DRG high set identification method, medium and equipment based on bill backtracking and causal imprint

Abstract

The application discloses a DRG high set identification method, medium and equipment based on bill backtracking and causal imprint, and the method comprises the following steps: acquiring medical charging detail data of a to-be-audited medical record, extracting diagnosis and treatment project features and mapping to a medical knowledge graph to obtain a feature atom set; inputting the feature atom set into a large language model to output a real disease phenotype as a pseudo label, and performing semantic comparison with a reported main diagnosis to identify main diagnosis high set; performing score sensitivity grading on other diagnoses reported, screening high-value complication entities, verifying resource consumption matching degrees of the high-value complication entities through a clinical path constraint library, and identifying fictitious complications; extracting main charging items and a low-value charging imprint rule library, identifying virtual increase of main trunk intervention projects through counterfactual causal comparison; and finally, calculating a comprehensive abnormality index by comprehensively considering deviation, matching degree and deviation, and outputting an audit report. The application automatically identifies DRG high set abnormal behaviors from multiple dimensions, and improves the accuracy and comprehensiveness of the audit.

Claims

1. A method for identifying high-level DRG (Dual Representation Group) transactions based on reverse invoice deduction and causal imprinting, characterized in that, The method includes: S1: Obtain the medical charge details data of the medical records to be audited, extract the treatment item features from the medical charge details data, and map the treatment item features to the semantic space of the medical knowledge graph to obtain the feature atom set; S2: Input the set of feature atoms into a large language model pre-trained on medical domain corpus, output the real disease phenotype of the medical record to be audited as a pseudo-label, and obtain the reported primary diagnosis corresponding to the medical record to be audited. The reported primary diagnosis is a diagnosis code declared by the medical institution for disease diagnosis-related grouping. Perform semantic comparison between the pseudo-label and the reported primary diagnosis, and output the deviation between the two. If the deviation exceeds the first preset threshold, mark the medical record to be audited as a suspected high-dose primary diagnosis medical record. S3: Obtain other reported diagnostic sequences corresponding to the medical record to be audited, and use a medical knowledge graph to classify the sensitivity of each diagnostic entity in the other reported diagnostic sequences by score, and identify the diagnostic entities with a score sensitivity exceeding the preset score as high-value complication entities; S4: Call the clinical pathway constraint library to obtain the set of treatment behavior constraints corresponding to the high-value complication entity, extract the resource consumption feature vector corresponding to the set of treatment behavior constraints from the medical charge details data, calculate the matching degree between the resource consumption feature vector and the set of treatment behavior constraints, if the matching degree is lower than the second preset threshold, determine that the high-value complication entity is fictitious, and mark the medical record to be audited as a suspected high-value complication medical record; S5: Extract the main billing items from the medical charge details data. The main billing items are billing items whose unit price exceeds a preset amount threshold. Obtain the low-value billing imprint rule library associated with the main billing items. The low-value billing imprint rule library contains at least one low-value billing item that needs to be executed when the main billing items are implemented, according to preset clinical diagnosis and treatment guidelines. S6: Extract the actual low-value billing trace features from the medical charge details data, compare the actual low-value billing trace features with the low-value billing items in the low-value billing trace rule base using counterfactual causality comparison, and obtain the deviation degree. If the deviation degree exceeds the third preset threshold, mark the medical record to be audited as a medical record with falsely inflated main intervention items. S7: Calculate the comprehensive anomaly index of the medical record to be audited based on the deviation degree output in step S2, the matching degree output in step S4, and the divergence degree output in step S6. When the comprehensive anomaly index exceeds the fourth preset threshold, output a structured audit report.

2. The DRG high-level matching identification method based on reverse invoice deduction and causal imprinting as described in claim 1, characterized in that, In step S1, the features of the diagnostic and treatment items are mapped to the semantic space of the medical knowledge graph to obtain a set of feature atoms, including: S11; Align each diagnostic and treatment item code in the diagnostic and treatment item features with the entity nodes in the medical knowledge graph; S12: Obtain the one-hop neighborhood subgraph of the aligned entity nodes in the medical knowledge graph, wherein the one-hop neighborhood subgraph includes drug entities, consumable entities, surgical operation entities and diagnostic entities that are directly associated with the treatment item code; S13: Encode all entity nodes in the one-hop neighborhood subgraph and the relationship types between them into the feature atom set.

3. The DRG high-level matching identification method based on reverse invoice deduction and causal imprinting as described in claim 1, characterized in that, In step S2, the large language model is a large language model that has been trained through supervised fine-tuning and feedback reinforcement learning using medical corpus. The large language model is configured as an autonomous auditing agent to perform the following steps: S21: After the autonomous auditing agent obtains the set of feature atoms, it calls the preset task decomposition prompt word template to decompose the task of outputting the real disease phenotype into at least one sub-task. S22: The autonomous auditing agent dynamically calls the corresponding data query interface or rule matching engine from the preset medical toolchain according to the sub-task. The medical toolchain includes a drug-diagnosis mapping query interface, a surgery-diagnosis association rule matching engine, and a test indicator-diagnosis logic reasoning module. S23: The autonomous auditing agent performs multi-hop inference based on the execution results of the data query interface or rule matching engine, and quantitatively evaluates the confidence of the current inference result during the multi-hop inference process; when the confidence of the inference result is lower than the preset confidence threshold, a re-inference process is triggered, which includes rolling back to the previous inference step or calling other data query interfaces again. S24: After completing all sub-tasks and the inference results converge, the autonomous auditing agent outputs a unique disease phenotype as the pseudo-label.

4. The DRG high-level document identification method based on reverse invoice deduction and causal imprinting as described in claim 1, characterized in that, In step S4, the resource consumption feature vector corresponding to the set of treatment behavior constraints is extracted from the medical charge details data, including: S41: Perform semantic parsing on each charge item in the medical charge details data, and extract the corresponding drug name, consumable name, treatment item name and charge amount for that charge item; S42: Map the drug name, consumable name, and treatment item name to a preset standard coding system to generate corresponding drug codes, consumable codes, and treatment item codes; normalize the charge amount and map it to a preset amount range code; concatenate the drug code, consumable code, treatment item code, and amount range code to generate the fingerprint feature of the charge item. S43: Based on each constraint in the set of treatment behavior constraints, select the billing items that match the constraint from the fingerprint features to form a subset of billing items corresponding to the constraint; S44: For each subset of charges, count the number of charges, the total amount of charges, and the number of charge types contained in the subset, and concatenate the number of charges, the total amount of charges, and the number of charge types into the resource consumption feature vector.

5. The DRG high-level matching identification method based on reverse invoice deduction and causal imprinting as described in claim 1, characterized in that, In step S7, based on the deviation degree output in step S2, the matching degree output in step S4, and the divergence degree output in step S6, the comprehensive abnormality index of the medical record to be audited is calculated, including: S71: Normalize the deviation, the matching degree, and the divergence degree respectively to obtain normalized deviation, normalized matching degree, and normalized divergence degree; S72: Stack the normalized deviation, the normalized matching degree, and the normalized divergence into a three-dimensional feature tensor matrix; S73: Input the three-dimensional feature tensor matrix into a pre-trained anomaly detection model, the anomaly detection model comprising an input layer, a multi-head self-attention layer, a fully connected layer and an output layer connected in sequence; The multi-head self-attention layer is used to calculate the correlation weights between different feature channels in the three-dimensional feature tensor matrix, and to perform weighted fusion of the three-dimensional feature tensor matrix based on the correlation weights to generate a fused feature vector; the fully connected layer is used to perform nonlinear transformation on the fused feature vector to generate the abnormal score of the medical record to be audited. S74: Map the abnormal scores to a preset index range to obtain the comprehensive abnormal index of the medical record to be audited.

6. The DRG high-level document identification method based on reverse invoice deduction and causal imprinting as described in claim 1, characterized in that, In step S2, the pseudo-label is semantically compared with the reported main diagnosis, and the deviation between the two is output. This also includes: S25: Obtain the medical charge details data of the medical record to be audited, and calculate the Shannon entropy of the cost amount of each medical item in the medical charge details data to obtain the medical resource entropy of the medical record to be audited; the medical resource entropy is used to characterize the uniformity of the consumption of each medical resource in the medical record to be audited. The higher the entropy value, the more dispersed the resource consumption is, and the lower the entropy value, the more concentrated the resource consumption is in a few items. S26: Based on the pseudo-label, query the corresponding standard complexity expected resource entropy from the standard treatment path database; wherein, the standard treatment path database pre-stores the mapping relationship between disease phenotype and standard complexity expected resource entropy; calculate the difference between the standard complexity expected resource entropy and the medical resource entropy to obtain the resource entropy deviation; S27: When the resource entropy deviation exceeds a preset deviation threshold, the resource entropy deviation is used as the incremental correction value of the deviation degree, and added to the deviation degree to obtain the updated deviation degree.

7. The DRG high-level document identification method based on reverse invoice deduction and causal imprinting as described in claim 1, characterized in that, Calculating the matching degree between the resource consumption feature vector and the set of diagnostic and treatment behavior constraints further includes: S45: Construct a four-dimensional heterogeneous knowledge graph. The entity nodes of the four-dimensional heterogeneous knowledge graph include doctor entities, patient entities, drug entities, and consumable entities. The edges are used to represent the relationships between entities. The relationships include doctor's professional qualifications and operating habits, patient's medical history and allergy information, drug indications and contraindications, and the scope of applicable surgeries for consumables. S46: Map the billing items corresponding to the resource consumption feature vectors in the medical records to be audited to the four-dimensional heterogeneous knowledge graph to obtain the graph subgraph of the medical records to be audited; S47: Run an association rule paradox mining algorithm on the graph subgraph to detect whether a preset paradox pattern exists in the graph subgraph; wherein, the association rule paradox mining algorithm is used to traverse the entity nodes and edges in the graph subgraph to identify subgraph structures that violate preset association rules; the preset paradox patterns include: The first paradox pattern is the mismatch between the physician's physical qualifications and the surgical procedure. The second paradox pattern is the pharmacological incompatibility or usage conflict between the drug entity and the consumable entity; The third paradox pattern is a logical contradiction between the temporal sequence and spatial location of diagnostic and treatment actions; S48: When the preset paradox pattern is detected, a paradox penalty coefficient for the medical record to be audited is generated based on the number and severity of the detected paradox patterns, and the paradox penalty coefficient is used as a correction factor for the matching degree to update the matching degree.

8. The DRG high-level document identification method based on reverse invoice deduction and causal imprinting as described in claim 1, characterized in that, In step S7, calculating the comprehensive abnormality index of the medical records to be audited further includes: S75: Obtain multiple historical medical records belonging to the same medical institution as the medical record to be audited, and construct a cross-medical record aggregated heterogeneous graph. The nodes of the aggregated heterogeneous graph include entities such as doctors, patients, diagnoses, drugs, consumables, and billing items. The edges of the aggregated heterogeneous graph are used to represent the co-occurrence relationship between the nodes of the aggregated heterogeneous graph. S76: Run a community detection algorithm based on modularity optimization on the aggregated heterogeneous graph to divide the aggregated heterogeneous graph into multiple subgraphs. The node connection density within each subgraph is higher than the connection density between subgraphs. Subgraphs with connection density exceeding a preset threshold are identified as abnormal subgraphs. S77: Perform structural similarity calculation on the anomalous subgraph, specifically including: using a graph kernel comparison method to compare the structural features of the anomalous subgraph with a preset normal subgraph, wherein the graph kernel comparison method is Weissfell-Lyman subgraph kernel comparison, generating a structural feature vector of the subgraph by iteratively aggregating the neighborhood labels of nodes, and calculating the similarity between the two subgraphs; based on the similarity, detecting whether the anomalous subgraph satisfies a preset collaborative anomalous pattern, wherein the preset collaborative anomalous pattern includes a homogenization anomalous pattern, a star-shaped radiating structure, and time-series synchronous oscillation; S78: When the preset collaborative anomaly pattern is detected, the collaborative anomaly coefficient of the medical record to be audited is generated according to the type of the preset collaborative anomaly pattern and the confidence level corresponding to the type, and the collaborative anomaly coefficient is used as a correction factor for the comprehensive anomaly index to update the comprehensive anomaly index.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the DRG high-level identification method based on invoice reverse deduction and causal imprint as described in any one of claims 1 to 8.

10. An electronic device having a computer program stored thereon, characterized in that, The device includes a processor and a storage medium, wherein a computer program is stored on the storage medium, and when executed by the processor, the computer program implements the DRG high-level identification method based on invoice reverse deduction and causal imprint as described in any one of claims 1 to 8.

Images