Medical charging rationality intelligent supervision system and method based on large model technology

Abstract

The invention relates to the technical field of large model technology application and natural language processing, in particular to a medical charging reasonability intelligent supervision system and method based on a large model technology. The unstructured text analysis unit extracts medical operation steps and corresponding duration through medical special entity recognition, regular matching and graph neural network association, and outputs unique steps and duration binding data streams through time sequence standardization and conflict resolution, and the effective step calculation unit queries a minimum effective duration threshold value by adopting a dynamic semantic mapping strategy, and calculates the effective step according to the minimum effective duration threshold value. A boundary compensation algorithm is combined to judge effective steps and accumulate the number, a charging compliance judgment unit constructs an operation time sequence chain, compensation time and step complexity weight factors are introduced to calculate the receivable fee, a sectional error tolerance strategy is adopted to be compared with actual charging, and an early warning report marking violation details is output. And the medical charging supervision accuracy and compliance are improved.

Description

Technical Field

[0001] This invention relates to the fields of large model technology applications and natural language processing technology, specifically to an intelligent monitoring system and method for the rationality of medical charges based on large model technology. Background Technology

[0002] Large model technology and natural language processing is an important technology. Specifically, it is applied to the medical insurance billing compliance assessment process. The core is to verify the validity of medical operation steps and determine the reasonableness of charges by accurately analyzing medical records and dynamically matching billing rules, thus meeting the core requirements of medical insurance supervision for accuracy, compliance and traceability. Current technologies for regulating the rationality of medical charges have significant limitations, making it difficult to meet the requirements of refined supervision. Existing systems struggle to accurately parse unstructured texts such as medical records, and their entity recognition accuracy for medical operation steps and duration descriptions is insufficient. They are susceptible to interference from differences in professional terminology and text redundancy, leading to misalignments between steps and their corresponding durations. Furthermore, they lack optimized threshold query mechanisms to address issues such as synonymous expressions of medical step names and lagging rule updates, resulting in threshold matching failures that affect the determination of valid steps. There is a lack of reasonable interruption identification mechanisms when duration falls below the threshold, rigidly deeming steps invalid and ignoring objective special circumstances in clinical operations. Charge calculations rely solely on the number of valid steps or a fixed proportion, failing to consider multi-dimensional factors such as step sequence correlation, interruption compensation, and operational complexity. This leads to a mismatch between receivable charges and actual clinical input. Simultaneously, the lack of traceable audit mechanisms for multi-stage data results in high misjudgment rates for violations and a lack of evidence for dispute resolution, hindering the efficient and accurate identification of overcharging and illegal billing. To address these issues, we provide an intelligent regulatory system and method for the rationality of medical charges based on large-scale model technology. Summary of the Invention

[0003] The purpose of this invention is to provide an intelligent monitoring system and method for the rationality of medical charges based on large model technology, so as to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, a medical billing rationality intelligent monitoring system based on large model technology is provided, comprising: The data input unit is used to receive medical records and medical insurance billing rules from medical institutions; The unstructured text parsing unit performs word segmentation and semantic analysis on the medical record text through named entity recognition and regular expression matching, identifies keywords of medical operation steps, locates the duration description adjacent to the keyword and step name, generates time unit keywords, and converts the time unit keywords and duration data in the medical record text into structured numerical format, outputting the single step name and the corresponding single step actual execution time. The threshold storage unit is used to store the predefined minimum effective duration threshold for a single step in the medical insurance billing rules. The minimum effective duration threshold for a single step is mapped to the name of each step in the form of key-value pairs. The effective step calculation unit couples the single step name with the corresponding actual execution time of the single step and the minimum effective time threshold of the single step. The dynamic rule engine executes the time threshold calibration logic, which includes querying the corresponding minimum effective time threshold of the single step for each single step name and the actual execution time of the single step, comparing the actual execution time of the single step with the minimum effective time threshold of the single step, and determining that the single step is valid only when the threshold condition is met, and accumulating the number of all valid steps as the final number of valid steps; The fee compliance determination unit calculates the fees due based on the final number of valid steps and the preset fee ratio rules, compares them with the actual fee data of the medical institution, determines whether the fees are compliant, and outputs a violation warning report.

[0005] The second objective of this invention is to provide a method for implementing an intelligent monitoring system for the rationality of medical charges based on large-scale model technology, including any one of the above-described features, comprising the following steps: S1. Receive medical record text, extract medical operation steps, time units and numerical description entities through a medical entity recognition model, trigger a sliding window scanning mechanism based on entity results, locate adjacent duration descriptions by combining dependency syntax tree parsing, calculate the association weight between step entities and duration descriptions using a graph neural network position association model, convert duration data into structured key-value pairs using a time series information standardization engine, and output a unique step and duration bound data stream through a conflict resolution mechanism. S2. Obtain the minimum effective duration threshold mapping table for a single step from the medical insurance rules. For each step name and actual execution duration, use a dynamic semantic mapping strategy to query the corresponding threshold. When the actual execution duration is lower than the threshold, activate the interruption factor analysis module to generate an interruption type label. If it is not within the scope of medical insurance exemption, the step is determined to be invalid. Accumulate the number of valid steps that meet the threshold conditions. S3. Construct an operation sequence chain based on the data stream bound by steps and duration, insert a compensation time weight coefficient by combining the interruption type label, and reconstruct the payable cost calculation formula by introducing a step complexity weight factor according to the final effective number of steps and medical insurance ratio rules, so that the actual execution time participates in the weighted calculation, and adopt a segmented error tolerance strategy to handle the exempted interruption type. S4. Compare the weighted calculated receivables with the actual receivables. If the deviation exceeds the tolerance threshold and there is no exemption coverage, generate a violation warning report. The report marks the violation step, the threshold deviation amount, and the conclusion of interrupted analysis.

[0006] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention accurately extracts medical operation steps and duration entities from medical records using a medical-specific entity recognition model. It combines graph neural networks for localization and association, solving the problem of misalignment in unstructured text parsing. A dynamic semantic mapping strategy enables accurate threshold queries, adapting to differences in step names and rule updates. A boundary compensation algorithm identifies reasonable interruptions, avoiding rigid judgments that could lead to misjudgments. Multi-dimensional calibration factors ensure that reimbursable charges align with clinical practice. A full-process audit log ensures traceability, improving the accuracy of medical charge supervision, reducing misjudgment rates, and efficiently identifying illegal charges. This meets the requirements of medical insurance supervision for compliance, accuracy, and traceability, balancing regulatory rigor with clinical operational flexibility. Attached Figure Description

[0007] Figure 1 This is an overall block diagram of the present invention; Figure 2 This is the overall flowchart of the present invention.

[0008] The meanings of the labels in the diagram are as follows: 1. Data input unit; 2. Unstructured text parsing unit; 3. Threshold storage unit; 4. Effective step calculation unit; 5. Fee compliance determination unit. Detailed Implementation

[0009] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0010] This invention provides an intelligent monitoring system for the rationality of medical charges based on large model technology. Please refer to [link / reference]. Figure 1 As shown, it includes: Data input unit 1 is used to receive medical records and medical insurance billing rules from medical institutions; The unstructured text parsing unit 2 performs word segmentation and semantic analysis on the medical record text through named entity recognition and regular expression matching, identifies keywords of medical operation steps, locates the duration description adjacent to the keyword and step name, generates time unit keywords, and converts the time unit keywords and duration data in the medical record text into structured numerical format, outputting the single step name and the corresponding single step actual execution time. Threshold storage unit 3 is used to store the predefined minimum effective duration threshold for a single step in the medical insurance charging rules. The minimum effective duration threshold for a single step is mapped to the name of each step in the form of key-value pairs. The effective step calculation unit 4 couples the single step name with the corresponding single step actual execution time and the single step minimum effective time threshold. It executes the time threshold calibration logic through the dynamic rule engine, including querying the corresponding single step minimum effective time threshold for each single step name and single step actual execution time, comparing the single step actual execution time with the single step minimum effective time threshold, and determining a single step as effective only when the threshold condition is met, and accumulating the number of all effective steps as the final number of effective steps; The fee compliance determination unit 5 calculates the fees due based on the final number of valid steps and the preset fee ratio rules, compares them with the actual fee data of the medical institution, determines whether the fees are compliant, and outputs a violation warning report.

[0011] Named entity recognition employs a medical-specific entity recognition model based on the fusion of bidirectional long short-term memory networks and conditional random fields. It learns the grammatical structure features of medical record texts through a pre-trained medical corpus and establishes a specialized entity tag library for the names of operation steps in the medical record texts. In the specific implementation, the input text is first encoded into character-level vectors. The bidirectional long short-term memory network captures the contextual semantic features and outputs the probability distributions of three types of entities, including medical operation steps, time units, and numerical descriptions. Then, the conditional random field outputs the optimal entity sequence based on the transition probability constraints. The specialized entity tag library covers all standard step names and their commonly used clinical synonyms in the radiology contrast-enhanced scanning protocol.

[0012] Regular expression matching is dynamically constructed based on named entity recognition results, and its matching rule base includes time-described syntax pattern templates: When a medical operation step entity is identified, a sliding window scanning mechanism centered on the location of the medical operation step entity is triggered. Within the window, time unit entities and numerical description entities are searched. For compound time expressions appearing within the scanning window, fuzzy matching rules are used to extract core values, and non-key modifiers are filtered through preset cleaning rules. Semantic analysis further introduces dependency syntax tree parsing. When no time entity is directly matched within the sliding window, the duration description clause that has a verb-object relationship with the medical operation step is traced according to the subject-verb-object relationship.

[0013] The duration description of the location keywords adjacent to the step name adopts a location association model based on graph neural network. The location association model uses medical operation step entities as anchors to construct a multi-dimensional feature vector that includes part-of-speech tagging, syntactic distance and semantic similarity. The trained graph neural network calculates the association weight between each time description entity and the medical operation step entity. Only when the weight exceeds a preset threshold is it determined to be a valid adjacent relationship. When generating time unit keywords, if there are multiple candidate duration descriptions for the same medical operation step, the duration description with the highest association weight is selected as the valid input, and the time unit keywords are automatically standardized into a unified measurement form.

[0014] The time unit keywords and duration data in the medical record text are converted into a structured numerical format, specifically through a time series information standardization engine: The time series information standardization engine has a built-in time unit conversion matrix. It performs unit normalization processing on time unit keywords. For non-standard numerical descriptions, it uses a pre-trained Chinese number conversion model to convert them into Arabic numerals. Finally, it outputs key-value pairs in a preset structured format, where the actual execution duration field is the pure numerical value after unit conversion.

[0015] When outputting the single-step name and its corresponding actual execution time, set up a conflict resolution mechanism: When multiple duration descriptions exist for the same medical operation step in structured data, the highest value is selected based on the association weight. If the difference in association weights among multiple duration descriptions is within the tolerance range, a rationality verification based on the clinical operation knowledge base is triggered, automatically removing outliers that exceed medical common sense. Finally, a data stream binding steps and durations with timestamps is generated to ensure that each step name is associated with only a unique and valid single-step actual execution time.

[0016] When querying the corresponding minimum effective duration threshold for a single step, a dynamic semantic mapping strategy is used: In threshold storage unit 3, the step name is precisely matched. If no match is found, the synonym backtracking of the special entity label library is initiated. The threshold corresponding to the alternative name is queried according to the synonym mapping relationship. When there is a change in the step name but the threshold is not updated, the medical insurance rule version management interface is called to obtain the latest threshold mapping table, and the threshold query path is recorded for audit tracing.

[0017] When comparing the actual execution time of a single step with the minimum effective time threshold for a single step, a boundary condition compensation algorithm is introduced: When the actual execution time is slightly lower than the threshold, the clinical operation interruption factor analysis module is automatically activated. The factor analysis module generates interruption type labels based on the extracted text interruption keywords. If the interruption type falls within the scope of exemption from medical insurance charging rules, the manual review process is triggered to suspend automatic judgment. Otherwise, the validity of the judgment steps is strictly followed according to the threshold comparison results, and the deviation between the interruption type and the threshold is recorded in the judgment log.

[0018] Based on the final number of effective steps and the preset charging ratio rules, calculate the amount due and construct a multi-dimensional calibration factor: First, the data stream is bound according to the steps and duration. An operation sequence chain is generated according to the execution order of the effective steps. Combined with the interruption type label, a compensation time weight coefficient is inserted for unexpected interruptions in the execution of continuous steps. Finally, the billing formula introduces the step complexity weight factor. The step complexity weight factor is defined by the medical insurance rules, so that the extracted actual execution time participates in the weighted calculation rather than a simple proportional division. When comparing with the actual billing data of medical institutions, a segmented error tolerance strategy is adopted, allowing a specific proportion of cost deviation for the triggered exempted interruption type.

[0019] Further explanation is needed. After the data input unit receives the medical record text from the medical institution, the primary task of the unstructured text parsing unit is to accurately extract three core entities: medical operation steps, time units, and numerical descriptions. Because medical record texts often contain highly specialized medical terminology and exhibit significant variations in expression, traditional entity recognition models are prone to insufficient accuracy. Therefore, named entity recognition requires a medical-specific entity recognition model based on the fusion of bidirectional long short-term memory networks (BSSNs) and conditional random fields (CRFs), combined with a pre-trained medical corpus and a specialized entity label library, to ensure the accuracy of entity extraction. The specific implementation method is as follows: This medical entity recognition model uses a bidirectional long short-term memory network (BSSN) as the core for feature extraction and a conditional random field (CRF) as the sequence optimization component, forming a two-level processing link of feature learning and sequence correction. The BSSN is a neural network that adds a backpropagation layer to the traditional BSSN network. It can capture contextual semantic features from both left to right and right to left directions of the text, avoiding the loss of semantic information caused by unidirectional propagation. For example, in a medical record, after intravenous injection of contrast agent, an enhanced scan is performed. Forward propagation can learn the association between intravenous injection and contrast agent, while backpropagation can capture the temporal relationship between the enhanced scan and the subsequent procedure, thus gaining a more complete understanding of the semantics of the operation steps.

[0020] Conditional Random Fields (CRF) are a probability-based sequence labeling model that learns the transition patterns between entity labels to correct the locally optimal results of bidirectional Long Short-Term Memory (LSTM) networks and obtain globally optimal entity sequences. Before training, the model needs to learn the grammatical structure features of medical record texts using a pre-trained medical corpus. This pre-trained medical corpus is an annotated dataset covering medical records from multiple departments. Each text is manually annotated with the boundaries and labels of three types of entities: medical operation steps, time units, and numerical descriptions. The grammatical structure of the text is also annotated. The pre-training process employs a transfer learning strategy. First, a bidirectional long short-term memory network is initially trained using a general medical corpus to enable it to grasp the basic grammatical rules of medical texts. Then, the pre-trained network parameters are used as initial parameters and adjusted using a constructed specialized medical record corpus. During adjustment, the cross-entropy loss function is used to evaluate the entity recognition accuracy. When the validation set accuracy is stable and exceeds 96%, training is stopped, the model parameters are saved, and the model can accurately learn the grammatical structure features of medical entities in medical records.

[0021] To further improve the accuracy of identifying entities in medical procedures, a specialized entity tag library needs to be established for the names of procedures in medical record texts. This specialized entity tag library is a structured database categorized by department, storing standard names of medical procedures and commonly used clinical synonyms. It primarily covers procedures explicitly stated in medical insurance billing rules, with a particular focus on radiology enhancement scan protocols. Radiology procedures are highly standardized, and there are significant differences in their terminology, making it a key area for medical billing supervision. The tag library is constructed as follows: We extracted all standard step names from the radiology enhancement scan protocol. By surveying a sample of medical records from the radiology departments of tertiary hospitals, we compiled common synonyms for each standard step, ensuring that each standard step corresponds to 3 to 5 synonyms, covering more than 98% of clinical record expressions. We assigned a unified entity label to each standard step and synonym, and recorded the medical insurance billing code corresponding to each step to facilitate subsequent association with the billing rules of the threshold storage unit. The label library supports the synchronous addition of step names every quarter through the medical insurance rule update interface to ensure the timeliness and completeness of the label library.

[0022] In the model inference stage, the input medical record text is first encoded using character-level vectorization. Character-level vectorization is a preprocessing operation that converts each character in the text into a fixed-dimensional vector. Unlike word-level vectorization, character-level encoding can handle the splitting and recognition of professional terms such as augmented scans, improving the model's generalization ability. Specifically, a medical text-specific character set is first constructed, including Chinese characters, numbers, common punctuation marks, and medical-specific symbols. A unique index is assigned to each character. The input text is then split into sequences according to the characters. Each character is converted into a one-dimensional vector based on the character set index. If the text length is less than the preset length, it is filled with the vector corresponding to the empty character index. If the length exceeds the preset length, it is truncated. Finally, a character-level vector matrix with a dimension of 200×3000 is generated as the input to the bidirectional long short-term memory network.

[0023] The vector matrix is ​​input into the trained bidirectional long short-term memory (LSS) network. The network captures contextual semantic features through bidirectional propagation. The forward LSS network layer learns the forward semantics of each character from the beginning to the end of the text, while the backward LSS network layer learns the backward semantics of each character from the end to the beginning of the text. Each layer of the bidirectional LSS network contains 128 hidden units to store intermediate semantic features. The semantic vectors output by the two layers (each 200×128 dimensions) are concatenated and fused into a comprehensive semantic vector of 200×256 dimensions. This vector can fully represent the semantic information of each character in its context. Subsequently, a classifier for three types of entities is set in the output layer of the bidirectional LSS network, corresponding to medical operation steps, time units, and numerical descriptions. The classifier uses the softmax function to convert the comprehensive semantic vector of each character into a probability distribution of the character belonging to the three types of entities, with values ​​ranging from 0 to 1, and the sum of the three probabilities being 1. Since the probability distribution output by the bidirectional LSS network is based on the local judgment of a single character, there may be a problem of discontinuous entity labels. Therefore, a conditional random field is needed to approximate the probability distribution based on the transition probability. The optimal entity sequence is output by the Conditional Random Field (CRF). The transition probability constraint is the transition pattern between entity labels learned by the CRF during training. During CRF optimization, the character-level entity probability distribution output by the bidirectional long short-term memory network is used as the basis. Combined with the transition probability of the transition matrix, the joint probability of all possible entity sequences is calculated. That is, the sum of the product of the entity probability of each character in the sequence and the transition probability of the adjacent character label. The sequence with the highest joint probability is selected as the optimal entity sequence. In this process, a specialized entity label library plays an auxiliary verification role. After the CRF outputs the optimal entity sequence, the system extracts the text fragments marked as medical operation steps and compares them with the standard step names and synonyms in the specialized entity label library. If the fragment has a match in the label library, whether it is a standard name or a synonym, the entity recognition result is confirmed to be valid. If no match is found, it is marked as a suspicious operation step, triggering a manual review process to further ensure the accuracy of medical operation step entity recognition. The label library covers all standard step names and clinically commonly used synonyms of the radiology enhanced scanning protocol, which can minimize the occurrence of suspicious operation steps and improve the recognition efficiency and reliability of the model.

[0024] After the medical entity recognition model has completed the initial extraction of medical operation steps, time units, and numerical description entities, it is necessary to further accurately locate the duration descriptions associated with the steps through regular expression matching. Since duration descriptions in medical record texts are often adjacent to operation steps but have different expressions, using fixed regular expression rules is prone to omissions or mismatches. Therefore, it is necessary to dynamically construct matching rules based on entity recognition results and combine sliding window scanning, compound expression processing, and syntactic parsing to ensure the accuracy of duration description location. The specific implementation method is as follows: First, regular expression matching is dynamically constructed based on named entity recognition results. Its matching rule base contains preset time-description syntax pattern templates. Unlike traditional fixed regular expression rules, the core of dynamic construction is: After the entity recognition model outputs the location information of the medical operation step entity, the system automatically uses that step entity as the matching anchor point, calls the time description template adapted to the medical operation scenario from the rule base, and generates a special regular expression for that step. This avoids irrelevant duration errors caused by global matching, greatly improving matching efficiency and accuracy. When a medical operation step entity is recognized, a sliding window scanning mechanism centered on the entity's location is immediately triggered. Within the window range, time unit entities and numerical description entities are searched. The sliding window scanning mechanism refers to a mechanism that uses the character position of the step entity in the text as the center to define a fixed-length text segment as the search range. The window size needs to be set according to the expression habits of the medical record text. Through statistics of 1,000 radiology medical records, it was found that more than 95% of the duration descriptions are located within 10 characters before and after the operation step entity. Therefore, the window is set to the character range from 10 characters before the start position of the step entity to 10 characters after the end position, which ensures coverage of the effective duration description while avoiding the introduction of too much irrelevant text.

[0025] During the scanning process, the system calls dynamically constructed regular expressions to match time unit entities and numerical description entities character by character within the window. For compound time expressions appearing in the scanning window, fuzzy matching rules are used to extract the core values, and preset cleaning rules are used to filter out non-key modifiers to avoid numerical extraction deviations caused by redundant expressions. Compound time expressions refer to duration descriptions containing modifiers and combinations of multiple units. In such expressions, modifiers and non-core units can interfere with the extraction of core values. Fuzzy matching rules use a preset modifier dictionary and unit priority to ignore redundant information and focus on the matching logic of core values. First, a modifier dictionary is constructed to include common non-key modifiers in the medical records. During fuzzy matching, these modifiers are located and marked using regular expressions. Then, for combinations of multiple units, the core values ​​that best reflect the operation duration are extracted according to the time unit priority. The preset cleaning rules are used to remove non-key information from the matching results: first, redundant words in the modifier dictionary are deleted; second, meaningless punctuation is filtered out; and third, duplicate expressions are merged. The final output is a concise duration description containing only core values ​​and time units, laying the foundation for subsequent structured conversion.

[0026] If no time entity is directly matched within the sliding window, dependency tree parsing is introduced through semantic analysis. Based on the subject-verb-object relationship, the duration description clauses that have a verb-object relationship with the medical operation steps are traced. Dependency tree parsing is a natural language processing technique that constructs a syntactic tree structure by analyzing the grammatical dependency relationships between words in a sentence. Its core function is to uncover the implicit semantic relationships in the text, rather than relying solely on character positions. In specific implementation: A dependency syntax tree is constructed from complete sentences containing medical operation steps using a syntax parsing tool. The nodes of the medical operation step entities in the syntax tree are located, and their corresponding core verbs are identified. Subsequent components with verb-object relationships with the core verb are traced along the syntax tree to verify the logical relationship between the duration description clause and the operation step. If the logical relationship is valid, 15 minutes is taken as the associated duration description of the operation step. This achieves accurate tracing of non-adjacent durations, avoiding duration omissions caused by expression intervals. It can accurately locate duration descriptions directly related to medical operation steps, and also cover implicit durations with expression differences or large intervals. This ensures that each valid operation step can be matched with unique and accurate duration information, providing reliable duration data support for subsequent time sequence information standardization and valid step determination.

[0027] After obtaining the entities of medical operation steps and their corresponding candidate durations through sliding window scanning and regular expression matching, relying solely on character position proximity is insufficient to accurately determine which duration is truly associated with the step. For example, in a radiology medical record, the arterial phase of an enhanced scan lasts 5 minutes, and the delayed scan lasts 3 minutes. The arterial phase of the enhanced scan step might be mistakenly associated with 3 minutes due to the close text interval. Therefore, a location association model based on graph neural networks is needed to quantify the association strength between steps and durations by comprehensively considering multi-dimensional features, while also resolving conflicts among multiple candidate durations and ensuring unit consistency. The specific implementation method is as follows: First, the location of duration descriptions adjacent to keywords and step names employs a graph neural network-based positional association model. This model models the potential associations between medical operation step entities and time description entities through graph structures. Leveraging the learning capabilities of graph neural networks for multi-dimensional features, it outputs the strength of the association (association weight). Unlike traditional methods that rely solely on character distance, this model integrates multi-dimensional information such as part-of-speech, syntax, and semantics, significantly improving association accuracy. The core logic of this model is to construct an analysis framework using the medical operation step entity as an anchor point. The anchor point is the core reference entity; the medical operation step entity is chosen as the anchor point because the step is the service object of the duration description, and all durations revolve around the step. The execution process unfolds around this central point to ensure the targeted nature of the correlation analysis. Around this anchor point, the model constructs a multi-dimensional feature vector including part-of-speech tagging, syntactic distance, and semantic similarity. Each feature dimension is designed specifically for the characteristics of medical text to ensure effective characterization of the relationship between steps and duration. Part-of-speech tagging is a feature that encodes the part-of-speech categories of step entities and time description entities, using the UniversalPOS (part-of-speech) tagging system. For example, medical operation step entities are mostly noun phrases (labeled NOUN), while the numerical part of the time description entity is labeled as numerals (labeled NUM), and the unit part is labeled as nouns (labeled NOUN). During encoding, the part-of-speech tags are converted into One-Hot vectors, reflecting the differences in part-of-speech tags. Entity type association, syntactic distance refers to the difference in node hierarchy between the step entity and the time description entity in the dependency syntax tree. The dependency syntax tree is a tree structure that describes the grammatical dependency relationships between words in a sentence. The calculation first locates the nodes of the core words of the two entities in the syntax tree, and then counts the number of edges of the shortest path between the two nodes. The smaller the syntactic distance, the stronger the grammatical association between the two entities. Semantic similarity is calculated using pre-trained medical word vectors. A medical-domain-adjusted BERT model is used to generate 768-dimensional semantic vectors for the step entity and the time description entity, respectively. Then, the cosine similarity between the two vectors is calculated, with a value of 0-1, where 1 represents complete similarity. For example, although the similarity between enhanced scan and 5 minutes is low, the semantic association judgment can be corrected by combining syntactic relationships. To avoid overlooking actual connections due to literal differences, these three types of feature vectors are concatenated into a multi-dimensional feature vector of length (One-Hot dimension + 1 + 1), which serves as the input to the graph neural network. The trained graph neural network calculates the association weights between each time-description entity and the medical procedure entity. Here, a graph attention network is chosen for the trained graph neural network, as it can assign differentiated weights to different features through an attention mechanism, more accurately capturing key association information. The model training process relies on a labeled dataset. First, a training set is constructed by collecting 10,000 radiology medical record texts, manually labeling the associations between medical procedure entities and their corresponding time descriptions. Then, a training graph is constructed. In the graph structure of each training sample…The model consists of one medical procedure step entity node and multiple time-description entity nodes. Edges represent the potential relationships between the step node and each duration node, and the edge features are the constructed multi-dimensional feature vectors. Next, training parameters are set, using cross-entropy loss as the loss function, Adam as the optimizer, and a learning rate of 0.001. The training run consists of 50 epochs. After each epoch, the accuracy is validated on the validation set. Training stops when the validation set accuracy consistently exceeds 95%, and the model parameters are saved.

[0028] During the inference phase, the graph structure (with edge features as multi-dimensional feature vectors) constructed from the step entity to be analyzed and the candidate duration entity is input into the trained GAT. The network weights the features through the attention layer and outputs the association weight of each duration entity to the step entity, with a value ranging from 0 to 1. The closer the weight is to 1, the stronger the association. Only when the association weight exceeds a preset threshold is it determined to be a valid adjacent relationship. The preset threshold is determined by analyzing the ROC curve of the training set. The thresholds between 0 and 1 are traversed, and the precision and recall corresponding to each threshold are calculated. The threshold with the highest F1 score is selected as the final judgment standard. Considering the rigor required for medical fee supervision, the final threshold is set to 0.7. That is, when the association weight is greater than or equal to 0.7, it is determined that there is a valid adjacent relationship between the time description entity and the medical operation step entity. If the weight is less than 0.7, it is determined that there is no association.

[0029] During the generation of time unit keywords, it often happens that the same medical operation step corresponds to multiple candidate duration descriptions. In this case, it is necessary to first filter out all candidate durations with a correlation weight greater than or equal to 0.7. If the weights are all less than 0.7, a signal indicating no valid duration is returned, triggering manual review and forming a candidate duration list. From the candidate duration list, the duration description with the highest correlation weight is selected as the valid input. Subsequently, the time unit keywords are automatically standardized to a unified measurement format. The unified measurement format is set to minutes based on the medical insurance billing rules. In the medical insurance rules, medical operation durations are mostly calculated in minutes. Standardization is then passed. The built-in time unit conversion table stores the conversion relationship between common time units (such as seconds, hours, minutes, and half an hour) and minutes (1 hour = 60 minutes, 1 quarter hour = 15 minutes, half an hour = 30 minutes, 1 minute = 60 seconds). For example, it converts 1.5 hours to 90 minutes, 30 seconds to 0.5 minutes, and 1 quarter hour to 15 minutes. For non-standard expressions, they are marked as non-standard units, triggering manual review to avoid conversion errors. After standardization, the output is a key-value pair of step name and effective duration (minutes) to ensure the consistency of units in subsequent effective step calculations and charging compliance judgments.

[0030] After determining the effective duration descriptions corresponding to medical operation steps through graph neural networks, these duration data may still have issues such as inconsistent units and non-standard numerical representations. If directly used for subsequent effective step determination and billing calculation, the chaotic format can easily lead to threshold comparison deviations or cost calculation errors. Therefore, it is necessary to transform them into a unified structured numerical format through a time-series information standardization engine to provide standardized data support for subsequent processes. The specific implementation method is as follows: First, the time series information standardization engine incorporates a time unit conversion matrix. The core function of this matrix is ​​to perform unit normalization processing on time unit keywords, ensuring that all duration data is uniformly measured in minutes. The time series information standardization engine is a dedicated module integrating unit conversion, numerical conversion, and format output functions, with a response time of less than or equal to 10 milliseconds, adapting to the rapid processing of batch duration data. The time unit conversion matrix is ​​a two-dimensional mapping table pre-stored in the engine's memory. The row dimension represents the original time unit, the column dimension represents the target unit, and the cell contains the corresponding conversion coefficient. In practice, the engine first extracts time unit keywords from the valid duration description using regular expressions, then uses these keywords as row indices to look up the corresponding conversion coefficients in the conversion matrix. The coefficients are then temporarily stored, and unit conversion is performed after the numerical conversion is complete, laying the foundation for subsequent unified measurement. For non-standard numerical descriptions commonly found in duration descriptions, the engine uses a pre-trained Chinese number conversion model to convert them to Arabic numerals, ensuring a unified numerical format. Non-standard numerical descriptions are common in medical texts. Due to differences in numerical representation caused by clinical recording habits, direct reading can easily lead to the problem of 15 not being recognized as 15. The pre-trained Chinese numeral conversion model is built based on a bidirectional long short-term memory network adapted for the medical field. The training process relies on a labeled dataset. First, 5000 samples of Chinese numeral representations from medical texts are collected, and the correspondence between Chinese numerals and Arabic numerals in the samples is labeled. Then, the Chinese numerals are split into single-character sequences, converted into vectors through a character-level embedding layer, and input into the bidirectional long short-term memory network to learn the combination rules between characters. The mean squared error loss function is used during training, the Adam optimizer is selected, and the learning rate is set to 0.001. Training stops when the conversion accuracy on the validation set is stable above 98%, and the model parameters are saved. In the inference stage, the engine extracts the numerical part from the effective duration description and inputs it into the trained model. The model outputs the corresponding Arabic numeral. If a compound expression is encountered, the model will first split it into two-quarters, convert it into 2 and 0.25 respectively, and then automatically sum them to output 2.25, ensuring the complete conversion of non-standard numerical values.

[0031] After completing the numerical conversion and unit coefficient query, the engine performs unit normalization calculation, multiplying the converted Arabic numerals by the corresponding conversion factor to obtain the pure numerical value in minutes. Subsequently, the engine outputs key-value pairs according to a preset structured format, which adopts JSON format for easy subsequent unit parsing and database storage. The key field includes a unique step ID, original duration description, standardized duration, numerical conversion record, and unit conversion factor. This structured data is synchronously pushed to the cache area of ​​the effective step calculation unit through the engine's output interface. The unit can directly read the pure numerical value of the standardized duration field and compare it with the minimum effective duration threshold for a single step, eliminating the need to handle format or unit issues. This significantly improves the processing efficiency and accuracy of subsequent processes. The time-series information standardization engine completely solves the problem of chaotic duration data formats in medical texts, ensuring both data uniformity and standardization, and achieving traceability by retaining the original records and conversion process. This provides crucial data support for the accuracy of medical charge reasonableness supervision.

[0032] Even after the time-series information standardization engine completes the unit normalization and numerical conversion of duration data, conflicts may still arise due to redundancy in the descriptions of medical record texts, resulting in multiple standardized duration descriptions for the same medical operation step. For example, in a radiology medical record, the venous phase of an enhanced scan lasts for 3 minutes, but is later extended to 3.5 minutes due to poor contrast enhancement. The enhanced scan venous phase step would correspond to two candidate durations: 3 minutes and 3.5 minutes. Directly using these would easily lead to confusion in subsequent determination of valid steps and billing calculations. Therefore, a conflict resolution mechanism needs to be set up. Through association weight ranking and verification using clinical common sense, a unique and valid actual execution time for a single step can be selected. The specific implementation method is as follows: First, when multiple duration descriptions exist for the same medical procedure step in the structured data, the highest value is prioritized based on the association weight. This association weight is a quantitative value calculated using a graph neural network, representing the degree of correlation between the duration description and the medical procedure step. This weight information is retained for each candidate duration description during the standardization process. Specifically, the time-series information standardization engine first extracts all candidate durations and corresponding association weights for the same step from the structured data, forming a list of duration-weight key-value pairs. These are then sorted in descending order of association weight, from highest to lowest, to facilitate quick identification of the most relevant duration. If only one candidate duration with the highest weight exists after sorting, that duration is directly identified as the duration for that step. The core logic behind this approach is that the higher the association weight, the closer the syntactic and semantic connection between the duration description and the step, and the more consistent it is with the clinical record's convention of stating the duration of the immediately adjacent step as the actual execution time. If, after sorting, the difference in association weights among multiple duration descriptions is found to be within the tolerance range (i.e., the difference between the highest and second-highest weights is less than or equal to the preset tolerance), then a rationality verification based on the clinical operation knowledge base is triggered, automatically eliminating outliers that exceed medical common sense. The tolerance range is a critical difference set to avoid misjudgment caused by minor weight differences. The clinical operation knowledge base is a pre-built structured database that includes the clinical standard duration range for each medical operation step. The specific process of rationality verification is as follows: First, the standard duration range for the current step is retrieved from the knowledge base. All candidate durations with association weights within the tolerance range are compared to the standard range. If a candidate duration exceeds the range, it is identified as an outlier exceeding medical common sense and automatically removed. From the remaining durations that meet the standard range, the deviation from the average standard duration is calculated, and the duration with the smallest deviation is selected as the valid duration. If the deviations are the same, the duration with slightly higher association weight is selected, ensuring that the verification results conform to clinical common sense while also considering the closeness of association. Finally, a data stream binding steps and durations with timestamps is generated, ensuring that each step name is associated with only a unique and valid single-step actual execution time. The timestamp indicates that the system handles time conflicts for that step. The current time is used for subsequent auditing and traceability. The step and duration binding data stream uses structured data in JSON format. To ensure unique association, after the system generates the data stream, it will store it in a dedicated step and duration binding database and set a unique step ID as the primary key. If the duration data of the same step is processed again later, the original record will be deleted before the new data is inserted to avoid multiple valid duration records for the same step. At the same time, the data stream will be synchronously pushed to the valid step calculation unit through the data interface, providing a unique and standardized data source for the unit to query the actual execution time of the step and compare it with the minimum valid duration threshold, thus completely solving the problem of subsequent process deviation caused by duration conflict.

[0033] After uniquely binding the name of a single step with its actual execution time, the effective step calculation unit needs to query the minimum effective time threshold for that single step, which is the minimum execution time that the step must achieve as stipulated in the medical insurance rules. If the execution time is less than this, it cannot be charged according to regulations. However, medical operation step names often have differences in expression or lag in rule updates. Therefore, a dynamic semantic mapping strategy needs to be adopted to ensure that the threshold query is accurate and adaptable to rule changes through multi-dimensional matching and rule synchronization. The specific implementation method is as follows: First, precise matching of step names is performed in threshold storage unit 3. Threshold storage unit 3 is a structured storage module specifically for storing predefined thresholds in medical insurance charging rules. It organizes data in key-value pairs, where the key is the standard name of the medical operation step, and the value is the corresponding minimum effective duration threshold for a single step. It also stores metadata such as the effective time of the threshold and the version number of the medical insurance rule, which facilitates subsequent traceability and updates. The specific process of precise matching is as follows: The effective step calculation unit extracts the step name from the step and duration bound data stream and uses it as the query key to call the query interface of the threshold storage unit. If a completely matching key exists in the unit, meaning the enhanced venous phase of the scan perfectly matches the stored standard name, the corresponding threshold is returned directly, completing the fast query. If the query result is empty, meaning no completely matching step name is found, the exact match is considered to have failed, and the synonym backtracking mechanism of the specialized entity tag library is immediately activated to avoid query interruption due to differences in expression. The specialized entity tag library is a structured database that covers the standard names of medical operation steps and commonly used clinical synonyms. Synonym backtracking refers to using the unmatched step name as input, querying all corresponding standard synonyms in the tag library, and then re-initiating the threshold query using the synonyms. In the value query process, the effective step calculation unit first calls the synonym query interface of the special entity tag library to obtain the standard synonym list corresponding to the step name. Then, according to the list order, each synonym is used as a new query key, and precise matching is performed in the threshold storage unit 3 in turn. If a synonym is successfully matched, the backtracking stops immediately, and the threshold is used as the effective threshold of the current step. If all synonyms are not successfully matched, the synonym backtracking is determined to be unsuccessful, and the next step of the rule update synchronization process is entered. When the synonym backtracking fails and it is determined that the step name has changed but the threshold has not been updated, that is, the step is an operation added or renamed after the medical insurance rule, and the old data in the threshold storage unit 3 has not been updated synchronously, the medical insurance rule version management interface needs to be called to obtain the latest threshold mapping table.

[0034] The medical insurance rule version management interface is a standardized interface provided by the medical insurance supervision platform. It is used to synchronize the latest medical insurance billing rules and threshold data. The interface supports filtering data by parameters such as the effective time of the rule for the operating department. It returns a threshold mapping table in JSON format, containing the latest step standard name and minimum effective duration threshold key-value pairs. In practice, the effective step calculation unit first initiates a data request through the interface parameters. After verifying the unit's identity, the interface returns the latest threshold mapping table from the radiology department. Upon receiving the mapping table, the unit first compares it with the data in the local threshold storage unit 3, updates the differences, and then uses the updated mapping table to re-query the threshold for the current step to ensure the latest valid data is obtained. If the interface call fails, the local cached emergency threshold, i.e., the average threshold of the operation category to which the step belongs, is triggered. Simultaneously, the interface call failure log is recorded, and synchronization is automatically retried after the network recovers, regardless of whether it is through exact matching or synonyms. Whether the threshold is obtained through backtracking or interface synchronization, the threshold query path must be recorded for audit tracing. The query path refers to a structured log that fully records the threshold query process, including fields such as step name, exact match result (success / failure), synonym backtracking list, interface call status (whether it was called, call result), final threshold obtained, threshold source (local storage / interface synchronization), and query timestamp. This log data is written to the audit database in real time, using an immutable time-series storage method to ensure that the medical insurance regulatory department can fully trace the threshold query process of each step during subsequent verification, verify the compliance of the threshold source, and avoid regulatory disputes caused by the lack of transparency in the query process. This not only solves the problem of threshold query failure caused by differences in the description of medical step names, but also ensures that the threshold data is synchronized with the latest medical insurance rules. At the same time, the audit log meets the regulatory traceability requirements, providing accurate and compliant threshold basis for subsequent determination of valid steps.

[0035] After obtaining the minimum effective time threshold for a single step through a dynamic semantic mapping strategy, the effective step calculation unit needs to compare the actual execution time of the single step with the threshold to determine the validity of the step. However, in clinical practice, there are often special cases where the actual execution time is slightly lower than the threshold but is interrupted due to reasonable reasons. If the rigid rule of invalidating steps below the threshold is applied directly, it is easy to misjudge compliant operations and harm the legitimate rights and interests of medical institutions and patients. Therefore, it is necessary to introduce a boundary condition compensation algorithm to distinguish between reasonable and unreasonable deviations through interruption factor analysis, ensuring that the judgment result is both in line with medical insurance rules and in line with the actual clinical scenario. The specific implementation method is as follows: First, it is necessary to clarify the definition of the actual execution time being slightly lower than the threshold to avoid the abuse of the algorithm due to the ambiguity of the slightly lower range. This range is determined based on the detailed rules of medical insurance charging and the calibration of 1,000 clinical operation deviation data: with the minimum effective time threshold of a single step as the benchmark, the absolute value of the difference between the actual execution time and the threshold is less than or equal to 5% of the threshold. When comparing the time, the effective step calculation unit first calculates the difference between the actual time and the threshold. If the difference is negative (the actual time is lower than the threshold) and the absolute value is less than or equal to 5% of the threshold, it is determined to be slightly lower than the threshold, and the clinical operation interruption factor analysis module is automatically activated.If the difference is less than or equal to 5% of the threshold (actual duration is much lower than the threshold) or greater than or equal to 0 (actual duration reaches the threshold), this module does not need to be triggered; the judgment is made directly according to the conventional logic (significantly lower is invalid, reaching the threshold is valid). The clinical operation interruption factor analysis module is an embedded analysis module specifically used to extract operation interruption information from medical records and classify and determine the nature of the interruption. Its core is to identify and locate reasonable interruption reasons through text keyword recognition. The module has a built-in clinical interruption keyword library, which is divided into three categories according to the interruption reason, covering common reasonable interruption scenarios recognized by medical insurance rules: first, patient physiological interruption; second, equipment operation interruption; and third, clinical process interruption. The keyword library supports the addition of new reasonable interruptions every quarter through the clinical case update interface. After the keyword module is activated, it first associates the corresponding medical record segment with the timestamp, and then uses regular expressions to match interruption keywords in the text segment. If a keyword is matched, an interruption type label is generated according to the keyword's category. For example, cough corresponds to a patient's physiological interruption label, equipment calibration corresponds to an equipment operation interruption label, and temporary confirmation of medical orders corresponds to a clinical process interruption label. If no interruption keyword is matched, a "no reasonable interruption" label is generated, providing core evidence for subsequent exemption determination. If the interruption type label falls within the scope of exemption under the medical insurance charging rules, a manual review process is triggered, and automatic determination is suspended to avoid misjudgment. The scope of exemption under the medical insurance charging rules refers to the slightly shorter duration due to reasonable interruption as clearly defined by the medical insurance regulatory department in the charging rules. The module determines the set of interruption types that are still considered valid at the threshold. This category is pre-stored in the module's exempted interruption type library, corresponding one-to-one with the clinical interruption keyword library. During the specific determination, the module compares the generated interruption type tag with the exempted interruption type library. If the tag is in the library, a manual review process is immediately triggered. The module sends a review request to the medical billing supervision platform via HTTP protocol. The request information includes the step's unique ID, step name, actual duration, threshold, deviation, interruption keyword, interruption type tag, and a screenshot of the corresponding medical record segment, while also specifying the reviewer. Upon receiving the request, the supervision platform automatically generates a review task in the reviewer's workbench and suspends the automatic validity determination process for that step. After logging into the platform, the reviewer checks... The system reviews all information attached to the request and assesses the reasonableness of the interruption based on clinical knowledge. If the review is successful, the step is manually marked as valid. Upon receiving the review result, the module resumes the judgment process and records the review conclusion. If the review is rejected, the system continues processing according to the standard threshold comparison logic. If the interruption type label does not fall under the exemption scope of medical insurance charging rules, the validity of the step is strictly determined according to the threshold comparison result, i.e., the step is deemed invalid and not included in the final count of valid steps. Simultaneously, the interruption type and threshold deviation are recorded in detail in the judgment log to ensure the traceability of the entire judgment process. The judgment log is a structured log storing information about the entire process of step validity judgment, using an immutable time-series storage format to ensure regulatory transparency and judgment compliance.

[0036] After determining the validity of all medical procedure steps and obtaining the final number of valid steps, the billing compliance determination unit needs to calculate the receivable amount based on this number and the preset billing ratio rules. However, in clinical practice, the execution of steps has a strict temporal correlation, and unexpected interruptions will affect the clinical value of the actual time spent. If the calculation is simply based on the number of valid steps multiplied by the fixed fee standard per step, it is easy to ignore the differences in operational complexity and the time value loss caused by interruptions, resulting in a mismatch between the receivable amount and the actual clinical input. Therefore, it is necessary to construct a multi-dimensional calibration factor, through temporal correlation, interruption compensation, complexity weighting, and segmented error tolerance, to ensure that the calculation of receivable amount not only complies with medical insurance rules but also accurately reflects clinical reality. The specific implementation method is as follows: First, the data stream is bound to steps and durations. An operation timeline chain is generated based on the execution order of the effective steps. This chain is a structured sequence formed by linking all effective steps in the order of their actual execution time. Its core function is to reconstruct the clinical workflow logic of the operation and avoid billing calculation errors caused by disordered step arrangement. In practice, the billing compliance determination unit first extracts the timestamp of each effective step from the step-duration-bound data stream. This timestamp represents the time when the step completes conflict resolution and determines its actual execution time. Using this timestamp as the sorting basis, the effective steps are sorted in ascending order (from earliest to latest). If steps with identical timestamps exist, the order is adjusted by referring to the preset step priority sorting table in the medical insurance rules. After sorting, an operation timeline chain is generated and stored as a JSON array. Each element in the array contains a unique step ID, step name, actual execution time (minutes), interruption type label, and timestamp. After generating the operation timeline chain, a compensation time weight coefficient is inserted for unexpected interruptions in the execution of consecutive steps, based on the interruption type label. An unexpected interruption refers to a step that has not been executed. For unplanned, brief pauses caused by objective factors during the process, where the interruption has been determined to fall within the scope of medical insurance exemption through a boundary condition compensation algorithm, the compensation time weight coefficient is a correction coefficient set according to the degree of impact of the interruption on the clinical value of the step. This coefficient is predefined by medical insurance rules and stored in the interruption compensation coefficient database. In practice, the billing compliance determination unit traverses each valid step in the operation sequence chain. If the interruption type label of a step belongs to unexpected interruption and falls within the exemption scope, it queries the interruption compensation coefficient database for the compensation coefficient corresponding to that label and applies the compensation coefficient to the actual single-step compensation of that step. The execution time is multiplied by the compensation coefficient to obtain the actual time after compensation. For example, if the actual time of the enhanced scan venous phase is 1.9 minutes, the interruption label is the patient's physiological interruption, the compensation coefficient is 0.95, and the actual time after compensation is 1.9 × 0.95 = 1.805 minutes. If the step is not interrupted or the interruption is not within the exemption scope, the compensation coefficient is set to 1.0, and the actual time after compensation is equal to the original actual execution time. Through this compensation, the time value loss caused by unexpected interruption is acknowledged, while avoiding the complete negation of the clinical value of the step due to interruption, ensuring that the time calculation is more in line with the actual input.

[0037] The final reimbursement calculation formula needs to incorporate a step complexity weighting factor. This factor is a quantitative coefficient defined by medical insurance rules based on factors such as the technical difficulty, equipment requirements, and personnel qualifications of medical operation steps. It is used to distinguish the differences in clinical value between different steps. This factor is pre-stored in a step complexity factor library, organized in key-value pairs by the operating department and step name. Unlike the traditional simple proportional segmentation (i.e., all steps are calculated using a uniform unit price × duration), after introducing this factor, the reimbursement calculation adopts a weighted summation logic. First, it queries the weighting factor corresponding to each valid step in the complexity factor library, then calculates the weighted cost contribution of each step. The formula is: actual time after compensation × step complexity weighting factor × basic unit price per step. Finally, the weighting factor is calculated. The final chargeable amount is calculated by adding the weighted cost contributions of each valid step. For example, a radiology procedure may involve two valid steps: Step 1, an enhanced scan with arterial phase compensation, lasts 2.1 minutes with a complexity factor of 1.2 and a base charge of 5 yuan / minute, resulting in a weighted cost contribution of 2.1 × 1.2 × 5 = 12.6 yuan; Step 2, an enhanced scan with venous phase compensation, lasts 1.805 minutes with a complexity factor of 1.1 and a base charge of 5 yuan / minute, resulting in a weighted cost contribution of 1.805 × 1.1 × 5 ≈ 9.93 yuan. The final chargeable amount is 12.6 + 9.93 ≈ 22.53 yuan. This calculation method uses the complexity factor to reflect the clinical value differences between different steps, avoiding the unreasonable practice of charging complex and simple steps the same way.

[0038] When comparing the calculated receivables with the actual charges of medical institutions, a segmented error tolerance strategy is adopted. This strategy allows a specific percentage of cost deviation for steps that trigger waiver interruption. The segmented error tolerance strategy involves setting different basic deviation tolerance percentages based on the range of receivables; the higher the amount, the lower the tolerance percentage, preventing large deviations in high-cost items. Simultaneously, the tolerance percentage is further relaxed for steps with waiver interruptions, balancing the rigor of medical insurance supervision with the flexibility of clinical operations. In specific implementation, three ranges and corresponding basic tolerance percentages are preset: 0-100 yuan (inclusive), basic tolerance percentage 5%; 100-500 yuan (inclusive), basic tolerance percentage 3%; and above 500 yuan, basic tolerance percentage 2%. These ranges and percentages are determined by analyzing historical medical insurance charge data. Then, the operation sequence chain is checked for steps with waiver interruption labels. If present, an additional 1% tolerance percentage is added to the basic tolerance percentage; otherwise, the basic tolerance percentage is maintained. Finally, the allowable deviation range is calculated using the formula: receivables × adjusted tolerance percentage. The tolerance ratio is calculated by comparing the absolute value of the difference between the actual charge and the chargeable amount. If the absolute value of the difference is less than or equal to the allowable deviation range, the charge is considered compliant. If the absolute value of the difference is greater than the allowable deviation range, there is a risk of charge violation. For example, if a procedure should have a charge of 22.53 yuan (in the range of 0-100 yuan), and there is an exemption for interruption, the adjusted tolerance ratio is 6%, and the allowable deviation range is 22.53 × 6% ≈ 1.35 yuan. If the medical institution actually charges 23.5 yuan, the difference of 0.97 yuan is less than or equal to 1.35 yuan, and the procedure is considered compliant. If the actual charge is 24 yuan... If the difference is 1.47 yuan, which is greater than 1.35 yuan, it is considered a violation. At the same time, the billing judgment log must record information such as the amount due, the actual amount charged, the amount of deviation, the adjusted tolerance ratio, the allowable deviation range, and whether there is an exemption interruption. For example, the amount due: 22.53 yuan; the actual amount charged: 23.5 yuan; the amount of deviation: 0.97 yuan; the adjusted tolerance ratio: 6%; the allowable deviation range: 1.35 yuan; the exemption interruption: yes (patient's physiological interruption); the judgment result: compliant. This ensures that the judgment process is traceable and provides a basis for generating subsequent violation warning reports.

[0039] In this invention, the data input unit receives medical record text and medical insurance charging rules. The unstructured text parsing unit extracts medical operation steps and their corresponding durations through medical-specific entity recognition, regular expression matching, and graph neural network association. After time-series standardization and conflict resolution, it outputs a unique data stream binding steps and durations. The effective step calculation unit uses a dynamic semantic mapping strategy to query the minimum effective duration threshold, combines a boundary compensation algorithm to determine the effective steps, and accumulates the number. The charging compliance determination unit constructs an operation time sequence chain, introduces compensation time and step complexity weight factors to calculate the chargeable fees, and uses a segmented error tolerance strategy to compare with the actual charges, outputting a warning report with marked violation details, thereby improving the accuracy and compliance of medical charging supervision.

[0040] The second objective of this invention is to provide a method for implementing an intelligent monitoring system for the rationality of medical charges based on large-scale model technology, including any of the above-mentioned features, comprising the following steps: S1. Receive medical record text, extract medical operation steps, time units and numerical description entities through a medical entity recognition model, trigger a sliding window scanning mechanism based on entity results, locate adjacent duration descriptions by combining dependency syntax tree parsing, calculate the association weight between step entities and duration descriptions using a graph neural network position association model, convert duration data into structured key-value pairs using a time series information standardization engine, and output a unique step and duration bound data stream through a conflict resolution mechanism. S2. Obtain the minimum effective duration threshold mapping table for a single step from the medical insurance rules. For each step name and actual execution duration, use a dynamic semantic mapping strategy to query the corresponding threshold. When the actual execution duration is lower than the threshold, activate the interruption factor analysis module to generate an interruption type label. If it is not within the scope of medical insurance exemption, the step is determined to be invalid. Accumulate the number of valid steps that meet the threshold conditions. S3. Construct an operation sequence chain based on the data stream bound by steps and duration, insert a compensation time weight coefficient by combining the interruption type label, and reconstruct the payable cost calculation formula by introducing a step complexity weight factor according to the final effective number of steps and medical insurance ratio rules, so that the actual execution time participates in the weighted calculation, and adopt a segmented error tolerance strategy to handle the exempted interruption type. S4. Compare the weighted calculated receivables with the actual receivables. If the deviation exceeds the tolerance threshold and there is no exemption coverage, generate a violation warning report. The report marks the violation step, the threshold deviation amount, and the conclusion of interrupted analysis.

[0041] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A medical fee rationality intelligent supervision system based on large model technology, characterized in that, Comprise: A data input unit (1) is used to receive medical records text and medical insurance charging rules of medical institutions; An unstructured text analysis unit (2) performs word segmentation and semantic analysis on the medical records text through named entity recognition and regular expression matching, identifies keywords of medical operation steps, locates time length descriptions adjacent to the keywords and step names, generates time unit keywords, and converts the time unit keywords and time length data in the medical records text into a structured numerical format, outputs single step names and corresponding single step actual execution time lengths; A threshold storage unit (3) is used to store pre-defined single step minimum effective time length thresholds in the medical insurance charging rules, which are mapped to each step name in the form of key-value pairs; An effective step calculation unit (4) couples single step names and corresponding single step actual execution time lengths with single step minimum effective time length thresholds, and executes time length threshold calibration logic through a dynamic rule engine, including querying corresponding single step minimum effective time length thresholds for each single step name and single step actual execution time length, comparing single step actual execution time length with single step minimum effective time length threshold, and determining valid single steps only when threshold conditions are met, and accumulating the number of all effective steps as the final effective step number; A charging compliance determination unit (5) calculates the charge based on the final effective step number and the pre-set charging ratio rule, compares it with the actual charging data of the medical institution, determines whether the charging is compliant, and outputs a violation warning report.

2. The medical charge rationality intelligent supervision system based on large model technology according to claim 1, characterized in that: The named entity recognition adopts a medical special entity recognition model based on the fusion of bidirectional long short-term memory network and conditional random field, learns the grammatical structure features of the medical records text through a pre-trained medical corpus, and establishes a special entity tag library for the operation step names of the medical records text. In specific implementation, first, the input text is character-level vectorized encoding, the context semantic features are captured through bidirectional long short-term memory network, and the probability distribution of three types of entities including medical operation steps, time units and numerical descriptions is outputted, and then the conditional random field outputs the optimal entity sequence according to the transition probability constraint, wherein the special entity tag library covers all standard step names and their clinical commonly used synonyms in the radiology enhanced scan protocol.

3. The medical charge rationality intelligent supervision system based on large model technology according to claim 2, characterized in that, The regular expression matching is dynamically constructed based on the named entity recognition results, and its matching rule library contains time description syntax pattern templates: When a medical operation step entity is identified, a sliding window scanning mechanism centered on the medical operation step entity position is triggered to search for time unit entities and numerical description entities within the window range. For the composite time expressions appearing in the scanning window, fuzzy matching rules are used to extract core numerical values, and non-key modifiers are filtered through pre-set cleaning rules. The semantic analysis further introduces dependency syntax tree analysis. When no time entity is directly matched within the sliding window, the time length description clause associated with the medical operation step is traced back according to the subject-predicate-object relationship.

4. The medical charge rationality intelligent supervision system based on large model technology according to claim 1, characterized in that: The time length description adjacent to the positioning keyword and the step name adopts a position association model based on a graph neural network, the position association model takes a medical operation step entity as an anchor point, and constructs a multi-dimensional feature vector containing a part-of-speech label, a syntactic distance and a semantic similarity; The association weight of each time description entity and the medical operation step entity is calculated through the trained graph neural network, and only when the weight exceeds a preset threshold, it is determined as an effective adjacent relationship, when generating the time unit keyword, in the case that there are multiple candidate time length descriptions for the same medical operation step, the time length description with the highest association weight is selected as the effective input, and the time unit keyword is automatically standardized into a unified measurement form.

5. The medical charge rationality intelligent supervision system based on large model technology according to claim 1, characterized in that: The time unit keyword and the time length data in the medical history record text are converted into a structured numerical format, which is specifically realized by a time sequence information standardization engine: The time sequence information standardization engine has a time unit conversion matrix, performs unit normalization on the time unit keyword, for non-standard numerical description, adopts a pre-trained Chinese number conversion model to convert it into Arabic numerals, and finally outputs a key-value pair in a preset structured format, wherein the actual execution time length field is a pure numerical value after unit conversion.

6. The medical charge rationality intelligent supervision system based on large model technology according to claim 5, characterized in that: When outputting the single-step name and the corresponding single-step actual execution time length, a conflict resolution mechanism is set: When there are multiple time length descriptions corresponding to the same medical operation step in the structured data, the highest value is selected according to the association weight order, if the difference between the association weights of multiple time length descriptions is within a tolerance range, a rationality verification based on a clinical operation knowledge base is triggered, abnormal values exceeding medical common sense are automatically removed, and finally step and time length binding data stream with timestamp markers are generated, ensuring that each step name is associated with only one valid single-step actual execution time length.

7. The medical charge rationality intelligent supervision system based on large model technology according to claim 1, characterized in that: When querying the corresponding single-step minimum effective time length threshold, a dynamic semantic mapping strategy is adopted: In the threshold storage unit (3), the step name is accurately matched, if the matching is unsuccessful, the synonym backtracking of the special entity label library is started, the threshold corresponding to the alternative name is queried according to the synonym mapping relationship, when there is a step name change but the threshold is not updated, the latest threshold mapping table is obtained by calling the medical insurance rule version management interface, and the threshold query path is recorded for audit tracking.

8. The medical charge rationality intelligent supervision system based on large model technology according to claim 1, characterized in that: When comparing the single-step actual execution time length with the single-step minimum effective time length threshold, a boundary condition compensation algorithm is introduced: When the actual execution time length is slightly lower than the threshold, the clinical operation interruption factor analysis module is automatically activated, the factor analysis module generates an interruption type label based on the extracted text interruption keywords, if the interruption type belongs to the exemption category of medical insurance charging rules, an artificial review process is triggered to suspend automatic determination, otherwise, the step validity is determined strictly according to the threshold comparison result, and the interruption type and threshold deviation are recorded in the determination log.

9. The medical charge rationality intelligent supervision system based on large model technology according to claim 1, characterized in that: When calculating the charge based on the final effective step number and the preset charging ratio rule, a multi-dimensional calibration factor is constructed: First, according to the step and duration binding data flow, generate operation timing chain in effective step execution order, combine interruption type label, insert compensation time weight coefficient for unexpected interruption in continuous step execution, finally introduce step complexity weight factor into the final receivable fee calculation formula, which is defined by medical insurance rules, so that the actual execution time participates in weighted calculation instead of simple proportional division. When comparing with the actual charging data of medical institutions, a segmented error tolerance strategy is adopted, and a specific proportion of fee deviation is allowed for the triggered exempted interruption type.

10. A method for implementing the intelligent supervision system for medical charge reasonableness based on large model technology according to any one of claims 1-9, characterized in that: The method comprises the following steps: S1, receiving a medical history record text, extracting medical operation steps, time units and numerical description entities through a medical special entity recognition model, triggering a sliding window scanning mechanism based on entity results, combining adjacent time length description through dependency syntax tree analysis, using a graph neural network position correlation model to calculate the correlation weight of step entity and time length description, using a time sequence information standardization engine to convert time length data into structured key-value pairs, and outputting a unique step and time length binding data flow through a conflict resolution mechanism; S2, obtaining a single-step minimum effective time threshold mapping table from medical insurance rules, querying the corresponding threshold for each step name and actual execution time using a dynamic semantic mapping strategy, activating the interruption factor analysis module to generate an interruption type label when the actual execution time is lower than the threshold, and determining that the step is invalid if it is not in the medical insurance exemption category, and accumulating the number of valid steps that meet the threshold condition; S3, constructing an operation timing chain based on the step and time length binding data flow, inserting a compensation time weight coefficient combined with the interruption type label, introducing a step complexity weight factor into the final receivable fee calculation formula according to the final number of valid steps and medical insurance proportion rules, so that the actual execution time participates in weighted calculation, and using a segmented error tolerance strategy to handle exempted interruption types; S4, comparing the weighted calculated receivable fee with the actual charging data, if the deviation exceeds the tolerance threshold and is not covered by exemption, generating a violation warning report, marking the violation steps, threshold deviation and interruption analysis conclusion in the report.

Images