Quality control feedback method and system for electrocardiogram data and diagnosis result

Abstract

The invention discloses a quality control feedback method and system for electrocardio data and diagnosis results, and the quality control feedback method for the electrocardio data and the diagnosis results comprises three steps. The method has remarkable advantages in the aspect of automatically feeding back and updating the electrocardiogram database, and due to the fact that an existing system mostly depends on manual database maintenance, the updating period is long, response is lagged, newest clinical diagnosis characteristics are difficult to reflect in time, and an effective quality control mechanism is lacked in the data storage process. According to the method, dynamic self-optimization and updating of the database can be realized. According to the method, text analysis and key content extraction are performed on a clinical diagnosis result of an electrocardiogram by using a deep learning model, and similarity matching is performed by combining electrocardiogram characteristics of an object with existing data in a database. And if the matching degree is lower than a set threshold value, the new diagnosis and the features thereof are automatically put in storage, so that real-time learning and continuous expansion are realized. The local database has the adaptive learning capability, and the intelligence and data timeliness of the system are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical data analysis, and in particular to a quality control feedback method and system for electrocardiogram data and diagnosis results. Background Art

[0002] In modern healthcare, a resting electrocardiogram (ECG) is a common type of electrocardiogram (ECG). It is a noninvasive examination method that records the heart's electrical activity while the patient is at rest. This is typically performed by placing electrodes on the patient's chest, limbs, and other areas to capture changes in the heart's electrical activity, thereby providing information about cardiac function. As a convenient, rapid, and noninvasive examination tool, the ECG is extremely important in clinical practice. It not only helps diagnose common heart conditions such as arrhythmias, myocardial ischemia, and myocardial infarction, but also assesses cardiac conduction function, monitors drug side effects, screens for underlying heart disease, and guides treatment. It is an indispensable tool in the diagnosis and treatment of heart disease and a crucial component of regular checkups and health management. However, traditional ECG analysis relies primarily on manual interpretation by medical staff. Manual ECG analysis often relies on the physician's experience and judgment. However, each physician is limited by their knowledge and experience, and diagnostic criteria may vary, resulting in highly subjective results. This can lead to misinterpretation and misdiagnosis or missed diagnosis of complex pathologies. Physicians' knowledge and experience can also affect their analytical skills, especially when faced with rare or atypical ECG changes. Experienced doctors can make accurate diagnoses, but beginners or inexperienced physicians may not be able to fully identify complex heart problems. ECG analysis methods and the latest research findings in cardiology can take a long time to be updated in doctors' practices, and manual analysis cannot integrate the latest diagnostic standards or guidelines in real time. Although existing technologies have made some progress in ECG analysis and disease diagnosis, they still face many challenges in practical application. A major challenge is the lack of timely updates to knowledge graphs. Many knowledge graphs use cumbersome update mechanisms that require manual intervention, resulting in delayed updates.

[0003] Currently, the traditional ECG diagnostic process relies primarily on physicians' manual interpretation and accumulated experience. Although some medical institutions have established information management systems for ECG data storage and query, these systems still have many deficiencies in practical application. For example, ECG data and clinical diagnostic information are often stored in a decentralized manner, lacking systematic integration. This leads to severe information silos, limiting in-depth data mining and intelligent applications. Existing databases mostly focus on raw signals and lack structured extraction and annotation of key electrophysiological features such as P waves, QRS complexes, T waves, and ST segments, making them difficult to support accurate auxiliary diagnosis. Furthermore, ECG databases are static and updated late, making it difficult to dynamically optimize them based on new clinical diagnostic results. This results in weak model generalization and makes it difficult for intelligent systems to adapt to the local needs of different hospitals. Medical research and clinical practice often advance at a rapid pace. If knowledge graphs are not updated promptly, they will not reflect the latest medical advances. While most existing systems can apply knowledge graphs to diagnostic recommendations, their infrequent update frequency makes it difficult to reflect the latest clinical data or research findings in real time. In the dynamic medical field, especially with emerging diseases and changes in treatments, knowledge graph updates need to be more timely and flexible. Many systems rely on manual review or periodic updates, and such update mechanisms often lag behind changes in actual clinical data.

[0004] Therefore, in view of the shortcomings of the existing technology, it is necessary to provide a quality control feedback method and system for ECG data and diagnostic results to solve the shortcomings of the existing technology. Summary of the Invention

[0005] The first object of the present invention is to overcome the shortcomings of the prior art and provide a quality control feedback method for ECG data and diagnostic results. The quality control feedback method for ECG data and diagnostic results can automatically complete the analysis of ECG signals and obtain a complete feature description text.

[0006] The above-mentioned purpose of the present invention is achieved through the following technical measures:

[0007] A quality control feedback method for ECG data and diagnostic results is provided, which is carried out in the following steps:

[0008] S1. Acquire multiple sets of historical data, each set of historical data is electrocardiogram (ECG) data and corresponding diagnostic information;

[0009] S2. Build a local database using the ECG data and diagnostic information from S1, where the local database includes the ECG, ECG data, and diagnostic information;

[0010] S3 obtains real-time diagnostic information and the corresponding electrocardiogram, performs quality control feedback on the real-time diagnostic information through the trained deep learning model, and updates S2's local database.

[0011] Preferably, the above S3 is performed by the following steps:

[0012] S3.1. Obtain real-time diagnostic information and the corresponding electrocardiogram, and extract the clinical diagnostic information through the trained deep learning model to obtain a diagnostic statement consisting of multiple keywords;

[0013] S3.2. Perform similarity matching on the diagnostic statement extracted in S3.1 and the diagnostic information in the local database. When the matching result is less than a set threshold, save the clinical diagnostic information and the clinical electrocardiogram to the local database, thereby updating the local database.

[0014] Preferably, the above-mentioned deep learning model is a Chinese named entity recognition model based on BERT layer, BiLSTM layer and CRF layer.

[0015] Preferably, the above-mentioned Chinese named entity recognition model is provided with an Inpu_Text layer, a BERT layer, two BiLSTM layers, a LayerNorm layer, a Dropout Activation layer and a CRF layer, and the Inpu_Text layer, the BERT layer, the BiLSTM layer, the BiLSTM layer, the LayerNorm layer, the Dropout Activation layer and the CRF layer are connected in sequence from front to back.

[0016] Preferably, the above-mentioned Inpu_Text layer is used to input the original Chinese sentence of the diagnostic information;

[0017] Preferably, the BERT layer has a built-in transformer structure, and the BERT layer segments the original Chinese sentence input by the Inpu_Text layer to obtain keywords, and semantically encodes the keywords to generate a context vector for each keyword.

[0018] Preferably, the BiLSTM layer is used to extract contextual dependencies in the sequence, thereby improving the ability to understand structural information.

[0019] Preferably, the LayerNorm layer is used to normalize the features of the keywords.

[0020] Preferably, the Dropout Activation layer is used to prevent overfitting and increase the expressive power of nonlinear data.

[0021] Preferably, the CRF layer is used to combine keywords to obtain a diagnostic statement, thereby outputting the most likely label sequence.

[0022] Preferably, the above S3.2 is performed by the following steps:

[0023] S3.2.1. Using an edit distance matching algorithm, transform each character string in the diagnostic statement obtained in S3.1 into the minimum number of edit operations required for the character string in the diagnostic information in the local database, and then perform normalization to ultimately obtain a similarity.

[0024] S3.2.2. When the similarity obtained in S3.2.1 is less than the set threshold, the real-time diagnostic information and the corresponding electrocardiogram are saved in the local database, thereby updating the local database.

[0025] Preferably, the training process of the above deep learning model is carried out by the following steps:

[0026] A1. Constructing a data set, wherein the data set is a text sequence annotated in BIO format and consists of multiple groups of diagnostic statements and conversion vectors corresponding to the diagnostic statements;

[0027] A2. Input the dataset into the deep learning model and output the extraction results;

[0028] A3: Optimize the deep learning model using the negative log-likelihood as the loss function, and then determine whether the early stopping mechanism conditions are met. If not, proceed to A4; if so, proceed to A5.

[0029] A4: Determine whether the maximum number of training times has been reached. If yes, proceed to A5; if no, return to A2.

[0030] A5. Use the current deep learning model as the trained deep learning model.

[0031] Preferably, the above-mentioned local database also includes an ECG knowledge graph constructed based on ECG data and diagnostic information.

[0032] Preferably, the process of establishing the above-mentioned ECG knowledge graph is carried out by the following steps:

[0033] S2.1. Treat the diagnostic information of each ECG as a diagnostic node for the ECG, and then extract the relationship data, condition data, description data, range data, measurement and calculation data, and computer language data from the corresponding ECG data, with the relationship data constituting the relationship node, the condition data constituting the condition node, the description data constituting the description node, the range data constituting the range node, the measurement and calculation data constituting the measurement and calculation node, and the computer language data constituting the computer language node;

[0034] S2.2. Input the diagnosis nodes, relationship nodes, condition nodes, description nodes, range nodes, measurement and calculation nodes, and computer language nodes corresponding to each electrocardiogram into the Neo4j graph database, so that the diagnosis nodes, relationship nodes, condition nodes, description nodes, range nodes, measurement and calculation nodes, and computer language nodes are connected in series from front to back to generate the electrocardiogram knowledge graph.

[0035] Preferably, the above relationship nodes are unlabeled relationships or labeled parallel relationships.

[0036] When a relationship node in the electrocardiogram has no labeled relationship, all conditions in the condition node are met at the same time.

[0037] When the relationship nodes in the electrocardiogram are labeled parallel relationships, only at least one condition in the condition node needs to be satisfied.

[0038] Preferably, the above-mentioned conditional node is type information of the electrocardiogram, and there is at least one conditional node.

[0039] Preferably, the above-mentioned description node is the specific manifestation and corresponding description of the condition node on the electrocardiogram.

[0040] Preferably, the above-mentioned range node is a characteristic range of electrocardiogram data.

[0041] Preferably, the above-mentioned measurement and calculation nodes are variables of electrocardiogram data.

[0042] Preferably, the above-mentioned computer language node is a form in which the corresponding range node is converted into computer processing form.

[0043] Preferably, the above variables are amplitude variables, frequency variables, amplitude difference variables, and time limit variables.

[0044] Preferably, the above-mentioned condition nodes are electrocardiogram amplitude, waveform, morphology, and waveform duration.

[0045] The second object of the present invention is to overcome the shortcomings of the prior art and provide a quality control feedback system for ECG data and diagnostic results. The quality control feedback system for ECG data and diagnostic results can automatically analyze the ECG signal and obtain a complete feature description text.

[0046] The above-mentioned purpose of the present invention is achieved through the following technical measures:

[0047] A quality control feedback system for ECG data and diagnostic results is provided, which executes the above-mentioned quality control feedback method for ECG data and diagnostic results.

[0048] The present invention provides a method and system for quality control feedback of ECG data and diagnostic results. The method comprises the following steps: S1. Acquiring multiple sets of historical data, each set of historical data consisting of ECG data corresponding to an ECG and corresponding diagnostic information; S2. Building a local database using the ECG data and diagnostic information in S1, wherein the local database includes the ECG, ECG data, and diagnostic information; S3. Acquiring real-time diagnostic information and the corresponding ECG, performing quality control feedback on the real-time diagnostic information using a trained deep learning model, and updating the local database in S2. Compared to existing technologies, the present invention offers significant advantages in automatically providing feedback and updating ECG databases. Existing systems often rely on manual database maintenance, resulting in long update cycles and delayed responses, making it difficult to timely reflect the latest clinical diagnostic features. Furthermore, they lack effective quality control mechanisms during data storage. The present invention enables dynamic self-optimization and updating of the database. Specifically, the present invention utilizes a deep learning model to perform text parsing and key content extraction on the ECG clinical diagnostic results, and then performs similarity matching based on the subject's ECG features with existing data in the database. If the match falls below a set threshold, the new diagnosis and its characteristics are automatically stored in the database, enabling real-time learning and continuous expansion. This innovative approach, unlike traditional methods that rely on manual comparison and updates, empowers the database with adaptive learning capabilities, significantly improving the system's intelligence and data timeliness, and providing more accurate, comprehensive, and up-to-date decision support for subsequent clinical auxiliary diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The present invention is further described with reference to the accompanying drawings, but the contents in the accompanying drawings do not constitute any limitation to the present invention.

[0050] Figure 1 The present invention is a flow chart of a quality control feedback method for electrocardiogram data and diagnosis results.

[0051] Figure 2 This is a structural diagram of the deep learning model of the present invention.

[0052] Figure 3 This is the electrocardiogram knowledge graph of Example 2.

[0053] Figure 4 Partial screenshots of diagnosis nodes, relationship nodes, condition nodes, description nodes, range nodes, measurement and calculation nodes, and computer language nodes in Excel spreadsheets. DETAILED DESCRIPTION

[0054] The technical solution of the present invention is further described with reference to the following examples.

[0055] Example 1

[0056] A quality control feedback method for ECG data and diagnostic results, such as Figure 1 As shown, the following steps are performed:

[0057] S1. Acquire multiple sets of historical data, each set of historical data is electrocardiogram (ECG) data and corresponding diagnostic information;

[0058] S2, builds a local database using the ECG data and diagnostic information of S1, the local database including the ECG, ECG data and diagnostic information;

[0059] S3 obtains real-time diagnostic information and the corresponding electrocardiogram, performs quality control feedback on the real-time diagnostic information through the trained deep learning model, and updates S2's local database.

[0060] Wherein, S3 of the present invention is carried out by the following steps:

[0061] S3.1. Obtain real-time diagnostic information and the corresponding electrocardiogram, and extract the clinical diagnostic information through the trained deep learning model to obtain a diagnostic statement consisting of multiple keywords;

[0062] S3.2. Perform similarity matching on the diagnostic statement extracted in S3.1 and the diagnostic information in the local database. When the matching result is less than the set threshold, save the clinical diagnostic information and the clinical electrocardiogram to the local database, thereby updating the local database.

[0063] like Figure 2 As shown in the figure, the deep learning model of the present invention is a Chinese named entity recognition model based on a BERT layer, a BiLSTM layer, and a CRF layer. The Chinese named entity recognition model is provided with an Inpu_Text layer, a BERT layer, two BiLSTM layers, a LayerNorm layer, a Dropout Activation layer, and a CRF layer. The Inpu_Text layer, the BERT layer, the BiLSTM layer, the BiLSTM layer, the LayerNorm layer, the Dropout Activation layer, and the CRF layer are connected in sequence from front to back.

[0064] The Inpu_Text layer inputs the original Chinese sentence containing diagnostic information. The BERT layer, with its built-in transformer architecture, segments the original Chinese sentence into keywords, semantically encodes the keywords, and generates a context vector for each keyword. The BiLSTM layer extracts contextual dependencies within the sequence, thereby improving the understanding of structural information. The LayerNorm layer normalizes the keyword features. The DropoutActivation layer prevents overfitting and improves the expressiveness of nonlinear data. The CRF layer combines keywords to generate diagnostic statements, outputting the most likely label sequence.

[0065] It should be noted that normalization of the LayerNorm layer in the deep learning model of the present invention can enhance stability.

[0066] S3.2 of the present invention is carried out by the following steps:

[0067] S3.2.1. Use the edit distance matching algorithm to transform each string in the diagnostic statement obtained in S3.1 into the minimum number of edit operations required for the string in the diagnostic information in the local database, and then perform normalization to finally obtain the similarity;

[0068] S3.2.2. When the similarity obtained in S3.2.1 is less than the set threshold, the clinical diagnostic information and the clinical electrocardiogram are saved in the local database, thereby updating the local database. The set threshold of the present invention can be 0.70, 0.75, 0.80, 0.85, 0.90, etc., which can be determined according to actual conditions. This embodiment sets a threshold.

[0069] It should be noted that the Levenshtein distance refers to the minimum number of editing operations required to transform one string into another, including character insertion, deletion, and replacement. The similarity of the present invention is calculated by the following formula:

[0070]

[0071] A is the source string, which is the string that needs to be changed, and B is the target string, which is the final string to be transformed.

[0072] For example, let's take the diagnosis statement "V2 has ST-segment elevation, amplitude 0.2mV" obtained in S3.1. The local database contains the following diagnostic information: A, V2 has ST-segment elevation, amplitude 0.2mV; B, V2 has ST-segment elevation, amplitude 0.3mV; and C, V1 has ST-segment depression, amplitude 0.1mV. The minimum number of edits required to transform the diagnosis statement "V2 has ST-segment elevation, amplitude 0.2mV" into A is 0, and the similarity is 1.00. The minimum number of edits required to transform it into B is 3, and the similarity is 0.82. The minimum number of edits required to transform it into C is 7, and the similarity is 0.71.

[0073] The training process of the deep learning model is carried out in the following steps:

[0074] A1. Construct a data set. The data set is a text sequence annotated in BIO format. The data set consists of multiple sets of diagnostic statements and conversion vectors corresponding to the diagnostic statements.

[0075] A2. Input the dataset into the deep learning model and output the extraction results;

[0076] A3: Optimize the deep learning model using the negative log-likelihood as the loss function, and then determine whether the early stopping mechanism conditions are met. If not, proceed to A4; if so, proceed to A5.

[0077] A4: Determine whether the maximum number of training times has been reached. If yes, proceed to A5; if no, return to A2.

[0078] A5. Use the current deep learning model as the trained deep learning model.

[0079] The early stopping mechanism condition is determined by checking whether the accuracy and loss have changed in the last N rounds of training. If there is no change, the early stopping mechanism condition is met. The accuracy of the present invention is the ratio of the number of correctly predicted labels to the total number of labels. N of the present invention is a positive number greater than or equal to 2 and can be determined according to actual conditions. In this embodiment, N is specifically 5. The maximum number of training times of the present invention can be 100, 200, 300, etc., and the maximum number of training times of this embodiment is specifically 200.

[0080] It should be noted that, taking the diagnostic information "ST-segment elevation 0.2mV in lead V2" as an example, the BIO format data for V2 is B-Lead, the BIO format data for Lead is I-Lead, the BIO format data for ST is B-Change, the BIO format data for Segment is I-Change, the BIO format data for Elevation is I-Change, the BIO format data for 0.2 is B-Value, and the BIO format data for mV is I-Value. When the diagnostic information "ST-segment elevation 0.2mV in lead V2" is input into the trained deep learning model, the keywords obtained by the BERT layer are "V2," "Lead," "ST," "Segment," "Elevation," "0.2," "m," and "V." The diagnostic statement output by the CRF layer is "ST-segment elevation of 0.2mV in lead V2."

[0081] Compared with the existing technology, the quality control feedback method of ECG data and diagnostic results of the present invention has significant advantages in automatic feedback and updating of ECG databases. Since the existing systems mostly rely on manual maintenance of the database, the update cycle is long and the response is delayed, it is difficult to reflect the latest clinical diagnostic features in a timely manner, and there is a lack of effective quality control mechanisms during the data storage process. The present invention can achieve dynamic self-optimization and updating of the database. Specifically, the present invention uses a deep learning model to perform text parsing and key content extraction on the clinical diagnostic results of the electrocardiogram, and combines the electrocardiogram features of the object with the existing data in the database for similarity matching. If the matching degree is lower than the set threshold, the new diagnosis and its features are automatically stored in the database, thereby achieving real-time learning and continuous expansion. This innovative method is different from the traditional method that relies on manual comparison and updating. It enables the database to have adaptive learning capabilities, significantly improves the intelligence of the system and the timeliness of data, and provides more accurate, comprehensive and up-to-date decision support for subsequent clinical auxiliary diagnosis.

[0082] Example 2

[0083] A quality control feedback method for ECG data and diagnostic results, which has the same other features as Example 1 and also has the following features: the local database also includes an ECG knowledge graph constructed based on the ECG data and diagnostic information.

[0084] The process of building an ECG knowledge graph consists of the following steps:

[0085] S2.1. Treat the diagnostic information of each ECG as a diagnostic node for the ECG, and then extract the relationship data, condition data, description data, range data, measurement and calculation data, and computer language data from the corresponding ECG data, with the relationship data constituting the relationship node, the condition data constituting the condition node, the description data constituting the description node, the range data constituting the range node, the measurement and calculation data constituting the measurement and calculation node, and the computer language data constituting the computer language node;

[0086] S2.2. Input the diagnosis nodes, relationship nodes, condition nodes, description nodes, range nodes, measurement and calculation nodes, and computer language nodes corresponding to each electrocardiogram into the Neo4j graph database, so that the diagnosis nodes, relationship nodes, condition nodes, description nodes, range nodes, measurement and calculation nodes, and computer language nodes are connected in series from front to back to generate an electrocardiogram knowledge graph.

[0087] It should be noted that the knowledge graph of this embodiment includes ECG types, judgment conditions, and ECG descriptions. In this graph, ECG types, judgment conditions, and descriptions serve as nodes, while "edges" connect these nodes, representing relationships between them. For example, there's a correspondence between a diagnosis and a specific judgment condition. In clinical applications, this ECG knowledge graph can be combined with a patient's real-time symptom data to push possible ECG types and their probabilities based on the patient's symptoms. This push system not only improves the efficiency of clinical diagnosis but also helps doctors make more accurate judgments in complex cases. The ECG knowledge graph created by this invention is a comprehensive and systematic medical knowledge representation model designed to integrate and organize the vast amount of knowledge in the field of electrocardiology, providing accurate and efficient decision support. The graph's design follows a structured approach and includes seven core node types: diagnosis nodes, relationship nodes, condition nodes, description nodes, range nodes, measurement and calculation nodes, and computer language nodes. Each node type carries rich medical information, covering various aspects of electrocardiology from multiple dimensions. Through the collaboration of relationship nodes, condition nodes, description nodes, range nodes, measurement and calculation nodes, a comprehensive analysis and diagnosis of the electrocardiogram is completed. The computer language node converts the quantitative information such as thresholds and intervals in the range node into a form that can be recognized and processed by the computer, thereby realizing the computer's automatic auxiliary diagnosis of the electrocardiogram.

[0088] The ECG knowledge graph of the present invention contains a total of 200 possible ECG types, and the diagnostic nodes within these 200 ECG types include normal ECG, approximately normal ECG, ST-T altered ECG, Brugada syndrome, J wave syndrome, and other diagnostic types. Through these multi-dimensional diagnostic types, the system can meticulously analyze each ECG feature and accurately match different cardiac pathological states, thereby improving the efficiency and accuracy of ECG interpretation. This provides clinicians with a comprehensive diagnostic basis, helping them to more accurately identify patients' ECG abnormalities and providing important reference for subsequent treatment decisions.

[0089] Among them, the relationship node is an unlabeled relationship or a labeled parallel relationship; when the relationship node in the electrocardiogram is an unlabeled relationship, all conditions in the condition node are met at the same time; when the relationship node in the electrocardiogram is a labeled parallel relationship, only at least one condition in the condition node needs to be met.

[0090] Taking the diagnosis of fascicular ventricular tachycardia as an example, there are two judgment conditions:

[0091] 1. Tachycardia originating from the right bundle branch: The QRS-T wave pattern shows a left bundle branch block pattern, that is, I, aVL, V5, and V6 show unidirectional R waves, and V1 and V2 show rS or QS types;

[0092] 2. Tachycardia originates from the left posterior fascicle: the QRS-T wave pattern shows right bundle branch block with left deviation.

[0093] When the ECG diagnosis node is fascicular ventricular tachycardia, the relationship node is marked as a parallel relationship, meaning that only one of the two judgment conditions needs to be met to complete the diagnosis. Therefore, the two judgment conditions are marked as a parallel relationship at the relationship node.

[0094] For example, if the diagnosis node of the electrocardiogram is the ST-T change type, and the judgment conditions of the ST-T change type include "ST segment change" and "T wave change", then the relationship node in the ST-T change type has no labeled relationship. When judging the ECG signal, both judgment conditions must be met at the same time to determine that the ECG signal is an ST-T change.

[0095] The conditional node is the type information of the electrocardiogram, and there is at least one conditional node. The conditional nodes are the electrocardiogram amplitude, waveform, morphology, and waveform duration.

[0096] It should be noted that conditional nodes refer to specific standards that need to be met when judging whether an electrocardiogram belongs to a certain type. These conditional nodes play an important role in the electrocardiogram knowledge graph. The conditional nodes are based on the changing patterns of the electrocardiogram. Through careful analysis of the electrocardiogram characteristics, combined with clinical symptoms and other examinations, they help doctors make more accurate diagnoses. These conditions usually include indicators such as the amplitude of the electrocardiogram, the shape of the waveform, the duration of the waveform, and the difference from the normal electrocardiogram. When these characteristics of the electrocardiogram match the standard conditions of a certain type of electrocardiogram, the doctor can preliminarily determine that the patient may have a specific heart disease or abnormality.

[0097] For example, the diagnosis of J wave syndrome requires the simultaneous fulfillment of the following conditions: "J point elevation greater than 0.1 mV", "duration greater than 20 ms", and "often accompanied by elevation of the origin or entire ST segment".

[0098] The diagnosis of retrograde atrioventricular reentrant tachycardia requires the simultaneous fulfillment of the following conditions: "heart rate between 150 and 250 beats / min", "regular RR interval", and "the P` wave is difficult to identify. If the retrograde P` wave can be identified, the RP` interval is greater than the P`R interval".

[0099] The diagnosis of junctional escape rhythm requires the simultaneous fulfillment of the conditions of "three or more consecutive junctional escape beats" and "heart rate between 40-60 beats / min".

[0100] The diagnosis of preexcitation syndrome requires the simultaneous fulfillment of the following conditions: "PR interval <0.12s", "QRS duration >0.12s, blunt QRS wave at the beginning, visible preexcitation wave-delta (delta) wave", "secondary ST-T changes" and "normal PJ interval".

[0101] These diagnostic conditions are based on the morphological characteristics and quantitative indicators of the electrocardiogram, and through the joint judgment of multiple conditions, the scientific nature and accuracy of the diagnosis are ensured.

[0102] The description node is the specific performance of the conditional node on the electrocardiogram and the corresponding description.

[0103] It should be noted that description nodes act as a bridge, providing detailed information on various signal features on the ECG. By transforming complex ECG data into easily understandable clinical information, they help doctors more intuitively and accurately understand and identify specific ECG manifestations. Each description node is closely related to its corresponding condition node, typically involving variables such as the ECG waveform, interval, and peak value. In the ECG knowledge graph of the present invention, a condition node may correspond to multiple description nodes. An ECG that meets any of the phenomena listed in the description node can be considered to meet the content of the condition node corresponding to its description node. For example, ST-T changes require meeting the condition of "T wave changes," which corresponds to multiple descriptions, including "peaked T wave," "inverted T wave," and "symmetric deep inversion of T wave." An ECG waveform that meets any of these characteristics can be considered to meet the condition of T wave changes. The role of description nodes is not only to qualitatively describe ECG signal characteristics but also to provide clinicians with clear and operational standards.

[0104] Among them, the range node is the feature range of the ECG data.

[0105] It should be noted that range nodes combine information from description nodes and measurement and calculation nodes to facilitate accurate measurement and judgment of ECG data and clarify the characteristic range of ECG signals. In clinical ECG analysis, range nodes primarily serve to standardize measurement results and identify abnormal signals. By setting reasonable boundaries for ECG characteristics (such as waveform duration, amplitude, and frequency), they provide doctors with accurate diagnostic evidence and facilitate the expression and learning of computer languages. For example, when a measured P wave frequency is 40-60 beats / minute, this phenomenon is classified as "bradycardia" in the description node. In this way, range nodes not only help define the normal and abnormal ranges of various ECG signals but also provide scientific and objective criteria for diagnosis.

[0106] The measurement and calculation nodes are variables of ECG data, including amplitude variables, frequency variables, amplitude difference variables, and time limit variables.

[0107] It's important to note that measurement and calculation nodes determine which ECG signal variables require precise measurement during ECG signal analysis to provide physicians with intuitive, quantitative data. Each measurement and calculation node typically corresponds to a description node. For example, when analyzing the "ST segment shortening" description node, the physician needs to locate and measure the duration of the ST segment in the ECG. Therefore, the measurement and calculation node corresponding to the "ST segment shortening" description node is "ST segment duration." Furthermore, measurement and calculation nodes aren't limited to duration measurement; in some ECG types, they also involve variables such as amplitude, frequency, and amplitude difference.

[0108] The computer language node converts the corresponding range node into a form for computer processing. By converting the standardized range of ECG features (such as the waveform's duration, amplitude, and frequency) with relevant data, this ensures that the computer can correctly read and analyze this information during ECG analysis. This node not only provides the necessary technical support for automatic ECG analysis but also enables the computer to follow predetermined standards and rules when processing and learning ECG data, thereby ensuring the accuracy and reliability of the analysis results. For example, when determining whether an ECG signal is a wandering rhythm within the sinoatrial node, the signal must meet the characteristic of a PP interval difference greater than 0.12s. This characteristic is expressed in computer language as (0.12,∞)s, allowing the computer to automatically identify and determine whether the ECG signal meets this standard. This automated identification and judgment greatly improves the efficiency and accuracy of ECG analysis.

[0109] In the process of establishing the ECG knowledge graph of the present invention, all diagnosis nodes, relationship nodes, condition nodes, description nodes, range nodes, measurement and calculation nodes and computer language nodes are stored in the form of an Excel table for subsequent processing and analysis. Each row in the Excel table represents the diagnosis node, relationship node, condition node, description node, range node, measurement and calculation node and computer language node corresponding to the same ECG, as shown in the following example: Figure 4 shown.

[0110] Existing medical knowledge graphs generally cover a wide range of diseases and symptoms. Although they can provide certain value for medical research, they lack in-depth analysis and precise classification in the field of electrocardiography.

[0111] The ECG knowledge graph of this embodiment focuses on the types and characteristic expressions of ECG signals, deeply focuses on professional content in the field of electrocardiology, and covers a variety of ECG-related diagnostic types, judgment conditions and waveform descriptions. The ECG knowledge graph integrates the latest electrocardiology research results and a large amount of clinical data, providing richer ECG signal type information and diagnostic clues. It can not only present the relationship between various ECG abnormalities and related characteristics in detail, but also help doctors interpret and diagnose ECGs more accurately through more refined structural data. The ECG knowledge graph is designed to focus on the characteristic data indicators of ECG signals. These characteristic data indicators are scientifically standardized and structured, which can not only efficiently extract key information from ECG signals, but also help accurately determine the type of ECG signals. Through this clear indicator system, the classification and analysis of ECG signals are clearer and more scientific, and can provide a more accurate reference basis for clinical diagnosis.

[0112] Example 3

[0113] A quality control feedback system for electrocardiogram data and diagnostic results executes the quality control feedback method for electrocardiogram data and diagnostic results of embodiment 1 or embodiment 2.

[0114] Compared with the existing technology, the quality control feedback system for ECG data and diagnostic results of the present invention has significant advantages in automatic feedback and updating of ECG databases. Since the existing systems mostly rely on manual maintenance of the database, the update cycle is long and the response is delayed, it is difficult to reflect the latest clinical diagnostic features in a timely manner, and there is a lack of effective quality control mechanisms during the data storage process. The present invention can achieve dynamic self-optimization and updating of the database. Specifically, the present invention uses a deep learning model to perform text parsing and key content extraction on the clinical diagnostic results of the electrocardiogram, and combines the electrocardiogram features of the object with the existing data in the database for similarity matching. If the matching degree is lower than the set threshold, the new diagnosis and its features are automatically stored in the database, thereby achieving real-time learning and continuous expansion. This innovative method is different from the traditional method that relies on manual comparison and updating. It enables the database to have adaptive learning capabilities, significantly improves the intelligence of the system and the timeliness of data, and provides more accurate, comprehensive and up-to-date decision support for subsequent clinical auxiliary diagnosis.

[0115] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the essence and scope of the technical solutions of the present invention.

Claims

1. A quality control feedback method for electrocardiogram data and diagnostic results, characterized in that: Proceed as follows: S1. Acquire multiple sets of historical data, each set of historical data is electrocardiogram corresponding electrocardiogram data and corresponding diagnostic information; S2. Build a local database using the ECG data and diagnostic information from S1, where the local database includes the ECG, ECG data, and diagnostic information; S3 obtains real-time diagnostic information and the corresponding electrocardiogram, performs quality control feedback on the real-time diagnostic information through the trained deep learning model, and updates S2's local database.

2. The quality control feedback method for ECG data and diagnostic results according to claim 1, characterized in that: The S3 is performed by the following steps: S3.

1. Obtain real-time diagnostic information and the corresponding electrocardiogram, and extract the clinical diagnostic information through the trained deep learning model to obtain a diagnostic statement consisting of multiple keywords; S3.

2. Perform similarity matching on the diagnostic statement extracted in S3.1 and the diagnostic information in the local database. When the matching result is less than a set threshold, save the clinical diagnostic information and the clinical electrocardiogram to the local database, thereby updating the local database.

3. The quality control feedback method for ECG data and diagnostic results according to claim 2, characterized in that: The deep learning model is a Chinese named entity recognition model based on BERT layer, BiLSTM layer and CRF layer; The Chinese named entity recognition model is provided with an Inpu_Text layer, a BERT layer, two BiLSTM layers, a LayerNorm layer, a Dropout Activation layer and a CRF layer, wherein the Inpu_Text layer, the BERT layer, the BiLSTM layer, the BiLSTM layer, the LayerNorm layer, the Dropout Activation layer and the CRF layer are sequentially connected from front to back; The Inpu_Text layer is used to input the original Chinese sentence of the diagnostic information; The BERT layer has a built-in transformer structure, and the BERT layer segmented the original Chinese sentence input by the Inpu_Text layer to obtain keywords, and semantically encoded the keywords to generate a context vector for each keyword; The BiLSTM layer is used to extract contextual dependencies in the sequence, thereby improving the ability to understand structural information; The LayerNorm layer is used to normalize the features of the keywords; The Dropout Activation layer is used to prevent overfitting and increase the expressive power of nonlinear data; The CRF layer is used to combine keywords to obtain a diagnostic statement, thereby outputting the most likely label sequence.

4. The quality control feedback method for ECG data and diagnostic results according to claim 2, characterized in that: The S3.2 is performed by the following steps: S3.2.

1. Using an edit distance matching algorithm, transform each character string in the diagnostic statement obtained in S3.1 into the minimum number of edit operations required for the character string in the diagnostic information in the local database, and then perform normalization to ultimately obtain a similarity. S3.2.

2. When the similarity obtained in S3.2.1 is less than the set threshold, the real-time diagnostic information and the corresponding electrocardiogram are saved in the local database, thereby updating the local database.

5. The quality control feedback method for ECG data and diagnostic results according to claim 2, characterized in that: The training process of the deep learning model is carried out by the following steps: A1. Constructing a data set, wherein the data set is a text sequence annotated in BIO format and consists of multiple groups of diagnostic statements and conversion vectors corresponding to the diagnostic statements; A2. Input the dataset into the deep learning model and output the extraction results; A3: Optimize the deep learning model using the negative log-likelihood as the loss function, and then determine whether the early stopping mechanism conditions are met. If not, proceed to A4; if so, proceed to A5. A4: Determine whether the maximum number of training times has been reached. If yes, proceed to A5; if no, return to A2. A5. Use the current deep learning model as the trained deep learning model.

6. The quality control feedback method for ECG data and diagnostic results according to any one of claims 1 to 5, characterized in that: The local database also includes an ECG knowledge graph constructed based on ECG data and diagnostic information.

7. The quality control feedback method for ECG data and diagnostic results according to claim 6, characterized in that: The process of establishing the ECG knowledge graph is carried out by the following steps: S2.

1. Treat the diagnostic information of each ECG as a diagnostic node for the ECG, and then extract the relationship data, condition data, description data, range data, measurement and calculation data, and computer language data from the corresponding ECG data, with the relationship data constituting the relationship node, the condition data constituting the condition node, the description data constituting the description node, the range data constituting the range node, the measurement and calculation data constituting the measurement and calculation node, and the computer language data constituting the computer language node; S2.

2. Input the diagnosis nodes, relationship nodes, condition nodes, description nodes, range nodes, measurement and calculation nodes, and computer language nodes corresponding to each electrocardiogram into the Neo4j graph database, so that the diagnosis nodes, relationship nodes, condition nodes, description nodes, range nodes, measurement and calculation nodes, and computer language nodes are connected in series from front to back to generate the electrocardiogram knowledge graph.

8. The quality control feedback method for ECG data and diagnostic results according to claim 7, characterized in that: The relationship nodes are either unlabeled or labeled parallel; When the relationship node in the electrocardiogram has no labeled relationship, all the conditions in the condition node are met at the same time; When the relationship nodes in the electrocardiogram are labeled parallel relationships, only at least one condition in the condition node needs to be satisfied.

9. The quality control feedback method for ECG data and diagnostic results according to claim 8, characterized in that: The conditional node is type information of the electrocardiogram, and there is at least one conditional node; The description node is the specific manifestation and corresponding description of the condition node on the electrocardiogram; The range node is a characteristic range of the electrocardiogram data; The measurement and calculation nodes are variables of electrocardiogram data; The computer language node is a form in which the corresponding range node is converted into computer processing; The variables are amplitude variables, frequency variables, amplitude difference variables, and time limit variables; The conditional nodes are electrocardiogram amplitude, waveform, morphology, and waveform duration.

10. A quality control feedback system for electrocardiogram data and diagnostic results, characterized by: The method for quality control and feedback of electrocardiogram data and diagnostic results according to any one of claims 1 to 9 is executed.

Images