A hospital medical record system based on artificial intelligence

Abstract

The application discloses a hospital medical record system based on artificial intelligence, which comprises a data fusion and management layer, an AI intelligent engine layer, a business application layer and a closed-loop feedback optimization layer. The application realizes AI-driven full-process from medical record generation, in-process quality control, intelligent coding, automatic archiving to data utilization, greatly liberates manpower, and through a medical record digital twin graph model and NLP understanding, manages from a form field level to a medical record connotation and diagnosis and treatment logic level, and realizes real connotation quality control and intelligent management.

Description

Technical Field

[0001] This invention relates to the field of medical information technology, and more specifically to a hospital medical record system based on artificial intelligence. Background Technology

[0002] Medical records are core medical documents that document the entire process of a patient's diagnosis and treatment, serving as crucial evidence for hospital medical quality assessment, performance evaluation, medical insurance reimbursement, and clinical research. With the explosive growth of medical data and the deepening of DRG / DIP medical insurance payment reforms, traditional medical record management models are no longer sufficient to meet the demands of modern hospitals' refined management needs.

[0003] Currently, hospital medical record management faces the following major pain points: First, outdated management methods. Most systems are still based on relational databases, making it difficult to support the deployment of complex artificial intelligence algorithms, resulting in difficulties in data indexing and management, and poor visualization capabilities. Second, inefficient quality control. Traditional quality control relies on post-event sampling with "human supervision," with an average sampling rate of only 5%-12%, far below national requirements. Furthermore, the accuracy of manual review is limited; for example, the accuracy rate of ICD coding in one hospital is only 50%-60%. Third, insufficient intelligence. Existing classification methods mostly rely on keyword matching or static rules, lacking intelligence and adaptive capabilities, and unable to effectively handle synonyms, ambiguities, and content changes in medical records. Finally, severe data silos exist between systems. Medical record data is scattered across different systems, making data sharing and big data intelligent analysis difficult, and hindering the formation of a value loop from quality control to management and research.

[0004] While some technological attempts have been made to improve this, such as converting medical records into graph structures for self-supervised learning to recall similar records, using deep learning models for automatic classification, and employing natural language processing (NLP) and knowledge graphs for quality control of medical record content, these solutions mostly focus on single stages (such as classification, quality control, or filing). They lack a unified, intelligent system framework covering the entire lifecycle of medical records—from generation to management, archiving, analysis, and optimization—and fail to organically connect these stages through a dynamic, data-driven closed-loop feedback model to achieve system self-iteration and continuous optimization.

[0005] Therefore, there is an urgent need for a hospital medical record system that can deeply integrate artificial intelligence technology to achieve intelligent management and automated archiving of medical records, and can continuously improve itself through a closed-loop feedback model. Summary of the Invention

[0006] The purpose of this invention is to provide an artificial intelligence-based hospital medical record system to solve the problems existing in the prior art, enabling intelligent management and archiving of medical records, and allowing the system to dynamically optimize its core AI model and management strategy based on actual application results.

[0007] To achieve the above objectives, the present invention provides the following solution: The present invention provides an artificial intelligence-based hospital medical record system, comprising:

[0008] The data fusion and governance layer is used to connect with the hospital's multi-source information system, collect medical record data, and construct a digital twin graph model of medical records. The graph model uses medical entities as nodes and diagnosis-treatment relationships as edges.

[0009] The AI ​​intelligent engine layer interacts with the data fusion and governance layer information; the AI ​​intelligent engine layer includes an intelligent archiving and classification engine, which is used to output multi-label classification results of medical records based on medical record graph data, through a target medical record classification model that includes graph embedding, graph coding and classification sub-models, to drive automated archiving.

[0010] The full-process quality control engine is used to perform three-level, six-category quality control on medical records documents based on natural language processing and medical knowledge graphs, covering form, logic, and content, to achieve automated quality control of the entire medical record.

[0011] Intelligent coding and homepage generation engine, used to generate ICD codes and generate medical record homepages with one click based on full medical record semantic analysis;

[0012] The business application layer interacts with the AI ​​intelligent engine layer; it is used for functional modules related to clinical practice, management, quality control management, data analysis, and the construction of disease-specific cohorts.

[0013] The closed-loop feedback optimization layer interacts with the business application layer and the AI ​​intelligent engine layer; it is used to collect user feedback and business performance data, dynamically optimize the model strategy of the AI ​​intelligent engine layer through a reinforcement learning framework, and automatically evolve the rule base of the full-process quality control engine through rule lifecycle management.

[0014] Preferably, the target medical record classification model in the intelligent archiving and classification engine has a graph coding sub-model based on a twin graph attention network that integrates text, numerical, and node type features.

[0015] Preferably, the three-level, six-category quality control of the full-process quality control engine includes: the first level is formal quality control based on management standards, the second level is logical quality control based on medical descriptive reasoning, and the third level is intrinsic quality control based on medical behavior supervision.

[0016] Preferably, the reward function of the reinforcement learning framework in the closed-loop feedback optimization layer is defined as:

[0017] ;

[0018] in, This is an indicator function for the adoption of quality control actions. It is a function that accurately indicates itself; This is an indicator function that constitutes interference; The function is an indicator of its own error; α, β, γ, δ are configurable weighting coefficients.

[0019] Preferably, the rule lifecycle management is evaluated by calculating rule health, and the calculation formula is as follows: ;

[0020] Where F is the trigger frequency, P is the accuracy rate, A is the clinical adoption rate, and w1, w2, and w3 are the weights; when the health score H value is lower than the preset threshold, the system automatically downgrades or archives the rule.

[0021] Preferably, the data fusion and governance layer collects data through any of the following methods: view docking, API interface, RPA robot, and performs post-structuring processing on unstructured medical record text.

[0022] Preferably, the intelligent AI archiving management subsystem in the business application layer includes functions such as one-click submission of medical records to generate PDFs, electronic archiving, historical version management, QR code-based tracking of paper medical record circulation, and digital extraction.

[0023] The present invention discloses the following technical effects:

[0024] Intelligent process throughout: It realizes AI-driven process from medical record generation, in-process quality control, intelligent coding, automatic archiving to data utilization, which greatly liberates manpower.

[0025] Deepening Management Transformation: Through the "digital twin" model of medical records and deep NLP understanding, management is moved from the level of form fields to the level of medical record content and diagnosis and treatment logic, realizing true content quality control and intelligent management. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0027] Figure 1 This is an overall architecture block diagram of the artificial intelligence-based hospital medical record system provided in an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the working principle of the closed-loop feedback optimization layer of this invention. Detailed Implementation

[0028] 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.

[0029] This invention discloses an artificial intelligence-based hospital medical record system, as shown in the attached figure. Figure 1 -Appendix Figure 2 The following are included:

[0030] The data fusion and governance layer is used to connect with the hospital's multi-source information system, collect medical record data, and build a digital twin graph model of medical records. The graph model uses medical entities as nodes and diagnosis and treatment relationships as edges.

[0031] The AI ​​intelligent engine layer interacts with the data fusion and governance layer; the AI ​​intelligent engine layer includes an intelligent archiving and classification engine, which is used to output multi-label classification results of medical records based on medical record graph data, through a target medical record classification model that includes graph embedding, graph coding and classification sub-models, to drive automated archiving.

[0032] The full-process quality control engine is used to perform three-level, six-category quality control on medical records documents based on natural language processing and medical knowledge graphs, covering form, logic, and content, to achieve automated quality control of the entire medical record.

[0033] Intelligent coding and homepage generation engine, used to generate ICD codes and generate medical record homepages with one click based on full medical record semantic analysis;

[0034] The business application layer interacts with the AI ​​intelligent engine layer; it is used for functional modules related to clinical practice, management, quality control management, data analysis, and the construction of disease-specific cohorts.

[0035] The closed-loop feedback optimization layer interacts with the business application layer and the AI ​​intelligent engine layer; it is used to collect user feedback and business performance data, dynamically optimize the model strategy of the AI ​​intelligent engine layer through a reinforcement learning framework, and automatically evolve the rule base of the full-process quality control engine through rule lifecycle management.

[0036] The target medical record classification model in the intelligent archiving and classification engine has a graph coding sub-model based on a twin graph attention network that integrates text, numerical, and node type features.

[0037] The three levels and six categories of quality control in the full-process quality control engine include: Level 1 formal quality control based on management standards, Level 2 logical quality control based on medical descriptive reasoning, and Level 3 intrinsic quality control based on medical behavior supervision.

[0038] In a reinforcement learning framework with a closed-loop feedback optimization layer, the reward function is defined as:

[0039] ;

[0040] in, This is an indicator function for the adoption of quality control actions. It is a function that accurately indicates itself; This is an indicator function that constitutes interference; The function is an indicator of its own error; α, β, γ, δ are configurable weighting coefficients.

[0041] Rule lifecycle management assesses rules by calculating their health status, using the following formula: ;

[0042] Where F is the trigger frequency, P is the accuracy rate, A is the clinical adoption rate, and w1, w2, and w3 are the weights; when the health score H value is lower than the preset threshold, the system automatically downgrades or archives the rule.

[0043] The data fusion and governance layer collects data through any means, including view docking, API interfaces, and RPA robots, and performs post-structuring processing on unstructured medical record texts.

[0044] The intelligent AI archiving management subsystem in the business application layer includes functions such as one-click submission of medical records to generate PDFs, electronic archiving, historical version management, QR code-based tracking of paper medical record circulation, and digital extraction.

[0045] In one embodiment of the present invention, a more specific data fusion and governance layer is used to interface with various hospital information systems and collect multi-source heterogeneous medical record data. Its core lies in constructing a unified "digital twin" model for medical records. This layer constructs a graph structure for the acquired medical record data, using entities such as patients, diagnoses, surgeries, medications, and examinations as nodes, and treating relationships, temporal relationships, and logical relationships as edges, forming target medical record graph data. Simultaneously, this layer integrates a natural language processing module to perform post-structural processing on unstructured text, extracting key medical entities, events, and relationships.

[0046] In one embodiment of the present invention, more specifically, the intelligent archiving and classification engine in the AI ​​intelligent engine layer receives medical record graph data from the data layer and processes it through a target medical record classification model. This model includes a graph embedding sub-model, a graph encoding sub-model, and a classification sub-model. The graph embedding sub-model maps different types of nodes (such as diagnoses and drugs) to a unified vector space to obtain a node type vector table. The graph encoding sub-model (such as a fused twin graph attention network GAT) encodes the graph structure and node vectors to obtain a medical record feature matrix that represents the global semantics of the medical record. The classification sub-model completes automatic multi-label classification of medical records based on this matrix (such as disease classification, DRG / DIP pre-grouping, research inclusion identifiers, etc.), and the classification results directly drive the archiving logic.

[0047] Furthermore, in the three-level, six-category quality control model:

[0048] The first level of formal quality control includes aspects such as timeliness and required fields.

[0049] The second level of logical quality control focuses on cross-document consistency and medical rationality.

[0050] The third level of quality control focuses on the completeness of key medical behavior records and the sufficiency of diagnostic and treatment evidence.

[0051] Furthermore, the full-process quality control engine covers more than 35 types of standard documents from admission records to discharge summaries, achieving 100% automated quality control of the entire medical record and replacing less than 20% of manual sampling.

[0052] Furthermore, the intelligent coding and homepage generation engine utilizes knowledge graphs and NLP models to automatically recommend or fill in ICD-10 / ICD-9-CM-3 codes, and performs logical verification on the selection of the primary diagnosis and the coding of surgical procedures. Simultaneously, it can perform semantic analysis on the entire medical record document, automatically extract key information, and "generate" the medical record homepage with a single click, significantly improving the quality and efficiency of the homepage data.

[0053] In one embodiment of the present invention, more specifically, the business application layer provides specific functional modules for different roles in the hospital's "medical-management-research" departments;

[0054] It provides clinicians with an intelligent writing assistant, in-process quality control reminders, diagnostic / surgical omission reminders, DRG / DIP grouping prediction and cost monitoring.

[0055] It provides intelligent AI archiving management, 100% end-of-life quality control of all medical records, automatic generation of quality control reports, digital medical record extraction, and medical record circulation tracking for medical record / quality control departments.

[0056] It provides hospital administrators with multi-dimensional data quality analysis, defect tracing, and medical efficiency and cost analysis dashboards to support management decisions.

[0057] For researchers, it provides flexible tools for building disease-specific cohorts, data asset licensing, and research marketplace services based on high-quality archived medical records, thus supporting clinical research.

[0058] In one embodiment of the present invention, a closed-loop feedback optimization layer constructs a model that enables the system to self-evolve. This layer continuously collects feedback signals from the business application layer, such as user interaction data, quality control result adoption rate, archiving accuracy, and management decision effectiveness. These signals are quantified into optimization objectives and used to dynamically adjust the model parameters and management rules of the AI ​​intelligent engine layer.

[0059] The reward function R is designed as follows:

[0060] ;

[0061] The formula is defined as follows: at time step t, the system observes the case status st and executes quality control actions.

[0062] (e.g., "The chief complaint is inconsistent with the present medical history"). Subsequently, the system receives a reward of rt+1 and a new status of st+1.

[0063] in, It is an indicator function; it is 1 if the doctor adopts the quality control recommendations, and 0 otherwise. This indicates whether the quality control reminder itself is correct; This indicates whether the alert has been marked as "invalid interference" by the doctor; α, β, γ, δ are configurable weighting coefficients.

[0064] Furthermore, the system updates its policy network using reinforcement learning algorithms (such as PPO and DQN).

[0065] The goal of πθ(a∣s) is to maximize the expected cumulative reward J(θ)=Eπθ[∑tγtrt], thereby achieving adaptive optimization of the quality control strategy, reducing invalid reminders, and improving the accuracy and adoption rate of suggestions.

[0066] Furthermore, the self-evolutionary direction of the management rule base is to maintain a dynamically updatable quality control and archiving rule base. The closed-loop feedback layer analyzes high-frequency defect patterns, newly emerging diagnostic and treatment terms, or changes in medical insurance policies. Rule candidates are automatically generated or optimized through machine learning, and then incorporated into the rule base after low-confidence manual review. Simultaneously, an effective rule evaluation index is established, the calculation formula of which is... ;

[0067] When the health H value is lower than the preset threshold, the system will automatically downgrade its weight or archive it to the historical database to ensure the timeliness and effectiveness of the rule base.

[0068] In one embodiment of the present invention, the system's medical record management method includes the following steps:

[0069] S1: Collect and process multi-source medical record data to construct a digital twin of medical records;

[0070] S2: Utilize an AI intelligent engine to intelligently classify and archive the medical record images, perform full-process quality control, and intelligently code and generate the first page;

[0071] S3: Provide intelligent services to users through business application modules;

[0072] S4: Collect user interaction data and business results in step S3, and dynamically adjust the AI ​​engine strategy and parameters in step S2 through closed-loop feedback optimization model.

[0073] In one embodiment of the present invention, an electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the above-described medical record management method.

[0074] In Embodiment 1 of the present invention, a large-scale tertiary-level general hospital is used as the application scenario, and the system construction includes:

[0075] Implementation of the Data Fusion and Governance Layer: This system integrates multiple existing information systems in the hospital, including HIS, EMR, LIS, PACS, and anesthesia management systems. It connects to an intelligent interface data collection platform, supporting data collection in various modes such as views, XML, APIs, and even RPA robots. The collected data is then processed in the data platform. The core task is to construct a "digital twin of medical records."

[0076] The digital twin of a pneumonia patient's medical record is as follows:

[0077] Node construction: Create patient nodes (attributes: ID, age, gender), diagnosis node "community-acquired pneumonia" (linked to the ICD-10 knowledge graph), drug node "levofloxacin", examination node "chest CT", surgery node, etc.

[0078] Edge construction: Create relationship edges such as "patient-confirmed diagnosis-diagnosis", "diagnosis-use-medication", "diagnosis-basis-examination", and "medication-effect on-patient", and assign timestamp attributes to reflect the time sequence of diagnosis and treatment.

[0079] Text processing: Simultaneously, the NLP module parses the medical records, extracting entities and symptoms such as "fever for 3 days" and "right lung moist rales," and associates them as attributes with the corresponding nodes. All unstructured text is transformed into post-structured data.

[0080] Implementation of the AI ​​Intelligent Engine Layer:

[0081] Intelligent Archiving and Classification Engine: A pre-trained target medical record classification model is deployed. When a new medical record image is input, the graph embedding sub-model maps different types of nodes, such as "pneumonia," "CT," and "levofloxacin," into vectors. The graph encoding sub-model (using the improved Siamese GAT network that integrates text, numerical values, and node types) aggregates the information of the entire graph and outputs a fixed-dimensional "medical record feature vector." The classification sub-model (a fully connected neural network) outputs multiple labels based on this vector: disease system classification (respiratory system), DRG pre-grouping (e.g., BR25), whether it meets the single-disease research inclusion criteria for "community-acquired pneumonia" (yes), nursing level, etc. Based on these labels, the system automatically archives the medical record to the "Respiratory System - Pneumonia" category and adds the corresponding research labels, completing intelligent archiving.

[0082] The end-to-end quality control engine operates in real-time while doctors are writing medical records. For example, if a doctor records "patient has no fever today," but the system uses NLP to read the day's temperature chart data as "38.5°C," it immediately triggers the second-level logical quality control, displaying a red warning in the sidebar: "Medical record 'no fever' conflicts with temperature chart data '38.5°C,' please verify." Before archiving, the final quality control module scans the entire medical record, checking for missing informed consent forms, logical consistency of key time points in surgical and anesthesia records, and compliance of antibiotic use with guidelines, among other intrinsic quality control points. All quality control results generate reports, requiring doctors or quality control personnel to implement closed-loop rectification.

[0083] Intelligent coding and homepage generation engine: After a doctor submits a diagnosis of "community-acquired pneumonia," the system automatically recommends the ICD-10 code "J15.9" from the knowledge graph and suggests that a more specific code may be needed. Clicking "Generate Homepage with One Click" allows the system to automatically extract information such as "Admission Condition," "Admission Route," "Discharge Method," and "Pathological Diagnosis" from admission records, medical history, discharge summary, and examination reports and fill them into the corresponding sections of the homepage. Doctors only need to make final confirmation and minor modifications.

[0084] Implementation of the business application layer:

[0085] Clinicians have integrated this system's plugin into the EMR, receiving real-time quality control alerts and DRG cost estimates. Medical records staff use a separate intelligent AI archiving management subsystem to batch process AI-pre-archived medical records for review, archiving, and borrowing management. Hospital administrators, through the management dashboard, can view real-time dashboards for indicators such as "Medical Record Grade A Rate," "Homepage Data Completeness Rate," and "DRG Grouping Abnormality Rate," and can drill down to view detailed deficiencies for specific departments and doctors. Researchers, through the research portal, can utilize the system's "Acute Exacerbation of Chronic Obstructive Pulmonary Disease" specialty database construction tool to quickly screen eligible patient cohorts from the past three years for retrospective studies.

[0086] Implementation of the closed-loop feedback optimization layer:

[0087] Three months after the system went live, the closed-loop feedback layer began to function. Data analysis revealed that the system frequently triggered quality control reminders to "please provide a more detailed description of abdominal signs" for patients with "abdominal pain," but the clinical adoption rate was less than 30%. The feedback layer captured this signal.

[0088] Rule Optimization: Analysis revealed that for patients in the emergency department who had been clearly diagnosed with "acute gastroenteritis" and prescribed simple medication, doctors believed that a detailed description of abdominal signs was unnecessary. Therefore, the system automatically generated an optimization rule candidate: "When the diagnosis includes 'acute gastroenteritis' and the prescription only includes oral medication, reduce the priority of the 'supplement abdominal signs' quality control rule." This rule took effect after being reviewed by the head of the medical records department with low-level access.

[0089] Model Optimization: Simultaneously, numerous negative feedback signals with low adoption rates were quantified as negative rewards in reinforcement learning (i.e., interference terms in the formula). The reinforcement learning agent gradually learned that in case states like "acute gastroenteritis + simple treatment," the expected reward for triggering this quality control action was very low, thus reducing the probability of selecting this action in the policy network. After several iterations, the system significantly reduced invalid prompts in this scenario, resulting in a significant improvement in doctor satisfaction.

[0090] Archive optimization: The system detected that the consistency rate between the AI ​​pre-grouping and the final grouping of medical insurance for medical records in the "BR25" DRG group was as high as 98%. Therefore, the system automatically increased the confidence weight of the grouping model and used it for more proactive cost prediction reminders. For some rare disease groups, the consistency rate was lower. The system automatically labeled these cases and used their feature vectors and the final grouping results as new training samples, periodically initiating incremental learning to update the classification model.

[0091] This embodiment, through the collaborative work of the above four levels, not only achieves high efficiency and accuracy in medical record management, but also realizes that the system becomes "smarter with use" through closed-loop feedback, continuously adapting to the personalized clinical practice and management needs of hospitals.

[0092] In Embodiment 2 of this invention, for in-depth optimization of quality control of specialized medical records, a scenario for quality control of oncology medical records is set up to elaborate on the specific implementation and effect of the reinforcement learning model in the closed-loop feedback optimization layer. Oncology medical records involve complex staging, subtyping, treatment plans (surgery, chemotherapy, targeted therapy, immunotherapy), and efficacy evaluation; the system construction includes:

[0093] Problem definition and state space construction:

[0094] The quality control process for a tumor disease course record is modeled as a Markov decision process.

[0095] State (st): Defined as the vectorized representation of the currently written paragraph and its context at time step t. This includes: the NLP embedding vector of the current paragraph, the patient's current diagnosis (e.g., "lung adenocarcinoma"), TNM stage (e.g., "T2N1M0"), recorded treatment history (e.g., "completed 2 cycles of chemotherapy"), and key entities already written in this medical record (e.g., "complaint of fatigue", "physical examination: PS score 2"), etc. All this information is fused into a high-dimensional state vector.

[0096] Action (at): Defines the type of quality control reminder that the system can execute. For example: {a1: "Prompt for supplementary assessment of adverse reactions after this chemotherapy", a2: "Prompt for tumor marker changes not described", a3: "Prompt for efficacy evaluation not recorded", a4: "No reminder"}.

[0097] Reward (rt): Designed according to the formula described in Example 1, but with a weighting bias towards "intrinsic accuracy". In the context of oncology, an accurate alert regarding "missing efficacy evaluation" has a much higher value (β value) than an alert regarding "incorrect date format".

[0098] Reinforcement learning model training and deployment:

[0099] Initial cold start: Supervised pre-training is performed using historically labeled tumor quality control medical record data (including document paragraphs, expert quality control opinions, and adoption status) to obtain an initial policy network.

[0100] πθinit(a∣s) and a value network.

[0101] Online Interaction and Learning: After the system goes live, it operates in an "exploration-exploitation" mode. For most situations, the system selects actions using the current optimal strategy; for a small number of situations, actions are randomly selected to explore new strategies. After each action is executed, the system collects feedback from doctors: adoption, ignoring, or marking it as "error"; this feedback is combined with subsequent expert sampling review results to calculate an immediate reward rt+1.

[0102] Policy Update: A proximal policy optimization algorithm is employed. Every certain time step (e.g., processing 1000 tumor case records), the policy gradient is calculated using the collected sequence data (state, action, reward, new state), and the policy network parameters θ are updated. The goal of the update is to make the system more inclined to select quality control actions that yield higher long-term rewards in specific case states.

[0103] Closed-loop optimization effect:

[0104] In the initial stages of system operation, it may frequently and indiscriminately remind doctors of all possible missing items, causing them to feel disturbed. For example, on the first day of a patient's illness after admission, it may remind them to "supplement with information on adverse reactions after chemotherapy," even though chemotherapy has not yet begun.

[0105] Through continuous optimization using reinforcement learning, the system gradually learned more sophisticated strategies:

[0106] In state s1 (diagnosis: stage IIIB lung adenocarcinoma, paragraph: initial course of illness, treatment history: planned neoadjuvant chemotherapy): the system learns that the long-term reward of choosing action a4 (without reminder) or only performing formal quality control is higher, because reminding about efficacy assessment is too early at this time.

[0107] In state s2 (diagnosis: stage IIIB lung adenocarcinoma, paragraph: disease course after 2 cycles of chemotherapy, written content: mentions "nausea and vomiting grade I", no mention of tumor markers): the system learned that the reward for choosing action a2 (prompting tumor markers) is higher than that for action a1 (prompting adverse reaction assessment) because adverse reactions have been recorded, and markers are crucial for judging efficacy.

[0108] In state s3 (diagnosis: postoperative lung adenocarcinoma, paragraph: first postoperative follow-up, content written: description of good general condition, no mention of imaging follow-up results): the system learns a strong tendency to select action a3 (suggesting efficacy evaluation / RECIST) because this is the key to postoperative assessment.

[0109] Specific quantification of the management formula: In this embodiment, the reward function can be specifically designed as follows:

[0110] ;

[0111] In this system, "error" refers to a system alert that is incorrect (such as misjudging a missing item), and it carries the highest penalty weight. "Interference" refers to alerts that are technically correct but are unimportant or unnecessary in the current clinical context. Through this reward design, the model is guided to make quality control decisions that are both accurate and timely.

[0112] After 3-6 months of closed-loop operation, the system has achieved a shift from "broad-based" to "precision-targeting" in the quality control of medical records in oncology departments. The clinical adoption rate of quality control reminders has increased from less than 40% initially to over 85%, and the workload of doctors reworking has been reduced by more than 50%. At the same time, it has ensured the integrity of key information records in cancer treatment, providing a high-quality data foundation for subsequent efficacy analysis, scientific research, and medical insurance payment.

[0113] This embodiment vividly demonstrates how a closed-loop feedback model enables AI systems to be deeply integrated into specialized clinical thinking, achieving a qualitative leap in the level of intelligence.

[0114] In the description of this invention, it should be understood that the terms "longitudinal", "lateral", "up", "down", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing this invention, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention.

[0115] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A hospital medical record system based on artificial intelligence, characterized in that, include: The data fusion and governance layer is used to connect with the hospital's multi-source information system, collect medical record data, and construct a digital twin graph model of medical records. The graph model uses medical entities as nodes and diagnosis-treatment relationships as edges. The AI ​​intelligent engine layer interacts with the data fusion and governance layer information; the AI ​​intelligent engine layer includes an intelligent archiving and classification engine, which is used to output multi-label classification results of medical records based on medical record graph data, through a target medical record classification model that includes graph embedding, graph coding and classification sub-models, to drive automated archiving. The full-process quality control engine is used to perform three-level, six-category quality control on medical records documents based on natural language processing and medical knowledge graphs, covering form, logic, and content, to achieve automated quality control of the entire medical record. Intelligent coding and homepage generation engine, used to generate ICD codes and generate medical record homepages with one click based on full medical record semantic analysis; The business application layer interacts with the AI ​​intelligent engine layer; it is used for functional modules related to clinical practice, management, quality control management, data analysis, and the construction of disease-specific cohorts. The closed-loop feedback optimization layer interacts with the business application layer and the AI ​​intelligent engine layer; it is used to collect user feedback and business performance data, dynamically optimize the model strategy of the AI ​​intelligent engine layer through a reinforcement learning framework, and automatically evolve the rule base of the full-process quality control engine through rule lifecycle management.

2. The hospital medical record system based on artificial intelligence according to claim 1, characterized in that: The target medical record classification model in the intelligent archiving and classification engine has a graph coding sub-model based on a twin graph attention network that integrates text, numerical, and node type features.

3. The hospital medical record system based on artificial intelligence according to claim 1, characterized in that, The three-level, six-category quality control of the full-process quality control engine includes: Level 1 formal quality control based on management standards, Level 2 logical quality control based on medical descriptive reasoning, and Level 3 intrinsic quality control based on medical behavioral supervision.

4. The hospital medical record system based on artificial intelligence according to claim 1, characterized in that: The reward function of the reinforcement learning framework in the closed-loop feedback optimization layer is defined as: ； in, This is an indicator function for the adoption of quality control actions. It is a function that accurately indicates itself; This is an indicator function that constitutes interference; The function is an indicator of its own error; α, β, γ, δ are configurable weighting coefficients.

5. A hospital medical record system based on artificial intelligence according to claim 1, characterized in that: The rule lifecycle management is evaluated by calculating rule health, and the calculation formula is as follows: ; Where F is the trigger frequency, P is the accuracy rate, A is the clinical adoption rate, and w1, w2, and w3 are the weights; when the health score H value is lower than the preset threshold, the system automatically downgrades or archives the rule.

6. The hospital medical record system based on artificial intelligence according to claim 1, characterized in that: The data fusion and governance layer collects data through any of the following methods: view docking, API interface, RPA robot, and performs post-structuring processing on unstructured medical record text.

7. A hospital medical record system based on artificial intelligence according to claim 1, characterized in that: The intelligent AI archiving management subsystem in the business application layer includes functions such as one-click submission of medical records to generate PDFs, electronic archiving, historical version management, QR code-based tracking of paper medical record circulation, and digital extraction.

Images