An intelligent standardization auditing system for medical procurement prices based on artificial intelligence and multi-source heterogeneous data

Abstract

The application discloses an intelligent standardization auditing system for medical procurement prices based on artificial intelligence and multi-source heterogeneous data, and relates to the technical field of intelligent auditing.The system comprises a presentation layer, an API gateway layer, a business service layer, an intelligent decision-making layer, a data fusion layer and a data persistence layer.The business service layer is provided with a standardization service module, an auditing service module, a rule service module and a data service module;the intelligent decision-making layer is provided with a large language model, a causal inductor and a rule determinator;the causal inductor performs abnormal price attribution by constructing a causal graph model of drug price influence factors and estimating the causal effect of each factor on the price;the rule determinator executes auditing rule determination;the large language model performs semantic coding on standardized drug data and outputs an auditing conclusion;the data fusion layer dynamically fuses and uniformly stores multi-source heterogeneous price reference data.The application realizes the automation, intelligentization and explainability of drug online price auditing.

Description

Technical Field

[0001] This invention relates to the field of intelligent auditing technology, and more specifically, to an intelligent standardized auditing system for pharmaceutical procurement prices based on artificial intelligence and multi-source heterogeneous data. Background Technology

[0002] In recent years, with the continuous advancement of the national centralized drug procurement policy and the increasing intensity of drug online price supervision, local pharmaceutical procurement platforms need to strictly review the prices submitted for listing of a large number of drugs. Currently, the review of drug online prices generally adopts a manual, item-by-item comparison method. Reviewers need to simultaneously consult more than nine types of heterogeneous data sources, including the national drug catalog, provincial online price lists, reference preparation catalogs, centralized procurement results, lists of drugs that have passed evaluation, and lists of drugs subject to price risk disposal, and manually complete the standardization of drug information and the determination of price compliance. This manual review model is not only inefficient, but the review standards are also easily influenced by personal experience, leading to inconsistent conclusions for the same drug in different batches of reviews.

[0003] The existing technology suffers from the following main problems: There is a lack of dedicated large language model application solutions for online price review scenarios, making it difficult to automatically generate review rules from unstructured policy texts; multi-source heterogeneous price reference data have varying formats, different update frequencies, and information conflicts, making it impossible for existing systems to achieve dynamic data fusion and unified utilization; review rules are configured statically, requiring manual modification of rule codes one by one after policy adjustments, failing to adapt to policy changes; review conclusions are output in a simple pass / fail format, lacking interpretability and audit traceability, making it difficult to meet the compliance audit requirements of pharmaceutical procurement supervision. Summary of the Invention

[0004] To overcome the problems of existing technologies, such as reliance on manual review of online prices, lack of dynamic data fusion and adaptive rule evolution, and lack of explainable traceability, this invention discloses an intelligent standardized review system for pharmaceutical procurement prices based on artificial intelligence and multi-source heterogeneous data, which can effectively solve the above-mentioned technical problems.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0006] A standardized intelligent auditing system for pharmaceutical procurement prices based on artificial intelligence and multi-source heterogeneous data includes:

[0007] The presentation layer is used to provide the user interface.

[0008] The API gateway layer is used to uniformly receive requests and perform identity authentication and interception.

[0009] The business service layer includes a standardization service module, an audit service module, a rule service module, and a data service module. The standardization service module is used to standardize drug data, the audit service module is used to call the intelligent decision-making layer to audit prices, the rule service module is used to manage audit rules, and the data service module is used to provide data access interfaces.

[0010] The intelligent decision-making layer includes a large language model, a causal inferrer, and a rule determiner. The causal inferrer is used to construct a causal graph model of drug price influencing factors and estimate the causal effect of each factor on the price to attribute abnormal prices. The rule determiner is used to execute the review rule determination. The large language model is used to semantically encode the standardized drug data output by the standardization service module and output the review conclusion.

[0011] The data fusion layer is used to dynamically semantically fuse multi-source heterogeneous price reference data and store them uniformly.

[0012] The data persistence layer includes a relational database and a vector database. The relational database is used to store structured business data, and the vector database is used to store semantic embedding vectors of drug data.

[0013] Furthermore, the causal inferrer includes:

[0014] The causal graph construction unit is used to model the factors affecting the listed price of drugs as a directed acyclic graph, wherein the factors include drug inherent attribute nodes, market environment nodes, and policy nodes;

[0015] The causal effect estimation unit is used to estimate the pure causal effect of each factor on price by employing a dual machine learning approach, with machine learning models controlling for confounding factors.

[0016] The anomaly attribution unit is used to locate the key factors causing the price anomaly based on the causal contribution of each factor when the declared price of a drug is determined to be abnormal, and output the attribution results.

[0017] Furthermore, the data fusion layer includes:

[0018] The unified semantic embedding unit is used to encode drug records from various data sources into embedding vectors in a unified semantic vector space through a large language model, and to train an alignment model through contrastive learning to make the semantic vectors of the same drug from different data sources closer to each other.

[0019] The dynamic weight allocation unit is used to allocate dynamic weights to each data source through a multilayer perceptron network according to different review scenarios.

[0020] The conflict resolution unit is used to automatically resolve conflicts when different data sources provide conflicting information about the same drug, based on the timeliness, authority, and consistency of the data sources.

[0021] Furthermore, the rule determiner includes:

[0022] The rule performance tracking unit is used to record the hit rate, false positive rate, manual review and correction rate, and performance differences of each review rule under different drug categories and price ranges.

[0023] The rule parameter auto-tuning unit is used to model rule parameter tuning as a black-box optimization problem and automatically search for the optimal parameter combination using the Bayesian optimization method.

[0024] The policy semantic parsing unit is used to automatically parse the policy text using a large language model when a new policy document is input, in order to extract the scope of drug categories involved in the policy, new price calculation rules or thresholds, policy effective time and applicable conditions, and automatically generate candidate review rules based on the parsing results;

[0025] The rule conflict detection unit is used to detect the conflict type between new rules and existing rules and output a conflict resolution solution.

[0026] Furthermore, the standardization service module includes:

[0027] The sensitivity assessment unit is used to evaluate the review sensitivity of the extracted drug fields in real time during the standardization process using a large language model. The standardization results of highly sensitive fields automatically trigger pre-review marking, while the standardization results of low-sensitivity fields are only recorded.

[0028] The feedback learning unit is used to automatically add erroneous standardized samples that lead to misjudgments in the audit due to standardization errors to the training set of the standardized model in order to achieve continuous model optimization.

[0029] Furthermore, the audit service module includes:

[0030] The decision traceability tree generation unit is used to generate a complete decision traceability tree for each review conclusion. The decision traceability tree records the input data, intermediate results, judgment basis, specific data source records referenced and their version numbers and import times for each level of rule judgment, as well as the key factors of causal inference and their contribution.

[0031] The audit log recording unit is used to record every data change, rule adjustment, and manual review operation during the audit process in a blockchain-like structure. Each operation block contains the hash value of the previous block.

[0032] The confidence score unit is used to output a data integrity score, rule certainty score, and model confidence score for each review conclusion, and then perform weighted fusion to obtain a comprehensive confidence score. When the comprehensive confidence score is lower than the threshold, it is automatically marked as a suggestion for manual review.

[0033] Furthermore, the standardization service module is also used to: for drug records whose standardization confidence of the large language model is lower than a preset threshold, ask the operator questions through multiple rounds of dialogue to obtain key information and complete the standardization based on the operator's feedback.

[0034] Furthermore, the data fusion layer is also used to manage the versioning of the fused data, with each data record containing a version number, effective time, and import time.

[0035] Furthermore, the relational database of the data persistence layer is a PostgreSQL database, and the vector database is used to store the embedding vectors generated by the large language model for encoding drug attributes to support semantic similarity retrieval.

[0036] Furthermore, the business service layer adopts the FastAPI asynchronous service framework, the API gateway layer uses Nginx reverse proxy and performs JWT authentication interception, and the presentation layer is built using Vue3 with TypeScript and TailwindCSS.

[0037] Compared with existing technologies, the beneficial effects of this invention are as follows: By setting up a causal inferrer in the intelligent decision-making layer, this invention constructs a directed acyclic graph of drug price influencing factors and uses a dual machine learning method to estimate the pure causal effect of each factor. When the declared price is judged to be abnormal, it can locate the key factors causing the price abnormality and their contribution, thus realizing a qualitative change in the review result from black-box judgment to causally explainable, providing a clear decision-making basis for manual review; By setting up a unified semantic embedding unit in the data fusion layer to encode multi-source heterogeneous data into a unified semantic vector space and train an alignment model through contrastive learning, combined with a dynamic weight allocation unit to automatically adjust the weights of each data source according to different review scenarios and a conflict resolution unit to automatically resolve contradictory information based on timeliness, authority, and consistency, this invention realizes dynamic processing of more than nine types of heterogeneous reference data. The integration and intelligent resolution enhance the completeness and accuracy of the review basis. By setting up a rule effect tracking unit in the rule judge to continuously monitor rule performance, an automatic rule parameter optimization unit to automatically search for the optimal parameter combination using Bayesian optimization, and a policy semantic parsing unit to parse new policy texts using a large language model and automatically generate candidate rules, continuous optimization and adaptive evolution of review rules are achieved, reducing system maintenance costs when policies change. By setting up a decision traceability tree generation unit in the review service module to generate a complete traceability chain for each review conclusion, an audit log recording unit to record all operations in a blockchain-like structure, and a confidence scoring unit to output multi-dimensional confidence scores and automatically suggest manual review when scores are below a threshold, the entire review process is made traceable, auditable, and risk-controllable, fully meeting the compliance requirements of pharmaceutical regulation. Attached Figure Description

[0038] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary. For those skilled in the art, other embodiments can be derived from the provided drawings without creative effort.

[0039] Figure 1 This is a structural diagram of a pharmaceutical procurement price intelligent standardization review system based on artificial intelligence and multi-source heterogeneous data, provided in an embodiment of this application. Detailed Implementation

[0040] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0041] It should be understood that the ordering terms such as "first" and "second" mentioned in the embodiments of this application are only used to distinguish different objects and are not used to describe a specific order. In addition, the terms "comprising" and "including" are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent in these processes, methods, products or devices.

[0042] The intelligent standardized review system for pharmaceutical procurement prices based on artificial intelligence and multi-source heterogeneous data provided in this application embodiment achieves intelligent standardized processing and automated review of drug procurement price data through a layered architecture design. This system dynamically fuses multi-source heterogeneous price reference data semantically, combining the semantic understanding capabilities of a large language model, the causal effect estimation capabilities of a causal inferrer, and the logical judgment capabilities of a rule-based decision-maker. While ensuring data consistency, it provides highly reliable and traceable price review conclusions. The specific implementation methods of this application embodiment are described in detail below with reference to the accompanying drawings.

[0043] Figure 1 A structural diagram of a pharmaceutical procurement price intelligent standardization review system based on artificial intelligence and multi-source heterogeneous data, provided in an embodiment of this application, is shown below. Figure 1 As shown, the system includes a presentation layer, an API gateway layer, a business service layer, an intelligent decision-making layer, a data fusion layer, and a data persistence layer.

[0044] The presentation layer provides the user interface. Optionally, the presentation layer is built using the Vue3 front-end framework, the TypeScript programming language, and the Tailwind CSS style library. The presentation layer organizes the user interface in a component-based manner, including a drug application information entry interface, a standardized results display interface, an audit conclusion display interface, a rule management interface, and a system management interface. Drug application data entered by the user in the presentation layer is sent to the API gateway layer via HTTP requests.

[0045] The API gateway layer is used to uniformly receive requests from the presentation layer and perform authentication and interception. The API gateway layer uses an Nginx reverse proxy server to implement request load balancing and routing. Before a request enters the business service layer, the API gateway layer verifies the validity of the JSONWebToken carried in the request, including the token signature, expiration time, and issuer identity. Only requests that pass authentication are forwarded to the corresponding microservice interface in the business service layer.

[0046] The business service layer comprises a standardization service module, an audit service module, a rules service module, and a data service module. It is built using the FastAPI asynchronous service framework, leveraging Python's asynchronous capabilities to handle concurrent requests. The standardization service module receives raw drug application data and standardizes fields such as drug name, specifications, dosage form, manufacturer, and approval number. The audit service module invokes the intelligent decision-making layer to audit prices and returns the audit results to the presentation layer. The rules service module manages the lifecycle of audit rules, including rule creation, editing, enabling, disabling, and deletion. The data service module provides a unified data access interface, encapsulating read and write operations to the data persistence layer.

[0047] The intelligent decision-making layer comprises a large language model, a causal inferrer, and a rule determiner. The large language model semantically encodes the standardized drug data output by the standardization service module, generating semantic embedding vectors for drug attributes and outputting a natural language description of the review conclusion. The causal inferrer constructs a causal graph model of drug price influencing factors, estimates the causal effect of each factor on the listed drug price, and performs attribution analysis when prices are abnormal. The rule determiner executes specific review rule determination logic, including rule matching, parameter tuning, and conflict detection.

[0048] The data fusion layer is used to dynamically semantically fuse and uniformly store price reference data from multiple heterogeneous sources. This heterogeneous data includes data from provincial centralized drug procurement platforms, historical hospital procurement data, pharmaceutical manufacturer price quotes, and third-party drug price monitoring data. Through unified semantic embedding, dynamic weight allocation, and conflict resolution mechanisms, the data fusion layer integrates records of the same drug from different data sources into consistent reference price information.

[0049] The data persistence layer comprises a relational database and a vector database. The relational database stores structured business data, including tables for basic drug information, review records, rule definitions, and user operation logs. The vector database stores semantic embedding vectors for drug data to support drug retrieval and comparison based on semantic similarity.

[0050] In one possible implementation, such as Figure 1 As shown, the causal inferencer includes a causal graph construction unit, a causal effect estimation unit, and an anomaly attribution unit.

[0051] The causal graph construction unit is used to model the factors influencing drug listing prices as a directed acyclic graph. Specifically, the causal graph construction unit first defines a set of nodes, which includes drug inherent attribute nodes, market environment nodes, and policy nodes. Drug inherent attribute nodes include the drug's generic name, chemical name, dosage form, specifications, packaging unit, manufacturer, approval number, quality level, and original drug identification. Market environment nodes include the number of drugs with the same generic name, regional market concentration, medical institution procurement volume, and number of distribution companies. Policy nodes include the national medical insurance catalog category, centralized volume-based procurement selection status, medical insurance payment standards, and provincial listing price limits. Based on domain expert knowledge and the conditional independence test results in historical data, the causal graph construction unit determines the directed edges between nodes and constructs a directed acyclic graph G=(V, E), where V represents the set of nodes and E represents the set of directed edges.

[0052] The causal effect estimation unit is used to estimate the pure causal effect of each factor on price using a dual machine learning approach. Specifically, for the target factor T (treatment variable), the listed drug price Y (outcome variable), and the confounding factor vector X, the causal effect estimation unit performs the following steps:

[0053] Step 1: Divide the sample data into a training set and a test set.

[0054] Step two: On the training set, use a machine learning model to fit the conditional expectation and conditional treatment probability of the outcome variable, respectively.

[0055]

[0056]

[0057] in, This represents the conditional expectation estimate of the outcome variable. This indicates a tendency to estimate the score.

[0058] Step 3: Based on the estimated values ​​obtained from the fitting, calculate the bias-reduced results on the test set:

[0059]

[0060] Where i is the sample index, representing the i-th observation sample in the test set.

[0061] Step S204: Estimate the causal effect value using the least squares method:

[0062]

[0063] in, This represents the estimated average treatment effect of the target factor on the listed price of drugs.

[0064] Optionally, the causal effect estimation unit adopts a cross-fitting strategy, dividing the sample into K folds, training the model using K-1 folds in turn, estimating the effect on the remaining 1 fold, and finally taking the average to reduce overfitting bias.

[0065] The anomaly attribution unit is used to identify key factors causing price anomalies when a drug's declared price is deemed abnormal, based on the causal contribution of each factor, and outputs the attribution results. Specifically, the anomaly attribution unit calculates the contribution of each factor to the difference between the declared price and the reference price:

[0066] in, This represents the contribution of the j-th factor. This represents the causal effect value of the j-th factor. This indicates the difference in the value of this factor between the drug under review and the reference drug. The anomaly attribution unit sorts each factor in descending order of its absolute contribution value and outputs the top N key factors and their contribution ratios as the attribution results.

[0067] In one possible implementation, such as Figure 1 As shown, the data fusion layer includes a unified semantic embedding unit, a dynamic weight allocation unit, and a conflict resolution unit.

[0068] The unified semantic embedding unit (SEM) encodes drug records from various data sources into embedding vectors in a unified semantic vector space using a large language model. Specifically, the SEM concatenates the text fields in each drug record to generate a descriptive text D, which includes the drug's generic name, specifications, dosage form, and manufacturer information. The SEM inputs the descriptive text into a pre-trained large language model, extracts the last hidden state as a text representation, and maps it to a fixed-dimensional semantic vector space through a linear projection layer.

[0069]

[0070] in, Let W represent the semantic embedding vector of the drug record, W represent the projection matrix, b represent the bias vector, and LLM(D) represent the encoding function of the large language model.

[0071] The unified semantic embedding unit also trains an alignment model through contrastive learning to bring the semantic vectors of the same drug from different data sources closer together. Specifically, the alignment model uses the InfoNCE loss function:

[0072]

[0073] in, The semantic vector representing the anchor sample. This represents the semantic vector of positive samples of the same drug from different data sources. Represents the semantic vector of a negative sample. Let sim represent the temperature coefficient, and cosine similarity function.

[0074]

[0075] u and v are two semantic embedding vectors, which are high-dimensional vector representations of drug records mapped to a unified semantic vector space after being encoded by a large language model.

[0076] The dynamic weight allocation unit is used to dynamically allocate weights to each data source through a multilayer perceptron network according to different review scenarios. Specifically, the dynamic weight allocation unit inputs the contextual features c of the current review scenario into the multilayer perceptron network and outputs the original weight scores of each data source:

[0077] z = MLP(c), where z = [z1, z2, ..., zK]^T, and K represents the number of data sources. The dynamic weight allocation unit normalizes the original weight scores into a probability distribution using the Softmax function:

[0078]

[0079] in, This represents the dynamic weight of the k-th data source, satisfying... , The original weight score for the k-th data source is output by a multilayer perceptron network (MLP) based on the context features of the current review scenario. The total number of data sources participating in the integration. Let J be the original weight score of the j-th data source, and j be the summation index. This process iterates through all K data sources. The context features are natural exponential functions, including the audit scenario type, drug category code, price range identifier, and medical institution level.

[0080] The conflict resolution unit is used to automatically resolve conflicts when different data sources provide contradictory information about the same drug, based on the timeliness, authority, and consistency of the data sources. Specifically, the conflict resolution unit calculates a comprehensive confidence score for each data source:

[0081]

[0082] in, Tk represents the overall confidence score of the k-th data source, Ak represents the timeliness score, Ck represents the authority score, and α, β, and γ represent preset weighting coefficients. The timeliness score is calculated based on the interval between the data update time and the current time; the authority score is determined based on the administrative level and certification status of the data source; and the consistency score is determined by calculating the standardized deviation between this data source and other data source records of the same category. The conflict resolution unit selects the data source record with the highest overall confidence score as the fusion result, or calculates the fused price reference value according to the weighted average formula.

[0083]

[0084] in, The dynamic weight of the k-th data source. This represents the price reference value provided by the k-th data source.

[0085] In one possible implementation, such as Figure 1 As shown, the rule determiner includes a rule effect tracking unit, a rule parameter automatic optimization unit, a policy semantic parsing unit, and a rule conflict detection unit.

[0086] The rule effectiveness tracking unit records the operational effectiveness metrics of each review rule. Specifically, the rule effectiveness tracking unit maintains an effectiveness tracking record for each rule. This record includes the rule identifier, rule name, cumulative hit count, cumulative false positive count, number of manual review corrections, a hit count matrix for different drug categories, and a false positive rate matrix for different price ranges. After each review is completed, the rule effectiveness tracking unit updates the corresponding rule's effectiveness tracking record based on the review conclusion and the results of manual review.

[0087] The rule parameter auto-tuning unit models rule parameter tuning as a black-box optimization problem, automatically searching for the optimal parameter combination using a Bayesian optimization method. Specifically, for an audit rule with a parameter vector θ, the rule parameter auto-tuning unit defines the optimization objective as minimizing the misjudgment rate on historical audit samples:

[0088]

[0089] in, The misclassification rate of the rule under parameter combination θ is obtained through evaluation using historical labeled data. The automatic rule parameter tuning unit uses a Gaussian process as a surrogate model to construct the acquisition function:

[0090]

[0091] in, This represents the set of observed parameter-performance samples. and Let represent the mean and standard deviation of the posterior distribution of the Gaussian process, respectively. This represents the exploration coefficient. The rule parameter auto-tuning unit iteratively selects the next evaluation point by maximizing the acquisition function until the preset maximum number of iterations or convergence condition is reached, and outputs the optimal parameter combination. .

[0092] The policy semantic parsing unit automatically parses policy text using a large language model when a new policy document is input. Specifically, the unit receives the policy document in natural language text format, segments the text, and inputs it into the large language model. It then guides the model to extract structured policy elements using pre-defined prompt templates. These structured policy elements include the scope of drug categories covered by the policy, new price calculation rules or thresholds, and the policy's effective date and applicable conditions. The unit converts the extracted structured policy elements into rule templates, fills them into the candidate rule library of the rule determiner, and waits for confirmation from rule administrators before taking effect.

[0093] The rule conflict detection unit detects the conflict types between newly added rules and existing rules and outputs conflict resolution solutions. Specifically, the rule conflict detection unit performs a logical formalization representation of each rule in the rule base, converting the condition part and action part of the rule into predicate logical expressions respectively. For the newly added rule Rnew and the existing rule Rexist, the rule conflict detection unit detects the following conflict types: redundant conflict, i.e., the condition parts of the two rules are logically equivalent and their action parts are the same; contradictory conflict, i.e., the condition parts of the two rules have overlap and their action parts contradict each other; subordinate conflict, i.e., the condition part of one rule implies the condition part of another rule. When a conflict is detected, the rule conflict detection unit outputs the conflict rule pair identifier, conflict type, and recommended resolution solution, which includes priority adjustment suggestions, condition refinement suggestions, and rule merging suggestions.

[0094] In one possible implementation, such as Figure 1 As shown, the standardization service module includes a sensitivity assessment unit and a feedback learning unit.

[0095] The sensitivity assessment unit is used to evaluate the review sensitivity of extracted drug fields in real time during the standardization process using a large language model. Specifically, the sensitivity assessment unit combines the extracted drug field names and values ​​into assessment text, inputs it into the large language model, and guides the large language model to output the sensitivity label for that field using binary classification prompts. Sensitivity labels are divided into two categories: high sensitivity and low sensitivity. High-sensitivity fields include the generic name, specifications, dosage form, manufacturer, approval number, and price information of the drug. When a field is marked as high sensitivity, the standardization service module automatically triggers a pre-review flag, sending the standardization result of that field to the review service module for pre-review. When a field is marked as low sensitivity, the standardization service module only records the standardization result to the standardization log and does not trigger any additional review processes.

[0096] The feedback learning unit automatically adds erroneous standardized samples that lead to misjudgments in audits due to standardization errors to the training set of the standardization model to achieve continuous model optimization. Specifically, the feedback learning unit establishes an error sample collection pipeline. When the audit conclusion of the audit service module is confirmed as a misjudgment by manual review, and the cause of the misjudgment is traced back to a standardization error in the standardization service module, the feedback learning unit combines the original drug record, the erroneous standardized result, and the correct standardized result into training samples. The feedback learning unit periodically adds the accumulated erroneous standardized samples to the fine-tuning dataset of the large language model, and incrementally trains the large language model using a low-rank adaptation method to update the model parameters to adapt to the new standardized patterns.

[0097] In one possible implementation, such as Figure 1 As shown, the audit service module includes a decision traceability tree generation unit, an audit log recording unit, and a confidence score unit.

[0098] The decision traceability tree generation unit is used to generate a complete decision traceability tree for each review conclusion. Specifically, the decision traceability tree generation unit records the complete decision path in the review process using a tree-like data structure. The root node of the decision traceability tree records the review request identifier and the standardized drug data input. The intermediate nodes of the decision traceability tree record the input data, intermediate results, and judgment basis for each level of rule determination. The leaf nodes of the decision traceability tree record the final review conclusion. The edges of the decision traceability tree record the logical dependencies between nodes. In addition, each node of the decision traceability tree is also associated with the specific data source record it references, its version number and import time, as well as the key factors for causal inference and their contribution.

[0099] The audit log recording unit is used to record every data change, rule adjustment, and manual review operation during the audit process in a blockchain-like structure. Specifically, the audit log recording unit records each operation as a block, and the block includes the operation type, operation timestamp, operator identifier, operation content digest, and hash value of the previous block. The hash value of the block is calculated using a cryptographic hash function.

[0100]

[0101] in, This represents the hash value of the nth block. This represents the hash value of the preceding block. Represents a timestamp. Indicates the operation content. Indicates operator identification. This represents a string concatenation operation. It creates an immutable chain structure by embedding the hash value of the previous block into the current block.

[0102] The confidence scoring unit outputs a data integrity score, a rule certainty score, and a model confidence score for each review conclusion, and then weights and fuses them to obtain a comprehensive confidence score. Specifically, the data integrity score is calculated based on the missing rate and field completion quality of the input drug field. The rule certainty score is calculated based on the confidence parameter of the hit rule and the rule's historical performance. The model confidence score is calculated based on the probability distribution entropy value of the output results of the large language model. The confidence scoring unit calculates the comprehensive confidence score using a weighted fusion formula:

[0103]

[0104] in, This indicates the overall confidence score. Indicates data integrity score, Indicates the rule-based deterministic score. This represents the model confidence score. , , This represents the preset weighting coefficient. When the overall confidence score is lower than the preset threshold, the confidence scoring unit automatically marks the review conclusion as requiring manual review.

[0105] In one possible implementation, the standardization service module is also used to ask questions of the operator through multi-turn dialogue to obtain key information for drug records whose standardization confidence score of the large language model is lower than a preset threshold, and to complete the standardization based on the operator's feedback. Specifically, the standardization service module calculates the confidence score of the large language model for the standardization result of the current drug record. When the confidence score is lower than the preset threshold, the standardization service module initiates an interactive standardization mode. In interactive standardization mode, the standardization service module generates natural language question text based on missing or ambiguous key fields in the current drug record, and displays it to the operator through the presentation layer. After the operator inputs answer information through the presentation layer, the standardization service module integrates the operator's feedback into the original drug record and re-standardizes it. This multi-turn dialogue process continues until the standardization confidence score reaches the preset threshold or the operator confirms that the standardization has been manually completed.

[0106] In one possible implementation, the data fusion layer is also used for version control of the fused data. Specifically, the data fusion layer assigns a globally unique version number to each fused data record, which adopts a semantic versioning format or a timestamp format. Each data record contains three version control fields: version number, effective time, and import time. The effective time indicates the point in time when the data record begins to participate in the review calculation, and the import time indicates the physical time when the data record enters the system. When writing a new version of fused data, the data fusion layer retains historical version data and updates the version number, effective time, and import time fields. When the review service module performs price review, it queries the data version within the corresponding effective time window based on the timestamp of the review request to ensure that the reference data used for review matches the review time.

[0107] In one possible implementation, the relational database for the data persistence layer is a PostgreSQL database. Specifically, the PostgreSQL database is used to store structured business data, including a basic drug information table, a drug price declaration record table, an audit conclusion record table, an audit rule definition table, a user account information table, and a system configuration parameter table. The data service module interacts with the PostgreSQL database via SQL, utilizing PostgreSQL's transaction and concurrency control mechanisms to ensure data consistency.

[0108] The vector database stores the embedding vectors generated by the large language model for encoding drug attributes to support semantic similarity retrieval. Specifically, the vector database uses an approximate nearest neighbor search algorithm to build a high-dimensional spatial index for the semantic embedding vector of each drug attribute. When the standardization service module or the review service module needs to query semantically similar drug records, it submits a query vector to the vector database. The vector database quickly retrieves the top K candidate records with the highest similarity to the query vector through the index structure and returns the corresponding drug identifier and similarity score.

[0109] In one possible implementation, the business service layer is built using the FastAPI asynchronous service framework. Specifically, the business service layer leverages FastAPI's asynchronous routing and dependency injection system to encapsulate the standardization service module, auditing service module, rule service module, and data service module as independent asynchronous processing endpoints. Each service endpoint implements non-blocking input / output operations using Python's async / await syntax, improving the system's concurrent processing capabilities.

[0110] The API gateway layer is implemented using an Nginx reverse proxy server, which performs JWT authentication interception. Specifically, the Nginx server is configured with an upstream backend service cluster, which distributes client requests to multiple FastAPI service instances in the business service layer according to a load balancing strategy. The Nginx server is configured with a JWT verification module, which parses and verifies the Authorization field in the HTTP request header. Requests that pass verification are forwarded to the backend service after appending user identity information, while requests that fail verification directly return a 401 status code.

[0111] In one possible implementation, the presentation layer is built using Vue3 in conjunction with TypeScript and TailwindCSS. Specifically, the presentation layer utilizes Vue3's composable API to organize front-end logic, leverages TypeScript's static type system to enhance code robustness, and utilizes TailwindCSS's atomic style classes to build a responsive user interface. The presentation layer communicates with the API gateway layer via the Axios library over HTTP to enable data interaction between the front-end and back-end.

[0112] The specific embodiments and / or examples described above are preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A standardized intelligent auditing system for pharmaceutical procurement prices based on artificial intelligence and multi-source heterogeneous data, characterized in that, include: The presentation layer is used to provide the user interface. The API gateway layer is used to uniformly receive requests and perform identity authentication and interception. The business service layer includes a standardization service module, an audit service module, a rule service module, and a data service module. The standardization service module is used to standardize drug data, the audit service module is used to call the intelligent decision-making layer to audit prices, the rule service module is used to manage audit rules, and the data service module is used to provide data access interfaces. The intelligent decision-making layer includes a large language model, a causal inferrer, and a rule determiner. The causal inferrer is used to construct a causal graph model of drug price influencing factors and estimate the causal effect of each factor on the price to attribute abnormal prices. The rule determiner is used to execute the review rule determination. The large language model is used to semantically encode the standardized drug data output by the standardization service module and output the review conclusion. The data fusion layer is used to dynamically semantically fuse multi-source heterogeneous price reference data and store them uniformly. The data persistence layer includes a relational database and a vector database. The relational database is used to store structured business data, and the vector database is used to store semantic embedding vectors of drug data.

2. The intelligent standardized review system for pharmaceutical procurement prices based on artificial intelligence and multi-source heterogeneous data as described in claim 1, characterized in that, The causal inferr includes: The causal graph construction unit is used to model the factors affecting the listed price of drugs as a directed acyclic graph, wherein the factors include drug inherent attribute nodes, market environment nodes, and policy nodes; The causal effect estimation unit is used to estimate the pure causal effect of each factor on price by employing a dual machine learning approach, with machine learning models controlling for confounding factors. The anomaly attribution unit is used to locate the key factors causing the price anomaly based on the causal contribution of each factor when the declared price of a drug is determined to be abnormal, and output the attribution results.

3. The intelligent standardized review system for pharmaceutical procurement prices based on artificial intelligence and multi-source heterogeneous data as described in claim 1, characterized in that, The data fusion layer includes: The unified semantic embedding unit is used to encode drug records from various data sources into embedding vectors in a unified semantic vector space through a large language model, and to train an alignment model through contrastive learning to make the semantic vectors of the same drug from different data sources closer to each other. The dynamic weight allocation unit is used to allocate dynamic weights to each data source through a multilayer perceptron network according to different review scenarios. The conflict resolution unit is used to automatically resolve conflicts when different data sources provide conflicting information about the same drug, based on the timeliness, authority, and consistency of the data sources.

4. The intelligent standardized review system for pharmaceutical procurement prices based on artificial intelligence and multi-source heterogeneous data as described in claim 1, characterized in that, The rule determiner includes: The rule performance tracking unit is used to record the hit rate, false positive rate, manual review and correction rate, and performance differences of each review rule under different drug categories and price ranges. The rule parameter auto-tuning unit is used to model rule parameter tuning as a black-box optimization problem and automatically search for the optimal parameter combination using the Bayesian optimization method. The policy semantic parsing unit is used to automatically parse the policy text using a large language model when a new policy document is input, in order to extract the scope of drug categories involved in the policy, new price calculation rules or thresholds, policy effective time and applicable conditions, and automatically generate candidate review rules based on the parsing results; The rule conflict detection unit is used to detect the conflict type between new rules and existing rules and output a conflict resolution solution.

5. The intelligent standardized review system for pharmaceutical procurement prices based on artificial intelligence and multi-source heterogeneous data as described in claim 1, characterized in that, The standardization service module includes: The sensitivity assessment unit is used to evaluate the review sensitivity of the extracted drug fields in real time during the standardization process using a large language model. The standardization results of highly sensitive fields automatically trigger pre-review marking, while the standardization results of low-sensitivity fields are only recorded. The feedback learning unit is used to automatically add erroneous standardized samples that lead to misjudgments in the audit due to standardization errors to the training set of the standardized model in order to achieve continuous model optimization.

6. The intelligent standardized review system for pharmaceutical procurement prices based on artificial intelligence and multi-source heterogeneous data as described in claim 1, characterized in that, The audit service module includes: The decision traceability tree generation unit is used to generate a complete decision traceability tree for each review conclusion. The decision traceability tree records the input data, intermediate results, judgment basis, specific data source records referenced and their version numbers and import times for each level of rule judgment, as well as the key factors of causal inference and their contribution. The audit log recording unit is used to record every data change, rule adjustment, and manual review operation during the audit process in a blockchain-like structure. Each operation block contains the hash value of the previous block. The confidence score unit is used to output a data integrity score, rule certainty score, and model confidence score for each review conclusion, and then perform weighted fusion to obtain a comprehensive confidence score. When the comprehensive confidence score is lower than the threshold, it is automatically marked as a suggestion for manual review.

7. The intelligent standardized review system for pharmaceutical procurement prices based on artificial intelligence and multi-source heterogeneous data as described in claim 1, characterized in that, The standardization service module is also used to: for drug records whose standardization confidence of the large language model is lower than a preset threshold, ask the operator questions through multiple rounds of dialogue to obtain key information and complete the standardization based on the operator's feedback.

8. The intelligent standardized review system for pharmaceutical procurement prices based on artificial intelligence and multi-source heterogeneous data as described in claim 1, characterized in that, The data fusion layer is also used to manage the versioning of the fused data. Each data record includes a version number, effective time, and import time.

9. The intelligent standardized review system for pharmaceutical procurement prices based on artificial intelligence and multi-source heterogeneous data as described in claim 1, characterized in that, The relational database of the data persistence layer is a PostgreSQL database, and the vector database is used to store the embedding vectors generated by the large language model for encoding drug attributes to support semantic similarity retrieval.

10. A pharmaceutical procurement price intelligent standardization review system based on artificial intelligence and multi-source heterogeneous data as described in any one of claims 1-6, characterized in that, The business service layer adopts the FastAPI asynchronous service framework, the API gateway layer uses Nginx reverse proxy and performs JWT authentication interception, and the presentation layer is built with Vue3, TypeScript and TailwindCSS.

Images