A method for searching electrocardiogram scientific research

Abstract

The application provides a kind of electrocardiogram scientific research search method, it is related to electrocardiogram data processing field.The method includes: generating search instruction according to the target electrophysiological phenomenon selected by user;From the standardization search rule set of the time sequence logic rule of clinical rule base, the occurrence time order that must be satisfied between each electrocardio event constituting target electrophysiological phenomenon is obtained;From electrocardio information management system, the time sequence feature extraction model of candidate electrocardio data set input is obtained, the occurrence time and time sequence correlation of each electrocardio event are extracted, and the structured time sequence relationship data is output;The structured time sequence relationship data is compared with time sequence logic rule, and electrocardiogram search result that meets time sequence logic rule is obtained.The application solves the problem that prior art cannot carry out accurate search for special electrophysiological phenomenon that depends on time sequence, and realizes efficient and accurate search of special electrophysiological phenomenon.

Description

Technical Field

[0001] This invention relates to the field of electrocardiogram (ECG) data processing, and in particular to a method for ECG research retrieval. Background Technology

[0002] In electrocardiography (ECG) research, clinical studies, and academic paper writing, ECG and cardiologists often need to retrieve ECG data exhibiting specific electrophysiological phenomena. The Witkinsky phenomenon is a typical example, and other phenomena include tetraphasic block, triphasic block, concealed atrioventricular conduction, and supernormal conduction. The core characteristics of these phenomena depend not only on dynamic indicators such as heart rate, cardiac cycle, and coupling interval, but also on the temporal sequence of occurrence of each characteristic node. They are dynamic, conditional, and time-series correlated myocardial electrophysiological abnormalities.

[0003] Current mainstream ECG retrieval methods essentially involve combining multiple measurements and diagnoses, i.e., combining multiple measurement parameters such as heart rate range, RR interval, and PR interval with multiple diagnostic keywords such as atrioventricular block and escape beats. However, such combined queries can only independently retrieve and simply combine individual feature nodes of electrophysiological phenomena, failing to capture the temporal order of occurrence between these feature nodes and unable to match the dynamic temporal logic of specific electrophysiological phenomena. Furthermore, phenomena such as the Wiekinsky phenomenon are descriptions of electrophysiological mechanisms and are not standard ECG diagnostic terms; most hospital ECG reports do not include these names, making them unavailable for direct keyword retrieval. Even with relevant measurement parameters and diagnostic keywords, it is impossible to accurately retrieve data on such specific electrophysiological phenomena with clear temporal relationships, resulting in numerous missed and false positives, failing to meet the accuracy and rigor requirements of scientific research retrieval. Summary of the Invention

[0004] The embodiments of the present invention provide a method for electrocardiogram (ECG) scientific research retrieval, which aims to solve the problem that existing technologies cannot accurately retrieve special electrophysiological phenomena that depend on the temporal sequence of multiple ECG events.

[0005] To achieve the above objectives, the present invention provides a method for electrocardiogram (ECG) research retrieval, comprising the following steps: Based on the target electrophysiological phenomenon selected by the user in the pre-defined search interface, a corresponding search instruction is generated. According to the search instruction, a standardized search rule set corresponding to the target electrophysiological phenomenon is retrieved from a predetermined clinical rule base; the standardized search rule set includes temporal logic rules, which define the chronological order of occurrence that must be satisfied among the various electrocardiographic events constituting the target electrophysiological phenomenon; Obtain the candidate electrocardiogram (ECG) dataset to be output from the ECG information management system. Each ECG in the candidate ECG dataset contains pre-labeled ECG metadata. The ECG metadata includes multiple pre-labeled ECG events and their corresponding time locations. The candidate electrocardiogram dataset is input into a predetermined time-series feature extraction model to extract the occurrence time and time-series correlation of each electrocardiogram event, and output structured time-series relationship data. The structured temporal relationship data is compared with the temporal logic rules to obtain at least one electrocardiogram retrieval result that conforms to the temporal logic rules.

[0006] Furthermore, the target electrophysiological phenomenon is any one of the following: Weijinsky phenomenon, tetraphasic block, triphasic block, concealed atrioventricular conduction, or supernormal conduction.

[0007] Furthermore, the standardized retrieval rule set also includes triggering conditions and exclusion conditions. The triggering conditions are used to define the heartbeat type, heartbeat cycle, and non-conducted or atrioventricular block P wave labeling features that the target electrophysiological phenomenon must meet. The exclusion conditions are used to define the sample characteristics that need to be excluded based on the reported diagnostic conclusion, including at least one of organic myocardial injury, drug effects, and electrolyte disturbances.

[0008] Furthermore, the temporal feature extraction model adopts a structure that combines an LSTM network with an attention mechanism; The LSTM network is used to process the input temporal feature vector frame by frame in chronological order. The temporal feature vector is generated by encoding the ECG event. The LSTM network captures the positive and negative temporal dependencies between ECG events by processing the temporal feature vector. The attention mechanism is used to calculate the correlation weight between the temporal feature vector at each time step and the target electrophysiological phenomenon, and to perform a weighted summation of the temporal feature vector output by the LSTM network to output the structured temporal relationship data.

[0009] Further, the structured temporal relationship data is compared with the temporal logic rules to obtain at least one electrocardiogram retrieval result that conforms to the temporal logic rules, specifically including: The structured temporal relationship data is compared with the temporal logic rules by a predetermined temporal feature matching model. Samples that do not meet the rule requirements in terms of the order of occurrence of ECG events or the relative time interval are eliminated, and at least one ECG retrieval result that conforms to the temporal logic rules is obtained. The temporal feature matching model adopts an encoder-decoder architecture; The encoder is used to perform hierarchical encoding of the structured temporal relationship data through a multi-head self-attention mechanism, and introduces ECG relative interval position encoding to convert the relative time interval or interval ratio between each ECG event into a position vector. The decoder is used to compare the encoded temporal feature vector with the temporal logic rules using a cross-attention mechanism, and to determine whether the temporal logic is consistent by using a matching threshold.

[0010] Furthermore, after obtaining at least one electrocardiogram retrieval result that conforms to the aforementioned temporal logic rules, the process further includes: The ECG retrieval results that conform to the time-series logic rules are further filtered by a decision tree classification model. Specifically, the ECG metadata, time-series relationship data and the matching confidence of the time-series logic rules, as well as the reported diagnostic conclusions, are input into the decision tree classification model. The decision tree classification model traverses the decision nodes layer by layer and compares the input data with the node thresholds to determine the category to which the sample belongs. After removing false positive samples, the ECG retrieval results are output.

[0011] Furthermore, the decision tree classification model is configured as a classification and regression tree (CART) model.

[0012] Furthermore, when obtaining candidate electrocardiogram datasets from the electrocardiogram information management system, a subset of samples that match the core feature combination of the target electrophysiological phenomenon is located using a predetermined multi-level inverted index method.

[0013] Furthermore, the multi-level inverted index method includes a first-level index, a second-level index, and a third-level index, wherein: The primary index uses the target electrophysiological phenomenon as the index key and pre-archives electrocardiogram data containing all types of electrocardiographic events required for the phenomenon. The secondary index uses the heartbeat type and heartbeat cycle in the triggering conditions as index keys to filter the subset of electrocardiogram data that meets the triggering conditions. The three-level index uses the unique number preset by the electrocardiogram as the index key to associate electrocardiogram data and its metadata.

[0014] Furthermore, it also includes: receiving feedback from doctors on the review of the search results, and automatically updating the rule parameters in the clinical rule base or the matching threshold of the temporal logic rule based on the review feedback.

[0015] The above technical solution has the following technical effects: This invention generates a corresponding search instruction based on the target electrophysiological phenomenon selected by the user in a predefined search interface. Based on the search instruction, it retrieves a standardized search rule set corresponding to the target electrophysiological phenomenon from a clinical rule base. This rule set includes temporal logic rules, which define the chronological order of occurrence that must be satisfied between the various electrocardiogram (ECG) events constituting the target electrophysiological phenomenon. A candidate ECG dataset is obtained from the ECG information management system. Each ECG contains pre-labeled ECG events and their corresponding time positions. The candidate ECG dataset is input into a temporal feature extraction model to extract the occurrence time and temporal correlation of each ECG event, outputting structured temporal relationship data. This structured temporal relationship data is compared with the temporal logic rules to obtain at least one ECG search result that conforms to the temporal logic rules. This invention solves the problem that existing technologies cannot accurately search for special electrophysiological phenomena that depend on the chronological order of multiple ECG events.

[0016] In a further embodiment, the present invention reuses the ECG metadata that has been analyzed and diagnosed by ECG diagnostic doctors in their daily practice, including each heartbeat type, heartbeat cycle, non-conducted or atrioventricular block P wave annotation, and complete report diagnostic conclusions. No secondary annotation is required throughout the process. This is different from the retrieval method of the prior art, which combines multiple measurement ranges and multiple diagnostic conclusions with fuzzy keywords, and greatly reduces the data annotation cost.

[0017] In a further embodiment, the present invention employs a multi-level retrieval architecture consisting of pre-screening, multi-level inverted indexing, temporal feature matching, and decision tree secondary screening. This architecture significantly improves retrieval efficiency while ensuring retrieval accuracy, making it suitable for research retrieval needs of large-scale raw electrocardiogram data and avoiding full data traversal. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating an embodiment of the electrocardiogram research retrieval method of the present invention.

[0019] Figure 2 This is a schematic diagram of the structure of an electrocardiogram research retrieval system according to an embodiment of the present invention. Detailed Implementation

[0020] To further illustrate the various embodiments, the present invention provides accompanying drawings. These drawings are part of the disclosure of the present invention, primarily used to illustrate the embodiments and to explain the operating principles of the embodiments in conjunction with the relevant descriptions in the specification. With reference to these drawings, those skilled in the art should be able to understand other possible implementations and the advantages of the present invention. Components in the drawings are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0021] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments.

[0022] Example 1: Figure 1 This is a flowchart illustrating an embodiment of the electrocardiogram (ECG) research retrieval method of the present invention. The method of this embodiment includes the following steps: Based on the target electrophysiological phenomenon selected by the user in the pre-selection search interface, a corresponding search instruction is generated.

[0023] In one specific implementation, the electrocardiogram (ECG) research retrieval system is integrated into the ECG information management system of a tertiary hospital. The drop-down menu on the search interface offers multiple preset options for target electrophysiological phenomena, including the Weijinsky phenomenon, four-phase block, three-phase block, concealed atrioventricular conduction, and supernormal conduction. When a user selects a target electrophysiological phenomenon, the system generates a corresponding search command in response. Users do not need to input any keywords or measurement ranges; a single selection triggers the entire subsequent search process.

[0024] According to the search instructions, the standardized search rule set corresponding to the target electrophysiological phenomenon is retrieved from the predetermined clinical rule base. The standardized search rule set includes temporal logic rules, which define the chronological order of occurrence that must be satisfied between the various electrocardiographic events that constitute the target electrophysiological phenomenon.

[0025] In one specific implementation, the system automatically retrieves a standardized retrieval rule set uniquely corresponding to the target electrophysiological phenomenon from the clinical rule base according to the retrieval command. Specifically, the rule set includes triggering conditions, exclusion conditions, and timing logic rules. Triggering conditions define the heartbeat type, heartbeat cycle, and non-conducted or atrioventricular block P-wave marking characteristics that the target electrophysiological phenomenon must satisfy. Exclusion conditions define the sample characteristics that need to be excluded based on the reported diagnostic conclusion, including at least one of organic myocardial injury, drug effects, and electrolyte disturbances. Timing logic rules define the chronological order of occurrence that must be satisfied among the various electrocardiographic events constituting the target electrophysiological phenomenon. Taking the Weijinsky phenomenon as an example, the timing logic rules are specifically as follows: high-degree atrioventricular block occurs first, then an escape beat occurs, then the first P wave after the escape beat is successfully conducted, then 1-3 consecutive P waves maintain conduction, and finally atrioventricular block reappears.

[0026] The candidate ECG dataset to be output is obtained from the ECG information management system. Each ECG in the candidate ECG dataset contains pre-labeled ECG metadata, which includes multiple pre-labeled ECG events and their corresponding time locations.

[0027] In one specific implementation, the system retrieves candidate electrocardiogram (ECG) datasets from an ECG information management system. This system stores 150,000 ECG data points accumulated over the years, each meticulously annotated by the diagnosing physician during routine diagnostic analysis. The annotations include the type of each heartbeat (e.g., sinus, atrial premature beats, junctional premature beats, ventricular premature beats, pacing, atrial escape beats, junctional escape beats, ventricular escape beats, ventricular fusion beats, atrial premature beats with aberrational conduction, including P waves from non-conducted atrial premature beats and P waves from atrioventricular block), the time location of each heartbeat, heart cycle values ​​such as RR interval and PR interval, and a complete diagnostic report. No secondary annotation is required throughout the process. For example, in a typical ECG sample, the doctor-annotated baseline sinus RR interval is 800ms, escape beat period is 1000ms, after four consecutive sinus P waves fail to conduct, a junctional escape beat appears, the first P wave after the escape beat successfully conducts, followed by two consecutive P waves that maintain conduction, and finally two consecutive P waves fail to conduct again. During acquisition, the system first eliminates samples with exclusion factors based on the reported diagnostic conclusions, such as those containing "acute myocardial infarction," "myocarditis," or "hyperkalemia." Then, based on hard thresholds of trigger conditions, samples that do not meet the basic metadata requirements are eliminated, such as samples without ≥3 consecutive unconducted P waves or without escape beats, completing the preliminary screening. After the preliminary screening, the system uses a multi-level inverted index to locate a subset of samples that match the core feature combination of the target electrophysiological phenomenon. This index adopts a three-level inverted index architecture with combined features at the forefront, relying entirely on the metadata features already completed by the ECG diagnosing doctor and the attributes of the original ECG data, requiring no secondary annotation. The core data sources for index construction include two aspects: first, the metadata features already annotated by doctors, including heartbeat type codes, non-transmitted or atrioventricular block P wave annotation types, and core trigger condition parameters; second, the attributes of the original electrocardiogram data, including the unique electrocardiogram number, patient ID, examination time, original waveform storage path, and report diagnostic conclusion code. A three-level index structure is constructed based on these data sources to achieve rapid location from coarse to fine.

[0028] The first-level index is a coarse-screening level, using the clinical rule base scenario code as the index key. It pre-archives sample sets containing all core metadata tags for different search scenarios such as the Weijinsky phenomenon and four-phase block, directly eliminating irrelevant samples. The second-level index is a fine-screening level, using the core trigger condition metadata combination as the index key within the first-level index's filtering results to quickly filter a subset of samples that meet the hard threshold of the trigger conditions, further narrowing the search scope. The third-level index is a retrieval level, using the unique ECG number as the joint index key, linking patient baseline data, the original waveform storage path, and the full information of the physician's metadata, achieving millisecond-level sample retrieval, while also adapting to the secondary filtering needs of research searches.

[0029] For index maintenance, the model supports an automatic incremental update mechanism. When a doctor completes a new ECG analysis annotation or modifies an existing annotation, the system automatically triggers an index update, updating only the newly added or modified annotations without requiring a full index reconstruction, ensuring consistency between the index and the doctor's annotation results and the original data. Through this three-level index, the system can retrieve target samples in milliseconds without traversing the entire dataset.

[0030] The candidate electrocardiogram (ECG) dataset is input into a predetermined time-series feature extraction model to extract the occurrence time and time-series correlation of each ECG event, and output structured time-series relationship data.

[0031] In one specific implementation, the system inputs candidate electrocardiogram (ECG) datasets into a temporal feature extraction model. This model employs an LSTM network combined with an attention mechanism. First, a metadata normalization mapping layer unifies the numerical encoding of the heartbeat types labeled by the doctor, directly extracting the heartbeat cycle values. This normalization mapping layer is based on the HL7 / FHIR ECG standard and the ISO / TS 22077 ECG terminology standard. The coordinates and labeling types of unrecorded or atrioventricular block P-wave labels are standardized. Simultaneously, a normalization preprocessing unit unifies the original ECG waveform data into a standard voltage-time series vector with sampling rates of 250Hz, 500Hz, or 1000Hz; complete rhythm segments are extracted 5 seconds before and after the doctor-labeled events. The two types of input data are then concatenated after normalization into a unified temporal feature vector, which is then input into the LSTM network. The LSTM network employs a bidirectional LSTM structure. The hidden layer dimension can be adjusted between 32, 64, and 128 dimensions depending on the sample size. It uses the tanh activation function, while the output layer uses the sigmoid activation function. An adaptive forgetting gate mechanism is implemented, dynamically adjusting the forgetting gate threshold (range 0.5-0.9) based on the total duration and number of heartbeats in the input samples. The LSTM network processes the input temporal feature vector frame by frame in chronological order. These temporal feature vectors are generated from ECG event encodings. The LSTM network captures both forward and backward temporal dependencies between ECG events by processing these feature vectors: the forward LSTM processes frame by frame from the beginning of the time series, capturing the forward temporal dependencies from the previous heartbeat to the current heartbeat and then to the next heartbeat; the backward LSTM processes frame by frame from the end of the time series, capturing the influence of subsequent heartbeats on preceding heartbeats. The LSTM network outputs the temporal feature vectors at each time step. The attention mechanism is used to calculate the correlation weights between the temporal feature vectors at each time step and the target electrophysiological phenomenon, and to perform a weighted summation of the temporal feature vectors output by the LSTM network, outputting structured temporal relationship data. Specifically, the attention mechanism employs a soft attention mechanism, configuring dynamic target feature templates linked to the clinical rule base. Template parameters are automatically extracted directly from the triggering conditions and temporal rules of the clinical rule base, requiring no manual pre-setting. The attention mechanism calculates the correlation weights between the temporal feature vectors at each time step and the target electrophysiological phenomenon. After normalization using the softmax function, the temporal feature vectors output by the LSTM network are weighted and summed, giving higher weights to feature regions related to the target electrophysiological phenomenon and weakening redundant features. After weighted summation, dimensionality reduction is performed using PCA principal component analysis, retaining more than 95% of the core features, and finally outputting structured temporal relationship data. This data includes the timestamps of each ECG event, the temporal correlation matrix of heartbeat types, the relative time intervals between ECG events, and the correlation features between untransmitted P-wave annotations and the heartbeat cycle, etc.The system also includes a feature weight visualization output unit, which directly maps the high-weight regions assigned by the attention mechanism to the metadata nodes annotated by the doctor and the positions of the original waveforms, and highlights them when displaying the results to ensure the interpretability of the algorithm.

[0032] The structured time-series relational data is compared with the time-series logic rules to obtain at least one electrocardiogram retrieval result that conforms to the time-series logic rules.

[0033] In one specific implementation, the system compares structured temporal relationship data with temporal logic rules using a temporal feature matching model. Samples whose ECG event sequence or relative time intervals do not meet the rule requirements are eliminated, resulting in at least one ECG retrieval result that conforms to the temporal logic rules. The temporal feature matching model employs an encoder-decoder architecture. The encoder performs hierarchical encoding of the structured temporal relationship data using a multi-head self-attention mechanism and introduces ECG relative interval position encoding, converting the relative time intervals or interval ratios between ECG events into position vectors. Specifically, the encoder uses eight attention heads to perform hierarchical encoding of the structured temporal relationship data, focusing on capturing the correlation features between untransmitted P-wave annotations and heartbeat types, as well as the temporal dependencies of each node. It abandons absolute timestamp encoding, converting the relative time intervals between ECG events into position vectors, thus adapting to ECG samples with different heart rates and sampling rates. The decoder uses a cross-attention mechanism to compare the encoded temporal feature vectors with the temporal logic rules, determining whether the temporal logic is consistent based on a matching threshold. Specifically, the decoder uses a cross-attention mechanism to compare the encoded temporal feature vector with temporal logic rules. An adjustable matching threshold, such as 0.85, is used to determine if the temporal logic is consistent. The temporal order and relative time intervals of each node of the target electrophysiological phenomenon are compared one by one. Samples with mismatched temporal logic or time intervals exceeding the threshold are marked and removed. After obtaining ECG retrieval results that conform to the temporal logic rules, the system further filters these results using a decision tree classification model. This includes inputting ECG metadata, temporal relationship data, the matching confidence score between the ECG metadata and the temporal logic rules, and the reported diagnostic conclusions into the decision tree classification model. The decision tree classification model uses the CART algorithm to construct a binary decision tree, with the classification objective being whether it is the target electrophysiological phenomenon. The Gini coefficient is used as the feature selection criterion. The model includes a decision layer for identifying other electrophysiological phenomena, used to prioritize distinguishing the target phenomenon from other easily confused electrophysiological scenarios such as abnormal conduction and gap phenomena. The decision tree classification model traverses decision nodes layer by layer. For example, the first layer determines if there are ≥3 consecutive untransmitted P waves; the second layer determines if the escape cycle is greater than or equal to 1.2 times the baseline sinus cycle; and the third layer determines if the report conclusion includes exclusion factors. It compares sample features with node thresholds to determine the sample's category, eliminating false positives. False positives include samples that meet the temporal logic but not the triggering conditions, samples that meet the triggering conditions but have exclusion factors, and samples that meet the basic conditions but represent other easily confused electrophysiological phenomena. The model outputs accurately matched ECG retrieval results. The decision tree classification model supports an incremental learning mechanism, automatically adjusting decision node thresholds and feature weights based on physician review feedback, without requiring secondary annotation during the optimization process.The system displays basic information and matching scores of all matched samples in a list format. When a user clicks on a sample, the details page displays the original electrocardiogram (ECG) waveform, metadata details, temporal matching logic, and rule matching details, providing end-to-end traceability. The original ECG waveform synchronously highlights the P wave and heartbeat type annotations completed by the doctor, along with the high-weighted matching areas of the model. The system also includes a closed-loop iterative optimization step: receiving feedback from doctors on the search results and automatically updating the rule parameters or temporal logic rule matching thresholds in the clinical rule base based on this feedback. Specifically, after the doctor reviews the search results, the system receives feedback and automatically updates the rule parameters in the clinical rule base, such as adjusting the escape cycle multiple threshold from 1.2 to 1.25 or the temporal logic rule matching threshold from 0.85 to 0.90. The optimization process requires no secondary annotation.

[0034] In this embodiment, the system successfully retrieved samples matching the target electrophysiological phenomenon from 150,000 electrocardiograms in the electrocardiogram information management system of a tertiary hospital. The retrieval process took about 3 seconds. The retrieval results were reviewed and confirmed by electrocardiogram experts, and the accuracy rate reached 100%. This fully demonstrates that the present invention can achieve accurate and efficient retrieval of special electrophysiological phenomena by automatically matching the temporal logic of multiple electrocardiogram events in the electrocardiogram without relying on the name of the report diagnosis conclusion.

[0035] Example 2: This embodiment provides a detailed description of the standardized retrieval rule set for various special electrophysiological phenomena pre-built in the clinical rule base. The rule set includes triggering conditions, exclusion conditions, and timing logic rules. All rules are strictly designed based on metadata already completed by electrocardiogram diagnostic physicians, including heartbeat type, heartbeat cycle, and P wave annotations for non-conducted or atrioventricular block, and contain no non-standard diagnostic terminology.

[0036] In one specific implementation, the clinical rule base contains a pre-defined set of standardized retrieval rules for five special electrophysiological phenomena: the Witkinsky phenomenon, four-phase block, three-phase block, concealed atrioventricular conduction, and supernormal conduction.

[0037] The triggering conditions for the Wejinsky phenomenon include six aspects. First, the presence of high-degree or third-degree atrioventricular block, specifically defined as the absence of conduction of ≥3 consecutive sinus P waves as indicated by the physician, with the block site limited to the atrioventricular node or His bundle level. This condition is used to screen ECG samples with severe conduction disorders. Second, the appearance of junctional or ventricular escape beats during the block interval, based on the physician's beat type annotation, is used to confirm that a backup pacemaker was activated during the conduction block. Third, the escape beat period is ≥1.2 times the baseline sinus cycle, where the escape beat period is calculated as the RR interval of the last conducted sinus beat before the escape beat, and the baseline sinus cycle reuses the cycle value indicated by the physician. This ratio condition is used to quantify the timing of the escape beat. Fourth, the first sinus P wave after the escape beat is successfully conducted, and this P wave falls within the stress period after the escape beat, used to confirm the occurrence of the Wejinsky facilitation effect. Fifth, the continuation of conduction of 1-3 consecutive sinus P waves after the facilitation effect is used to confirm the persistence of the Wejinsky effect. Sixth, the recurrence of atrioventricular block with ≥2 consecutive P waves failing to conduct after the effect has ended is used to confirm that the conduction state has returned to the block baseline. Exclusion criteria include ruling out organic myocardial injury, drug effects, electrolyte disturbances, and other easily confused electrophysiological phenomena based on the reported diagnostic conclusion, to avoid misdiagnosing similar manifestations caused by other etiologies as the Weijinsky phenomenon. The temporal logic rule is defined as the strict chronological order of the above six electrocardiographic events: high-degree atrioventricular block occurs first, followed by junctional or ventricular escape beats, then the first P wave following the escape beat, then 1-3 consecutive P waves maintaining conduction, and finally the recurrence of atrioventricular block.

[0038] For quadriphase block, the triggering conditions include three aspects. First, a normal PR interval at a slow baseline heart rate, based on cardiac cycle and PR interval markings, is used to establish a normal conduction baseline at a slow heart rate. Second, atrioventricular block occurs after the heart rate increases, based on heartbeat type and cardiac cycle markings, to identify conduction abnormalities induced by heart rate changes. Third, atrioventricular block disappears after the heart rate recovers slowly, based on heartbeat type and cardiac cycle markings, to confirm the reversibility of the conduction abnormality disappearing as the heart rate recovers. Exclusion conditions include excluding myocardial ischemia, drug effects, and cardiomyopathy based on the reported diagnostic conclusion, to exclude other causes that may lead to heart rate-related conduction abnormalities. The temporal logic rule is: a slow baseline heart rate with normal PR occurs first, then atrioventricular block occurs with an increased heart rate, and finally the block disappears with a decreased heart rate.

[0039] For triphasic block, the triggering conditions include three aspects. First, atrioventricular block occurs at a moderate heart rate, based on heart rate type and heart cycle labeling, to identify conduction abnormalities that occur only within a specific heart rate range. Second, atrioventricular conduction returns to normal when the heart rate is too fast or too slow, based on heart rate type and heart cycle labeling, to confirm that the conduction abnormality occurs only within a moderate heart rate range. Third, there is no fixed conduction relationship between the P wave and QRS complex during the block, based on non-conducting or atrioventricular block P wave labeling and heart rate type labeling, to confirm that the nature of the block is a functional block at the atrioventricular node level rather than a fixed proportion block. Exclusion conditions include excluding electrolyte disturbances, cardiomyopathy, and acute myocardial infarction based on the reported diagnosis, to exclude other causes that may lead to heart rate-related conduction abnormalities. The temporal logic rule is: normal conduction occurs first at slow heart rates, then atrioventricular block occurs at moderate heart rates, and finally normal conduction occurs at fast heart rates.

[0040] For concealed atrioventricular conduction, the triggering conditions include three aspects. First, the presence of a non-conducted P wave, based on non-conducted or atrioventricular block P wave markings, is used to identify the initiation event of concealed conduction. Second, the non-conducted P wave affects the conduction of subsequent P waves, based on non-conducted or atrioventricular block P wave markings and cardiac cycle markings. Specifically, although a non-conducted P wave itself fails to excite the ventricle, its excitation is partially conducted within the atrioventricular node, interfering with the conduction of subsequent excitations. Third, the PR interval of subsequent P waves is prolonged or fails to conduct again, based on non-conducted or atrioventricular block P wave markings and PR interval markings, used to quantify the degree of interference caused by concealed conduction to subsequent conduction. Exclusion conditions include excluding acute myocardial infarction, drug effects, and electrolyte disturbances based on the reported diagnostic conclusion, in order to exclude other causes that may lead to PR interval prolongation. The temporal logic rule is: a non-conducted P wave occurs first, followed by PR prolongation or failure to conduct again in subsequent P waves.

[0041] For abnormal conduction, the triggering conditions include three aspects. First, atrioventricular block is present, based on the heartbeat type and the P wave marking of non-conducted or atrioventricular block, used to establish the baseline state of conduction abnormality. Second, the P wave can conduct normally within a specific time window, based on the P wave marking of non-conducted or atrioventricular block and the cardiac cycle marking. This time window is usually located in a specific phase of the cardiac cycle, such as a narrow window after the T wave and before the next P wave. Third, the P wave outside this time window is still blocked, based on the P wave marking of non-conducted or atrioventricular block and the cardiac cycle marking, used to confirm that the conduction improvement is limited to this specific time window. Exclusion conditions include excluding drug effects, electrolyte disturbances, and the Weijinsky phenomenon based on the reported diagnosis, to exclude other electrophysiological phenomena that may cause temporary conduction improvement. The temporal logic rule is: atrioventricular block occurs first, then the P wave conducts normally within the specific time window, and finally the P wave is blocked again outside the time window.

[0042] For each of the above-mentioned special electrophysiological phenomena, the retrieval and matching order is uniformly as follows: First, a preliminary coarse screening is performed, eliminating samples with exclusion factors based on exclusion criteria and eliminating samples that do not meet the basic metadata requirements based on trigger condition hard thresholds; then, through a retrieval acceleration model, a three-level inverted index is established by combining the scene encoding of the target phenomenon and the core metadata to quickly locate relevant original electrocardiogram samples; next, through a temporal feature matching model, the relative temporal relationship of the samples is compared with the temporal logic rules in the rules, eliminating samples whose temporal order or relative time interval does not meet the requirements; finally, through a decision tree classification model, based on the metadata annotated by the doctor, the report diagnosis conclusion, and the temporal matching results, a second screening is performed to eliminate false positive samples and output accurate matching results.

[0043] In this embodiment, the rule base configurations for the five special electrophysiological phenomena are strictly based on the metadata design already completed by the ECG diagnostic physicians, enabling accurate retrieval without needing to write the phenomenon names in the diagnostic report. The rule parameters can be flexibly adjusted according to the physician's clinical research needs and are linked in synergy with the algorithm model of the retrieval engine module. When the rule base parameters are adjusted, the algorithm comparison threshold is automatically updated synchronously to ensure retrieval accuracy.

[0044] Example 3 Figure 2 This is a schematic diagram of the structure of an electrocardiogram (ECG) research retrieval system according to an embodiment of the present invention. Figure 2 As shown, the system includes a data structuring module, a clinical rule base module, a search engine module, and a search result display and tracing module, among which: The data structuring module is used to reuse the ECG metadata and report diagnostic conclusions already completed by doctors, transform them into structured data, and extract the temporal relationships of each feature node; The clinical rule base module stores a standardized set of retrieval rules built upon physician-annotated metadata. The retrieval engine module is connected to the data structuring module and the clinical rule base module, respectively, and is used to perform precise retrieval with temporal logical matching based on structured data and standardized retrieval rule sets; The search results display and traceability module binds search results with raw electrocardiogram (ECG) data and doctor-annotated metadata, providing end-to-end traceability. The search engine module further includes a time-series feature matching model, a decision tree classification model, and a search acceleration model. During system operation, the data structuring module connects to the ECG information management system to obtain the raw ECG data and metadata analyzed by doctors, transforming it into structured time-series relational data. The search engine module, based on the rule set provided by the clinical rule base module, sequentially performs preliminary screening, index positioning, time-series matching, and secondary screening. The search results display and traceability module outputs matching results and provides traceability information. The system automatically updates rule parameters based on doctor review feedback, forming a closed-loop optimization.

[0045] Although the invention has been specifically shown and described in conjunction with preferred embodiments, those skilled in the art should understand that various changes in form and detail may be made to the invention without departing from the spirit and scope of the invention as defined in the appended claims, all of which shall be within the scope of protection of the invention.

Claims

1. A method for research retrieval of electrocardiograms, characterized in that, Includes the following steps: Based on the target electrophysiological phenomenon selected by the user in the pre-defined search interface, a corresponding search instruction is generated. The standardized retrieval rule set corresponding to the target electrophysiological phenomenon is retrieved from a predetermined clinical rule base according to the retrieval instruction. The standardized retrieval rule set includes temporal logic rules, which define the chronological order of occurrence that must be satisfied among the various electrocardiographic events constituting the target electrophysiological phenomenon. The standardized retrieval rule set also includes triggering conditions and exclusion conditions. The triggering conditions are used to define the heartbeat type, heartbeat cycle, and non-conducted or atrioventricular block P wave labeling features that the target electrophysiological phenomenon must satisfy. The exclusion conditions are used to define the sample characteristics that need to be excluded based on the reported diagnostic conclusion, including at least one of organic myocardial injury, drug effects, and electrolyte disturbances. Obtain the candidate electrocardiogram (ECG) dataset to be output from the ECG information management system. Each ECG in the candidate ECG dataset contains pre-labeled ECG metadata. The ECG metadata includes multiple pre-labeled ECG events and their corresponding time locations. The candidate electrocardiogram dataset is input into a predetermined time-series feature extraction model to extract the occurrence time and time-series correlation of each electrocardiogram event, and output structured time-series relationship data. The structured temporal relationship data is compared with the temporal logic rules to obtain at least one electrocardiogram retrieval result that conforms to the temporal logic rules; After obtaining at least one electrocardiogram retrieval result that conforms to the aforementioned temporal logic rules, the process further includes: The ECG retrieval results that conform to the time-series logic rules are further filtered by the decision tree classification model. Specifically, the ECG metadata, time-series relationship data and the matching confidence of the time-series logic rules, as well as the reported diagnostic conclusions, are input into the decision tree classification model. The decision tree classification model traverses the decision nodes layer by layer and compares the input data with the node thresholds to determine the category to which the sample belongs. After removing false positive samples, the ECG retrieval results are output. When obtaining candidate electrocardiogram datasets from the electrocardiogram information management system, a subset of samples that match the core feature combination of the target electrophysiological phenomenon is located using a predetermined multi-level inverted index method. The multi-level inverted index method includes first-level indexes, second-level indexes, and third-level indexes, wherein: The primary index uses the target electrophysiological phenomenon as the index key and pre-archives electrocardiogram data containing all types of electrocardiographic events required for the phenomenon. The secondary index uses the heartbeat type and heartbeat cycle in the triggering conditions as index keys to filter the subset of electrocardiogram data that meets the triggering conditions. The three-level index uses the unique number preset by the electrocardiogram as the index key to associate electrocardiogram data and its metadata.

2. The electrocardiogram research retrieval method according to claim 1, characterized in that, The target electrophysiological phenomenon is any one of the following: Weijinsky phenomenon, tetraphasic block, triphasic block, concealed atrioventricular conduction, or supernormal conduction.

3. The electrocardiogram research retrieval method according to claim 1, characterized in that, The temporal feature extraction model adopts a structure that combines an LSTM network with an attention mechanism; The LSTM network is used to process the input temporal feature vector frame by frame in chronological order. The temporal feature vector is generated by encoding the ECG event. The LSTM network captures the positive and negative temporal dependencies between ECG events by processing the temporal feature vector. The attention mechanism is used to calculate the correlation weight between the temporal feature vector at each time step and the target electrophysiological phenomenon, and to perform a weighted summation of the temporal feature vector output by the LSTM network to output the structured temporal relationship data.

4. The electrocardiogram research retrieval method according to claim 1, characterized in that, The structured time-series relation data is compared with the time-series logic rules to obtain at least one electrocardiogram retrieval result that conforms to the time-series logic rules, specifically including: The structured temporal relationship data is compared with the temporal logic rules by a predetermined temporal feature matching model. Samples that do not meet the rule requirements in terms of the order of occurrence of ECG events or the relative time interval are eliminated, and at least one ECG retrieval result that conforms to the temporal logic rules is obtained. The temporal feature matching model adopts an encoder-decoder architecture; The encoder is used to perform hierarchical encoding of the structured temporal relationship data through a multi-head self-attention mechanism, and introduces ECG relative interval position encoding to convert the relative time interval or interval ratio between each ECG event into a position vector. The decoder is used to compare the encoded temporal feature vector with the temporal logic rules using a cross-attention mechanism, and to determine whether the temporal logic is consistent by using a matching threshold.

5. The electrocardiogram research retrieval method according to claim 1, characterized in that, The decision tree classification model is configured as a classification and regression tree model.

6. The electrocardiogram research retrieval method according to claim 1, characterized in that, Also includes: Receive feedback from doctors on the search results, and automatically update the rule parameters in the clinical rule base or the matching threshold of the temporal logic rule based on the feedback.

Images