A system for screening patients with chronic obstructive pulmonary disease in pneumology

Abstract

This invention discloses a screening system for COPD patients in respiratory medicine, specifically in the field of computer-aided diagnosis, to address the technical problems of low screening efficiency, low patient cooperation, and difficulty in large-scale promotion at the grassroots level in existing screening methods. The system acquires medical records of the target population containing drug delivery and disease diagnosis codes to construct standardized patient medical event sequences; identifies characteristic medication and diagnostic patterns prior to COPD diagnosis using sequence pattern mining technology; constructs a medical knowledge graph using these patterns; learns feature embedding vectors representing COPD medical trajectories through graph neural networks to establish high-risk digital profiles; inputs the medical sequences of the individuals to be screened into a feature extraction model trained on the same knowledge graph; and generates individualized risk probability prediction values ​​by calculating the matching degree between the medical sequence and the digital profile; automatically generates screening result reports, achieving efficient and non-invasive early risk screening.

Description

Technical Field

[0001] This invention relates to the field of computer-aided diagnostic technology, and more specifically, to a screening system for COPD patients in respiratory medicine. Background Technology

[0002] Chronic obstructive pulmonary disease (COPD) is a common respiratory disease that seriously endangers public health. Its early symptoms are insidious, and most patients are already in the middle stage when they are diagnosed, with irreversible damage to lung function, missing the best time for intervention.

[0003] Currently, in primary healthcare institutions, large-scale physical examinations, and health management settings, initial screening for COPD heavily relies on questionnaires (such as the CAT questionnaire) and pulmonary function tests. However, questionnaires are highly subjective and have low sensitivity; while pulmonary function tests require expensive specialized equipment, professional operators, and require patients to perform forceful exhalation with a high degree of cooperation. This is difficult to implement in elderly, frail, or comorbid populations, severely limiting the accessibility and coverage of screening.

[0004] In their daily operations, healthcare information systems at all levels continuously generate massive amounts of accurate and objective structured data, including detailed medication delivery records and disease diagnosis records for each patient. This data provides a temporal overview of an individual's treatment, potentially containing medication patterns and diagnostic trajectories reflecting early airway function decline. However, within the current technological framework, this valuable medical data is primarily used for record management and financial settlement, completely neglecting its early warning value for COPD risks. Existing technologies lack an effective means to automatically extract characteristic patterns closely related to COPD pathological development from this fragmented, multi-source temporal data and transform them into quantitative risk assessments, thus creating a technological bottleneck resulting in low efficiency for early screening in resource-constrained scenarios. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, embodiments of the present invention provide a screening system for COPD patients in respiratory medicine to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A screening system for COPD patients used in respiratory medicine includes:

[0008] The data extraction module is used to obtain the medical diagnosis records of the target population, which includes COPD patients of different age groups, and the medical diagnosis records include drug delivery records and lung disease diagnosis records.

[0009] The data processing module is used to construct a medical event sequence for each COPD patient according to the medical diagnosis record time, standardize the medical event sequence, and construct a standardized time-series event sequence.

[0010] The pattern recognition module is used to identify common medication and diagnosis sequence patterns in the historical data of COPD patients based on standardized time-series event sequences.

[0011] The profile building module is used to construct a medical knowledge graph based on medication and diagnosis sequence patterns. It uses graph neural networks to learn the feature embedding vectors of the medical trajectories of COPD patients and builds a digital profile of high-risk COPD patients.

[0012] The risk prediction module is used to input the standardized time-series event sequence of the population to be screened into the feature extraction model, calculate the matching degree with the digital profile of high COPD risk, and generate an individualized COPD risk probability prediction value.

[0013] The screening output module is used to generate a screening risk report when the predicted risk probability value continuously exceeds the preset risk threshold.

[0014] In a preferred embodiment, the data extraction module specifically includes obtaining the medical diagnosis records of the target population:

[0015] Extract drug delivery records from electronic prescriptions in the medical information database, including the generic name of the drug, delivery time, dosage, and frequency of use.

[0016] Extract lung disease diagnosis records from electronic medical records in the medical information database, including diagnosis name, diagnosis time, and diagnosis result information;

[0017] By linking and matching medication delivery records with lung disease diagnosis records using patient identification numbers, preliminary medical records are established for each patient.

[0018] Based on the preset age stratification criteria, the target population is divided into multiple age subgroups. Data of COPD patients who meet the requirements of each age subgroup are selected from the preliminary medical records to establish a medical diagnosis record set for the target population.

[0019] In a preferred embodiment, the data processing module constructs a medical event sequence for each COPD patient according to the medical diagnosis record time, and performs standardization processing on the medical event sequence. The construction of a standardized time-series event sequence specifically includes:

[0020] Arrange each patient's medical diagnosis records in chronological order of diagnosis to form a medical event sequence;

[0021] Standardize the coding process for drug information in medical event sequences, and uniformly convert the generic names of drugs into codes based on anatomical, therapeutic, and chemical classification systems.

[0022] The diagnostic results in the sequence are standardized and coded, and the codes of diagnostic results under different standards are uniformly mapped to the disease classification standard codes.

[0023] Establish a standardized timestamp format to convert the occurrence time of all medical events into a unified time representation.

[0024] Based on standardized drug codes, diagnostic codes, and timestamps, a standardized medical event sequence containing complete time-series information is constructed.

[0025] In a preferred embodiment, the pattern recognition module, based on a standardized time-series event sequence, identifies medication and diagnostic sequence patterns that co-occur within a preset time period before diagnosis in the historical data of a COPD patient population. Specifically, this includes:

[0026] Using confirmed cases as the dividing point, extract all standardized time-series event sequences within a specified observation period before diagnosis from the historical data of COPD patients;

[0027] Sequence pattern mining algorithms were used to analyze event sequences and identify recurring specific event combinations in the COPD patient population.

[0028] The event combinations include consecutive drug delivery record sequences, consecutive lung disease diagnosis record sequences, and combinations of alternating drug delivery records and lung disease diagnosis records.

[0029] The identified event combination sequences are subjected to significance testing, and event sequences whose frequency reaches a preset significance threshold are selected and marked as characteristic sequence patterns.

[0030] The selected characteristic sequence patterns are categorized according to their frequency of occurrence and event combinations to establish a sequence pattern library related to COPD.

[0031] In a preferred embodiment, the profile building module constructs a medical knowledge graph based on medication and diagnosis sequence patterns, and uses a graph neural network to learn the feature embedding vector of the medical trajectory of COPD patients to build a high-risk digital profile of COPD, specifically including:

[0032] A medical knowledge graph was constructed based on a sequence pattern library related to COPD. The graph node types include drug nodes, diagnosis nodes, and temporal relationship nodes.

[0033] Define the multidimensional relationships between nodes, including the compatibility relationships between drugs, the developmental relationships between diagnoses, and the therapeutic relationships between drugs and diagnoses;

[0034] A first graph neural network is constructed to perform global embedding learning on a medical knowledge graph, and the relationship information between nodes is captured through multi-layer graph convolution operations.

[0035] The medical event sequences of COPD patients are mapped to a knowledge graph, and the topological features of the sequences are extracted using a graph traversal algorithm.

[0036] By combining temporal relationships and topological features, a high-dimensional feature vector space reflecting the patient's medical trajectory is constructed, forming a digital profile of high-risk COPD.

[0037] In a preferred embodiment, the multidimensional relationships between defined nodes, including drug compatibility relationships, diagnostic development relationships, and therapeutic relationships between drugs and diagnostics, specifically include:

[0038] The compatibility relationships between the drugs include synergistic effects, antagonistic effects, and adverse reaction relationships;

[0039] The developmental relationships between the diagnoses include complication relationships, outcome relationships, and progression relationships;

[0040] The therapeutic relationship between the drug and the diagnosis is established based on the drug's instructions, defining the indications and contraindications.

[0041] In a preferred embodiment, the risk prediction module inputs a standardized time-series event sequence of the population to be screened into a feature extraction model, calculates the matching degree with the digital profile of high-risk COPD, and generates an individualized COPD risk probability prediction value, specifically including:

[0042] The standardized temporal event sequence of the patients to be screened is mapped onto the knowledge graph to form a subgraph structure of the patients to be screened;

[0043] A second graph neural network is constructed as a feature extraction model. The subgraph structure of the patient to be screened is input to obtain a high-dimensional feature vector of the patient's medical trajectory.

[0044] The second graph neural network is constructed based on the model structure and model parameters of the first graph neural network, and the model parameters remain fixed during the feature extraction process;

[0045] Calculate the distance metric between the high-dimensional feature vector of the medical trajectory of the patient to be screened, which is based on the output of the feature extraction model, and the feature vector of the known digital profile of COPD patients in the spatial dimension.

[0046] The distance metric is converted into a risk probability prediction value, and the result is an individualized prediction of the risk of developing COPD.

[0047] In a preferred embodiment, the screening output module generates a screening risk report when the predicted risk probability value continuously exceeds a preset risk threshold, specifically including:

[0048] Establish a risk trend monitoring mechanism to continuously track changes in the predicted risk probability within a specified time window;

[0049] When the predicted risk probability value continuously exceeds the set threshold within a continuous monitoring period, a COPD screening risk report is triggered.

[0050] The screening risk report is linked to the corresponding patient identification information and transmitted to the designated respiratory medicine department's medical management platform through a standard data interface.

[0051] The technical effects and advantages of the COPD patient screening system for respiratory medicine of the present invention are as follows:

[0052] By leveraging underutilized structured data such as drug delivery and disease diagnosis codes in existing medical information systems, a COPD screening system based on time-series patterns and a medical knowledge graph was constructed. Its technical advantages lie in systematically achieving low-cost, non-invasive, and efficient early risk warning; by serializing fragmented medical events into standardized time-series data and identifying characteristic medication and diagnostic patterns related to COPD, the objectivity and proactiveness of screening are significantly improved; furthermore, by constructing a medical knowledge graph and applying graph neural network technology, medical prior knowledge is deeply integrated with individual patient medical trajectories to generate highly discriminative feature vectors and digital profiles. This eliminates reliance on single questionnaires or lung function tests requiring high patient cooperation for risk assessment, greatly enhancing accessibility and coverage in primary healthcare and large-scale population screening scenarios; the fixed parameters of the feature extraction model ensure fair measurement of all individuals to be screened within a unified feature space, resulting in stable and reliable calculations; finally, by dynamically monitoring risk probabilities and automatically generating screening results, it provides a precise and efficient decision support tool for early intervention and hierarchical diagnosis and treatment in respiratory medicine. Attached Figure Description

[0053] Figure 1 This is a schematic diagram of the structure of a COPD patient screening system for respiratory medicine according to the present invention. Detailed Implementation

[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0055] Example 1

[0056] Figure 1 The present invention discloses a screening system for COPD patients in respiratory medicine, comprising:

[0057] The data extraction module is used to obtain the medical diagnosis records of the target population, which includes COPD patients of different age groups, and the medical diagnosis records include drug delivery records and lung disease diagnosis records.

[0058] The data processing module is used to construct a medical event sequence for each COPD patient according to the medical diagnosis record time, standardize the medical event sequence, and construct a standardized time-series event sequence.

[0059] The pattern recognition module is used to identify common medication and diagnosis sequence patterns in the historical data of COPD patients based on standardized time-series event sequences.

[0060] The profile building module is used to construct a medical knowledge graph based on medication and diagnosis sequence patterns. It uses graph neural networks to learn the feature embedding vectors of the medical trajectories of COPD patients and builds a digital profile of high-risk COPD patients.

[0061] The risk prediction module is used to input the standardized time-series event sequence of the population to be screened into the feature extraction model, calculate the matching degree with the digital profile of high COPD risk, and generate an individualized COPD risk probability prediction value.

[0062] The screening output module is used to generate a screening risk report when the predicted risk probability value continuously exceeds the preset risk threshold.

[0063] The data extraction module obtains the medical diagnosis records of the target population.

[0064] The data field scope of electronic prescriptions in the medical information database is clearly defined, and these fields serve as the source of drug delivery records. Specifically, electronic prescriptions are typically automatically generated by hospital pharmacies or drug delivery centers after a patient's visit, and include the generic name of the drug, the specific delivery time, dosage, and frequency of use. During data extraction, prescription records are retrieved sequentially according to unified search criteria. Each prescription record is uniquely located using the patient's identification number to ensure that the extracted data is not overlapping or omitted. For generic drug names, a standardized mapping is performed based on the National Pharmacopoeia or WHO drug standards, converting any common or brand names that may exist in different hospitals or prescription systems into universally accepted standard names to avoid inconsistencies in subsequent analysis. Dosage and frequency of use are recorded completely according to the original prescription data, and corresponding structured fields are established to ensure accurate reflection of patients' medication habits during subsequent time-series modeling. Delivery times are recorded using a unified timestamp format.

[0065] After the medication delivery records are extracted, electronic medical record data related to lung diseases is retrieved from the medical information database. Electronic medical records are typically generated by the attending physician during outpatient visits, hospitalizations, or follow-up examinations, and contain key information such as the diagnosis name, diagnosis time, and diagnosis result. During the extraction process, search criteria for lung diseases should be established, such as using diagnostic entries belonging to chronic obstructive pulmonary disease and its related complications from the national standard disease classification catalog (e.g., ICD coding) as the filtering scope, thereby limiting the boundaries of data extraction. When extracting diagnosis names, standardized medical names should be used consistently to ensure that synonyms or abbreviations used in different hospitals do not cause subsequent data matching discrepancies. The diagnosis time uses the medical record generation time field, which needs to be converted to a unified standardized timestamp format, ensuring that the time zone is consistent with the medication delivery records. The diagnosis result information needs to fully record the doctor's confirmed, suspected, or follow-up results to reflect the patient's disease progression at different stages. When saving the data, each diagnosis record is indexed by the patient's identification number and automatically arranged in chronological order for subsequent matching with medication delivery records. This process generates a raw dataset of lung disease diagnoses for each patient, ensuring the integrity of the data source and the consistency of medical terminology.

[0066] After extracting medication delivery records and lung disease diagnosis records, the two types of data are correlated and matched using the patient's identification number to establish a preliminary medical record for each patient. During implementation, the patient's identification number is used as a unique index to perform a bidirectional match between medication delivery data and corresponding lung disease diagnosis data. Missing data is filled with empty fields based on the identification number to ensure the integrity of the record structure remains unaffected. For cases where the same patient has multiple diagnoses and medications at different times, these are integrated chronologically to form a sequence set containing multiple records. The medical record data structure generally includes three main parts: a basic identification index, a medication delivery sub-table, and a diagnosis record sub-table. The medication sub-table records all extracted medication delivery information for the patient, and the diagnosis sub-table records all extracted diagnosis information for the patient. Both are sorted using a unified timestamp to ensure direct chronological output during retrieval. To avoid cross-patient data mixing, each record undergoes identification verification during data construction to ensure uniqueness and consistency.

[0067] The target population is segmented according to pre-defined age stratification criteria, and data on COPD patients meeting the criteria are screened to ultimately establish a medical diagnosis record set for the target population. During implementation, clearly defined age stratification intervals are first established. For example, following commonly used international epidemiological grouping methods, the population is divided into strata such as "40-49 years old," "50-59 years old," "60-69 years old," and "70 years old and above," with each stratum being a closed interval without overlap or omissions. Subsequently, the birth date or age field for each patient is extracted from the preliminary medical records and compared with the current analysis time to calculate the actual age, thereby determining their stratum. For records with incomplete birth dates, the birth year is supplemented based on the ID card number or medical card number to ensure no omissions in the screening due to missing data. After stratification, patients within each stratum are further screened, retaining only data from patients who meet the criteria for a confirmed COPD diagnosis or multiple related diagnoses, and removing records due to misdiagnosis or other diseases. Ultimately, the medical diagnostic record set constructed according to the stratification criteria is indexed by stratification labels. Each stratification corresponds to a data subset, which includes the medical records of all eligible patients within that age group.

[0068] In the data processing module, a medical event sequence for each COPD patient is constructed according to the medical diagnosis record time, and the medical event sequence is standardized to construct a standardized time-series event sequence.

[0069] The medical diagnosis records for each patient are arranged chronologically according to the time of diagnosis to form a medical event sequence. To achieve this, all diagnosis and medication information for each patient needs to be compiled, with the time field used as the sole sorting criterion. The time field typically originates from the electronic prescription issuance time and the electronic medical record diagnosis time. The time recording formats of these raw data may differ; for example, some records may use "year-month-day," while others may use "year-month-day hour:minute:second." Before sorting, a unified format must be used to ensure time accuracy down to the second. Subsequently, using the patient's identification number as an index, all events occurring during the patient's entire treatment process are fully retrieved and arranged chronologically to form a continuous event trajectory. During this process, if multiple events occur at the same time, a priority sorting principle needs to be established, such as prioritizing diagnosis events before medication delivery events, to avoid confusion in the order of events at the same time point.

[0070] After chronologically arranging the event sequence, the drug information and diagnostic results in the sequence are standardized and coded to eliminate discrepancies caused by different medical institutions or recording methods. First, generic drug names are standardized and uniformly converted to Anatomical, Therapeutic, and Chemical Classification (ATC) codes. During implementation, a drug mapping table should be established based on authoritative pharmacopoeias or international coding directories. For example, different names such as "aminophylline" and "theophylline sustained-release tablets" should be uniformly mapped to the same ATC code. For drugs without a direct mapping relationship, supplementary mappings are made based on pharmacological effects and clinical usage, and the original records are retained in the mapping table to ensure data traceability. Second, diagnostic results are standardized, mapping different expressions of diagnosis in the original medical records to a unified disease classification code, such as ICD-10 codes. For example, "COPD," "chronic obstructive pulmonary disease," and "COPD" are all classified under the same standard code item. During implementation, a mapping rule base needs to be constructed to include common synonyms, abbreviations, and regional expressions to ensure accurate standardization of diagnostic data from different sources.

[0071] The occurrence time of events is standardized and formatted to form a standardized timestamp representation. During implementation, a unified time format specification should be established, such as adopting internationally accepted time formats, and stipulating that all time records use the 24-hour clock to avoid confusion due to regional differences. Simultaneously, for records lacking specific times, they should be supplemented in a reasonable manner; for example, if only a date field is available, the time should be supplemented to "00:00:00" of the current day, and the data annotation should indicate that the record has date-level precision. After standardizing the timestamp format, the coded drug events, diagnostic events, and timestamps are combined to form a standardized medical event sequence with complete temporal information. In the specific construction process, each event consists of three parts: drug or diagnostic code, event occurrence timestamp, and patient identity index. After construction, the sequence is sorted primarily by timestamp to ensure that the event occurrence trajectory can be completely reconstructed when retrieved.

[0072] In the pattern recognition module, based on a standardized sequence of time events, medication and diagnosis sequence patterns that occur together within a preset time period before diagnosis are identified in the historical data of COPD patients.

[0073] The system processes historical medical event data for COPD patients, using their diagnosis event as a dividing point. The diagnosis event is determined based on the diagnosis record time field in the electronic medical record, such as the time when a doctor explicitly diagnoses "chronic obstructive pulmonary disease" during an outpatient or inpatient visit. This time point serves as the dividing line on the timeline, dividing the patient's medical event sequence into pre-diagnosis and post-diagnosis parts. Within the pre-diagnosis data, an observation period is further defined. The method for setting the observation period should be clear and universally applicable, such as 24 or 36 months prior to diagnosis. This setting can be determined based on the common time span between pre-COPD medication use and symptom onset in clinical studies. This setting ensures that the extracted data range covers the patient's complete medical trajectory before the onset of the disease while avoiding excessively long timeframes that could lead to data noise. In practice, the system traces back the set observation period length based on each patient's diagnosis timestamp, extracting all standardized time-series event records within that time period to form a patient-specific observation dataset.

[0074] After extracting data from the pre-diagnosis observation period, sequence pattern mining was performed on these standardized time-series event sequences. Specifically, using each patient's medical event sequence as input, the medication delivery records and diagnostic records were traversed chronologically, and a pattern mining algorithm was used to identify recurring event combinations across multiple patients. The event combinations were categorized into three main types: first, sequences composed of continuous medication delivery records, such as repeated prescriptions of the same type of bronchodilator at multiple consecutive time periods; second, sequences composed of continuous lung disease diagnostic records, such as repeated diagnoses of "chronic bronchitis" or "asthmatic bronchitis" at different consultation times; and third, sequences where medication delivery records and lung disease diagnostic records alternate, such as a diagnosis followed by a period of medication prescriptions, and then a follow-up diagnosis. During the identification process, clear definition criteria for event combinations were established, such as requiring at least two events for continuity, and ensuring that alternating sequences strictly adhered to the medication-diagnosis or diagnosis-medication order. Through traversal and comparison, recurring specific event combinations could be screened from a large amount of historical data, and these combinations were labeled as candidate patterns. After obtaining candidate event combination sequences, a significance test is performed to confirm their representativeness and stability in the COPD patient population. A significance threshold is set, and the method for setting this threshold should be clearly defined, for example, using a frequency higher than 20% or 30% of the total sample patients as the significance criterion. Based on this threshold, frequency statistics are performed on the candidate event combinations, eliminating combinations below the threshold and retaining only those combinations that appear multiple times in a large sample to ensure that the selected patterns are statistically significant. During this process, the repetition of event combinations needs to be considered; for example, multiple repeated combinations appearing in the same patient's sequence are only counted once to avoid individual patient data excessively influencing the overall statistical results. The final selected event combinations are labeled as characteristic sequence patterns. These patterns are characterized by high frequency of occurrence during the pre-diagnosis observation period and can truly reflect the common characteristics of the patient population.

[0075] These characteristic patterns were categorized and organized to establish a COPD-related sequence pattern library. Specifically, the classification criteria included both the frequency and structure of event combinations. Grouping was primarily based on frequency, with high-frequency patterns grouped into one category and medium-frequency patterns into another, ensuring the library reflected different levels of importance. Secondly, a secondary classification based on combination structure was used, for example, establishing category labels for pure drug sequences, pure diagnostic sequences, and drug-diagnostic alternation sequences. During the organization process, each pattern entry required detailed recording of its specific combination content, frequency of occurrence, involved drug or diagnostic codes, and the number of people covered in the sample. The final sequence pattern library was presented in database or structured table format.

[0076] In the profile building module, a medical knowledge graph is constructed based on the medication and diagnosis sequence pattern, and a graph neural network is used to learn the feature embedding vector of the medical trajectory of COPD patients to establish a high-risk digital profile of COPD.

[0077] The previously constructed COPD-related sequence pattern library serves as the foundation for knowledge extraction, thereby establishing the overall framework of a medical knowledge graph. In practice, characteristic event combinations are read sequentially from the pattern library, and the drug codes, diagnostic codes, and temporal relationships between events are decomposed and mapped to the basic nodes and edges of the graph. Node types are divided into three categories: drug nodes, diagnostic nodes, and temporal relationship nodes. Drug nodes are derived from standardized codes of generic drug names, with each drug generating a separate node; diagnostic nodes are derived from standardized disease classification codes, with each lung disease diagnosis establishing a corresponding node; temporal relationship nodes are generated from the sequential relationships between events, used to clarify the order of occurrence of different medical events. After establishing the nodes, the time span information of events in the sequence pattern needs to be incorporated into the attributes of the temporal relationship nodes, such as labeling "interval of one week" or "interval of one month," so that it not only represents the sequence relationship but also includes the time length feature. Subsequently, drug nodes and diagnostic nodes are connected in a directed manner according to the order of appearance in the original pattern and associated through temporal relationship nodes, thereby accurately reflecting the common drug and diagnostic interaction paths of the patient group before diagnosis in the graph structure. The resulting medical knowledge graph contains rich node types and edge attributes, and can comprehensively represent the temporal structure and semantic relationships of COPD-related medical behaviors.

[0078] After the knowledge graph's node system is constructed, multi-dimensional relationships are defined between different types of nodes to ensure that the graph not only reflects a single temporal sequence but also embodies complex connections in medical semantics. First, compatibility relationships are established between drug nodes. These relationships are further subdivided into three categories: synergistic effects, antagonistic effects, and adverse reactions. For example, bronchodilators and inhaled corticosteroids often have synergistic effects clinically, while β-receptor agonists and certain β-blockers have antagonistic relationships. Drug combinations with known interactions leading to increased side effects are marked as adverse reaction relationships. Second, developmental relationships are defined between diagnostic nodes, including complication relationships, outcome relationships, and progression relationships. For example, "chronic bronchitis" and "emphysema" are marked as progression relationships, "COPD" and "pulmonary heart disease" are marked as complication relationships, and changes in some diseases during recovery or disease reversal are marked as outcome relationships. Finally, treatment relationships are established between drug nodes and diagnostic nodes. These relationships are based on clear records in drug instructions and clinical indications and are divided into indication relationships and contraindication relationships. For example, an indication relationship is established between aminophylline and "chronic obstructive pulmonary disease," while a contraindication relationship may exist between certain drugs containing specific components and "acute exacerbation of bronchial asthma." Through the construction of the above multidimensional relationships, the knowledge graph not only possesses information in the time dimension, but also has multiple semantic levels of clinical treatment, disease progression, and drug compatibility.

[0079] A first graph neural network is constructed based on an established medical knowledge graph to perform global embedding learning on the entire graph, capturing multi-level association information between drug nodes, diagnosis nodes, and temporal relationship nodes. During construction, the overall network architecture is determined, using multi-layer graph convolution operations as the basic unit. Each convolution layer takes the adjacency relationship of a node as input and outputs the updated feature vector of the node. Specifically, the adjacency matrix and initial feature vectors of nodes in the medical knowledge graph are used as input, and convolution operations are performed layer by layer. The convolution kernel function completes the update by aggregating the features of neighboring nodes. To ensure the integrity of information propagation, at least three convolutional layers are required. The first layer focuses on capturing local neighborhood features, while the second and third layers gradually expand to a larger graph structure to form an embedding representation under multi-hop neighbor relationships. During training, all nodes are fully traversed to ensure that drug, diagnosis, and temporal nodes are updated simultaneously, avoiding isolated subgraphs. After embedding, each node in the graph is mapped to a high-dimensional vector space, which comprehensively reflects the node's own attributes and its structural relationships with other nodes.

[0080] After completing the global embedding of the knowledge graph, the medical event sequences of COPD patients are mapped into the knowledge graph to extract the topological features corresponding to each sequence in the graph structure. Specifically, using the standardized medical event sequence of the patient as input, the corresponding entities for each drug node and diagnostic node in the sequence are found in the knowledge graph. Based on the chronological order of the sequence, a directed path is formed within the graph from the first event node to the last event node. During the mapping process, if an event node does not have a corresponding node in the knowledge graph, it is expanded according to preset mapping rules, adding the node to the graph and establishing an edge connection with existing nodes. After mapping, the path is structurally analyzed using a graph traversal algorithm. During traversal, not only are the adjacency relationships of nodes on the path extracted, but also the temporal relationships are considered to ensure that the chronological order of the sequence is reflected simultaneously with the topological structure. When extracting features, the embedding vectors of nodes, edge type information, and the length and branching of the path are collected and combined into a set of topological features for the sequence.

[0081] By fusing temporal relationships with topological features, a high-dimensional feature vector space is constructed that comprehensively reflects a patient's medical trajectory, ultimately forming a digital profile of high-risk COPD. Specifically, the topological features obtained through graph traversal are combined with corresponding temporal attributes from the patient's event sequence. These temporal attributes include the time interval between events, the duration of events, and the frequency of events; these attributes are concatenated with the topological feature vector as numerical vectors. Subsequently, the concatenated feature set is normalized to ensure that features of different dimensions can be uniformly calculated within the same vector space. After normalization, the combined features are input into the high-dimensional feature space through a feature mapping function, forming a comprehensive representation with a dimension greater than the original node embedding vector. This high-dimensional feature vector not only contains the medical attributes of individual nodes but also integrates temporal evolution and structural patterns, thus comprehensively reflecting the patient's medical trajectory characteristics before diagnosis. Finally, this vector is used as the main expression of the high-risk COPD digital profile and stored in a feature vector library.

[0082] In the risk prediction module, the standardized time-series event sequence of the population to be screened is input into the feature extraction model to calculate the matching degree with the high-risk digital profile of COPD and generate an individualized COPD risk probability prediction value.

[0083] The standardized temporal event sequence of the patient to be screened is mapped onto a medical knowledge graph, thus forming a subgraph structure corresponding to the patient. In practice, the standardized medical event sequence of the patient is used as input. The sequence contains uniformly coded drug information and diagnostic information, along with standardized timestamps. Drug codes in the sequence are mapped one by one to drug nodes in the knowledge graph, and diagnostic codes are mapped to diagnostic nodes. Directed connections consistent with the sequence are established within the graph according to the chronological order of the timestamps. During the mapping process, if a node does not yet exist in the knowledge graph, the graph is expanded according to preset mapping rules, and the node is added and connected with related nodes by edges, ensuring the integrity and continuity of the mapping. Subsequently, the path corresponding to the patient's event sequence is extracted and defined as a subgraph structure. This subgraph contains drug nodes, diagnostic nodes, and temporal relationship nodes related to the patient's medical events, while maintaining the chronological order of events and semantic connections within the graph.

[0084] After obtaining the subgraph structure, a second graph neural network is constructed as a feature extraction model to process the subgraph as input and output a high-dimensional feature vector. Specifically, the structure of the second graph neural network is built based on the architecture and parameters of the first graph neural network, maintaining consistency in the number of convolutional layers, aggregation method, and neighborhood sampling strategy. The difference lies in that the first network is used for global graph embedding training, while the second network is only used for subgraph feature extraction, and its parameters remain fixed and are not updated. In practice, the subgraph structure is used as input, sequentially fed into multiple layers of graph convolutional units. In each layer, features of nodes in the subgraph are aggregated through adjacency relationships, generating new embedding representations layer by layer. Because the parameters are fixed, the features output by the second network maintain consistency with the space learned globally by the first network, thus ensuring feature comparability. After multiple convolutional and aggregation operations, a high-dimensional feature vector representing the patient's medical trajectory is finally obtained, which simultaneously contains the temporal relationship and topological semantic information of the sequence.

[0085] After obtaining the high-dimensional feature vector of the patient to be screened, it is necessary to calculate the distance metric between this vector and the feature space of known COPD patient digital profiles, and generate a risk probability prediction value based on this. In practice, a feature vector space is first established, which stores the feature vectors of high-risk COPD digital profiles generated from confirmed patient data. The high-dimensional vector of the patient to be screened is then used to calculate the distance between each known profile vector in this space. Common distance definitions include Euclidean distance, cosine similarity, and Mahalanobis distance; the specific method used needs to be uniformly set before implementation, for example, choosing Euclidean distance as the sole metric. After calculation, the resulting distance values ​​are normalized and converted into probability values. Normalization can be achieved by setting maximum and minimum distance boundaries, linearly mapping the values ​​to the interval between 0 and 1. If the patient's feature vector is close to the high-risk profile, the mapping result shows a higher probability value; if the distance is far, the probability value is lower. The final output risk probability prediction value serves as the patient's individualized COPD risk result and is linked to the patient's identity for risk management or screening result generation.

[0086] In the screening output module, a screening risk report is generated when the predicted risk probability value continuously exceeds the preset risk threshold.

[0087] A risk trend monitoring mechanism is established to continuously track the predicted risk probability values ​​of patients to be screened. In practice, a fixed time window is set, such as one week or one month as the monitoring cycle length, and the time window setting must be consistent with the actual clinical follow-up frequency. Using the time window as a unit, the predicted risk probability values ​​output by the prediction model for each patient are stored and updated, and the predicted values ​​at each time point are arranged in chronological order to form a risk probability change sequence. Subsequently, the monitoring mechanism needs to perform trend calculations on this sequence to determine its fluctuations in different periods, such as continuous increases, sustained high levels, or stable low levels. During implementation, it is ensured that all patient data are collected with the same temporal precision, for example, using a uniform day as the smallest unit, to avoid affecting trend analysis due to incomplete data. To ensure the accuracy of the tracking results, an abnormal data removal process is established to exclude fluctuations caused by temporary abnormal visits or short-term data gaps.

[0088] After continuously tracking the risk probability trend, a threshold determination is made for the predicted values ​​during the continuous monitoring period. In practice, a clear risk threshold is pre-set, such as 0.6 or 0.7 as the high-risk cutoff point for COPD. The threshold setting must reference clinical trial data or epidemiological statistics to ensure the scientific validity and consistency of the determination criteria. Subsequently, in each monitoring period, the predicted risk probability value for each patient is compared with the set threshold. When a patient's predicted value remains above the threshold for two or more consecutive periods, the patient is determined to be in a persistently high-risk state. Upon triggering this determination condition, the system immediately generates a screening risk report. The report must include the patient's basic identity information, a numerical record of the risk probability, a trend curve, and the time period exceeding the threshold, ensuring that the report content clearly reflects the risk status. During report generation, the data also needs to be integrated according to a predefined report format, such as presenting it in the form of text descriptions combined with charts, to facilitate subsequent review by medical personnel.

[0089] After the screening risk report is generated, it is bound to the patient's identification information and transmitted to the designated respiratory medicine medical management platform via a standard data interface. In implementation, the patient's identification number or medical card number is used as a unique index to bind the risk report to this index field. After binding, the report data is structured and encapsulated, for example, storing information such as the report title, risk value, time series, and trend description in a unified message format. Subsequently, the standardized data interface is invoked to complete cross-platform transmission. The interface protocol can adopt the HL7 or FHIR standards commonly used in the medical industry to ensure data consistency and resolvability during transmission. Before transmission, the report undergoes integrity verification and timestamp signing to ensure that no data is omitted or tampered with during transmission. After transmission, the medical management platform can receive and update the report in the patient's file in real time, allowing clinicians to directly access and view it during follow-up or screening phases.

[0090] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0091] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. The semiconductor medium can be a solid-state drive.

[0092] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0093] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0094] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0095] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0096] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0097] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0098] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0099] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A system for screening of patients with chronic obstructive pulmonary disease for respiratory medicine, characterized in that, The method comprises the following steps: a data extraction module is used to obtain medical diagnosis records of a target population, wherein the target population comprises patients with chronic obstructive pulmonary disease (COPD) of different age distributions, and the medical diagnosis records comprise drug dispensing records and pulmonary disease diagnosis records; a data processing module is used to construct a medical event sequence of each COPD patient according to the time of the medical diagnosis records, and to perform standardization processing on the medical event sequence to construct a standardized time sequence event sequence; a pattern recognition module is used to identify a drug and diagnosis sequence pattern that commonly appears within a preset time period before diagnosis in historical data of the COPD patient population based on the standardized time sequence event sequence; an image establishment module is used to construct a medical knowledge graph based on the drug and diagnosis sequence pattern, learn a feature embedding vector of a medical trajectory of a COPD patient by using a graph neural network, and establish a high-risk digital image of COPD; a risk prediction module is used to input a standardized time sequence event sequence of a to-be-screened population into a feature extraction model, calculate a matching degree with the high-risk digital image of COPD, and generate an individualized COPD risk probability prediction value; a screening output module is used to generate a screening risk report when the risk probability prediction value continuously exceeds a preset risk threshold. In the image establishment module, the medical knowledge graph is constructed based on the drug and diagnosis sequence pattern, the feature embedding vector of the medical trajectory of the COPD patient is learned by using the graph neural network, and the high-risk digital image of COPD is established, which specifically comprises: a medical knowledge graph is constructed based on a sequence pattern library related to COPD, and the graph node types include drug nodes, diagnosis nodes and time sequence relationship nodes; a multi-dimensional association relationship between nodes is defined, including compatibility relationships between drugs, development relationships between diagnoses, and treatment relationships between drugs and diagnoses; a first graph neural network is constructed to perform global embedding learning on the medical knowledge graph, and the association information between nodes is captured through multi-layer graph convolution operations; the medical event sequence of the COPD patient is mapped into the knowledge graph, and the topological structure features of the sequence are extracted through a graph traversal algorithm; combining the time sequence relationship and the topological features, a high-dimensional feature vector space reflecting the medical trajectory of the patient is constructed, and a high-risk digital image of COPD is formed; In the risk prediction module, the standardized time sequence event sequence of the to-be-screened population is input into the feature extraction model, the matching degree with the high-risk digital image of COPD is calculated, and an individualized COPD risk probability prediction value is generated, which specifically comprises: the standardized time sequence event sequence of the to-be-screened patient is mapped into the knowledge graph to form a subgraph structure of the to-be-screened patient; a second graph neural network is constructed as a feature extraction model, the subgraph structure of the to-be-screened patient is input, and a high-dimensional feature vector of the medical trajectory of the to-be-screened patient is obtained; the second graph neural network is constructed based on the model structure and model parameters of the first graph neural network, and the model parameters remain fixed in the feature extraction process; the distance measurement value between the high-dimensional feature vector of the medical trajectory of the to-be-screened patient output based on the feature extraction model and the feature vector space of the known digital image of the COPD patient is calculated; the distance measurement value is converted into a risk probability prediction value, and an individualized COPD incidence risk prediction result is output.

2. The COPD screening system for pneumology use according to claim 1, characterized in that, The data extraction module, the medical diagnosis record of the target population specifically comprises: Extracting drug dispensing records from electronic prescriptions in the medical information database, including drug generic name, dispensing time, dosage specification and medication frequency information; Extracting lung disease diagnosis records from electronic medical records in the medical information database, including diagnosis name, diagnosis time and diagnosis result information; Associating and matching the drug dispensing records with the lung disease diagnosis records through the patient identification number to establish a preliminary medical archive of patients as a unit; Based on the pre-set age stratification standard, the target population is divided into multiple age subgroups, and the data of COPD patients meeting the requirements of each age subgroup is screened from the preliminary medical archive to establish the medical diagnosis record set of the target population.

3. The COPD screening system for pneumology use according to claim 1, characterized in that, In the data processing module, the medical event sequence of each COPD patient is constructed according to the medical diagnosis record time, and the medical event sequence is standardized, and the standardization of the time sequence event sequence specifically comprises: Arranging the medical diagnosis records of each patient in the order of diagnosis time to form a medical event sequence; Standardizing and coding the drug information in the sequence, and uniformly converting the drug generic name into the coding of the anatomical, therapeutic and chemical classification system; Standardizing and coding the diagnosis results in the medical event sequence, and uniformly mapping the diagnosis result coding under different standards to the disease classification standard coding; Establishing a standard timestamp format to convert the occurrence time of all medical events into a unified time representation form; Based on the standardized drug coding, diagnosis coding and timestamp, a standardized medical event sequence containing complete time sequence information is constructed.

4. The COPD screening system for pulmonology as claimed in claim 1 wherein, In the pattern recognition module, based on the standardized time sequence event sequence, the medication and diagnosis sequence patterns commonly appearing within a pre-set time period before diagnosis are identified in the historical data of the COPD patient population, specifically comprising: Taking the diagnosis event as a segmentation point, extracting all standardized time sequence event sequences within a specified observation period before diagnosis from the historical data of COPD patients; Using sequence pattern mining algorithm to analyze the event sequence and identify specific event combinations that repeatedly appear in the COPD patient population; The event combination includes a continuous drug dispensing record sequence, a continuous lung disease diagnosis record sequence, and a combination sequence in which drug dispensing records and lung disease diagnosis records appear alternately; Significance test is performed on the identified event combination sequence, and the event sequence with an appearance frequency reaching a pre-set significance threshold is screened and labeled as a characteristic sequence pattern; The characteristic sequence patterns screened are classified according to the appearance frequency and event combination to establish a sequence pattern library related to COPD.

5. The COPD screening system for pulmonology as claimed in claim 1 wherein, The multi-dimensional association relationship between the nodes is defined, including the compatibility relationship between drugs, the development relationship between diagnoses, and the treatment relationship between drugs and diagnoses, specifically comprising: The compatibility relationship between drugs includes synergistic effect, antagonistic effect and adverse reaction relationship; The development relationship between diagnoses includes complication relationship, outcome relationship and progression relationship; The treatment relationship between drugs and diagnoses is the indication relationship and contraindication relationship between drugs and diagnoses based on drug instructions.

6. The COPD screening system for pulmonology as claimed in claim 1 wherein, The screening output module, when the risk probability prediction value continuously exceeds the preset risk threshold, specifically includes: Establish a risk trend monitoring mechanism to continuously track the changes of the risk probability prediction value within a specified time window; When the risk probability prediction value continuously exceeds the set threshold in the continuous monitoring period, trigger the slow lung screening risk report generation; Bind the screening risk report with the corresponding patient identification information and transmit it to the designated respiratory medicine medical management platform through the standard data interface.

Images