A DRG reverse high coding batch self-checking method based on multi-modal evidence deep audit and responsibility chain reconstruction

Abstract

This invention discloses a DRG reverse high-level coding batch self-inspection method and system based on multimodal evidence deep auditing and responsibility chain reconstruction. It is aimed at medical insurance fund supervision and hospital internal coding compliance management scenarios. It addresses the problems of existing auditing methods, such as single evidence dimension, high rate of missed and false detection, weak batch processing capability, and difficulty in tracing responsibility. It constructs a full-process self-inspection system from data collection, anomaly identification, logical auditing, closed-loop review to model optimization. The system connects to heterogeneous data sources such as HIS, EMR, PACS, and LIS, collecting multimodal data including structured diagnostic codes, medical record texts, image reports, and laboratory test results. It then performs cleaning, standardization, and timeline alignment. Based on deep learning and attention mechanisms, it achieves cross-modal evidence fusion, using anomaly scores $S_{anomaly}$ to initially screen high-risk cases. Next, it uses the medical knowledge base and DRG rule base to reverse-audit the coding basis, dynamically reconstructing the chain of responsibility for diagnosis and treatment, and identifying weak evidence points and responsible parties. A closed loop is formed through machine initial screening, manual review, evidence supplementation, and machine re-evaluation. Finally, a visualized audit report is generated, and the model is incrementally trained or fine-tuned using the closed-loop labeled data, thereby continuously improving the accuracy and efficiency of DRG reverse high-risk identification and auditing.

Description

Technical Field

[0001] This invention relates to a DRG reverse high-level batch self-inspection method and system based on multimodal evidence deep auditing and responsibility chain reconstruction. More specifically, it is an automated detection technology and system for medical insurance fund violations that integrates multimodal medical data deep analysis and full-process responsibility traceability for DRG payment supervision scenarios. It is mainly applied in the fields of medical insurance fund supervision and medical service quality control.

[0002] DRG (Diagnosis Related Groups) payment is the core model of current medical insurance payment reform. It achieves refined management of the medical insurance fund by categorizing cases into corresponding groups based on diagnosis and treatment methods. However, in practice, some medical institutions may engage in reverse DRG over-compilation, i.e., by falsifying diagnostic information, exaggerating disease descriptions, or using unreasonable coding, to assign cases to higher-weighted DRG groups with higher payment standards, thereby misappropriating more medical insurance funds. This not only exacerbates the pressure on the medical insurance fund's revenue and expenditure but also undermines the fairness and standardization of medical services. Traditional manual review models rely on the professional experience of reviewers, resulting in low efficiency, limited coverage, and high rates of missed diagnoses and misjudgments, making it difficult to cope with the regulatory needs of massive amounts of medical data.

[0003] This invention leverages multimodal evidence deep auditing technology to integrate structured and unstructured medical data, including medical records, diagnostic codes, laboratory reports, medical orders, and billing details. Utilizing natural language processing, deep learning, and knowledge graph technologies, it performs correlation analysis and in-depth mining of the multimodal data to accurately identify DRG-related high-level coding issues, such as mismatches between case diagnosis and treatment, and discrepancies between coding and patient condition. This allows for precise localization of violations. Simultaneously, the system possesses the capability to reconstruct the chain of responsibility. By recording the entire process of case generation, coding, review, and submission, along with data modification logs, it can trace the responsible parties and operational details at each stage after identifying high-level coding issues. This clarifies the occurrence of violations and the attribution of responsibility, providing clear evidence for subsequent rectification and accountability.

[0004] This method and system support batch self-testing, which can process large-scale case data simultaneously. Compared with manual review, it significantly improves testing efficiency and coverage. Moreover, the testing standards are unified, which can effectively avoid deviations caused by human factors. It provides medical insurance regulatory departments and hospital managers with an efficient and accurate DRG reverse high-level coding testing tool, helps to standardize the operation of DRG payment, and safeguard the security of medical insurance funds and the healthy development of medical services.

[0005] Background Art Related to the Invention

[0006] Firstly, the main problems that the DRG reverse high-level coding batch self-inspection method and system aim to solve are the current pain points in DRG high-level coding review, such as the single dimension, high rate of missed and false detections, insufficient batch processing capacity, difficulty in tracing responsibility, poor human-machine collaboration, and the inability of the model to adapt to the dynamic changes in medical rules. The core objective is to achieve accurate and efficient DRG reverse high-level coding batch self-inspection, identify the specific links and responsible parties of suspected high-level coding, form a closed-loop self-inspection process and continuously optimize the review capabilities, and ultimately standardize medical coding behavior and reduce the loss of medical insurance funds.

[0007] The core idea of ​​this patent is to construct a complete batch self-inspection system based on the deep fusion of multimodal evidence and the reverse reconstruction of the chain of responsibility. Specifically, the process involves: first, collecting multimodal data such as structured diagnostic codes, unstructured medical records, imaging reports, and test results from multiple heterogeneous data sources within the hospital; then, using natural language processing and feature parsing techniques to clean, standardize, and align the data with the timeline, forming a unified multimodal evidence dataset; next, inputting the standardized data into a deep learning model with an attention mechanism to achieve deep fusion and cross-validation of multidimensional evidence, automatically learning the complex mapping relationship between case features and DRG codes, and identifying inconsistencies within the chain of evidence. The system initially screens high-risk cases based on contradictory or abnormal patterns. Then, for these abnormal cases, it utilizes the built-in medical knowledge base and DRG rule base to trace back each coded node of the treatment pathway, dynamically reconstructing the complete chain of responsibility. Key nodes with weak evidence or those not conforming to guidelines are marked, identifying suspected high-risk links and responsible parties. The responsibility chain report is then pushed to reviewers via an interactive interface, enabling human-machine collaborative judgment, supporting evidence supplementation, and forming a closed loop of "machine initial screening - human review - evidence supplementation - machine re-evaluation." Finally, a visual audit report is generated, and the results of human confirmation are used as incremental labeled data to train the model, achieving adaptive optimization of the system. Its highlights lie in the deep fusion of multimodal evidence rather than simple splicing, the accurate tracing achieved through reverse responsibility chain reconstruction, the human-machine collaborative closed-loop self-checking, and the adaptive iteration of the model.

[0008] Currently, there are four main technical solutions and shortcomings in the field of DRG reverse overwriting self-checking: First, the single structured data review method, which relies solely on structured information such as diagnostic codes and billing data for rule matching. Its drawback is that it completely ignores unstructured medical records, images, and other crucial evidence, failing to identify overwriting behavior where "codes do not match clinical reality," resulting in a very high false negative rate. Furthermore, it only supports single-point review and does not support batch processing. Second, the rule engine static review method, which checks through preset fixed code matching rules. Its drawback is that the rule coverage is limited, unable to cope with complex and ever-changing clinical scenarios, and prone to errors in judgments requiring multiple pieces of evidence, such as those concerning complications and comorbidities. The methods used have several drawbacks: misjudgment, high cost of rule updates, and inability to adapt to dynamic adjustments in DRG rules; third, manual sampling review, which relies on manually selecting a small number of cases for review, is extremely inefficient, cannot achieve comprehensive self-inspection of batch cases, has strong subjectivity leading to inconsistent review standards, high missed detection rate, and cannot trace the problem nodes of the entire diagnosis and treatment process, making it difficult to define responsibility; fourth, simple multimodal splicing review, which simply splices together multimodal data such as text and codes for analysis, has the drawback of not achieving deep integration and cross-validation of multiple evidences, being unable to identify logical contradictions within the evidence chain, such as cases where medical records describe mild symptoms but code severe symptoms, resulting in low identification accuracy.

[0009] This patent offers significant advantages over other methods: It addresses the shortcomings of single-structured data review by achieving comprehensive collection and standardization of multimodal data, covering all dimensions of diagnostic and treatment evidence; it addresses the deficiencies of static rule-based review by employing deep learning models to learn complex clinical relationships, and uses an attention mechanism to achieve cross-validation of multiple pieces of evidence, making it more flexible and accurate than fixed rules, and capable of adaptive updates; it addresses the inefficiency of manual sampling by supporting automated initial screening and review of batch cases, with human-machine collaboration significantly improving efficiency and reducing labor costs; it addresses the limitations of simple multimodal data splicing by achieving deep fusion of multimodal evidence, accurately identifying contradictions in the evidence chain, and significantly reducing false positives and false negatives; simultaneously, it solves the problem of difficulty in tracing responsibility through chain of responsibility reconstruction, and closed-loop self-checking and model iteration enable continuous system optimization.

[0010] When using the DRG reverse high-level coding batch self-inspection method and system disclosed in this paper, based on multimodal evidence deep auditing and responsibility chain reconstruction, hospital medical insurance management departments can achieve batch accurate self-inspection of massive numbers of cases, significantly reducing the missed and false detection rates of DRG high-level coding, clarifying the specific links and responsible parties of suspected high-level coding, improving review efficiency while standardizing medical coding behavior; medical insurance regulatory departments can use this system to achieve efficient supervision of hospitals within their jurisdiction, reducing unreasonable expenditures of medical insurance funds; the system can also continuously adapt to changes in medical rules through adaptive iteration, maintaining the accuracy of review in the long term. Therefore, there is an urgent need to further utilize the capabilities of this method and system to solve the practical problems of high missed and false detection rates, low batch processing efficiency, difficulty in tracing responsibility, and inability of review capabilities to dynamically adapt to rule changes in DRG reverse high-level coding self-inspection. This invention is precisely to meet this practical need. Detailed Implementation

[0011] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the present invention.

[0012] This invention discloses a method and system for batch self-inspection of DRG reverse high-resolution code based on multimodal evidence deep auditing and chain of responsibility reconstruction. The steps of this method include:

[0013] [Step 1]: Construct a multimodal medical data acquisition and standardization module. The system automatically collects multimodal data related to the target case from multiple heterogeneous data sources, including hospital information systems, electronic medical records, medical image archiving systems, and laboratory information systems. This data includes structured diagnostic codes, unstructured text medical records, image reports, and laboratory test results. For data from different sources and formats, the system applies natural language processing technology to perform entity recognition and information extraction on text medical records, key feature parsing on image reports, and cleaning, denoising, and standardizing all collected data to form a unified, time-aligned multimodal evidence dataset, laying the data foundation for subsequent in-depth analysis. The specific implementation method of this step is as follows:

[0014] The system establishes connections with heterogeneous data sources such as HIS, EMR, PACS, and LIS through a secure data interface adapter deployed on the hospital's intranet. For structured data (such as ICD-10 diagnostic codes and test values), a predefined mapping table is used. Perform instant standardization, the mapping relationship is as follows ,in The original value, Identify the data source type. This standardizes medical terminology or unit values. This process ensures the comparability of similar data from different systems, laying the foundation for subsequent multimodal alignment.

[0015] For unstructured text data (such as medical records and imaging reports), the system employs a natural language processing pipeline based on a pre-trained language model. First, text cleaning is performed to remove irrelevant symbols and formatting noise. Then, a fine-tuned BERT model is used for named entity recognition, identifying medical entities such as diseases, symptoms, medications, and examination items. The entity recognition model is defined as follows: ,in Given a sequence of text as input, For model parameters, output This is a collection of entities and their types. Next, a relationship extraction module based on rules and attention mechanisms is used. The system extracts medical relationships such as temporal and causal relationships between entities, for example, "Drug A relieved symptom B". For the image report text, it specifically analyzes key feature descriptions, such as "nodule size approximately 1.5cm" and "ground-glass opacity," and quantifies them into structured feature vectors. .

[0016] All collected data items must be aligned with a unified timeline. Each data record is associated with a timestamp. The system employs a time normalization function. Convert the original timestamp to a reference time point. Standardized time (e.g., the patient's first admission time) Finally, a timeline-aligned multimodal evidence dataset is constructed for each target case:

[0017]

[0018] in, Indicates modal type (such as text, encoding, numeric). This dataset is a cleaned and standardized data set in a unified format. Its well-structured architecture supports subsequent time-series analysis and multimodal fusion.

[0019] [Step 2]: Multimodal Evidence Fusion and Anomaly Pattern Recognition Based on Deep Learning. The system inputs the standardized multimodal evidence data processed in Step 1 into a pre-trained deep learning model. This model uses techniques such as attention mechanisms to deeply fuse and cross-validate multi-dimensional evidence, including text, codes, and numerical data, rather than simply splicing them together. The model automatically learns the complex mapping relationship between case features and diagnosis-related group codes, and identifies anomalies such as inconsistencies, contradictions, or insufficient support within the evidence chain. For example, it identifies cases where the severity of the medical record description does not match the assigned high-risk diagnostic code, or where a key examination result is missing but still coded as a complication, thus initially screening out cases that may have a high coding risk. The specific implementation method of this step is as follows:

[0020] The standardized multimodal evidence data output from step one, including text feature vectors, is used to... Encoding feature vectors Sum of numerical eigenvectors As input, the model first projects features from different modalities onto a unified latent space through independent feature mapping layers, obtaining... , , ,in and These are trainable parameters. For deep fusion, the model employs a multi-head cross-attention mechanism. Cross-attention is calculated using the text modality as the query and the encoded and numerical modalities as the key and value, respectively. The core of this process is calculating the attention weights.

[0021]

[0022] in:

[0023]

[0024] This mechanism enables the model to dynamically assess the relevance between different pieces of evidence, such as focusing on laboratory test results and diagnostic codes most relevant to the current medical record description, rather than treating all features equally. This effectively improves the representational power of the fused features. The fused features are represented as follows:

[0025]

[0026] Similarly, using encoding and numerical modalities as queries, cross-attention with other modalities is calculated, and finally all fused features are concatenated. .

[0027] One of the core tasks of the model is to learn the mapping relationship between case features and Diagnosis Related Group (DRG) codes, and to identify abnormal patterns. This involves fusing features. Given a multilayer perceptron, predict the probability distribution of cases belonging to each DRG code. Simultaneously, the model incorporates an anomaly scoring module, which quantifies the degree of anomaly in the evidence chain through reconstruction error and consistency metrics. First, the model attempts to analyze fused features... Reconstruct the original high-level features of each modality and calculate the reconstruction loss:

[0028]

[0029] Large reconstruction errors may indicate that some modal information has been distorted during the fusion process or that there is a fundamental conflict with other evidence. Secondly, the predicted DRG coding probabilities are calculated. Compared to the expected coding distribution derived from a single strong piece of evidence (such as the key surgical code). KL divergence between .

[0030] Finally, the model outputs a comprehensive anomaly score:

[0031]

[0032] in and To balance hyperparameters, a threshold is set. ,when At that time, the case was initially screened as a suspected case with a high risk of being misidentified. For example, when the medical record text describes "mild pneumonia" but the assigned DRG code corresponds to "pneumonia with severe complications," the model, through cross-attention, will find that the association weight between the text features and the high-risk coded features is extremely low, leading to increased reconstruction error and KL divergence, thus generating a high anomaly score and labeling it. This deep learning-based deep fusion and cross-validation method can automatically capture the complex nonlinear interactions and contradictions between multimodal evidence, significantly improving the accuracy and automation level of identifying coding inconsistencies and evidence chain defects.

[0033]

Step 3

[0034] The system identifies each labeled abnormal case. Construct a time series representation vector ,in Indicates the first Timestamps for key medical events (such as diagnosis, surgical procedures, and complication records). Also, timestamps for each event. Extract its corresponding feature vector Includes diagnostic codes Operation codes Severity score and related evidence index The system queries and matches data from an updatable knowledge base. and Relevant clinical pathway templates This template defines a standard sequence of clinical events and their logical constraints. The audit process traces back the events by calculating the alignment between the actual event chain of a case and the standard path. First, events are defined. Events in the standard path Matching degree Diagnostic coding similarity Time logic rationality and the degree of support of evidence Weighted calculation:

[0035]

[0036] in For adjustable weights, The calculation is based on the completeness of medical records and the matching degree of examination results. The core of dynamically reconstructing the chain of responsibility is to find the optimal event alignment sequence. The system solves for the path with the maximum matching degree through dynamic programming, and the objective function is:

[0037]

[0038] in and These represent the actual cases and the first case in the standard pathway, respectively. Index of an alignment event, Penalize logical jumps in the sequence (such as missing necessary pre-checks). This is the penalty coefficient. This optimization process can systematically identify weak points in evidence, such as when... Below the threshold When, corresponding event Marked. For nodes that do not conform to clinical guidelines, the system calculates their correlation with the set of prohibited or non-recommended operations in the knowledge base. Conflict score:

[0039]

[0040] in This is an indicator function. Finally, the system outputs an annotated chain of responsibility graph. The nodes represent medical events, and the edges represent time and logical relationships, satisfying the following conditions: , Value too high or The nodes are highlighted, and based on the information of the departments and personnel associated with the event, potential responsible parties (such as coders and clinicians) are initially identified. This automated auditing method based on dynamic programming and multi-dimensional matching degree enables in-depth tracing of the coding decision basis. Compared with static rules that only verify the final code, it can more accurately reveal the specific logical breakpoints suspected of being over-coded, providing clear targets for subsequent manual review.

[0041]

Step 4

[0042] The system is based on the doubt matrix output in step three. Build an interactive task queue. (The following is an example of the task queue.) This represents the total number of cases pending review. For each case, the corresponding dimension of suspicious features, matrix elements Indicates the first The first case in The system generates a batch push list based on the confidence score of doubt on each rule or node in the chain of evidence. This score is determined by a pre-defined push strategy (e.g., based on score thresholds, department, or disease priority).

[0043]

[0044] in For case identifiers, The push threshold is set. After the reviewer logs into the system, the interface is displayed in the form of a list and a details panel. The report includes case studies and corresponding chains of responsibility for suspected issues. The report visually presents the location of problematic nodes within the evidence graph and the triggered audit rules. And the basis for system reasoning, such as the difference in posterior probabilities calculated based on Bayesian networks. ,here Represents the target diagnosis or coding. For the set of observed evidence, This constitutes the complete set of evidence required by the rules. Reviewers can conduct collaborative human-machine assessments of each point of contention, and the results are recorded as manual labels. These represent "Issue Confirmed," "Requires Supplementary Evidence," and "Dispute Rejected," respectively. The system records the triplet. .

[0045] for For suspected cases, the system will process them according to the rules. The structured definition is used to parse out the set of missing or unconfirmed evidence items. The interface provides support for... Each item The system provides an entry point for initiating supplementary investigations. After the reviewer initiates the process, the system generates a structured investigation task sheet, which is automatically assigned to relevant clinical departments or document management personnel via an internal messaging interface. This process ensures the accuracy and traceability of supplementary evidence requests. Supplementary evidence returned from external sources... It is received by the system, formatted and stored, and associated with the original case data.

[0046] The system is based on human feedback and new evidence This triggers automatic updates and reassessments of the case status. The update logic is handled by a state transition function. Define, where The current status of the case (e.g., "Pending review," "Pending supplementation," "Closed loop"). If or If so, the case status is directly marked as closed loop. and The evidence has been acquired. The system will integrate the old and new evidence and re-evaluate using the core reasoning model from step two. After evidence fusion, the new observation set will be... Input it into the reasoning model and recalculate the doubt score. ,in This is the scoring function for the model. This step achieves self-optimization based on incremental evidence, if... If the closed-loop threshold is reached, the system automatically marks the suspicious point as "resolved" and updates the overall case assessment status. All interactions, judgments, supplementary evidence, and status change logs are persisted, forming a complete audit trail and providing a data foundation for closed-loop quality analysis and model iteration. This closed-loop mechanism of "machine initial screening - human review - evidence supplementation - machine re-evaluation" significantly improves the accuracy and credibility of audit results, while effectively embedding expert knowledge into system rules and models.

[0047]

Step 5

[0048] The system constructs a closed-loop audit data warehouse to store all case data tuples that have completed the manual review process:

[0049]

[0050] in As a characteristic of the case, This is the initial set of anomaly labels predicted by the model. This is a set of audit labels that are ultimately confirmed by human review. As supplementary evidence document vectors, This repository contains case metadata, including department, disease type, and risk level. Based on this repository, the system performs multi-dimensional aggregate queries and statistical analyses. For high-risk distributions, it calculates the proportion of high-risk cases under each Diagnosis Related Group (DRG) or Major Diagnostic Category (MDC):

[0051]

[0052] in The index set of all cases belonging to a specific group, This is an indicator function. The system automatically generates interactive visualization reports in the form of bar charts, heatmaps, pie charts, etc., intuitively displaying departments with clustered risks, frequently occurring problem types (such as diagnostic omissions, missing surgical procedures), and common patterns of missing evidence (such as the missing rate of specific examination reports).

[0053]

[0054] Provide data insights for management decisions.

[0055] At the same time, the system will review the high-quality data generated in the closed-loop process. and the corresponding evidence context After desensitization and standardization, it was added to the model training sample library. The sample database employs versioned management, recording the source batch and review timestamp for each data entry. To achieve adaptive iteration of the model, the system is configured to perform periodic (e.g., monthly) or trigger-based (e.g., when the accumulated sample size reaches a threshold) iterations. The training task involves using incremental learning or fine-tuning strategies with the pre-trained model parameters from step two. As initialization, in the new hybrid dataset Training is performed on the above. The loss function, based on the original cross-entropy, incorporates a constraint term to enhance the audit logic, specifically addressing sensitivity to evidence-missing patterns. For example:

[0056]

[0057] in For the evidence relevance assessment subnetwork, To balance hyperparameters, the model parameters are adjusted through periodic retraining. It can be updated to make its predictions High-quality labels generated by continuously approaching the latest audit standards and practices This design enables the system's anomaly detection capabilities to keep pace with updates to medical coding rules and changes in clinical practice, achieving continuous self-optimization and accuracy improvement. This solves the problem of performance degradation in traditional static models due to rule lag. The entire iterative process is automated, ensuring the system's robustness and evolutionary capabilities in long-term operation.

[0058] Key points (points to be protected)

[0059] [1]: Construct a multimodal medical data acquisition and standardization processing module, which connects to multiple heterogeneous medical systems such as hospital information systems and electronic medical records through a secure data interface adapter, and collects multimodal data such as structured diagnostic codes, unstructured text medical records, image reports, and test results, and performs cleaning and noise reduction processing.

[0060] [2]: For unstructured medical text data, a fine-tuned BERT model is used to complete named entity recognition. The timing, causal and other medical relationships between entities are extracted by combining rules and attention mechanisms. At the same time, the key feature descriptions in the image report are quantified into structured feature vectors.

[0061] [3]: Achieve timeline alignment of multimodal data, convert the original timestamps of each data item into standardized times with reference to the patient's first admission time, and construct a regular timeline aligned multimodal evidence dataset for each case to support subsequent time series analysis and fusion.

[0062] [4]: A multimodal evidence deep fusion model based on multi-head cross-attention is adopted. Different modal features are projected onto a unified latent space and cross-modal cross-validation is performed to dynamically evaluate the correlation between various types of evidence and improve the representation ability of fused features.

[0063] [5]: By combining the abnormal scoring mechanism of reconstruction error and KL divergence, the degree of abnormality of the evidence chain is quantified, and high-risk patterns such as mismatch between medical record description and DRG coding, missing key examinations but coding complications are identified, and suspected cases are preliminarily screened.

[0064] [6]: Conduct reverse DRG coding logic audit for abnormal cases, call up an updatable medical knowledge base and DRG grouping rule base, and solve the optimal alignment between the actual diagnosis and treatment path and the standard path through dynamic programming to accurately identify key nodes with weak evidence and logical jumps.

[0065] [7]: Dynamically reconstruct the complete chain of responsibility for diagnosis and treatment from admission to discharge, mark nodes that do not conform to clinical guidelines, and preliminarily locate the responsible parties suspected of being responsible based on the information of departments and personnel related to the event, providing clear targets for manual review.

[0066] [8]: Construct a batch case interactive self-inspection closed-loop mechanism, push cases according to the confidence level of doubt, support reviewers to view problem nodes and system reasoning basis, initiate a precise evidence supplementation process, and realize a complete closed loop of "machine initial screening - manual review - evidence supplementation - machine re-evaluation".

[0067] [9]: Generate multi-dimensional visual audit reports, statistically display the distribution of high-risk groups and departments in DRG, as well as common problem types, evidence missing patterns, etc., to provide intuitive data support for medical management decisions.

[0068]

[10] : Based on the high-quality manually labeled data generated by closed-loop auditing, the deep learning model is incrementally trained or fine-tuned on a regular basis, and evidence relevance constraints are added to enable the model to adapt to the dynamic changes of medical coding rules and clinical practice, so as to achieve the self-evolution of the system and continuous improvement of accuracy. Attached Figure Description

[0069] Figure 1 This is a flowchart of a DRG reverse high-resolution batch self-inspection method and system based on multimodal evidence deep auditing and chain of responsibility reconstruction.

Claims

1. A method and system for batch self-inspection of DRG reverse high-resolution code based on multimodal evidence deep auditing and chain of responsibility reconstruction, characterized in that, Includes the following steps: 1) Construct a multimodal medical data acquisition and standardization processing module: Automatically collect multimodal data such as structured diagnostic codes, unstructured text medical records, image reports, and test results of target cases from multiple heterogeneous data sources such as hospital information systems, electronic medical records, medical image archiving systems, and laboratory information systems. Use natural language processing technology to complete text entity recognition and information extraction, and key feature parsing of image reports. Then, clean, denoise, and standardize all data and align them to a unified timeline to form a well-organized multimodal evidence dataset. 2) Deep learning-based multimodal evidence fusion and abnormal pattern recognition: Standardized multimodal evidence data is input into a pre-trained deep learning model. An attention mechanism is used to achieve deep fusion and cross-validation of multi-dimensional evidence such as text, code, and numerical data. The complex mapping relationship between case features and DRG codes is automatically learned to identify abnormal patterns that are inconsistent, contradictory, or lack support within the evidence chain and to initially screen high-risk cases. 3) Automated DRG coding logic auditing and dynamic reconstruction of the responsibility chain: For abnormal cases, the built-in updatable medical knowledge base and DRG grouping rule base are invoked to trace the decision basis of each coding node such as diagnosis, operation, and complications in reverse. By analyzing the time sequence, logical relationship and corresponding evidence of medical behavior, the complete diagnosis and treatment responsibility chain is dynamically reconstructed, and key nodes in the chain with weak evidence, logical jumps or non-compliance with clinical guidelines are marked to identify suspected high coding links and responsible parties. 4) Batch interactive self-inspection and evidence supplementation closed loop: The responsibility chain report marked with doubts is pushed to the reviewers in batches through an interactive interface. The reviewers can view the problem nodes and reasoning basis and make human-machine collaborative judgments. For doubtful cases, a supplementary investigation process for specific evidence can be initiated. The system automatically updates the case evaluation status based on human feedback and new evidence, forming a closed-loop self-inspection process of "machine initial screening - human review - evidence supplementation - machine re-evaluation". 5) Visual audit report generation and adaptive model iteration optimization: Multi-dimensional summary analysis is performed on cases that have completed the closed loop to generate a visual audit report containing high-risk distribution, common problem types, evidence missing patterns, etc. At the same time, the manual confirmation results, supplementary evidence and final judgment conclusions are stored as labeled data in the training sample library. New samples are used regularly to incrementally train or fine-tune the deep learning model to achieve self-evolution of the model to adapt to changes in medical practice and coding rules.

2. The DRG reverse high-resolution batch self-inspection method and system based on multimodal evidence deep auditing and chain of responsibility reconstruction as described in claim 1, characterized in that, The multimodal data acquisition and standardization process described in step (1) specifically includes: establishing connections with various heterogeneous data sources through a secure data interface adapter deployed on the hospital intranet; standardizing structured data using a predefined mapping table to ensure the comparability of similar data; performing entity recognition on unstructured text data using a finely tuned pre-trained language model, extracting temporal and causal medical relationships between entities through rules and attention mechanisms, parsing key features of image report text and quantifying them into structured feature vectors; converting the original timestamps of all data into standardized times with reference to the patient's first admission time, and constructing a timeline-aligned case-specific multimodal evidence dataset.

3. The DRG reverse high-resolution batch self-inspection method and system based on multimodal evidence deep auditing and chain of responsibility reconstruction according to claim 1, characterized in that, The multimodal fusion and anomaly identification of the deep learning model described in step (2) specifically includes: projecting features of different modalities into a unified latent space, using a multi-head cross-attention mechanism, using any one modality as the query object and the other modalities as keys and values ​​to calculate attention weights, dynamically evaluating the correlation between various pieces of evidence and completing deep fusion; learning the mapping relationship between case features and DRG codes through a multilayer perceptron, and simultaneously using a weighted combination of reconstruction error and distribution difference to calculate a comprehensive anomaly score, setting a threshold to screen high-risk cases, and accurately identifying contradictions and inconsistencies within the evidence chain.

4. The DRG reverse high-resolution batch self-inspection method and system based on multimodal evidence deep auditing and chain of responsibility reconstruction according to claim 1, characterized in that, The dynamic reconstruction and logical audit of the responsibility chain in step (3) specifically includes: constructing time-series representations and feature vectors of medical events for abnormal cases, retrieving corresponding diagnosis and treatment path templates from the knowledge base; calculating the matching degree between actual events and standard paths by weighting diagnostic coding similarity, temporal logical rationality, and evidence support, using dynamic programming to solve the path with the maximum matching degree and penalizing logical jumps; calculating the conflict score between events and the set of prohibited operations, and finally outputting a responsibility chain diagram with highlighted marks to locate highly suspicious nodes and potential responsible parties.

5. The DRG reverse high-resolution batch self-inspection method and system based on multimodal evidence deep auditing and chain of responsibility reconstruction according to claim 1, characterized in that, The interactive self-inspection and evidence supplementation closed loop described in step (4) specifically includes: generating a batch push list based on the confidence score of the doubts; the interface displays the doubts of the case, the triggering rules and the reasoning basis in the list and details panel; the reviewers can mark the doubts with three types of tags: "confirmed problem", "needs to supplement evidence" and "rejected doubts"; for cases that need to supplement evidence, the system automatically generates a structured investigation task sheet and assigns it to the corresponding personnel. After receiving and associating the supplementary evidence, the system calls the deep learning model to re-evaluate the incremental evidence, updates the case status and doubt score, and forms a complete and traceable audit trajectory.

6. The DRG reverse high-resolution batch self-inspection method and system based on multimodal evidence deep auditing and chain of responsibility reconstruction according to claim 1, characterized in that, The adaptive model iterative optimization described in step (5) specifically includes: constructing a closed-loop audit data warehouse to store case data tuples that have completed the audit, and using versioning to manage the training sample library; periodically or triggerably starting incremental model training tasks, using the original model parameters as initialization, training on a mixed dataset, and adding an evidence relevance constraint term to the loss function based on cross-entropy, so that the model prediction results continuously approach the latest audit standards and clinical practice experience, thereby achieving a dynamic improvement in the system's ability to identify anomalies.

7. The DRG reverse high-resolution batch self-inspection method and system based on multimodal evidence deep auditing and chain of responsibility reconstruction according to claim 1, characterized in that, The multimodal fusion of the deep learning model differs from simple feature splicing. It dynamically focuses on evidence related to the core features of the case through an attention mechanism, captures the nonlinear interaction relationship between multimodal data, significantly improves the accuracy of abnormal pattern recognition, and solves the problem that traditional multimodal splicing cannot identify logical contradictions in the evidence chain.

8. The DRG reverse high-resolution batch self-inspection method and system based on multimodal evidence deep auditing and chain of responsibility reconstruction according to claim 1, characterized in that, The DRG coding logic audit adopts a reverse tracing method, which differs from the static rule audit that only verifies the final code. By reconstructing the complete diagnosis and treatment responsibility chain, it accurately locates the specific links suspected of high coding, thus solving the pain point of difficult responsibility tracing in traditional methods.

9. The DRG reverse high-resolution batch self-inspection method and system based on multimodal evidence deep auditing and chain of responsibility reconstruction according to claim 1, characterized in that, The system supports automated processing of batch cases, from multimodal data collection and initial screening of anomalies to chain of responsibility auditing and batch push. Combined with human-machine collaborative closed loop, it greatly improves review efficiency and solves the problems of low efficiency and high false negative rate of manual sampling review.

10. The DRG reverse high-resolution batch self-inspection method and system based on multimodal evidence deep auditing and chain of responsibility reconstruction according to claim 1, characterized in that, The visualized audit report covers multiple dimensions, including DRG group risk ratio, departmental risk distribution, high-frequency problem types, and common patterns of missing evidence. It provides data insights for hospital medical insurance management and supervision departments. At the same time, the continuous updating of the sample library ensures that the model adapts to changes in medical rules and solves the problem of performance decay of traditional static models.

Images