data, the intermediate identification results are processed to adjust their probability distribution, generating the final target identification result.

[0140] Here is a specific example:

[0141] In a practical application scenario, suppose we have a pre-trained deep fusion learning model capable of identifying various types of cardiac abnormalities. To further improve detection accuracy, the system introduces an additional machine learning model. This model is trained using comprehensive feature representations (such as local morphological features, time-series features, and key features) as well as enhanced feature representations (such as statistical features, frequency domain features, time domain features, and nonlinear features). These enhanced features provide more dimensions of information, helping the model better understand and classify complex electrocardiogram (ECG) data.

[0142] For each time window (e.g., a 5-second ECG data segment), the deep fusion learning model first outputs preliminary identification results, including possible abnormality types and their confidence scores. This preliminary result is then used as the first layer of feature input for a stacked generalization technique. The machine learning model then predicts whether each time window contains an abnormal signal and its corresponding abnormality type and confidence score based on the combined and enhanced feature representations; these predictions serve as the second layer of feature input. Through this two-layer input structure, the system generates more accurate intermediate identification results.

[0143] Subsequently, the system constructs a Bayesian network based on the patient's detailed information (such as medical records, historical electrocardiogram (ECG) records, and clinical symptoms). This network not only considers the current ECG data but also incorporates the patient's historical health records and personal characteristics to generate context-aware inference results. For example, if the patient has a history of heart disease, the Bayesian network will pay special attention to abnormal patterns similar to historical records, thereby improving the accuracy and personalization of the diagnosis.

[0144] Finally, based on the inference results generated by the Bayesian network, and combined with the patient's individual differences and historical cardiac data, the system further processed the intermediate identification results. This step adjusted the probability distribution of the intermediate identification results, making the final target identification 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 correspondingly increase its confidence score and mark that time window as abnormal.

[0145] This method allows doctors to obtain a detailed report of abnormalities, supporting them in making more accurate diagnostic decisions. This process not only improves the accuracy of abnormality detection but also provides personalized diagnostic information, significantly enhancing the quality and efficiency of healthcare services. Furthermore, 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.

[0146] To address the limitations of potential omissions or uncertainties in preliminary identification results and further improve the accuracy and personalization of abnormal signal detection, the aforementioned optional solution's "based on the inference results, combined with the acquired individual differences and historical cardiac data of the patient, performing inference processing on the intermediate identification results to adjust the probability distribution corresponding to the intermediate identification results and generate target identification results" specifically includes:

[0147] Based on the individual differences and historical cardiac data of the patients, the probability distribution of abnormal signals in each time window of the intermediate identification results is adjusted to obtain an adjusted probability distribution; based on the inference results, the possible omissions or uncertain parts in the intermediate identification results are supplemented to enhance the reliability of the intermediate identification results, resulting in an optimized intermediate identification result; the adjusted concept distribution and the optimized intermediate identification results are integrated to generate a target identification result.

[0148] In this approach, individual differences refer to the differences in physiological characteristics between different patients, such as age, gender, and medical history. These differences affect the presentation of electrocardiogram signals and therefore need to be taken into special consideration when detecting abnormalities.

[0149] Historical cardiac data includes a patient's previous electrocardiogram (ECG) records, treatment responses, and other relevant medical information. This data provides additional contextual information, helping to more accurately assess current ECG signals.

[0150] A probability distribution represents the likelihood of a certain type of anomaly occurring within each time window. The probability distribution predicted by the model reflects the probability of that anomaly occurring and forms the basis for subsequent confidence assessments.

[0151] The adjusted probability distribution refers to the probability distribution adjusted for individual patient differences and historical cardiac data. This adjustment makes anomaly detection more personalized and improves diagnostic accuracy.

[0152] The inference results are generated by a Bayesian network using context-aware information based on the patient's detailed information. This inference not only considers the current electrocardiogram data but also incorporates the patient's individual differences and historical cardiac data, providing a more comprehensive diagnostic basis.

[0153] The optimized intermediate identification result refers to the intermediate identification result that has been supplemented with possible omissions or uncertainties through reasoning, thereby enhancing its reliability. This step ensures the accuracy and completeness of the final identification result.

[0154] In this embodiment, firstly, based on the acquired individual differences and historical cardiac data of the patient, the probability distribution of abnormal signals within each time window of the intermediate identification results is adjusted to obtain an adjusted probability distribution. Then, based on the inference results, any potentially omitted or uncertain parts of the intermediate identification results are supplemented to enhance their reliability, resulting in an optimized intermediate identification result. Finally, the adjusted probability distribution and the optimized intermediate identification results are integrated to generate the final target identification result.

[0155] Here is a specific example:

[0156] In a practical application scenario, suppose we have a pre-trained system capable of identifying various types of cardiac abnormalities and generating preliminary intermediate identification results. To further improve detection accuracy, the system incorporates personalized adjustment and optimization steps.

[0157] The system first acquires the patient's individual differences (such as age, gender, and medical history) and historical cardiac data (such as previous electrocardiogram records and treatment responses). This information is used to adjust the probability distribution of abnormal signals within each time window in the intermediate identification results.

[0158] For example, if a patient has a history of atrial fibrillation, the system will increase the probability score for that abnormality type across all time windows. For a young patient with no history of heart disease, the system may decrease the probability score for certain common abnormalities, thus focusing more on less common abnormal patterns.

[0159] The adjusted probability distribution better reflects the actual situation of patients and improves the personalization of abnormality detection.

[0160] Next, the system uses the inference results generated by the Bayesian network to supplement any potentially missing or uncertain parts of the intermediate identification results. The Bayesian network is constructed based on detailed patient information (such as medical records, historical electrocardiogram records, and clinical symptoms) to generate context-aware inference results.

[0161] For example, if an abnormal signal has a low confidence level within a certain time window, 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.

[0162] This supplement enhances the reliability and completeness of intermediate identification results, reducing the possibility of misdiagnosis and missed diagnosis.

[0163] Finally, the system integrates the adjusted probability distribution and optimized intermediate identification results to generate the final target identification result. This result not only includes the abnormality type, location, and severity within each time window, but also incorporates individual patient differences and historical cardiac data, providing more comprehensive and reliable diagnostic information. Doctors can view these detailed abnormality reports through a customized visualization interface, supporting them in making more accurate diagnostic decisions.

[0164] This method allows doctors to obtain a detailed report of abnormalities, supporting them in making more accurate diagnostic decisions. This process not only improves the accuracy of abnormality detection but also provides personalized diagnostic information, significantly enhancing the quality and efficiency of healthcare services. Furthermore, 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.

[0165] To address the potential issue of insufficient time-series consistency in the initial identification 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 further includes: optimizing the time-series consistency of the target identification results based on the long-term trends and short-term fluctuations in the patient's historical cardiac data.

[0166] In this approach, long-term trends refer to changes in cardiac activity patterns over extended periods within a patient's historical cardiac data. These trends reflect the evolution of the patient's cardiac health and are crucial for understanding the context of the current electrocardiogram signal.

[0167] Short-term fluctuations refer to changes in cardiac activity over a relatively short period of time (such as hours to days). These fluctuations may be caused by specific events (such as exercise or emotional changes) or may indicate the onset of acute cardiac problems.

[0168] Time series consistency refers to the coherence and logic of target identification results over time. Optimizing time series consistency means ensuring that detected anomalous signals not only conform to the characteristics of the current time window, but also match the long-term trends and short-term fluctuations in historical data, thereby providing more reliable diagnostic information.

[0169] In this embodiment, the system not only adjusts the probability distribution of intermediate identification results based on individual patient differences and historical cardiac data, but also optimizes the temporal consistency of target identification results according to the long-term trends and short-term fluctuations in the patient's historical cardiac data. This step ensures 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, providing more reliable and coherent diagnostic information.

[0170] Here is a specific example:

[0171] In a practical application scenario, suppose we have a pre-trained system capable of identifying various types of cardiac abnormalities and generating preliminary target identification results. To further improve detection accuracy and consistency, the system introduces a time-series consistency optimization step based on historical cardiac data.

[0172] The system first analyzes long-term trends in the patient's historical cardiac data. For example, if the patient has a long history of atrial fibrillation, the system will check whether the current electrocardiogram data shows a similar abnormal pattern.

[0173] Long-term trend analysis helps the system understand the evolution of a patient's cardiac health status, ensuring that detected abnormal signals are not isolated events, but consistent with historical records.

[0174] 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) in the past few days, the system will pay special attention to abnormal signals during these time periods.

[0175] Short-term volatility assessment helps the system detect potential acute problems and ensures that these signals are correctly flagged and reported.

[0176] Finally, the system optimizes the time-series consistency of target identification results based on long-term trends and short-term fluctuations. This step ensures that detected anomalous signals not only conform to the characteristics of the current time window but also match the long-term trends and short-term fluctuations in historical data. For example, if an anomalous signal within a certain time window has a low confidence level, but the Bayesian network infers that the signal is highly correlated with the patient's medical history, the system will correspondingly increase its confidence score and mark that time window as anomalous. Simultaneously, the system also checks whether the anomalous signal matches the long-term trends and short-term fluctuations in historical data to ensure time-series consistency.

[0177] After optimization, the system generates a final diagnostic report. This report not only includes the type, location, and severity of the abnormality within each time window, but also integrates the patient's individual differences, historical cardiac data, and the results of time series consistency optimization.

[0178] Doctors can view these detailed anomaly reports through a customized visual interface, enabling them to make more accurate diagnostic decisions. Furthermore, the system can automatically categorize and store all generated reports according to the hospital's internal coding system, ensuring the standardization and consistency of data management.

[0179] This application addresses the fact that, in order to more comprehensively and accurately assess a patient's cardiac health status, traditional methods rely solely on the detection results of abnormality types and confidence scores, neglecting the influence of local features, time-series features, and key features. Furthermore, existing methods fail to adequately consider individual patient differences and the long-term trends and short-term fluctuations in historical cardiac data. Therefore, this application proposes a novel cardiac health status scoring formula. This formula is based not only on the abnormality type and its corresponding confidence score of ECG data segments within each time window, but also incorporates local feature representations, time-series feature representations, and key feature representations, while simultaneously considering the influence of individual patient differences and additional features. In this way, the system can generate a more comprehensive and reliable cardiac health status score, supporting physicians in making more accurate diagnostic decisions.

[0180] The plan includes:

[0181] Based on the abnormality type and corresponding confidence score of the electrocardiogram data segment within each time window, and combined with the local feature representation, time series feature representation, and key feature representation within that time window, the patient's cardiac health status score is assessed; wherein, the cardiac health status score is calculated using the following formula:

[0182] ;

[0183] in, It is a heart health status score. It is the number of time windows. It is the index of the time window; It is the number of exception types. It is an index of an exception type; It is the first Within the first time window The confidence score for each anomaly type refers to the reliability score of the prediction result of the deep fusion learning model for each anomaly type. It is an index factor for the confidence score, used to adjust the degree of influence of the confidence score, and is set... ,in It is the maximum value among all confidence scores; It is the first The weights of different abnormality types reflect the importance of different abnormality types to cardiac health status; It is the first The influence factor of local feature representation within a time window is calculated using the standardized score of the local feature representation. It is the first The influence factors represented by time series features within a time window are calculated using the standardized scores of the time series feature representations. It is the first The influence factor of key feature representation within a time window is calculated by the standardized score of key feature representation; These are the weighting coefficients for local feature representation, time series feature representation, and key feature representation, respectively, used to adjust the influence of each feature representation, and determined according to the importance and contribution of each feature; These are the index factors for local feature representation, time series feature representation, and key feature representation, respectively, used to adjust the degree of influence of each feature representation, and determined according to the sensitivity and importance of each feature;

[0184] It is a Gaussian kernel function, used to smooth the influence of confidence scores, where It is the average of all confidence scores. It is the standard deviation of the confidence score;

[0185] It is the first The comprehensive impact factor for each time window is defined as:

[0186] ;

[0187] in, This refers to the number of additional features, which include at least the patient's age, gender, and medical history. It is the first The first time window and the first The relevance score of the additional features; It is the first The weights of additional features are used to adjust the degree of influence of these additional features.

[0188] The parameters involved in the above formulas are explained in detail below:

[0189] Cardiac Health Status Score. This score comprehensively reflects the patient's overall cardiac health and is a weighted sum of abnormality types and their characteristic representations across all time windows.

[0190] Number of time windows. Each time window represents a fixed-length segment of ECG data, used for segment-by-segment analysis of the ECG signal.

[0191] The index of the time window. Used to identify a specific time window.

[0192] The number of abnormal types. Each abnormal type corresponds to a specific pattern of cardiac activity, such as atrial fibrillation, premature rolls, etc.

[0193] : Index of the exception type. Used to identify a specific exception type.

[0194] : No. Within the first time window Confidence scores for each anomaly type. These scores are obtained through predictions using a deep fusion learning model and reflect the model's predictive reliability for each anomaly type.

[0195] : The index factor for confidence scores. Set to This is used to adjust the influence of confidence scores, enhancing the importance of high-confidence scores. Maximum confidence score. It is the maximum value among all confidence scores.

[0196] : No. The weights of different abnormality types are determined based on their importance to cardiac health, typically grounded in clinical experience and statistical data.

[0197] : No. The influence factors are represented by local features within a time window. The scores are extracted and standardized using a convolutional neural network to reflect the impact of local morphological features (such as changes in QRS complexes, P waves, and subsequent waves) on cardiac health within that time window.

[0198] : No. The influence factors are represented by time series features within a time window. A recurrent neural network is used to capture and standardize scores to calculate the impact of cardiac rhythm variation patterns and long-term dependencies across multiple time windows on cardiac health.

[0199] : No. Influence factors representing key features within a time window are identified. An attention mechanism is used to identify and standardize score calculations, reflecting the impact of the key features most likely indicating abnormalities on cardiac health.

[0200] These are the weighting coefficients for local feature representation, time series feature representation, and key feature representation, respectively. They are determined based on the importance and contribution of each feature and are used to balance the influence of different feature representations.

[0201] These are the exponential factors for local feature representation, time series feature representation, and key feature representation, respectively. They are determined based on the sensitivity and importance of each feature and are used to further adjust the influence of each feature representation.

[0202] The Gaussian kernel function is used to smooth the influence of confidence scores. It is the average of all confidence scores. This is the standard deviation of the confidence score. This function ensures that fluctuations in the confidence score within a certain range do not produce excessively large differences in impact.

[0203] : No. The comprehensive impact factor of each time window. Defined as:

[0204] ;

[0205] in It refers to the number of additional characteristics, including at least the patient's age, gender, and medical history; It is the first The first time window and the first The relevance score of an additional feature, calculated through statistical analysis or machine learning methods, reflects the strength of the association between the data and the specific additional feature within that time window. For example, this association can be quantified using correlation coefficients, mutual information, or other suitable statistical measures. It is the first The weighting of additional features. This factor takes into account the influence of individual patient differences and historical cardiac data.

[0206] The following is a brief explanation of the rationale behind the design of each term in the above formula:

[0207] This sub-item combines the weights of anomaly types with the influence factors of the three feature representations. Anomaly type weights This reflects the importance of different abnormality types, while the influence factors representing the features assess the characteristics of cardiac activity from different perspectives. These factors are multiplied by their respective weighting coefficients. and exponential factors This ensures the reasonable allocation of different feature representations in the scoring.

[0208] This section uses a Gaussian kernel function to smooth the influence of confidence scores. This helps reduce the excessive influence of extreme confidence scores and ensures stability when scores fluctuate within a certain range.

[0209] This sub-item uses an index factor. Adjust for the influence of confidence scores. Index factor. The design enhances the importance of high confidence scores, thereby better highlighting those more reliable anomaly detection results.

[0210] This sub-item takes into account individual patient differences and the influence of historical cardiac data. By introducing additional features (such as age, gender, and medical history), the scoring becomes more personalized and comprehensive.

[0211] The above formula, by multiplying the confidence score, the influence factor of the feature representation, the Gaussian kernel function, and the smoothing factor, ensures that abnormal signals within each time window are comprehensively evaluated. The multiplication operation centrally reflects the specific impact of each factor on the cardiac health status score, while preventing any single factor from dominating the score result. By summing the scores from all time windows, comprehensive information from the entire electrocardiogram recording can be summarized. The addition operation ensures the overall consistency and logic of the scoring, reflecting a holistic view of cardiac health status.

[0212] The overall design of this formula aims to provide a comprehensive, accurate, and personalized assessment method for evaluating a patient's cardiac health. It considers not only the type of abnormality detected and confidence scores but also the influence of multiple feature representations and individual patient differences. In this way, the system can generate a more reliable cardiac health score, supporting physicians in making more accurate diagnostic decisions. Furthermore, the formula's design takes into account the consistency and coherence of time series data, reducing the possibility of misdiagnosis and missed diagnosis, and significantly improving the quality and efficiency of medical services. .

[0213] Here is a specific example:

[0214] 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 electrocardiogram (ECG) data and divides it into multiple fixed-length time windows (e.g., each window is 5 seconds long). We will use the formula described above to calculate the patient's cardiac health status score. .

[0215] First, set the parameters:

[0216] Number of time windows (12 windows per minute, for a total of 24 hours).

[0217] Number of exception types (Atrial fibrillation, premature ventricular contractions, conduction block).

[0218] Confidence score Assume the confidence score output by the model is This corresponds to three exception types.

[0219] Weighting coefficient : respectively This reflects the importance of different anomaly types.

[0220] Influence factors of local feature representation Influence factors of time series feature representation Influence factors represented by key features Assume the standardized score is .

[0221] Weighting coefficient Set as follows .

[0222] Exponential factor Set as follows .

[0223] Gaussian kernel function parameters: Assume .

[0224] Comprehensive Impact Factor Additional feature relevance score and weight Assumption They are respectively They are respectively , 0.2.

[0225] Secondly, calculate the index factor of the confidence score. :

[0226] ;

[0227] Assuming the maximum confidence score is 0.8, then .

[0228] Further calculate the influence factors of local feature representation, time series feature representation, and key feature representation:

[0229] ;

[0230] Further application of the Gaussian kernel function:

[0231] ;

[0232] Further calculation of the comprehensive impact factor :

[0233] ;

[0234] Finally, substitute the above values ​​into the calculation to determine the score:

[0235] ;

[0236] Assuming that the final cardiac health status score is calculated after combining all window and abnormality types, The score was 85 out of 100. This score indicates that the patient's heart health is relatively good, but there is a certain risk of abnormality.

[0237] Specifically, a higher score indicates that abnormal signals were reliably detected most of the time window, and that these abnormal signals were of lower severity. The comprehensive influencing factors take into account additional characteristics such as the patient's age, gender, and medical history, further reflecting the impact of individual patient differences on cardiac health. The score also considers long-term trends and short-term fluctuations in historical cardiac data, ensuring the time-series consistency of the diagnostic results. Physicians can use this detailed score report, combined with other clinical information, to develop more personalized treatment plans or preventative measures. This not only improves diagnostic accuracy but also provides a comprehensive assessment of cardiac health.

[0238] This application addresses the potential for omissions or uncertainties in preliminary identification of electrocardiogram (ECG) abnormalities, particularly when faced with individual differences and complex historical cardiac data. Traditional single models often fail to adequately capture the influence of these factors, leading to inaccurate and unpersonalized diagnostic results. To overcome these issues, a method is needed that comprehensively considers individual patient differences, historical cardiac data, and context-aware reasoning to improve the accuracy and reliability of abnormal signal detection.

[0239] Therefore, this application proposes a novel formula scheme aimed at optimizing intermediate identification results by adjusting the probability distribution of abnormal signals within each time window and combining the inference results provided by a Bayesian network, ultimately generating more accurate target identification results. This scheme not only improves detection accuracy but also provides personalized diagnostic information, significantly enhancing the quality and efficiency of medical services. The scheme includes:

[0240] The step of adjusting the probability distribution of abnormal signals within each time window of the intermediate identification results based on the acquired individual differences and historical cardiac data of the patient to obtain the adjusted probability distribution includes:

[0241] The adjusted probability distribution is calculated using the following formula:

[0242] ;

[0243] in, It is the first Within the first time window Adjusted probability distribution of the anomaly types It is the first Within the first time window Initial probability distribution of the anomaly types; It is the first The influencing factors of individual patient differences within a time window are calculated based on the patient's age, gender, and medical history. It is the average of the individual difference influencing factors across all time windows; It is the standard deviation of the factors influencing individual differences across all time windows; It refers to the quantity of historically specific characteristics, which include specific indicators in the patient's historical cardiac data, including historical electrocardiogram records and treatment response. It is a historically specific index; It is the first The first time window and the first The relevance score of a specific historical feature; It is the first The weights of historically specific features are used to adjust the degree of influence of those features.

[0244] The parameters of the above formula are explained in detail below:

[0245] It is the first Within the first time window The adjusted probability distribution for each abnormality type. This value reflects the model's most recent assessment of the presence of a specific abnormality type after considering individual patient differences and the influence of historical cardiac data. It is derived from the initial probability distribution. Calculated using a series of adjustment factors, this ensures that the test results are more personalized and accurate.

[0246] It is the first Within the first time window The initial probability distribution of the abnormality types. This value is directly predicted by the deep fusion learning model and represents the model's preliminary assessment of the abnormal signal without considering individual patient differences and historical cardiac data. It is obtained by processing ECG data segments for each time window using a trained deep learning model and outputting the corresponding abnormality type and its confidence score.

[0247] It is the first The influencing factor of individual patient differences within a time window. This factor comprehensively considers information such as the patient's age, gender, and medical history to quantify the impact of these factors on the ECG data of the current time window. It is obtained by extracting relevant information from the patient's basic medical records (such as electronic medical records) and calculating the individual difference influencing factor for each time window using a predefined formula or algorithm.

[0248] This is the average of the individual difference impact factors across all time windows. This value is used to standardize the individual difference impact factors within each time window, making comparisons between different time windows more reasonable. It is obtained by calculating over all time windows. The mean of the values ​​is used to obtain a benchmark value that represents the overall level of individual differences.

[0249] This is the standard deviation of the individual difference influencing factors across all time windows. The standard deviation reflects the dispersion of these factors across all time windows and is used to measure data volatility. It is obtained by calculating over all time windows. The standard deviation is used to facilitate subsequent Gaussian kernel function smoothing, ensuring that the adjusted probability distribution is more stable.

[0250] This refers to the number of historically specific features, which include specific indicators from the patient's historical cardiac data, such as historical ECG recordings and treatment response. Each historically specific feature provides additional contextual information, helping to more accurately assess the current ECG data. The acquisition method involves extracting relevant historical cardiac data from the patient's electronic medical record and determining which indicators are most important for the current analysis.

[0251] It is the first The first time window and the first A correlation score for a specific historical feature. This score reflects the strength of the association between the current time window's ECG data and a particular historical feature. It is obtained by calculating the similarity or correlation between the data in each time window and the specific historical feature using statistical analysis or machine learning methods, thus providing a basis for adjusting the probability distribution.

[0252] It is the first Weights are assigned to historically specific features to adjust for their influence. These weights reflect the importance of different historically specific features in current anomaly detection and are typically determined based on clinical experience and statistical data. They are obtained through expert evaluation or data analysis, assigning appropriate weights to each historically specific feature to ensure they play a suitable role in adjusting the probability distribution.

[0253] By leveraging the combined effects of the aforementioned parameters and adjusting the probability distribution of abnormal signals within each time window, the final identification result is based not only on current ECG data but also fully considers individual patient differences and historical cardiac data, thereby improving diagnostic accuracy and personalization. This method ensures more reliable abnormal signal detection, supporting physicians in making more precise clinical decisions.

[0254] The process of supplementing the intermediate identification results with potentially omitted or uncertain parts based on the reasoning results enhances the reliability of the intermediate identification results, resulting in optimized intermediate identification results, including:

[0255] The optimized intermediate recognition result is calculated using the following formula:

[0256] ;

[0257] in, It is the first Intermediate recognition results after time window optimization It represents the number of exception types; It is the first Within the first time window Adjusted probability distribution of various anomaly types; The first in the context-aware reasoning results provided by the Bayesian network Within the first time window Confidence scores for different types of anomalies.

[0258] The parameters of the above formula are explained in detail below:

[0259] It is the first The intermediate identification result is optimized for each time window. This value comprehensively considers the adjusted probability distribution and the context-aware inference results provided by the Bayesian network, aiming to enhance the reliability and accuracy of the intermediate identification results. The result calculated by this formula can more comprehensively reflect the existence and severity of abnormal signals within each time window, providing a solid foundation for the final target identification result.

[0260] This refers to the number of abnormality types, specifically the total number of different cardiac abnormality types the system can identify. Common abnormality types might include atrial fibrillation, premature ventricular contractions (PVCs), and conduction blocks. The acquisition method is determined based on clinical needs and the abnormality types covered during model training. This parameter ensures that the formula can handle all possible abnormalities, thereby improving the comprehensiveness of the detection.

[0261] It is the first Within the first time window The adjusted probability distribution for each abnormality type. This value reflects the model's most recent 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 calculating using the formula described in the previous step, adjusting the initial probability distribution using factors influencing individual patient differences and specific historical characteristics. This ensures that test results are more personalized and accurate.

[0262] The first in the context-aware reasoning results provided by the Bayesian network Within the first time window The confidence score for each abnormality type is generated by a Bayesian network based on detailed patient information (such as medical records, historical ECG records, and clinical symptoms). This provides context-aware information, supplementing potentially missed or uncertain parts of the initial identification results. The process involves constructing and training a Bayesian network, taking detailed patient information as input, and outputting a confidence score for each abnormality type within each time window, thereby enhancing the reliability of the identification results.

[0263] Indicates the first The maximum value among the adjusted probability distributions of all anomaly types within each time window. This value is used to select the most likely anomaly type and serves as a basis for enhancing the reliability of intermediate identification results. It is obtained by comparing the adjusted probability distributions of all anomaly types within each time window. Then, the maximum value is selected. This step ensures that even if the probability of some anomalies is low, the most likely scenario is prioritized.

[0264] The above formula The purpose is to optimize the first The intermediate identification results within each time window are used to make them more reliable and accurate. Specifically:

[0265] Maximum probability choice: Pass Select the most likely anomaly type to ensure that no important anomaly signals are missed.

[0266] Bayesian Network Supplement: Through Adding context-aware inference results provided by Bayesian networks supplements the parts that may have been missed or uncertain in the initial identification results, thereby enhancing the reliability of the identification results.

[0267] This method not only improves the accuracy of anomaly detection but also provides personalized diagnostic information, significantly enhancing the quality and efficiency of healthcare services. By combining the adjusted probability distribution with the inference results of Bayesian networks, doctors can obtain more comprehensive and reliable anomaly reports, supporting them in making more accurate diagnostic decisions.

[0268] The process of integrating the adjusted probability distribution and the optimized intermediate recognition results to generate target recognition results includes:

[0269] The target recognition result is calculated using the following formula:

[0270] ;

[0271] in, It is the first Target recognition results within a time window It is the first The result corresponding to the maximum adjusted probability distribution within each time window; It is the first Optimized intermediate recognition results for each time window; It is the weighting coefficient between the adjusted probability distribution and the optimized intermediate recognition result, used to balance the influence of the two.

[0272] The parameters of the above formula are explained in detail below:

[0273] It is the first The target identification results for each time window. This value comprehensively considers the adjusted probability distribution and optimized intermediate identification results, aiming to generate the final abnormal signal detection result. The result calculated by this formula can more comprehensively reflect the presence and severity of abnormal signals within each time window, providing doctors with the most accurate and reliable diagnostic information.

[0274] It is the first The result corresponds to the maximum adjusted probability distribution within each time window. This value reflects the most probable outcome in the model's most recent assessment of the presence of a specific abnormality type, after considering individual patient differences and the influence of historical cardiac data. It is obtained by comparing the... Adjusted probability distribution of all anomaly types within a time window Select the maximum value among them. This step ensures that even if the probability of certain anomalies is low, the most likely scenario is prioritized.

[0275] It is the first The optimized intermediate recognition result for each time window. This value comprehensively considers the adjusted probability distribution and the context-aware inference results provided by the Bayesian network, enhancing the reliability and accuracy of the intermediate recognition result. It is obtained through the formula in the previous step. The calculations ensured that the identification results were based not only on current ECG data, but also on individual patient differences and historical cardiac data, thereby improving the accuracy and personalization of the diagnosis.

[0276] This is a weighting coefficient between the adjusted probability distribution and the optimized intermediate recognition results, used to balance the influence of the two. This coefficient determines the adjusted probability distribution. and optimized intermediate recognition results In the final target recognition result The relative importance of each component. This is determined through expert evaluation or experimental verification to establish appropriate weight values. Typically, The range of values ​​is within The specific values ​​can be adjusted according to the needs of different application scenarios. For example, if more emphasis is placed on the model's initial predictions, a larger value can be selected. Value; conversely, if more emphasis is placed on the context-aware reasoning results provided by the Bayesian network, a smaller value can be selected. value.

[0277] The above formula The purpose is to integrate the first The final target recognition result is generated by combining the adjusted probability distribution of each time window and the optimized intermediate recognition results. Specifically:

[0278] Adjusted probability distribution: via By introducing an adjusted probability distribution, we ensure that the model's initial assessment of anomalous signals still accounts for a certain proportion.

[0279] Optimized intermediate recognition results: Through The optimized intermediate recognition results are introduced to supplement the context-aware inference results provided by the Bayesian network, thereby enhancing the reliability of the recognition results.

[0280] Weighting 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.

[0281] This method not only improves the accuracy of anomaly detection but also provides personalized diagnostic information, significantly enhancing the quality and efficiency of healthcare services. By combining the adjusted probability distribution with optimized intermediate identification results, doctors can obtain more comprehensive and reliable anomaly reports, supporting them in making more accurate diagnostic decisions. The final target identification result... This provides a solid foundation for clinical applications, ensuring the scientific rigor and reliability of the diagnosis.

[0282] Here is a specific example of the complete solution described above:

[0283] Let's assume we have a specific application scenario involving a 50-year-old male patient with a history of hypertension. The system collected 24 hours of electrocardiogram (ECG) data and divided it into multiple fixed-length time windows (e.g., each window is 5 seconds long). We will use the formula described above to calculate the patient's cardiac health score and interpret the results.

[0284] First, set the parameters:

[0285] Number of time windows (12 windows per minute, for a total of 24 hours);

[0286] Number of exception types (Atrial fibrillation, premature ventricular contractions, conduction block);

[0287] Initial probability distribution Assume the initial probability distribution of the model output is as follows: This corresponds to three exception types.

[0288] Influencing factors of individual differences Assuming the calculation is based on the patient's age, gender, and medical history. They are respectively .

[0289] Average value of individual difference influencing factors ;

[0290] Standard deviation of factors affecting individual differences ;

[0291] Number of historical specific features (Historical electrocardiogram records, treatment response, medication use);

[0292] Correlation score Assume they are respectively ;

[0293] Weight of historical specific features : respectively ;

[0294] Confidence scores provided by Bayesian networks Assume they are respectively ;

[0295] Weighting coefficient ;

[0296] Next, calculate the adjusted probability distribution:

[0297] a. Gaussian kernel function:

[0298] ;

[0299] For example, for The calculation result is:

[0300] ;

[0301] b. The influence of specific historical characteristics:

[0302] ;

[0303] c. Adjusted probability distribution:

[0304] ;

[0305] Assumption ,but:

[0306] ;

[0307] Further calculations were performed on the optimized intermediate recognition results:

[0308] a. Maximum adjusted probability distribution:

[0309] ;

[0310] b. Bayesian Network Supplement:

[0311] ;

[0312] c. Optimized intermediate recognition results:

[0313] ;

[0314] Further calculation of target recognition results:

[0315] a. The value corresponding to the maximum adjusted probability distribution:

[0316] ;

[0317] b. Target recognition results:

[0318] ;

[0319] The final calculated target recognition result The score is 1.12. This score indicates that, after considering individual patient differences and the influence of historical cardiac data, the system provides a more accurate assessment of the presence and severity of abnormal signals within this time window. Specifically: in ECG abnormality detection, the target identification result... The numerical value itself is a comprehensive score, reflecting the first... The presence and severity of abnormal signals within a specific time window. To draw specific conclusions from this value, we need to interpret it in conjunction with clinical thresholds, historical data, and the professional judgment of physicians.

[0320] The following is a detailed explanation of the steps and conclusions:

[0321] First, based on the established clinical thresholds, we assume the normal range is... The warning area is The abnormal range is .because Falling within the warning range indicates that there may be underlying heart problems within that time window, but they have not yet reached a serious level.

[0322] Secondly, if the patient has a history of atrial fibrillation and has repeatedly shown similar high scores recently, special attention is needed. If this is the first time a high score has appeared, it is recommended to further monitor data from other time periods to confirm whether there is a persistent problem.

[0323] Finally, the doctor can... A comprehensive assessment should be conducted based on the values ​​and other clinical information (such as blood pressure, heart rate, etc.). The conclusion may be: there is a mild to moderate risk of cardiac abnormalities within this time window, and further observation and necessary examinations (such as Holter monitoring, echocardiography, etc.) are recommended.

[0324] Based on the above analysis, the target recognition results This indicates that there may be an underlying heart problem within this time window, but it has not yet reached a serious stage. Specifically:

[0325] Mild to moderate abnormality: This score falls within the warning range, indicating a certain risk of abnormality.

[0326] Further observation: It is recommended to continue monitoring data from other time periods to confirm whether there is a persistent problem.

[0327] Personalized diagnosis: By combining individual patient differences with historical cardiac data, doctors can develop more personalized diagnostic and treatment plans.

[0328] By introducing an adjusted probability distribution, optimized intermediate identification results, and a final target identification result, this scheme effectively enhances the accuracy and personalization of ECG abnormality detection. This method ensures more reliable abnormal signal detection, supporting physicians in making more accurate clinical decisions. The final target identification result provides a solid foundation for clinical applications, ensuring the scientific rigor and reliability of the diagnosis.

[0329] Figure 2 This application provides a schematic diagram of the structure of a deep learning-based ECG data anomaly device, as shown in the embodiment of this application. Figure 2 As shown, the device includes:

[0330] The acquisition and processing module 21 is used to acquire the patient's raw electrocardiogram (ECG) signal and perform a multi-level preprocessing procedure to obtain the target ECG signal. The multi-level preprocessing procedure includes at least: using a bandpass filter to remove high-frequency noise and power frequency interference from the raw ECG signal; applying wavelet transform or adaptive filtering techniques to suppress electromyographic artifacts and low-frequency noise from the raw ECG signal; and using a differential operational amplifier to enhance the raw ECG signal.

[0331] Training module 22 is used to train a deep fusion learning model using the target electrocardiogram signal. The deep fusion learning model combines a convolutional neural network and a recurrent neural network. 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. At the same time, by introducing an attention mechanism, the deep fusion learning model can focus on the key parts indicating abnormal signals to improve the sensitivity of abnormal information identification.

[0332] The identification output module 23 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 a pre-trained machine learning model, perform hybrid model fusion processing on the preliminary identification result through a stacked generalization fusion method, and obtain the inference result of the preliminary identification result through a Bayesian network constructed from the patient's detailed information to generate the target identification result;

[0333] The generation and display module 24 is used to generate an abnormal report based on the target recognition result using a rule engine. The abnormal report includes the abnormal type, location, and severity of the target electrocardiogram signal. The abnormal report is classified and stored according to predefined standards and displayed through a customized visualization interface.

[0334] Figure 2 The aforementioned deep learning-based ECG data anomaly identification device can perform... Figure 1 The implementation principle and technical effects of the deep learning-based ECG data anomaly identification method described in the illustrated embodiment will not be repeated here. The specific operation methods of each module and unit in the deep learning-based ECG data anomaly identification device described in the above embodiments have been described in detail in the embodiments related to this method, and will not be elaborated upon here.

[0335] In one possible design, Figure 2 The deep learning-based ECG data anomaly identification device of 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;

[0336] The storage component 31 stores one or more computer instructions, wherein the one or more computer instructions are invoked and executed by the processing component 32.

[0337] The processing component 32 is used to: acquire the patient's original electrocardiogram (ECG) signal and perform a multi-level preprocessing procedure to obtain a target ECG signal. The multi-level preprocessing procedure includes at least: using a bandpass filter to remove high-frequency noise and power frequency interference from the original ECG signal; applying wavelet transform or adaptive filtering techniques to suppress electromyographic artifacts and low-frequency noise from the original ECG signal; using a differential operational amplifier to enhance the original ECG signal; and training a deep fusion learning model using the target ECG signal. The deep fusion learning model combines a convolutional neural network (CNN) and a recurrent neural network (RNN). The CNN is used to extract local features of the target ECG signal, and the RNN is used to capture the time-series features of the target ECG signal. An attention mechanism is introduced to enable the deep fusion learning model to focus on abnormal signals in the target ECG signal. Abnormal signals in the target electrocardiogram signal are identified using the deep fusion learning model to obtain preliminary identification results. These preliminary identification results are then combined with a pre-trained machine learning model, and a hybrid model fusion process is performed using a stacked generalization fusion method. Furthermore, a Bayesian network constructed from the patient's detailed information is used to obtain inference results from the preliminary identification results, thereby generating a target identification result. Based on the target identification result, an anomaly report is generated using a rule engine. This report includes the type, location, and severity of the abnormality in the target electrocardiogram signal. The anomaly report is categorized and stored according to predefined standards and displayed through a customized visualization interface.

[0338] 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-described method. Alternatively, the processing component may be implemented as 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-described method.

[0339] 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 storage, flash memory, magnetic disk, or optical disk.

[0340] Of course, computing devices may also include other components, such as input / output interfaces, display components, communication components, etc.

[0341] Input / output interfaces provide interfaces between processing components and peripheral interface modules, which can be output devices, input devices, etc.

[0342] The communication components are configured to facilitate wired or wireless communication between computing devices and other devices.

[0343] The computing device can be a physical device or an elastic computing host provided by a cloud computing platform. In this case, the computing device can refer to a cloud server, and the aforementioned processing components, storage components, etc., can be basic server resources rented or purchased from the cloud computing platform.

[0344] This application also provides a computer storage medium storing a computer program, which, when executed by a computer, can perform the above-described functions. Figure 1 The embodiment shown is a method for detecting abnormal electrocardiogram data based on deep learning.

[0345] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0346] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0347] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0348] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for identifying abnormalities in electrocardiogram (ECG) data based on deep learning, characterized in that, include: The patient's raw electrocardiogram (ECG) signal is acquired and a multi-level preprocessing procedure is performed to obtain the target ECG signal; A deep fusion learning model is trained using the target electrocardiogram signal, which combines convolutional neural networks and recurrent neural networks; the deep fusion learning model is then used to identify abnormal signals in the target electrocardiogram signal to obtain preliminary identification results. The preliminary identification results are combined with a pre-trained machine learning model, and the preliminary identification results are fused using a stacked generalization fusion method. The inference results of the preliminary identification results are obtained through a Bayesian network constructed from the patient's detailed information to generate the target identification results. Based on the target identification results, an anomaly report is generated using a rule engine, the anomaly report is classified and stored according to predefined standards, and the anomaly report is displayed through a customized visualization interface; The process of obtaining the inference results from the preliminary identification results to generate target identification results includes: adjusting the probability distribution of abnormal signals within each time window of the intermediate identification results based on the individual differences and historical cardiac data of the patient, to obtain an adjusted probability distribution. The intermediate identification results are obtained by using the preliminary identification results output by the deep fusion learning model as the first layer of feature input and the prediction results of the machine learning model as the second layer of feature input, utilizing a stacked generalization technique. Based on the inference results, the process supplements potentially omitted or uncertain parts of the intermediate identification results to enhance their reliability, resulting in optimized intermediate identification results. Finally, the adjusted concept distribution and the optimized intermediate identification results are integrated to generate the target identification results. The adjusted probability distribution is calculated using the following formula: ； in, It is the first Within the first time window Adjusted probability distribution of the anomaly types It is the first Within the first time window Initial probability distribution of the anomaly types; It is the first The influencing factors of individual patient differences within a time window are calculated based on the patient's age, gender, and medical history. It is the average of the individual difference influencing factors across all time windows; It is the standard deviation of the factors influencing individual differences across all time windows; It refers to the quantity of historically specific characteristics, which include specific indicators in the patient's historical cardiac data, including historical electrocardiogram records and treatment response. It is a historically specific index; It is the first The first time window and the first The relevance score of a specific historical feature; It is the first The weights of specific historical features are used to adjust the degree of influence of those features. The optimized intermediate recognition result is calculated using the following formula: ； in, It is the first Intermediate recognition results after time window optimization It represents the number of exception types; It is the first Within the first time window Adjusted probability distribution of various anomaly types; The first in the context-aware reasoning results provided by the Bayesian network Within the first time window Confidence scores for different anomaly types; The target recognition result is calculated using the following formula: ； in, It is the first Target recognition results within a time window It is the first The result corresponding to the maximum adjusted probability distribution within each time window; It is the first Optimized intermediate recognition results for each time window; It is the weighting coefficient between the adjusted probability distribution and the optimized intermediate recognition result, used to balance the influence of the two.

2. The method according to claim 1, characterized in that, The step of identifying abnormal signals in the target electrocardiogram signal through the deep fusion learning model to obtain preliminary identification results includes: The target electrocardiogram signal is segmented into time windows to obtain multiple time windows of fixed length. Each time window represents an electrocardiogram data segment within a time period. All the electrocardiogram data segments corresponding to the time windows can form the target electrocardiogram signal. The electrocardiogram (ECG) data segments within each time window are fed into the convolutional neural network in the deep fusion learning model to extract local morphological features of the ECG data within each time window through the convolutional neural network, thereby obtaining a local feature representation that characterizes the local properties of cardiac activity. The local feature representation includes at least the morphological changes of the QRS complex, P wave, and T wave, as well as heart rate variability. Electrocardiogram data segments within each time window are fed into the recurrent neural network in the deep fusion learning model to capture time series features between multiple time windows through the recurrent neural network, and to determine the change pattern and long-term dependency of heart rhythm based on the time series features, thereby obtaining a time series feature representation. An attention mechanism is introduced into the deep fusion learning model so that when processing ECG data segments in each time window, the deep fusion learning model can identify and highlight the key features most likely to indicate abnormalities from the ECG data segments and generate key feature representations with weights, the weights being used to reflect the importance of each key feature for abnormal signal detection. The local feature representation, the time series feature representation, and the key feature representation are integrated into a comprehensive feature representation, and the comprehensive feature representation is used as input to train the deep fusion learning model so that the deep fusion learning model can predict anomaly types and confidence levels. A multi-task learning framework is applied in the trained deep fusion learning model to predict the abnormality type and confidence level of ECG data segments within each time window. The abnormality types include atrial fibrillation, premature ventricular contractions, and conduction block. Based on the abnormality type and confidence level of ECG data segments within each time window, abnormal signals with confidence levels higher than a preset threshold are identified, and the abnormality type, location, and severity of the abnormal signals are marked. The patient's cardiac health status score was assessed based on the abnormality type and confidence level of all electrocardiogram data segments within all time windows. Preliminary identification results are generated based on the abnormal signals corresponding to all time windows and the patient's cardiac health status score.

3. The method according to claim 2, characterized in that, A multi-task learning framework is applied to the trained deep fusion learning model to predict the abnormality type and confidence level of ECG data segments within each time window. Based on the abnormality type and confidence level of the ECG data segments within each time window, abnormal signals with a confidence level higher than a preset threshold are identified, including: Using 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. The probability distribution corresponding to each abnormality type is used to represent the possibility of the abnormality type existing within each time window. The confidence level of each anomaly type is evaluated to generate a confidence score, which refers to the reliability score of the prediction results of the deep fusion learning model for each anomaly type. The system filters out time windows with confidence scores higher than a preset threshold, identifies abnormal signals within the time windows, marks the time windows identified as abnormal, and records the abnormality 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 process of combining the preliminary identification results with a pre-trained machine learning model, performing hybrid model fusion processing on the preliminary identification results using a stacked generalization fusion method, and obtaining the inference results of the preliminary identification results through a Bayesian network constructed from the patient's detailed information to generate target identification results includes: 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 abnormality 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 intermediate recognition results. Based on the detailed information of the patient obtained, a Bayesian network is constructed, and a context-aware inference result is generated through the Bayesian network for the preliminary identification, so as to provide additional contextual information for the detection of abnormal signals through the inference result. The detailed information includes medical records, historical electrocardiogram records and clinical symptoms. Based on the reasoning results, and combined with the acquired individual differences and historical cardiac data of the patients, the intermediate identification results are processed to adjust the probability distribution corresponding to the intermediate identification results and generate target identification results.

5. The method according to claim 4, characterized in that, The method further includes: The temporal 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 patient's cardiac health status score is assessed based on the abnormality type and confidence level of electrocardiogram data segments across all time windows, including: Based on the abnormality type and corresponding confidence score of the electrocardiogram data segment within each time window, and combined with the local feature representation, time series feature representation, and key feature representation within that time window, the patient's cardiac health status score is assessed; wherein, the cardiac health status score is calculated using the following formula: ； in, It is a heart health status score. It is the number of time windows. It is the index of the time window; It is the number of exception types. It is an index of an exception type; It is the first Within the first time window The confidence score for each anomaly type refers to the reliability score of the prediction result of the deep fusion learning model for each anomaly type. It is an index factor for the confidence score, used to adjust the degree of influence of the confidence score, and is set... ,in It is the maximum value among all confidence scores; It is the first The weights of different abnormality types reflect the importance of different abnormality types to cardiac health status; It is the first The influence factor of local feature representation within a time window is calculated using the standardized score of the local feature representation. It is the first The influence factors represented by time series features within a time window are calculated using the standardized scores of the time series feature representations. It is the first The influence factor of key feature representation within a time window is calculated by the standardized score of key feature representation; These are the weighting coefficients for local feature representation, time series feature representation, and key feature representation, respectively, used to adjust the influence of each feature representation, and determined according to the importance and contribution of each feature; These are the index factors for local feature representation, time series feature representation, and key feature representation, respectively, used to adjust the degree of influence of each feature representation, and determined according to the sensitivity and importance of each feature; It is a Gaussian kernel function, used to smooth the influence of confidence scores, where It is the average of all confidence scores. It is the standard deviation of the confidence score; It is the first The comprehensive impact factor for each time window is defined as: ； in, This refers to the number of additional features, which include at least the patient's age, gender, and medical history. It is the first The first time window and the first The relevance score of the additional features; It is the first The weights of additional features are used to adjust the degree of influence of these additional features.

7. A deep learning-based electrocardiogram (ECG) data anomaly identification device, used to execute the deep learning-based ECG data anomaly identification method according to any one of claims 1-6, characterized in that, include: The acquisition and processing module is used to acquire the patient's raw electrocardiogram (ECG) signal and perform a multi-level preprocessing procedure to obtain the target ECG signal; The training module is used to train a deep fusion learning model using target electrocardiogram signals, the deep fusion learning model combining convolutional neural networks and recurrent neural networks; The 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 a pre-trained machine learning model, perform hybrid model fusion processing on the preliminary identification result through a stacked generalization fusion method, and obtain the inference result of the preliminary identification result through a Bayesian network constructed from the patient's detailed information to generate the target identification result; The generation and display module is used to generate anomaly reports based on the target recognition results using a rule engine, classify and store the anomaly reports according to predefined standards, and display the anomaly reports through a customized visualization interface.

8. A computing device, characterized in that, It includes a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are invoked and executed by the processing component to implement the deep learning-based ECG data anomaly identification method as described in any one of claims 1 to 6.

9. A computer storage medium, characterized in that, The device contains a computer program that, when executed by a computer, implements a deep learning-based method for identifying abnormal electrocardiogram data as described in any one of claims 1 to 6.

Images