Drug template construction and review method based on disease prediction, medium and device

Abstract

The application discloses a kind of based on disease prediction's medication template construction and review method, medium and equipment, the method includes: obtaining medical ticket set and carrying out standardization preprocessing;Standardized data is input into the disease prediction model based on pseudo-label mechanism training, obtains the predicted disease of each ticket;For each predicted disease, aggregate its ticket sample and construct the drug co-occurrence network with point mutual information as weight based on drug co-occurrence relationship;Based on the network, mainstream medication template is automatically induced using graph analysis algorithm;Extract the drug set of the ticket to be reviewed, and compare the mainstream medication template of the corresponding disease in multidimensional such as core drug missing rate, abnormal drug existence, structural similarity and semantic conflict;Comprehensive comparison result calculates rationality comprehensive score, and outputs the structured review result containing risk level.The application can automatically construct medication knowledge system from unlabeled ticket data, realize intelligent and refined review on ticket medication compliance.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and medical invoice processing technology, specifically to a method, medium, and device for constructing and reviewing medication templates based on disease prediction. Background Technology

[0002] External invoice management platforms and similar systems typically only obtain structured data from invoices (such as drug lists and amounts), lacking actual disease diagnoses or medical insurance enrollment information. This "data labeling deficiency" makes it difficult to directly apply traditional supervisory drug use rationality review models that rely on clearly defined disease labels.

[0003] To overcome this limitation, existing technologies attempt to use machine learning methods to predict the possible diseases associated with invoices based on their content (such as medications and departments), thus generating pseudo-labels. However, obtaining only a single disease prediction result is insufficient to support in-depth analysis and judgment of the rationality of medication regimens. A complete review of a medication regimen requires not only knowing "what the possible disease is," but also "what is a conventional, reasonable, and complete medication template for this disease," and "what specific differences or risks exist between the current invoice's medication regimen and this conventional model." Currently, there is a lack of a systematic method that can automatically mine and construct such a disease-related, structured medication knowledge template from unlabeled or weakly labeled invoice data, and use this template for refined and interpretable review of medication rationality. Summary of the Invention

[0004] In view of the above problems, the present invention provides a technical solution for the construction and review of medication templates based on disease prediction, in order to solve the technical problem of how to automatically construct medication knowledge templates associated with diseases in the context of medical invoice data lacking real disease labels, and to realize the automated and intelligent review of the compliance of medication invoices based on this template.

[0005] To achieve the above objectives, in a first aspect, the present invention provides a method for constructing and reviewing medication templates based on disease prediction, the method comprising:

[0006] S1: Obtain a set of medical invoices, perform standardized preprocessing on the structured data in the invoices, and generate standardized invoice data containing a normalized sequence of drug names. The structured data includes a drug list, information on the department visited, and patient profile information.

[0007] S2: Input the standardized invoice data into the disease prediction model trained based on the pseudo-label mechanism to obtain the predicted disease corresponding to each invoice;

[0008] S3: For each predicted disease, aggregate all invoice samples predicted for that disease, and construct a weighted drug co-occurrence network based on drug co-occurrence relationships. The drug co-occurrence network uses Point Mutual Information (PMI) as edge weights.

[0009] S4: Based on the drug co-occurrence network, a graph analysis algorithm is used to automatically summarize the mainstream drug templates. The mainstream drug templates include a core drug set determined by the PageRank algorithm, an auxiliary drug set determined by the community detection algorithm, and an abnormal drug set determined by node centrality analysis.

[0010] S5: Extract the set of drugs from the invoices to be reviewed and perform a multidimensional comparison with the mainstream drug templates for the corresponding predicted diseases. The multidimensional comparison includes core drug missing rate calculation, abnormal drug presence index calculation, structural similarity calculation, and semantic conflict detection.

[0011] S6: Based on the multidimensional comparison results, the comprehensive reasonableness score is calculated by combining the core drug missing rate, abnormal drug presence index, structural similarity and semantic conflict detection results, and the structured review results are output. The structured review results include determining the risk level of the invoice to be reviewed based on the comparison results of the comprehensive reasonableness score and the scoring threshold. The risk level includes any one of reasonable invoice, suspicious invoice and high-risk invoice.

[0012] Furthermore, in step S2, the disease prediction model trained based on the pseudo-label mechanism is trained in the following manner:

[0013] Based on a dynamic reasoning knowledge base for medical coding, rule-based reasoning is performed on unlabeled invoice data to generate a first set of pseudo-labeled diseases;

[0014] Based on a large language model pre-trained with medical text, semantic understanding and reasoning are performed on the same unlabeled ticket data to generate a second set of pseudo-labeled diseases.

[0015] The first set of pseudo-labeled diseases and the second set of pseudo-labeled diseases are merged to generate a fused pseudo-label as the training target;

[0016] A multi-label disease classification model is trained using the fused pseudo-labels as the target, resulting in a trained disease prediction model.

[0017] Furthermore, in step S3, constructing a weighted drug co-occurrence network based on drug co-occurrence relationships includes:

[0018] Statistical prediction of the co-occurrence frequency of drug pairs among all invoices belonging to the same disease category;

[0019] The point mutual information (PMI) value is calculated based on the co-occurrence frequency of the drug pairs. The PMI value is denoted as... The calculation formula is as follows:

[0020] ;

[0021] in, Indicates medicine and medicines The estimated joint probability of simultaneous occurrence of the same document in the same document. and They represent medicines. and medicines An estimate of the marginal probability of a single ticket appearing alone, the probability estimate being calculated based on all tickets whose predictions belong to the same disease category;

[0022] Will As edge weights in the drug co-occurrence network, a weighted undirected graph network is constructed to obtain the drug co-occurrence network.

[0023] Furthermore, the core drug set is determined by the following method: based on the PMI edge weights of the drug co-occurrence network, the PageRank authority score of each drug node is calculated, and drugs with authority scores higher than a first preset threshold are included in the core drug set.

[0024] The auxiliary drug set is determined by performing community detection on the drug co-occurrence network to identify multiple drug sub-clusters; for each drug sub-cluster, drugs within the drug sub-cluster that are not included in the core drug set are included in the auxiliary drug set.

[0025] The abnormal drug set is determined by: calculating the degree centrality of each drug node in the drug co-occurrence network and counting the frequency of each drug in the medical bill set; and including drugs with a degree centrality lower than a second preset threshold and a frequency of occurrence in the medical bill set lower than a third preset threshold into the abnormal drug set.

[0026] Furthermore, the calculation of the core drug missing rate includes: calculating the proportion of the number of missing core drugs in the pending invoices to the total number of core drugs in the core drug set;

[0027] The calculation of the abnormal drug index includes: counting the number of abnormal drugs in the invoices to be reviewed;

[0028] Structural similarity calculation includes: calculating the similarity between the set of drugs invoices to be reviewed and mainstream drug templates based on Jaccard similarity;

[0029] Semantic conflict detection includes: based on a drug pharmacology knowledge base, detecting whether there are drug combinations with pharmacological semantic conflicts in the invoices to be reviewed, and outputting a quantitative index of semantic conflict.

[0030] Furthermore, the formula for calculating the comprehensive rationality score is as follows:

[0031] ;

[0032] in, The overall score indicates the reasonableness of the assessment. Indicates the missing rate of core drugs. Indicates the quantity of abnormal drugs. Indicates structural similarity. This represents the no-conflict confidence score calculated based on the semantic conflict detection results, with a value range of [0,1], where 1 indicates that no conflict was detected. , , and These represent the weighting coefficients.

[0033] Furthermore, step S3 includes:

[0034] S31: For each predicted disease, based on the key physiological and pathological indicators in the patient profile information, all ticket samples predicted to be that disease are clustered into multiple patient subgroups.

[0035] S32: For each patient subgroup, construct a drug co-occurrence network with time or stage labels based on the treatment stage identifier or medication sequence information in the invoice;

[0036] Step S4 includes:

[0037] S41: For each patient subgroup, based on the tagged drug co-occurrence network corresponding to that patient subgroup, summarize the mainstream medication template for each patient subgroup; wherein, the mainstream medication template for each patient subgroup includes the core drug set applicable to that patient subgroup, the staged auxiliary drug subset associated with different stages of diagnosis and treatment, and the abnormal drug set;

[0038] In step S5, the multidimensional comparison includes:

[0039] Based on the patient profile information of the invoices to be reviewed, the corresponding patient subgroup template is matched, and the core drug missing rate, abnormal drug presence index, and auxiliary drug structure similarity with the patient subgroup template are calculated in sequence.

[0040] Furthermore, the semantic conflict detection in step S5 is achieved through deep conflict detection based on the drug knowledge graph, including: deep conflict detection based on the drug knowledge graph, specifically including: by querying the pre-built drug knowledge graph, identifying and quantifying the pharmacological conflict, adverse reaction superposition risk and metabolic pathway competition relationship between drugs in the prescription to be reviewed, and generating a conflict score;

[0041] Step S6 also includes:

[0042] S61: Calculate a comprehensive rationality score based on the core drug missing rate, abnormal drug presence index, phased auxiliary drug structural similarity, and the conflict score of the deep conflict detection.

[0043] S62: Determine the risk level of the bill to be reviewed based on the comparison results of the comprehensive reasonableness score and the scoring threshold;

[0044] S63: Generate a structured review result and an interpretability report, wherein the interpretability report shall at least list the key comparison items that led to the score change, the drug involved, the template rules or knowledge graph relationships cited, and indicate the specific conflict type or missing content.

[0045] In a second aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for constructing and reviewing medication templates based on disease prediction as described in the first aspect of the present invention.

[0046] In a third aspect, the present invention provides an electronic device having a computer program stored thereon, including a processor and a storage medium, wherein the computer program is stored on the storage medium, and when the computer program is executed by the processor, it implements the method for constructing and reviewing medication templates based on disease prediction as described in the first aspect of the present invention.

[0047] Unlike existing technologies, the above-mentioned technical solution involves a method, medium, and device for constructing and reviewing medication templates based on disease prediction, aiming to solve the problem of difficulty in automating the review of the rationality of medication use in medical invoices when there is a lack of real disease labels. The method includes: acquiring a set of medical invoices and performing standardized preprocessing; inputting the standardized data into a disease prediction model trained based on a pseudo-label mechanism to obtain the predicted disease for each invoice; for each predicted disease, aggregating its invoice samples and constructing a drug co-occurrence network with point mutual information as weights based on drug co-occurrence relationships; based on this network, automatically summarizing mainstream medication templates using graph analysis algorithms, the templates including a core drug set determined by PageRank, an auxiliary drug set determined by community discovery, and an abnormal drug set determined by node centrality; extracting the drug set of the invoices to be reviewed and performing multi-dimensional comparisons with the mainstream medication templates of the corresponding diseases, including core drug missing rate, abnormal drug presence, structural similarity, and semantic conflict; finally, calculating a comprehensive rationality score based on the comparison results and outputting a structured review result including risk level. This invention can automatically construct a medication knowledge system from unlabeled invoice data, enabling intelligent and refined review of medication compliance on invoices.

[0048] The above description of the invention is merely an overview of the technical solution of the present invention. In order to enable those skilled in the art to better understand the technical solution of the present invention and to implement it based on the description and drawings, and to make the above-mentioned objectives and other objectives, features and advantages of the present invention easier to understand, the following description is provided in conjunction with the specific embodiments and drawings of the present invention. Attached Figure Description

[0049] The accompanying drawings are only used to illustrate the principles, implementation methods, applications, features, and effects of specific embodiments of the present invention and other related contents, and should not be considered as limitations on the present invention.

[0050] In the accompanying drawings of the instruction manual:

[0051] Figure 1 This is a flowchart of the method for constructing and reviewing medication templates based on disease prediction as described in the first exemplary embodiment of the present invention;

[0052] Figure 2 This is a flowchart of the method for constructing and reviewing medication templates based on disease prediction as described in the second exemplary embodiment of the present invention;

[0053] Figure 3 This is a flowchart of the method for constructing and reviewing medication templates based on disease prediction as described in the third exemplary embodiment of the present invention;

[0054] Figure 4 This is a flowchart of the method for constructing and reviewing medication templates based on disease prediction as described in the fourth exemplary embodiment of the present invention;

[0055] Figure 5 This is a schematic diagram of an electronic device according to an exemplary embodiment of the present invention;

[0056] The reference numerals used in the above figures are explained as follows:

[0057] 10. Electronic equipment; 101. Processor; 102. Storage medium. Detailed Implementation

[0058] To explain in detail the possible application scenarios, technical principles, specific feasible solutions, and the objectives and effects that can be achieved by this invention, the following detailed description is provided in conjunction with the listed specific embodiments and accompanying drawings. The embodiments described herein are only used to more clearly illustrate the technical solutions of this invention, and are therefore only examples, and should not be used to limit the scope of protection of this invention.

[0059] In the first aspect, such as Figure 1 As shown, this invention provides a method for constructing and reviewing medication templates based on disease prediction, the method comprising:

[0060] S1: Obtain a set of medical invoices, perform standardized preprocessing on the structured data in the invoices, and generate standardized invoice data containing a normalized sequence of drug names. The structured data includes a drug list, information on the department visited, and patient profile information.

[0061] S2: Input the standardized invoice data into the disease prediction model trained based on the pseudo-label mechanism to obtain the predicted disease corresponding to each invoice;

[0062] S3: For each predicted disease, aggregate all invoice samples predicted for that disease, and construct a weighted drug co-occurrence network based on drug co-occurrence relationships. The drug co-occurrence network uses Point Mutual Information (PMI) as edge weights.

[0063] S4: Based on the drug co-occurrence network, a graph analysis algorithm is used to automatically summarize the mainstream drug templates. The mainstream drug templates include a core drug set determined by the PageRank algorithm, an auxiliary drug set determined by the community detection algorithm, and an abnormal drug set determined by node centrality analysis.

[0064] S5: Extract the set of drugs from the invoices to be reviewed and perform a multidimensional comparison with the mainstream drug templates for the corresponding predicted diseases. The multidimensional comparison includes core drug missing rate calculation, abnormal drug presence index calculation, structural similarity calculation, and semantic conflict detection.

[0065] S6: Based on the multidimensional comparison results, the comprehensive reasonableness score is calculated by combining the core drug missing rate, abnormal drug presence index, structural similarity and semantic conflict detection results, and the structured review results are output. The structured review results include determining the risk level of the invoice to be reviewed based on the comparison results of the comprehensive reasonableness score and the scoring threshold. The risk level includes any one of reasonable invoice, suspicious invoice and high-risk invoice.

[0066] In this embodiment, the medical invoice set refers to the electronic invoice data set generated in scenarios such as medical insurance settlement, outpatient charges, and inpatient settlement. It includes structured fields such as invoicing institution, drug list, treatment items, charge amount, department of visit, and patient basic information. It is the only original data source for the present invention to perform disease prediction, template construction, and rationality review, without relying on external data such as medical record cover page, diagnosis record, and medical insurance enrollment results.

[0067] Standardized preprocessing refers to the process of uniformly cleaning, normalizing, mapping, deduplicating, and aligning the unstructured and semi-structured text and numerical information in medical invoices. The purpose is to eliminate data heterogeneity issues such as drug aliases, abbreviations, common names, non-standard naming, unit differences, and inconsistent project codes, so as to ensure the consistency and accuracy of subsequent model input and statistical analysis.

[0068] Normalized drug name sequence refers to the ordered text sequence formed by converting drugs in the invoice into the national pharmacopoeia standard generic name according to their appearance order and removing interference from dosage form and specification. It is the core input feature for the construction of disease prediction model and drug co-occurrence network, and can preserve the contextual association and treatment combination information between drugs to the greatest extent.

[0069] The disease prediction model trained based on the pseudo-label mechanism refers to a deep learning classification model that generates pseudo-labels through rule reasoning, semantic understanding, and weakly supervised learning in a scenario where there is a complete lack of real disease labels, medical insurance diagnostic data, and medical record annotation data. This model can then be trained to output standard disease codes (ICD-10 / ICD-11 / local medical insurance disease codes) based on drug lists and treatment items.

[0070] Drug co-occurrence network refers to an undirected weighted graph structure constructed with standardized drugs as nodes, drugs appearing together in the same invoice as edges, and point mutual information (PMI) as edge weights. It is used to accurately express the correlation strength between drugs in the treatment of the same disease and to filter out noise interference from high-frequency common drugs.

[0071] The mainstream medication template refers to a standardized medication reference system that is automatically generated for a specific predicted disease by aggregating a large number of invoices for the same disease, statistically analyzing drug characteristics, and mining graph structural patterns. It consists of three parts: a core drug set, an auxiliary drug set, and an abnormal drug set. It is the only criterion for automatically comparing the rationality of medication use.

[0072] Multidimensional comparison refers to simultaneously performing quantitative calculations and logical judgments from four dimensions: the integrity of core drugs, the existence of abnormal drugs, the overall similarity of drug use structure, and the consistency of pharmacological semantics, to achieve a full-dimensional review from the surface drug list to the deep treatment logic.

[0073] The comprehensive rationality score refers to the standardized score in the range of 0 to 1 obtained by merging the results of multidimensional comparison through a weighted formula. It is used to objectively, uniformly and reproducibly measure the rationality of medication use on a single invoice, and automatically classifies it into three risk levels: reasonable, doubtful and high-risk, to support the automated processing of the medical insurance audit system.

[0074] In step S1, a batch of medical invoices is first obtained. The structured data contained in the invoices, such as drug lists, department information, and patient profile information, is parsed, denoised, formatted, and standardized. The original heterogeneous invoices are converted into standardized invoice data with consistent structure, standardized fields, and normalized drug name sequences. This provides stable and high-quality input for subsequent disease prediction and avoids model bias and statistical distortion caused by non-standard data.

[0075] In step S2, the preprocessed standardized invoice data is input into the trained disease prediction model. This model builds a training set based on a pseudo-label mechanism, does not rely on internal hospital medical records or real diagnostic data from the medical insurance bureau, and can complete inference solely based on the information of the invoice itself. It outputs the predicted disease and confidence level corresponding to each invoice, realizing automatic disease classification in unsupervised scenarios and providing a grouping basis for subsequent aggregation of medication templates by disease.

[0076] In step S3, all tickets are aggregated according to the predicted disease. Aggregation is performed on all ticket samples with the same predicted disease. The co-occurrence relationship between drugs is statistically analyzed. Using Point Mutual Information (PMI) as edge weights, a weighted drug co-occurrence network that can truly reflect the strength of drug treatment association is constructed. This network can effectively weaken the statistical interference brought by common drugs such as saline, glucose, and vitamins, and highlight the core drug combination that is truly used for the treatment of this disease.

[0077] In step S4, based on the constructed drug co-occurrence network, various graph analysis algorithms are used to automatically mine and summarize drug templates: the PageRank algorithm is used to calculate the authority score of drugs in the treatment network to determine the core drug set; the community detection algorithm is used to cluster the network to identify clinically significant drug sub-clusters to determine the auxiliary drug set; and node centrality analysis is used to identify low-frequency, weakly associated, and treatment-irrelevant drugs to determine the abnormal drug set; finally, a well-structured mainstream drug template that can be directly used for comparison is formed.

[0078] In step S5, the invoices to be reviewed are analyzed for drugs, and a normalized set of drugs is extracted. Based on the predicted disease, the corresponding mainstream drug template is located, and four core comparison calculations are performed: core drug missing rate, abnormal drug presence index, drug structure similarity, and pharmacological semantic conflict detection. Deep verification is completed from four levels: completeness, non-compliance, structure, and logic.

[0079] In step S6, the above multidimensional comparison results are substituted into the comprehensive scoring formula, and a comprehensive reasonableness score is obtained by weighted calculation. Based on the comparison results between the score and the preset threshold, the invoices are automatically classified into one of the following: reasonable invoices, suspicious invoices, and high-risk invoices. The structured review results containing invoice identifier, predicted disease, score, risk level, abnormal details, and system suggestions are output, thereby realizing the full automation of the medical insurance audit process.

[0080] The aforementioned solution, for the first time, achieves automated review of medication rationality within an external invoice management platform and third-party monitoring system that completely lacks authentic disease labels, medical insurance enrollment results, and high-quality medical record data. This overcomes the strong reliance of traditional supervised learning methods on high-value labeled data, significantly reducing the implementation threshold and data costs of the medical insurance monitoring system. The fully automated design, from drug lists to disease prediction, from aggregation statistics to template construction, and from multidimensional comparison to risk grading, significantly reduces the workload of manual review, shortens the audit cycle, and improves regulatory efficiency. Medication template mining based on drug co-occurrence networks and graph algorithms replaces manual rule formulation, making review standards closer to real clinical medication patterns and reducing misjudgment and omission rates. Multidimensional comparison and quantitative scoring mechanisms ensure that review results are objective, consistent, reproducible, and auditable, effectively addressing the pain points of strong subjectivity, inconsistent standards, and difficulty in large-scale promotion of manual review. It can accurately identify violations such as irrational drug use, over-provisioning of disease groups, and fictitious invoicing, providing efficient and reliable technical support for the security of the medical insurance fund.

[0081] In some embodiments, such as Figure 2 As shown, in step S2, the disease prediction model trained based on the pseudo-label mechanism is trained in the following manner:

[0082] S201: Based on the dynamic reasoning knowledge base used for medical coding, rule reasoning is performed on unlabeled invoice data to generate the first set of pseudo-labeled diseases;

[0083] S202: Based on a large language model pre-trained with medical text, semantic understanding and reasoning are performed on the same unlabeled ticket data to generate a second set of pseudo-labeled diseases;

[0084] S203: Merge the first set of pseudo-labeled diseases and the second set of pseudo-labeled diseases to generate a fused pseudo-label as the training target;

[0085] S204: Train a multi-label disease classification model using the fused pseudo-labels as the target to obtain a trained disease prediction model.

[0086] In this embodiment, the pseudo-label mechanism refers to a weakly supervised learning method that generates approximately credible labels for training deep learning models in scenarios with unlabeled or poorly labeled medical data, without relying on manual annotation and real diagnostic results. This method generates labels through rule reasoning, model semantic understanding, knowledge fusion, etc., and is the core technology of this invention to achieve high-accuracy disease prediction under data-constrained conditions.

[0087] The medical coding dynamic reasoning knowledge base refers to a dynamically updated knowledge base that incorporates ICD-10, ICD-11, national medical insurance disease codes, drug-disease correspondence, treatment item-disease correspondence, and department-disease association rules. It can perform deterministic rule reasoning based on drug combinations, treatment items, and departments in the invoices and output highly reliable disease pseudo-labels.

[0088] Medical text pre-trained large language model refers to a large model that has been pre-trained on a large amount of medical literature, drug instructions, clinical guidelines and electronic medical records. It has the ability to understand medical semantics, text representation and related reasoning. It can extract deep semantic information from unstructured drug lists and generate pseudo-labels for diseases that are difficult to cover by rules.

[0089] Fusion pseudo-labels refer to high-precision weakly supervised labels obtained by fusing the first set of pseudo-labels generated by rule-based reasoning with the second set of pseudo-labels generated by a large language model through conflict resolution, deduplication, weighted voting, and intersection / union. They combine the rigor of rules with the generalization ability of large models.

[0090] Multi-label disease classification models refer to classification models that support the simultaneous output of multiple candidate diseases on a single ticket (adapting to clinical comorbidities, complications, and comorbidities). The output results include disease codes and confidence levels, which are more in line with real medical scenarios.

[0091] This embodiment specifies the training method for the disease prediction model and constructs a dual-path pseudo-label fusion training framework to solve the problem that high-precision classification models cannot be trained with unlabeled invoice data.

[0092] During the model training phase, the model is first based on unlabeled medical invoice data. Through a medical coding dynamic reasoning knowledge base, deterministic rule matching reasoning is performed based on information such as drug combinations, treatment items, and departments visited to generate the first set of pseudo-labeled diseases. This set is characterized by logical rigor, no semantic ambiguity, and high credibility.

[0093] Meanwhile, the same batch of unlabeled invoice data is input into a large language model pre-trained with medical text. By leveraging its semantic understanding and knowledge reasoning capabilities in the medical field, the drug list is subjected to deep semantic analysis to generate a second set of pseudo-labeled diseases. This set can cover rare diseases, complex combinations, and new drug templates not included in the rule base and has strong generalization capabilities.

[0094] The first and second pseudo-label sets are then fused: consistent labels are directly retained, while conflicting labels are resolved through confidence comparison, medical knowledge verification, and voting decisions, ultimately generating high-precision fused pseudo-labels as the model training targets.

[0095] Using standardized invoice data as model input and fused pseudo-labels as training supervision signals, a multi-label disease classification model is trained, enabling it to stably and accurately predict diseases based solely on invoice information output during the inference phase, thus providing a reliable grouping basis for subsequent medication template construction.

[0096] The above solution employs a dual-path pseudo-label fusion training method, enabling the model to maintain high classification accuracy even without any real labels, thus addressing the industry pain point that external invoice platforms cannot obtain medical record data, preventing model training. Rule-based reasoning and large-scale model semantic understanding complement each other, ensuring both the rigor and compliance of the labels while enhancing the model's adaptability to complex medications, rare diseases, and novel treatment options. The multi-label classification structure closely aligns with real-world clinical comorbidities, avoiding mis-aggregation of medication templates caused by single labels and improving the accuracy and practicality of subsequent mainstream medication templates. The entire training process is automated, scalable, and iterative, requiring no manual annotation, significantly reducing model building costs and timelines, making it suitable for large-scale medical insurance regulatory deployments.

[0097] In some embodiments, in step S3, constructing a weighted drug co-occurrence network based on drug co-occurrence relationships includes:

[0098] Statistical prediction of the co-occurrence frequency of drug pairs among all invoices belonging to the same disease category;

[0099] The point mutual information (PMI) value is calculated based on the co-occurrence frequency of the drug pairs. The PMI value is denoted as... The calculation formula is as follows:

[0100] ;

[0101] in, Indicates medicine and medicines The estimated joint probability of simultaneous occurrence of the same document in the same document. and They represent medicines. and medicines An estimate of the marginal probability of a single ticket appearing alone, the probability estimate being calculated based on all tickets whose predictions belong to the same disease category;

[0102] Will As edge weights in the drug co-occurrence network, a weighted undirected graph network is constructed to obtain the drug co-occurrence network.

[0103] In this embodiment, co-occurrence frequency refers to the total number of times two different drugs appear simultaneously on the same medical bill, which is a basic statistical measure to determine whether there is a therapeutic relationship between drugs.

[0104] Estimate of joint probability Medicines With medicine The frequency of two drugs appearing together in the same invoice relative to the total number of invoices for the same disease reflects the likelihood that the two drugs will be used together for treatment.

[0105] Marginal probability and They refer to medicines or medicine The frequency of a drug appearing alone on a bill relative to the total number of bills for the same disease reflects the prevalence of the drug's use.

[0106] Point mutual information (PMI) is an indicator used to measure the strength of the true association between two drugs. By performing a logarithmic transformation on the ratio of joint probability to marginal probability, it eliminates the statistical bias caused by frequently used drugs. A higher PMI value indicates that the two drugs have a more real therapeutic relevance, rather than random co-occurrence.

[0107] Weighted undirected graph networks refer to graph structures with drugs as nodes, co-occurrence relationships as undirected edges, and PMI values ​​as edge weights. They can intuitively and quantitatively express the degree of therapeutic association between drugs, providing a standard data structure for graph algorithms to mine drug use templates.

[0108] This embodiment precisely defines the construction method of the drug co-occurrence network to ensure the accuracy and anti-interference capability of drug association mining. During the construction process, the system first iterates through all tickets predicted to be for the same disease, enumerates all drug pairs, and counts the co-occurrence frequency of each pair. Based on this, the joint probability is calculated. With marginal probability and Then substitute the points into the mutual information calculation formula: This formula amplifies the weights of truly relevant drug pairs through ratio and logarithmic transformations, compressing the random co-occurrence weights caused by high-frequency, commonly used drugs. This accurately preserves clinically significant drug combinations and suppresses noise interference. Finally, using drugs as nodes, co-occurrence relationships as undirected edges, and PMI values ​​as edge weights, a weighted undirected graph network is constructed. This forms a drug co-occurrence network that can be directly used for graph calculations such as PageRank ranking, community detection, and centrality analysis, providing a high-quality data structure for the subsequent automatic summarization of mainstream drug usage templates.

[0109] The above scheme uses PMI as edge weights, significantly improving the quality of the drug co-occurrence network and solving the problems of traditional co-occurrence statistics being easily affected by common drugs and inaccurate drug template extraction. The PMI weighting mechanism makes the network more focused on real treatment-related drug combinations, improving the accuracy of core and auxiliary drug identification and reducing the false positive rate of abnormal drugs. The undirected weighted graph structure has a high degree of standardization, is compatible with various graph analysis algorithms, has good scalability and compatibility, and can seamlessly connect to subsequent template generation processes, ensuring the stable and efficient operation of the entire system.

[0110] In this embodiment, the core drug set is determined by: calculating the PageRank authority score of each drug node based on the PMI edge weights of the drug co-occurrence network, and including drugs with authority scores higher than a first preset threshold into the core drug set;

[0111] The auxiliary drug set is determined by performing community detection on the drug co-occurrence network to identify multiple drug sub-clusters; for each drug sub-cluster, drugs within the drug sub-cluster that are not included in the core drug set are included in the auxiliary drug set.

[0112] The abnormal drug set is determined by: calculating the degree centrality of each drug node in the drug co-occurrence network and counting the frequency of each drug in the medical bill set; and including drugs with a degree centrality lower than a second preset threshold and a frequency of occurrence in the medical bill set lower than a third preset threshold into the abnormal drug set.

[0113] PageRank authority score refers to the ranking algorithm score based on graph link analysis, which simulates the clinical treatment logic of "mutual recommendation and mutual support" among drugs. The higher the score, the more core and irreplaceable the drug is in the treatment network of that disease.

[0114] The first preset threshold is the cutoff value used to screen out the drug with the highest PageRank score and the most important therapeutic position from all drugs. It can be flexibly configured according to disease type, department, and region.

[0115] Community detection algorithms are graph clustering algorithms used to divide complex networks into several sub-clusters with tight internal connections and sparse external connections, such as Louvain and Infomap. They can automatically identify drug combination modules with clear clinical therapeutic significance.

[0116] Degree centrality is a comprehensive indicator that measures the number of connections and the weight of connections of a drug node in a co-occurrence network. It reflects the importance of a drug in the treatment network. The lower the degree centrality, the weaker the association between the drug and the mainstream drug template.

[0117] The second preset threshold refers to the degree centrality lower limit threshold. A value below this threshold indicates that the drug hardly participates in the mainstream treatment network.

[0118] The third preset threshold refers to the lower limit of the frequency of a drug appearing in the invoice set. If the frequency is below this threshold, it indicates that the drug is rarely used to treat the disease.

[0119] The above scheme clearly defines the rules for classifying the three categories of drugs in the mainstream drug template, thereby achieving fully automated, standardized, and interpretable template construction.

[0120] The core drug set is determined using the PageRank algorithm. Specifically, it involves calculating the authority score of each drug node based on the PMI edge weights of the drug co-occurrence network, and including drugs with scores higher than a first preset threshold in the core drug set. These drugs are the most important, most commonly used, and most indispensable treatment drugs for the disease and are the primary basis for rationality review.

[0121] The set of auxiliary drugs is determined by a community discovery algorithm, which specifically includes: performing clustering on the drug co-occurrence network to divide it into several clinically significant drug sub-clusters; excluding nodes that have been included in the core drug set within each sub-cluster; and classifying the remaining drugs as adjunctive treatment, symptomatic treatment, and supportive treatment drugs into the set of auxiliary drugs.

[0122] The abnormal drug set is determined by a combination of node centrality and frequency. Specifically, drugs that simultaneously meet the conditions of having a degree centrality lower than the second preset threshold and appearing more frequently in the disease-related invoice set than the third preset threshold are considered to be drugs that are unrelated to mainstream treatment, rarely used, and are likely to be mistakenly, incorrectly, or abused, and are thus included in the abnormal drug set.

[0123] Through the above three-tiered rules, the system automatically generates mainstream drug use templates that are clearly structured, logically rigorous, and clinically relevant, providing a unified benchmark for subsequent multidimensional comparisons and rationality scoring.

[0124] The above solution uses graph algorithms to automatically classify medication templates, replacing manual rule formulation and making the standards more objective, stable, and closer to real clinical medication patterns. The clear and intuitive three-tiered structure of core / auxiliary / abnormal levels facilitates rapid identification of medication issues, improving review efficiency and interpretability. Configurable thresholds allow the templates to adapt to medication differences across various diseases, departments, regions, and hospital levels, providing high flexibility and broad applicability.

[0125] In some embodiments, the calculation of the core drug missing rate includes: calculating the proportion of the number of missing core drugs in the pending invoices to the total number of core drugs in the core drug set;

[0126] The calculation of the abnormal drug index includes: counting the number of abnormal drugs in the invoices to be reviewed;

[0127] Structural similarity calculation includes: calculating the similarity between the set of drugs invoices to be reviewed and mainstream drug templates based on Jaccard similarity;

[0128] Semantic conflict detection includes: based on a drug pharmacology knowledge base, detecting whether there are drug combinations with pharmacological semantic conflicts in the invoices to be reviewed, and outputting a quantitative index of semantic conflict.

[0129] In this embodiment, the core drug missing rate refers to a key indicator used to quantitatively assess whether the core treatment drugs in the treatment plan for the target disease in the invoice under review are complete and whether there is any "missing prescriptions" issue. Essentially, it is the calculated missing rate obtained by comparing the core drugs actually included in the invoice under review with the standard core drug set defined in the mainstream drug template. This rate directly reflects the completeness and standardization of the main treatment plan and is an important basis for medical insurance supervision to determine whether there are behaviors such as insufficient treatment, over-provisioning of disease groups, or issuing false invoices.

[0130] The abnormal drug presence indicator refers to the statistical value used to directly identify whether there are drugs on the invoice that are completely unrelated to the treatment logic of the disease, rarely used in clinical practice, or even mismatched, abused, or illegally prescribed across different diseases. This indicator uses a pre-defined set of abnormal drugs in the mainstream drug use template as the judgment benchmark. Once there is an intersection, it means that the medication has deviated significantly from clinical practice. The larger the value, the more significant the suspicion of medication violation.

[0131] Structural similarity is a quantitative indicator that measures the degree of matching between the set of drugs in the invoice under review and the mainstream rational drug use template for that disease, from the perspective of the overall drug combination. It does not judge whether a single drug is reasonable, but rather judges whether the entire drug combination's medication path conforms to clinical practice based on the overall structure, correlation, and co-occurrence patterns. It can effectively identify hidden violations such as scattered drug use, piecemeal drug use, and illogical combinations.

[0132] Semantic conflict detection refers to in-depth compliance testing that goes beyond simple set comparison and delves into the level of clinical pharmacology logic. Based on mature professional knowledge of pharmacology, drug combination contraindications, treatment guidelines, and indications, it infers and judges the combination relationships, mechanisms of action, applicable scenarios, and dosage form logic among multiple drugs listed on the invoice, identifying deep-seated unreasonable issues such as pharmacological contradictions, opposite effects, contraindicated combinations, dosage form conflicts, and off-label use.

[0133] The pharmacology knowledge base refers to a professional structured knowledge base built by authoritative medical institutions, pharmacopoeia standards, clinical guidelines, and medical insurance regulatory rules. It contains comprehensive information such as generic drug name, pharmacological classification, indications, contraindications, interactions, adverse reactions, common dosage forms, usage and dosage, and medical insurance coverage limits.

[0134] In calculating the core drug missing rate, the system first reads the mainstream medication template corresponding to the invoice under review after disease prediction, extracts the core drug set already identified in the template, and uses it as the benchmark for the standard treatment plan for that disease. Then, the system performs drug analysis and normalization on the invoice under review to obtain the actual medication set. Through item-by-item matching and comparison, the system counts the number of core drug entries that do not appear in the actual medication, and divides the number of missing entries by the total number of core drug items to obtain the core drug missing rate. This calculation method can accurately identify the problem of missing primary treatment drugs, such as pneumonia invoices without antibiotics or hypertension invoices without first-line antihypertensive drugs, providing the most basic and crucial basis for reasonableness judgment.

[0135] During the calculation of abnormal drug indicators, the system uses the abnormal drug set defined in the mainstream medication template as a blacklist. It performs a strict intersection operation between the actual medication used in the invoice under review and the abnormal drug set, directly counting the number of abnormal drugs detected. Since the abnormal drug set is automatically filtered based on rules such as low frequency, weak correlation, and lack of treatment logic, once a match is found, it can be directly determined that the medication use is clearly unreasonable. For example, the presence of completely unrelated medications such as anti-tumor drugs or psychotropic drugs in a pneumonia invoice can be quickly identified.

[0136] In the structural similarity calculation process, the system uses Jaccard similarity as a quantification tool. The calculation object is the complete set of reasonable drug use, consisting of the drug set of the invoice to be reviewed and the core drugs plus auxiliary drugs in the mainstream drug template. The number of drugs in the intersection of the two sets is divided by the number of drugs in the union, yielding a continuous score between 0 and 1. This method effectively measures the overall similarity of drug combinations, avoiding the bias caused by judging a single drug, and identifying hidden anomalies such as "missing core drugs, abuse of auxiliary drugs, and scattered drug combinations."

[0137] During semantic conflict detection, the system invokes a pre-built pharmacology knowledge base to perform combinatorial reasoning verification on all drugs in the invoice. On one hand, it detects whether there are issues such as opposing or antagonistic pharmacological effects, additive toxicity, or enhanced adverse reactions among the drugs. On the other hand, it detects whether there are dosage form conflicts, such as the simultaneous prescription of the same drug in oral and injectable forms, or unreasonable combination of external and internal dosages. Simultaneously, it combines diagnostic and treatment logic to determine whether there are cases of off-label use or cross-departmental drug use violations, and outputs corresponding semantic conflict quantification indicators based on the severity of the conflict. Through the independent calculation and comprehensive judgment of the above four indicators, this invention achieves a three-tiered progressive review from "whether the drug appears" to "whether the combination is reasonable" and then to "whether the pharmacology is compliant," making the rationality judgment more comprehensive, rigorous, and closer to clinical practice.

[0138] The aforementioned scheme breaks down the abstract judgment of medication rationality into four calculable, comparable, and traceable quantitative indicators, completely eliminating the subjectivity, inconsistencies, and risks of omissions inherent in manual review. This standardizes, automates, and makes the review process more transparent. The core drug missing rate directly verifies the completeness of the main treatment plan; the abnormal drug presence indicator quickly identifies obvious violations; structural similarity assesses whether the medication template deviates from the norm at an overall level; and semantic conflict detection enhances the depth and professionalism of the review from a pharmacological mechanism perspective. These four indicators complement each other and progress step-by-step, ensuring both review efficiency and significantly improving accuracy. They effectively identify key issues in medical insurance supervision such as irrational drug use, excessive use of disease groups, and fictitious medical services, providing solid and reliable technical support for automated auditing.

[0139] In some embodiments, the formula for calculating the comprehensive reasonableness score is:

[0140] ;

[0141] in, The overall score indicates the reasonableness of the assessment. Indicates the missing rate of core drugs. Indicates the quantity of abnormal drugs. Indicates structural similarity. This represents the no-conflict confidence score calculated based on the semantic conflict detection results, with a value range of [0,1], where 1 indicates that no conflict was detected. , , and These represent the weighting coefficients.

[0142] In this embodiment, a comprehensive rationality score is used. This invention provides a final quantitative score for uniformly measuring the rationality of medication use on a single medical invoice. The score is preferably limited to between 0 and 1. A higher score indicates that the medication use is more in line with clinical practice, closer to mainstream medication templates, and less likely to involve violations. A lower score indicates that the medication use is more scattered, deviates more from the norm, and is more likely to involve violations. This comprehensive score is a weighted fusion of multi-dimensional comparison results and is interpretable, reproducible, auditable, and configurable.

[0143] The core drug missing rate, which ranges from 0 to 1, indicates that the core drug is missing more severely and the main treatment plan is less complete. It is the most weighted core indicator in the rationality score.

[0144] The number of abnormal drugs is a non-negative integer. The larger the value, the more illegal, mismatched, or irrelevant drugs appear in the invoice, which directly penalizes the reasonableness score.

[0145] Structural similarity, calculated using the Jaccard formula, ranges from 0 to 1 and reflects the overall matching degree between the drug combination and the mainstream pattern. A higher score indicates a more standardized combination.

[0146] The no-conflict confidence score is assigned based on the semantic conflict detection results. It takes the maximum value of 1 when there is no conflict, takes the value in the range of 0.5 to 0.9 when there is a slight conflict, and takes the value in the range of 0 to 0.4 when there is a moderate to severe conflict. It is used to accurately reflect the degree of conformity between the medication logic and the pharmacological norms.

[0147] W1, W2, W3, and W4 are the core matching weight, anomaly penalty weight, structural similarity weight, and logical consistency weight, respectively. All are non-negative configurable parameters, and the sum of the four weights is 1. The weights can be flexibly adjusted according to disease characteristics, department type, regional medical practices, and medical insurance regulatory priorities, making the scoring system highly adaptable.

[0148] In this embodiment, the comprehensive scoring formula is designed following the principles of "positive bonus, negative penalty, adjustable weight, and clear logic." The first term in the formula... The core drug missing rate is inversely converted into a core completeness score; the fewer the missing drugs, the higher the score, highlighting the dominant role of core therapeutic drugs in the rationale assessment; the second item A penalty mechanism is established for abnormal drug use; the more abnormal drugs are detected, the lower the penalty score. This mechanism automatically deducts points for obvious violations of drug use, strengthening the supervision of indiscriminate, incorrect, and excessive drug prescriptions. (The third item...) The original structural similarity score is directly used to reflect the overall closeness of the drug combination to the mainstream pattern, while also taking into account the standardization of the drug structure; Item 4 This reflects the compliance of medication use in terms of pharmacology and clinical logic. A perfect score is awarded for no conflicts, while deductions are made according to the severity of any conflicts, thus enhancing the professionalism and rigor of the review. The four scores are multiplied by their respective weights and then summed to obtain the final comprehensive score S(T).

[0149] In practical applications, the weights can be dynamically adjusted according to regulatory needs: for example, increasing the weight of W1 in scenarios where the focus is on monitoring treatment integrity; increasing the weight of W2 in scenarios where the focus is on cracking down on the misprescription of high-priced or abnormal drugs; increasing the weight of W3 in scenarios where the standardization of drug combinations is emphasized; and increasing the weight of W4 in scenarios where pharmacological safety and clinical compliance are valued. After the comprehensive scoring is completed, the system classifies the invoices into three risk levels according to preset thresholds: invoices with a score of 0.85 to 1 are considered reasonable and are automatically approved by the system; invoices with a score of 0.60 to 0.85 are considered suspicious and are marked for manual review; invoices with a score below 0.60 are considered high-risk and are directly pushed to the medical insurance audit queue for key review.

[0150] The aforementioned solution organically integrates multi-dimensional review indicators, upgrading the reasonableness judgment from multiple scattered rules to a single continuous score, facilitating automated system processing, risk classification, and batch auditing. The scoring process is transparent, explainable, reproducible, and auditable, fully meeting the standardization and normalization requirements of medical insurance supervision. The configurable weight feature allows the scoring system to adapt to the differentiated needs of different diseases, departments, regions, and regulatory levels, balancing universality and flexibility. The three-level risk classification is directly linked to the system's automatic handling actions, significantly improving the efficiency of medical insurance auditing, reducing labor costs, and achieving precise, focused, and efficient supervision.

[0151] In some embodiments, such as Figure 3 As shown, step S3 includes:

[0152] S31: For each predicted disease, based on the key physiological and pathological indicators in the patient profile information, all ticket samples predicted to be that disease are clustered into multiple patient subgroups.

[0153] S32: For each patient subgroup, construct a drug co-occurrence network with time or stage labels based on the treatment stage identifier or medication sequence information in the invoice;

[0154] Step S4 includes:

[0155] S41: For each patient subgroup, based on the tagged drug co-occurrence network corresponding to that patient subgroup, summarize the mainstream medication template for each patient subgroup; wherein, the mainstream medication template for each patient subgroup includes the core drug set applicable to that patient subgroup, the staged auxiliary drug subset associated with different stages of diagnosis and treatment, and the abnormal drug set;

[0156] In step S5, the multidimensional comparison includes:

[0157] Based on the patient profile information of the invoices to be reviewed, the corresponding patient subgroup template is matched, and the core drug missing rate, abnormal drug presence index, and auxiliary drug structure similarity with the patient subgroup template are calculated in sequence.

[0158] In this embodiment, patient profile information refers to structured information extracted from medical bills and related medical records that can comprehensively reflect the individual characteristics of patients, including but not limited to age, gender, weight, allergy history, underlying diseases, liver and kidney function, major physiological and pathological indicators, high-risk factors, etc., which is an important basis for individualized and precise clinical medication.

[0159] Patient subgroups refer to homogeneous patient groups obtained by clustering patients under the same predicted disease according to key physiological and pathological indicators in their patient profiles. For example, the same pneumonia disease can be divided into adult subgroups, pediatric subgroups, elderly subgroups, subgroups with underlying diseases, and severe subgroups, etc. There are significant differences in medication patterns, drug selection, and dosage guidelines among different subgroups.

[0160] Treatment stage markers are labels used to mark the treatment period a patient is in throughout the entire disease treatment process, including the acute phase, remission phase, recovery phase, consolidation phase, and follow-up phase. The medication focus, drug combination, and adjuvant treatment methods are significantly different for different treatment stages.

[0161] Temporal drug co-occurrence networks are weighted graph networks constructed by adding medication time sequence and treatment stage attributes to the traditional drug co-occurrence network. They can accurately reflect the drug use patterns, switching logic and combination modes at different treatment stages, and are closer to the real clinical treatment path than ordinary co-occurrence networks.

[0162] A phased subset of auxiliary medications refers to a set of auxiliary medications that are divided into stages of treatment within the mainstream medication template. These medications are used only within the corresponding stage. For example, during the acute phase, the focus is on auxiliary medications for reducing fever, expectorating, and fighting infection, while during the recovery phase, the focus is on medications for conditioning, support, and improving symptoms. This makes the medication template more refined and accurate.

[0163] In this embodiment, by introducing patient subgroup clustering and treatment stage division, the traditional coarse-grained disease-level medication template is upgraded to a fine-grained, individualized, and staged precision medication template system, which significantly improves the relevance and accuracy of the review.

[0164] In step S31, after completing disease prediction and categorizing tickets by disease, the system does not directly construct a unified template. Instead, it further uses a clustering algorithm based on key physiological and pathological indicators in the patient profile, such as age segmentation, liver and kidney function status, presence or absence of underlying diseases, and severity of illness, to divide tickets for the same disease into multiple patient subgroups. Through subgrouping, the system can group patients with similar clinical characteristics into one category, avoiding misjudgments and omissions caused by using the same template for adult and pediatric medications, or for mild and severe medications.

[0165] In step S32, for each patient subgroup, the system further reads the treatment stage identifier from the invoice or infers treatment stage information from the medication sequence and drug type, and constructs a drug co-occurrence network with time sequence or stage labels based on this. This network not only preserves the co-occurrence relationships and PMI weights between drugs, but also additionally records the stage attributes of drug appearance, so that drug associations are deeply bound to the treatment process, accurately restoring the clinical medication patterns at different stages.

[0166] In step S41, the system, based on the time-tagged drug co-occurrence network corresponding to each patient subgroup, summarizes and generates mainstream drug templates specific to that subgroup. The template structure includes: a core drug set applicable to that subgroup, a phased subset of auxiliary drugs corresponding to different treatment stages, and a unified set of abnormal drugs for the entire subgroup. This forms a three-tiered refined template system of "disease type—patient subgroup—treatment stage," making the templates more closely aligned with real clinical practice.

[0167] In step S5, during the multidimensional comparison process, the system first accurately matches the patient profile information of the invoice to be reviewed to the corresponding patient subgroup template; then, based on the treatment stage information in the invoice, it locates the auxiliary drug subset of the corresponding stage in the subgroup template; subsequently, it calculates the core drug missing rate, abnormal drug presence index, and auxiliary drug structure similarity under the current treatment stage in sequence, so that each comparison strictly fits the individual characteristics of the patient and the current treatment stage, greatly improving the accuracy of the review.

[0168] The aforementioned solution, by introducing refined modeling from two dimensions—patient subgroups and treatment stages—completely addresses the shortcomings of traditional single-prescription templates, which cannot adapt to different populations, disease courses, and clinical scenarios, significantly reducing misjudgment and missed judgment rates. The review rules have been upgraded from a "one-size-fits-all" approach to one that is "individualized, precise, and stage-based," better aligning with clinical treatment logic and the requirements of refined medical insurance supervision. The template system is scalable, updatable, and adaptable to the medication habits of different regions, hospitals, and departments, possessing strong feasibility and practicality. Simultaneously, it maintains full-process automation, without increasing labor costs, and can operate stably and efficiently in large-scale invoice supervision scenarios.

[0169] In some embodiments, the semantic conflict detection in step S5 is achieved through deep conflict detection based on a drug knowledge graph, including: deep conflict detection based on a drug knowledge graph, specifically including: identifying and quantifying the pharmacological conflict, adverse reaction superposition risk and metabolic pathway competition relationship between drugs in the prescription to be reviewed by querying a pre-built drug knowledge graph, and generating a conflict score;

[0170] like Figure 4 As shown, step S6 further includes:

[0171] S61: Calculate a comprehensive rationality score based on the core drug missing rate, abnormal drug presence index, phased auxiliary drug structural similarity, and the conflict score of the deep conflict detection.

[0172] S62: Determine the risk level of the bill to be reviewed based on the comparison results of the comprehensive reasonableness score and the scoring threshold;

[0173] S63: Generate a structured review result and an interpretability report, wherein the interpretability report shall at least list the key comparison items that led to the score change, the drug involved, the template rules or knowledge graph relationships cited, and indicate the specific conflict type or missing content.

[0174] In this embodiment, the drug knowledge graph refers to a large-scale structured medical knowledge network constructed through a graph structure, with drugs as the core entity and pharmacological effects, targets, metabolic pathways, indications, contraindications, contraindications for combined use, adverse reactions, and medical insurance coverage as relational edges. It possesses capabilities for associative reasoning, path retrieval, and deep mining, and can support the identification and quantification of complex pharmacological conflicts. It is an authoritative knowledge carrier for drug safety review in the field of medical AI.

[0175] Deep conflict detection is an upgraded approach compared to traditional simple rule-based conflict detection. It goes beyond the comparison of drug names or categories, delving into the underlying logic of pharmacological mechanisms, metabolic pathways, toxicological effects, indications, and contraindications to use in combination. It identifies deep-seated medication risks that are difficult to detect with traditional rules, such as metabolic competition, toxicity superposition, antagonism, off-label use, and contraindications to use in combination.

[0176] The conflict score is a quantitative deduction indicator calculated based on the results of in-depth conflict detection, taking into account the conflict type, severity level, number of drugs involved, and degree of harm to patients. The higher the score, the more serious the conflict and the greater the risk of medication use, which directly affects the overall rationality score.

[0177] The interpretability report is a structured explanatory document generated by this invention to meet the requirements of traceability, auditability, and verifiability in medical insurance supervision. The report fully lists the risk level, scoring basis, key anomalies, involved drugs, referenced template rules, knowledge graph relationships, specific conflict types, or missing drug information, providing direct, clear, and complete evidence for manual auditing.

[0178] This embodiment further upgrades semantic conflict detection to deep conflict detection based on a drug knowledge graph, and adds interpretable output, making the review more in-depth, authoritative, and transparent. In the semantic conflict detection stage of step S5, the system no longer relies solely on a basic pharmacology knowledge base for simple rule matching, but instead calls upon a pre-built large-scale drug knowledge graph to perform a comprehensive and in-depth association analysis of all drugs within the invoice. The specific working principle is as follows:

[0179] Through graph retrieval and reasoning, we can identify deep-seated risks among drugs, such as opposing pharmacological effects, competing metabolic pathways, overlapping toxicities, contraindications, off-label use, duplicate use, and unreasonable dosage form combinations. At the same time, by combining information such as drug dosage, cost, and frequency of use, we can comprehensively judge the severity of the conflict and generate a standardized conflict score.

[0180] In step S61, the system weights and integrates four indicators: core drug missing rate, abnormal drug presence index, phased auxiliary drug structural similarity, and deep conflict score, to calculate a more accurate and rigorous comprehensive rationality score. The higher the conflict score, the greater the deduction from the overall score, thereby strengthening the identification and early warning of high-risk drug use.

[0181] In step S62, the system determines the risk level according to the preset scoring threshold, while maintaining the three-level classification standard of reasonable invoices, suspicious invoices, and high-risk invoices. This ensures seamless integration with the existing medical insurance audit process, system interface, and business rules, without changing the original regulatory path and reducing the difficulty of system transformation.

[0182] In step S63, the system automatically generates a structured review result and an interpretability report. The report not only outputs the risk level and score, but also lists in detail the key comparison items that led to the score change, the names of the drugs involved, the mainstream medication template rules on which they were based, the authoritative relationships in the drug knowledge graph, the specific conflict type, or the missing content of core drugs. This allows auditors to quickly locate the problem, understand the cause, and verify the basis without having to repeatedly trace back the original data and model logic.

[0183] This embodiment utilizes a drug knowledge graph to achieve deep conflict detection, elevating drug review from superficial compliance to the pharmacological mechanism and safety level. It can identify hidden drug risks that are difficult to detect using traditional methods, further enhancing anomaly identification capabilities and regulatory accuracy. The interpretable report ensures a complete traceability chain for review results, meeting the stringent requirements of medical insurance supervision, medical auditing, and administrative verification for compliance, transparency, and traceability. This significantly reduces the cost of manual review and improves audit efficiency and credibility. Simultaneously, it maintains the system's automation, standardization, and scalability, adapting to the needs of various entities such as medical insurance bureaus, health departments, medical institutions, and third-party invoice platforms, demonstrating broad application prospects and promotional value.

[0184] After obtaining disease prediction results from a large number of medical invoices, this invention performs aggregation analysis on invoice samples predicted to be the same disease. The aim is to extract typical medication templates corresponding to that disease and construct a "mainstream medication template" for subsequent rationality review. The specific steps include the following:

[0185] First, the set of medical bills for all diseases predicted as disease b is defined as:

[0186] ;

[0187] Each ticket It contains a collection of medicines The definition is as follows:

[0188] ;

[0189] Preferably, to improve accuracy, a confidence threshold for disease prediction can be set, and only ticket samples with a confidence level higher than the threshold (e.g., softmax probability > 0.8) can be retained.

[0190] Secondly, for the sample set of bills Statistical analysis was performed on all drug-related invoices, and the following characteristics were extracted:

[0191] Frequency statistics: Statistical analysis of the frequency of each drug (p) in... Frequency of occurrence ;

[0192] Coverage: Calculates the percentage of drug items appearing on invoices;

[0193] Co-occurrence frequency: The number of times drugs co-occur on the same invoice, used for subsequent construction of drug co-occurrence maps;

[0194] TF-IDF value: measures the importance of a drug to a disease and suppresses interference from generic drugs.

[0195] Then, a drug co-occurrence network is constructed, specifically based on the co-occurrence relationships between drugs in the invoices, to build a drug co-occurrence graph. , where: node set Each node Representative diseases A standardized drug appears in the relevant invoices. (Edge set) Indicates if medicine and On the same bill If they appear together, then at the node and Establish a co-occurrence edge between them Weighting, in order to quantify the correlation strength between drugs, this invention not only considers co-occurrence frequency but also introduces pointwise mutual information (PMI) as an edge weighting factor to eliminate noise interference from all-purpose drugs such as "saline solution". The calculation formula is as described above.

[0196] After constructing the weighted network, this invention employs the following multidimensional graph analysis algorithm to automatically summarize the medication template, specifically including:

[0197] Node Centrality Analysis: By calculating degree centrality or betweenness centrality, this analysis identifies the core nodes in the network, forming a core set of nodes. Alternative nodes.

[0198] PageRank ranking algorithm: simulates the "mutual voting" mechanism among drugs. The higher the PageRank value, the stronger the authority of the drug in the treatment pathway (i.e., the greater the possibility of it being the primary drug).

[0199] Community detection algorithms use Louvain or Infomap algorithms to cluster the network and automatically divide it into different "drug-using sub-clusters." For example, for the disease "pneumonia," the algorithm can automatically identify a high-confidence representative subgraph (i.e., the mainstream drug combination) of "antibiotics + cough suppressants and expectorants + antipyretics."

[0200] Based on statistical and graphical analysis, a template for mainstream medications is constructed for each disease b. Based on different weights and occurrence characteristics, drugs are divided into the following three categories:

[0201] Core drug collection Drugs that are frequently and strongly co-occurring in this disease category and have high TF-IDF values ​​are usually the primary treatment drugs; meeting the following criteria:

[0202] ;

[0203] in , These are the set frequency and specificity thresholds, respectively.

[0204] Auxiliary drug collection A collection of drugs that appear with moderate frequency, or drugs that appear only in specific receipts, used for supplementary treatment, auxiliary diagnosis, etc.

[0205] Collection of abnormal drugs A set of drugs that appears very infrequently, or that contains drugs that have almost no connection with disease b in the co-occurrence graph, may indicate that the drugs were mistakenly or excessively prescribed.

[0206] The mainstream drug use template structure is represented as follows:

[0207] ;

[0208] This mainstream medication template will serve as a standard set for comparing the rationality of medications under disease category b, supporting subsequent anomaly detection and scoring analysis.

[0209] The following description uses "pneumonia" as the target disease to further explain the method involved in this invention:

[0210] Suppose that, through a disease prediction model, the system identifies "pneumonia" as the predicted disease for a batch of medical invoices, forming a medical invoice set T.肺炎 A total of 1200 items;

[0211] The drug field of all receipts in this medical receipt collection is aggregated and statistically analyzed. The results are shown in Table 1 below:

[0212] Table 1. Aggregated statistical results of the drug field in all medical invoices in the medical invoice collection.

[0213]

[0214] After graph structure analysis and statistical classification, the mainstream drug template M (pneumonia) for the disease "pneumonia" is generated as follows:

[0215] Core drug group C (pneumonia): Cephalosporin antibiotics (such as ceftriaxone sodium), expectorants (such as ambroxol), and antipyretic analgesics (such as ibuprofen);

[0216] Supportive medications (A) for pneumonia: antiviral drugs (such as oseltamivir), traditional Chinese medicine, intravenous vitamins, etc.

[0217] Abnormal drug category E (pneumonia): anti-tumor drugs (such as tegafur capsules) and psychotropic drugs (such as clozapine).

[0218] When a medical insurance bill T (predicted disease as pneumonia) is detected to have any of the following conditions: no core drug is present (e.g., no antibiotics), or there is a clearly abnormal drug (e.g., clozapine), or the drug combination deviates significantly from the overall pattern (e.g., only auxiliary drugs + abnormal drugs are present), it can be automatically marked as "unreasonable drug combination" and enter the high-risk audit queue to improve the efficiency of medical insurance audit.

[0219] After constructing the mainstream drug template corresponding to disease type b Subsequently, for each invoice T to be reviewed, the present invention determines whether its medication content is reasonable by matching and analyzing differences with the template. This process constitutes the core discrimination module of the present invention, used for intelligently identifying abnormal invoice behavior.

[0220] Assuming that the invoice T to be reviewed has been predicted as disease b by the disease prediction module, the set of drugs extracted from this invoice is denoted as . ;

[0221] The mainstream drug template for disease type b is known. These are categorized into core drug sets, auxiliary drug sets, and abnormal drug sets.

[0222] This invention employs the following comparison rules to comprehensively assess the risk of the documents under review:

[0223] (1) Judgment of missing core drugs: If Not included For certain high-weight drugs, there may be instances of "prescriptions not being issued when required." The missing percentage is calculated as follows:

[0224] ;

[0225] If the missing percentage exceeds a preset percentage threshold (e.g., 30%), the invoice to be reviewed is determined to be a core missing invoice.

[0226] (2) Judgment of the presence of abnormal drugs: If Appeared in Drugs listed are considered abnormal; it is supported to set severity levels for rare drugs, such as psychiatric, oncology, and chemotherapy drugs; the formula for judging the presence of indicators for abnormal drugs is as follows: If the indicator is greater than or equal to 1, a risk flag will be triggered.

[0227] (3) Judgment of deviation in combined structure: and ∪ The overall similarity of the resulting ideal combination is calculated. Jaccard similarity can be used for the similarity calculation, and the formula is as follows:

[0228] ;

[0229] Cosine similarity (based on word vectors or frequency vectors) can also be used when calculating similarity.

[0230] If the calculated similarity between the two is lower than the preset similarity threshold (e.g., 0.4), it indicates that the overall structure is abnormal.

[0231] (4) Semantic conflict and logical inconsistency detection: Specifically, semantic reasoning judgment is performed on drug combinations. For example, if the same invoice contains "antiviral drug + antipsychotic drug" but there is no auxiliary explanation, oral + injection + external use exist at the same time, or the dosage of the drug is extremely small but the amount is abnormally high, such semantic judgment can be combined with the rule engine, such as using template-driven logic detection.

[0232] Based on the above comparison results, the comprehensive scoring function S(T) for the rationality of this invention is calculated using the formula described above. Finally landed Based on this range, the system outputs the risk conclusion for the current invoice under review, that is, mapping the comprehensive reasonableness score to the final risk level of the invoice under review, as shown in Table 2 below:

[0233] Table 2. Mapping between the comprehensive reasonableness score and the final risk level of the invoices under review.

[0234]

[0235] The weighting of each option indicator is as follows:

[0236] ;

[0237] ;

[0238] ;

[0239] ;

[0240] Taking a certain invoice T (predicted disease as "pneumonia") under review as an example, its drug list is as follows:

[0241] (1) Basic data preparation for invoices to be reviewed The collection of medicines: The system generates mainstream drug templates for specific diseases. as follows:

[0242] Core Drug Collection (3 items in total).

[0243] Abnormal Drug Collection .

[0244] The standard combination (core + auxiliary) is assumed to have a size of 6 items.

[0245] (2) Calculation of core drug matching score for each dimension. The core drugs included in the invoice are "ibuprofen" and "ambroxol", but "ceftriaxone" is missing. The multidimensional comparison results are as follows:

[0246] Core drug shortage rate The score is .

[0247] Drug purity score: If the abnormal drug "clozapine" is found in the invoice to be reviewed, the quantity of the abnormal drug... Then the purity score of the medicine. .

[0248] Structural similarity score: The intersection of the actual drug and the standard combination is... (2 items); the union assumption of the actual drug and the standard combination is 6 items. Structural similarity score

[0249] .

[0250] Logical consistency score: The system detected that "antipyretic analgesic (ibuprofen)" and "antipsychotic (clozapine)" appeared simultaneously without any supporting diagnostic evidence (serious logical deviation), which was judged as a serious logical conflict and scored accordingly. .

[0251] Based on a comprehensive score of reasonableness The calculation formula is obtained by substituting the above scores into the formula:

[0252] ;

[0253] Based on the comprehensive score of reasonableness obtained from the calculation The system automatically identified the invoice as a "high-risk invoice." It automatically marked the invoice with a "serious violation of medication logic" tag and pushed the invoice and its details to the medical insurance audit specialist queue for focused review.

[0254] Preferably, after comparing each ticket, the following structured results can be output for further processing by the system or manually:

[0255] {

[0256] "Invoice ID": "T_20250701_001",

[0257] "Predicted Disease": "Pneumonia"

[0258] Reasonableness score: 0.47

[0259] "Missing core drug": ["Ceftriaxone"],

[0260] “Abnormal Drug”: [“Clozapine”],

[0261] Structural similarity: 0.33

[0262] Risk Level: High Risk

[0263] The system suggests: "Enter the medical insurance audit queue."

[0264] }

[0265] This output structure can be directly integrated into medical insurance audit systems, invoice management systems, or used for generating audit reports and providing intelligent customer service feedback.

[0266] Preferably, in some embodiments, the "missing sensitivity" of different drugs can be adjusted according to the disease; a "weighting coefficient" or "severity level" can be set for some drugs, such as psychiatric drugs and anticancer drugs; and custom thresholds can be configured by department / unit / region to achieve differentiated regulatory strategies.

[0267] This invention proposes a method for constructing mainstream medication templates based on disease prediction results in the absence of real disease labels and medical insurance data. Through a disease prediction module, the drug list and treatment items in medical invoices are transformed into disease prediction inputs, thereby constructing a pseudo-label dataset and training a deep learning model (such as TextCNN, BiLSTM, or Transformer) to achieve disease classification. This method overcomes the limitation of traditional supervised learning relying on real labels, providing a foundation for subsequent invoice validity review.

[0268] This invention, after predicting the disease type, performs aggregate analysis on invoice samples for the same predicted disease, extracting multi-dimensional statistical features such as drug frequency, coverage, co-occurrence frequency, and TF-IDF value. By constructing a drug co-occurrence network and using graph analysis algorithms to extract representative drug combinations, a mainstream medication template for the disease is formed. This template classifies drugs into three categories: core drugs, auxiliary drugs, and abnormal drugs, providing detailed comparison standards for reviewing the reasonableness of invoices.

[0269] This invention designs a comprehensive set of comparison rules, including core drug missing detection, abnormal drug presence detection, combination structure deviation detection, and semantic conflict and logical inconsistency detection. Using these rules, combined with indicators such as Jaccard similarity and cosine similarity, the medication combination of each invoice is quantitatively scored, and risk levels are assigned based on the scores. This mechanism can automatically identify abnormal invoice behavior, improving the efficiency and accuracy of medical insurance auditing.

[0270] This invention offers high flexibility and scalability. For example, it allows for adjusting the "missing sensitivity" of drugs based on disease type, setting "weighting coefficients" or "severity levels" for certain drugs, and configuring custom thresholds by department, unit, or region to achieve differentiated regulatory strategies.

[0271] This invention can be seamlessly integrated into existing medical insurance auditing systems and invoice management systems to automate the review of invoice compliance. By outputting structured reasonableness scores and risk level results, it provides a basis for further processing by the system or by humans, significantly improving the efficiency and intelligence level of medical insurance supervision.

[0272] In a second aspect, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for constructing and reviewing medication templates based on disease prediction as described in the first aspect of the present invention.

[0273] The computer-readable storage medium may be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory.

[0274] The non-volatile memory may be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), a magnetic random access memory (FRAM), a flash memory, a magnetic surface memory, an optical disc, or a compact disc read-only memory (CD ROM); the magnetic surface memory may be a disk storage device or a magnetic tape storage device.

[0275] The volatile memory may be random access memory (RAM), which serves as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM), synchronous static random access memory (SSRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synclink dynamic random access memory (SLDRAM), and direct memory bus random access memory (DRRAM). The computer-readable storage media described in the embodiments of the present invention are intended to include these and any other suitable types of memory.

[0276] like Figure 5 As shown, in a third aspect, the present invention provides an electronic device 10, including a processor 101 and a storage medium 102, wherein a computer program is stored on the storage medium, and when the computer program is executed by the processor, it implements the method for constructing and reviewing medication templates based on disease prediction as described in the first aspect of the present invention.

[0277] In some embodiments, the processor may be implemented by software, hardware, firmware, or a combination thereof, and may be a circuit, one or more of an application-specific integrated circuit (ASIC), a digital signal processor (DSP), a digital signal processing device (DSPD), a programmable logic device (PLD), a field-programmable gate array (FPGA), a central processing unit (CPU), a controller, a microcontroller, or a microprocessor, thereby enabling the processor to execute some or all of the steps or any combination thereof in the disease-based prediction-based medication template construction and review method described in the various embodiments of the present invention.

[0278] Finally, it should be noted that although the above embodiments have been described in the description and drawings of this invention, this should not limit the scope of patent protection of this invention. Any technical solutions that are based on the essential concept of this invention, utilize the content described in the description and drawings of this invention to make equivalent structural or procedural substitutions or modifications, as well as the direct or indirect application of the technical solutions of the above embodiments to other related technical fields, are all included within the scope of patent protection of this invention.

Claims

1. A method for constructing and reviewing medication templates based on disease prediction, characterized in that, The method includes: S1: Obtain a set of medical invoices, perform standardized preprocessing on the structured data in the invoices, and generate standardized invoice data containing a normalized sequence of drug names. The structured data includes a drug list, information on the department visited, and patient profile information. S2: Input the standardized invoice data into the disease prediction model trained based on the pseudo-label mechanism to obtain the predicted disease corresponding to each invoice; S3: For each predicted disease, aggregate all invoice samples predicted for that disease, and construct a weighted drug co-occurrence network based on drug co-occurrence relationships. The drug co-occurrence network uses Point Mutual Information (PMI) as edge weights. S4: Based on the drug co-occurrence network, a graph analysis algorithm is used to automatically summarize the mainstream drug template. The mainstream drug template includes a core drug set determined by the PageRank algorithm, an auxiliary drug set determined by the community detection algorithm, and an abnormal drug set determined by node centrality analysis. S5: Extract the set of drugs from the invoices to be reviewed and perform a multidimensional comparison with the mainstream drug templates for the corresponding predicted diseases. The multidimensional comparison includes core drug missing rate calculation, abnormal drug presence index calculation, structural similarity calculation, and semantic conflict detection. S6: Based on the multidimensional comparison results, the comprehensive reasonableness score is calculated by combining the core drug missing rate, abnormal drug presence index, structural similarity and semantic conflict detection results, and the structured review results are output. The structured review results include determining the risk level of the invoice to be reviewed based on the comparison results of the comprehensive reasonableness score and the scoring threshold. The risk level includes any one of reasonable invoice, suspicious invoice and high-risk invoice.

2. The method for constructing and reviewing medication templates based on disease prediction as described in claim 1, characterized in that, In step S2, the disease prediction model trained based on the pseudo-label mechanism is trained in the following manner: Based on a dynamic reasoning knowledge base for medical coding, rule-based reasoning is performed on unlabeled invoice data to generate a first set of pseudo-labeled diseases; Based on a large language model pre-trained with medical text, semantic understanding and reasoning are performed on the same unlabeled ticket data to generate a second set of pseudo-labeled diseases. The first set of pseudo-labeled diseases and the second set of pseudo-labeled diseases are merged to generate a fused pseudo-label as the training target; A multi-label disease classification model is trained using the fused pseudo-labels as the target, resulting in a trained disease prediction model.

3. The method for constructing and reviewing medication templates based on disease prediction as described in claim 1, characterized in that, In step S3, constructing a weighted drug co-occurrence network based on drug co-occurrence relationships includes: Statistical prediction of the co-occurrence frequency of drug pairs among all invoices belonging to the same disease category; The point mutual information (PMI) value is calculated based on the co-occurrence frequency of the drug pairs. The PMI value is denoted as... The calculation formula is as follows: ； in, Indicates medicine and medicines The estimated joint probability of simultaneous occurrence of the same document in the same document. and They represent medicines. and medicines An estimate of the marginal probability of a single ticket appearing alone, the probability estimate being calculated based on all tickets whose predictions belong to the same disease category; Will As edge weights in the drug co-occurrence network, a weighted undirected graph network is constructed to obtain the drug co-occurrence network.

4. The method for constructing and reviewing medication templates based on disease prediction as described in claim 1, characterized in that, The core drug set is determined by the following method: based on the PMI edge weights of the drug co-occurrence network, the PageRank authority score of each drug node is calculated, and drugs with authority scores higher than a first preset threshold are included in the core drug set. The auxiliary drug set is determined by performing community detection on the drug co-occurrence network to identify multiple drug sub-clusters; for each drug sub-cluster, drugs within the drug sub-cluster that are not included in the core drug set are included in the auxiliary drug set. The abnormal drug set is determined by: calculating the degree centrality of each drug node in the drug co-occurrence network and counting the frequency of each drug in the medical bill set; and including drugs with a degree centrality lower than a second preset threshold and a frequency of occurrence in the medical bill set lower than a third preset threshold into the abnormal drug set.

5. The method for constructing and reviewing medication templates based on disease prediction as described in claim 1, characterized in that, The calculation of the core drug missing rate includes: statistically analyzing the proportion of the number of missing core drugs in the pending invoices to the total number of core drugs in the core drug set; The calculation of the abnormal drug index includes: counting the number of abnormal drugs in the invoices to be reviewed; Structural similarity calculation includes: calculating the similarity between the set of drugs invoices to be reviewed and mainstream drug templates based on Jaccard similarity; Semantic conflict detection includes: based on a drug pharmacology knowledge base, detecting whether there are drug combinations with pharmacological semantic conflicts in the invoices to be reviewed, and outputting a quantitative index of semantic conflict.

6. The method for constructing and reviewing medication templates based on disease prediction as described in claim 5, characterized in that, The formula for calculating the comprehensive rationality score is as follows: ； in, Reasonableness score Indicates the missing rate of core drugs. Indicates the quantity of abnormal drugs. Indicates structural similarity. This represents the no-conflict confidence score calculated based on the semantic conflict detection results, with a value range of [0,1], where 1 indicates that no conflict was detected. , , and These represent the weighting coefficients.

7. The method for constructing and reviewing medication templates based on disease prediction as described in claim 1, characterized in that, Step S3 includes: S31: For each predicted disease, based on the key physiological and pathological indicators in the patient profile information, all ticket samples predicted to be that disease are clustered into multiple patient subgroups. S32: For each patient subgroup, construct a drug co-occurrence network with time or stage labels based on the treatment stage identifier or medication sequence information in the invoice; Step S4 includes: S41: For each patient subgroup, based on the tagged drug co-occurrence network corresponding to that patient subgroup, summarize the mainstream medication template for each patient subgroup; wherein, the mainstream medication template for each patient subgroup includes the core drug set applicable to that patient subgroup, the staged auxiliary drug subset associated with different stages of diagnosis and treatment, and the abnormal drug set; In step S5, the multidimensional comparison includes: Based on the patient profile information of the invoices to be reviewed, the corresponding patient subgroup template is matched, and the core drug missing rate, abnormal drug presence index, and auxiliary drug structure similarity with the patient subgroup template are calculated in sequence.

8. The method for constructing and reviewing medication templates based on disease prediction as described in claim 7, characterized in that, The semantic conflict detection in step S5 is achieved through deep conflict detection based on the drug knowledge graph, including: deep conflict detection based on the drug knowledge graph, specifically including: by querying the pre-built drug knowledge graph, identifying and quantifying the pharmacological conflict, adverse reaction superposition risk and metabolic pathway competition relationship between drugs in the prescription to be reviewed, and generating a conflict score; Step S6 also includes: S61: Calculate a comprehensive rationality score based on the core drug missing rate, abnormal drug presence index, phased auxiliary drug structural similarity, and the conflict score of the deep conflict detection. S62: Determine the risk level of the bill to be reviewed based on the comparison results of the comprehensive reasonableness score and the scoring threshold; S63: Generate a structured review result and an interpretability report, wherein the interpretability report shall at least list the key comparison items that led to the score change, the drug involved, the template rules or knowledge graph relationships cited, and indicate the specific conflict type or missing content.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method for constructing and reviewing medication templates based on disease prediction as described in any one of claims 1 to 8.

10. An electronic device having a computer program stored thereon, characterized in that, The device includes a processor and a storage medium, wherein a computer program is stored on the storage medium, and when executed by the processor, the computer program implements the method for constructing and reviewing medication templates based on disease prediction as described in any one of claims 1 to 8.

Images