Real-time intelligent auxiliary and generative quality control method and system for electronic medical records

Abstract

The invention particularly relates to a real-time intelligent auxiliary and generative quality control method and system for electronic medical records, and relates to the technical field of medical information. An intelligent differential transmission module; a server context reconstruction and management module; a real-time intelligent auxiliary generation module; and a generative quality control and feedback module. According to the invention, the client interaction module supports multi-modal input and fine-grained snapshot, so that a doctor can input more conveniently, differential transmission is combined with an anchor point mechanism to greatly reduce the data volume, stable transmission can be realized under a weak network, medical record editing gets rid of network constraints, and average editing time consumption is reduced; intelligent auxiliary generation is based on a chapter accurate matching model, continuous writing is pushed in real time, suggestions are complemented, doctors focus on clinical decisions, and repeated input is reduced; according to the generation type quality control, multi-dimensional real-time error correction from integrity, accuracy and consistency is achieved, post rectification is changed into pre-prevention, the medical record quality is guaranteed, and medical resource waste caused by error backtracking is avoided.

Description

Technical Field

[0001] This invention relates to the field of medical information technology, and in particular to a method and system for real-time intelligent assistance and generative quality control of electronic medical records. Background Technology

[0002] Electronic medical records, as the core carrier of medical informatization, play a crucial role in improving diagnostic and treatment efficiency and standardizing medical practices. However, existing systems still have significant pain points. On the one hand, doctors bear a heavy burden of data entry, needing to frequently switch between structured fields and unstructured text. The system suffers from poor adaptability to multimodal input (such as voice and handwriting), low accuracy in recognizing professional terminology, and widespread duplication of redundant information. This results in lengthy medical record writing processes, encroaching on clinical diagnosis and treatment time.

[0003] On the other hand, traditional electronic medical record systems have shortcomings in data transmission and quality control. Full data transmission requires high network bandwidth, which can easily lead to delays and interruptions in primary healthcare institutions or in scenarios with fluctuating network speeds, affecting real-time collaboration. Quality control relies heavily on post-treatment manual review, which not only lags behind the treatment process but also struggles to cover in-depth issues such as clinical logical consistency (e.g., symptom-diagnosis matching) and the standardization of professional terminology, potentially creating hidden medical security risks. Furthermore, the lack of error-tolerance mechanisms for network transmission further exacerbates the instability of medical record data.

[0004] Therefore, a real-time intelligent assistance and generative quality control method and system for electronic medical records is needed to address the problems mentioned above. Summary of the Invention

[0005] The purpose of this invention is to provide a real-time intelligent assistance and generative quality control method and system for electronic medical records in order to solve the above-mentioned problems.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A real-time intelligent assistance and generative quality control system for electronic medical records includes: The client interaction and snapshot management module is configured to ensure snapshot management through editing and context capture, relying on the fine-grained division, incremental update mechanism and local caching of the snapshot engine; The intelligent differential transmission module is configured to achieve a balance between real-time performance and efficiency by minimizing Delta and simplifying identifiers, leveraging the dynamic prioritization of transmission scheduling strategies and network adaptive anti-interference. The server-side context reconstruction and management module is configured as a data integration engine to merge multi-source data and handle conflicts, and achieves efficient resource management through multi-level caching and elastic scaling. The real-time intelligent auxiliary generation module is configured to implement model routing and optimization based on AI, generate continuation suggestion content and push it in a non-intrusive manner; The generative quality control and feedback module is configured such that the multi-dimensional quality control engine achieves multi-dimensional quality control through a dual-drive of rule engine and AI, and the feedback and iteration mechanism corrects errors in real time during the editing process through hierarchical feedback and closed-loop optimization.

[0007] Preferably, the client interaction and snapshot management module specifically includes: It is based on an open-source framework and supports medical-specific formats, and integrates keyboard shortcuts. Multimodal input adaptation includes support for speech-to-text, handwriting trajectory recognition, and real-time conversion of input content into text snapshots; Monitor and record all editing actions, including text manipulation, structure manipulation, and metadata changes; The medical record is split into four levels, each text block is assigned a unique ID, and the snapshot only records the changed text block ID and content hash; When the content of a text block changes, only the hash value of that block is updated and the operation type is marked; a baseline snapshot is generated after each transmission, and subsequent Delta is calculated based on the latest baseline. The client caches the most recent preset number of snapshots locally for use during offline editing.

[0008] Preferably, the intelligent differential transmission module specifically includes: Extracting precise differences based on an optimized version of the Myers algorithm: For nested data, the modified cell is located by table ID, row number, and column number, and only the changed value and unit are transmitted; Identifiers are encapsulated using TLV encoding format, and the total length of all identifiers is strictly controlled within a preset number of bytes. Prioritization is based on the degree of impact of the operation on subsequent AI assistance, the amount of data, and the frequency of the doctor's operation. The transmitted content is encrypted, and the key is dynamically allocated by the server; corresponding layered compression is performed based on different types of data.

[0009] Preferably, the network adaptive anti-interference includes: By periodically sending a preset number of byte probe packets, the round-trip time and packet loss rate are calculated to dynamically classify network levels; The transmission state table is maintained based on the session ID. If the transmission is interrupted, the client will only send unacknowledged Delta fragments after reconnecting.

[0010] Preferably, the method further includes enhancing data robustness through multi-dimensional anchor information: Each data packet sent carries three levels of anchor information, including primary anchor, secondary anchor, and semantic anchor; Main anchor point: formed by concatenating the chapter ID with the starting block hash; Secondary anchor point: formed by concatenating the hash of the predecessor block and the hash of the successor block; Semantic anchors: These are combinations of entities that are closest to and have the strongest semantic relationship among the encoded medical entities. The network level is determined by the data returned by the byte probe packet, and the relevant parameters of the anchor point are dynamically adjusted accordingly. Create an anchor mapping table to clarify the relationships between primary anchors, secondary anchors, and semantic anchors; assign a context-based weight value to each anchor; and adopt a phased transmission method for anchors based on network conditions and the weight value of the anchors. When network congestion or instability occurs, anchors with larger weight values ​​are transmitted first.

[0011] Preferably, the server-side context reconstruction and management module specifically includes: When querying basic patient data, limit the fields returned to reduce the amount of data. For specific departments, supplement specialized data; retrieve previously edited content cached in Redis using the session ID; If multiple terminals are editing simultaneously, conflicts are resolved based on timestamps and version numbers: the latest changes are retained, conflicting parts are marked and notified to the client; CPU / memory usage is monitored in real time, and cache nodes are automatically expanded when the number of sessions exceeds the threshold.

[0012] Preferably, the real-time intelligent auxiliary generation module specifically includes: Model routing automatically matches models based on chapter type, including present medical history / progress records and diagnostic conclusions; The model optimization employs a strategy of pre-filtering with a small model and deep generation with a large model, including: first using a lightweight model to determine whether suggestions need to be generated; calling a large model when necessary, with only the current chapter context as input; and generating subsequent content based on the current input. Push notifications should be sent to the client in a non-intrusive manner. For lightweight texts, it is recommended to display an underlined prompt below the text; for long texts, it is recommended to display them in a floating window on the right.

[0013] Preferably, the generative quality control and feedback module specifically includes: Rigid rules are executed based on a rule engine, and AI quality control is based on medical knowledge graphs, clinical guidelines, and large-scale model training; the quality control dimensions include completeness, accuracy, and consistency. The tiered feedback system is adapted to different levels of problem severity: the prompt level provides optimization suggestions and respects the doctor's judgment; the warning level focuses on safety risks and requires confirmation before implementation; and the error level addresses the core issues and forces corrections to ensure the bottom line. Closed-loop optimization records doctors' adoption rates, allowing low-value suggestions to undergo manual review, and updates the rule base and AI model monthly.

[0014] A method for real-time intelligent assistance and generative quality control of electronic medical records includes: Client-side interaction and snapshot building: Medical records are split into four levels: document, chapter, text block, and sentence. A unique ID is assigned to the changed text block, a hash value is calculated, the operation type is marked, and fine-grained snapshots are generated and cached locally. Intelligent differential transmission and anchor point enhancement: Based on an optimized version of the Myers algorithm, extract the differential identifiers from medical record text and compress them for transmission; add primary anchor points, secondary anchor points, and semantic anchor points to data packets; Server-side context reconstruction and management: After receiving Delta and anchor information, the server queries the patient's basic and specialty data, combines the pre-edited content cached in Redis, and integrates multi-source data; the context is managed through multi-level caching. Real-time intelligent auxiliary generation: Automatically matches models according to medical record chapter types; generates auxiliary content and pushes it to the client in a non-intrusive manner; Generative quality control and feedback loop: Relying on the rule engine to execute rigid quality control rules, and using AI quality control to handle the flexible logic of clinical practice, medical records are scanned from the dimensions of completeness, accuracy and consistency; problems are fed back in a graded manner.

[0015] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. This invention supports multimodal input and fine-grained snapshots through a client-side interaction module, making data entry more convenient for doctors. Differential transmission combined with an anchor point mechanism significantly reduces data volume and ensures stable transmission even under weak network conditions, freeing medical record editing from network constraints and reducing average editing time. Intelligent auxiliary generation uses a chapter-based model to accurately match data and pushes continuation and completion suggestions in real time, allowing doctors to focus on clinical decision-making and reducing repetitive data entry. Generative quality control corrects errors in real time from multiple dimensions, including completeness, accuracy, and consistency, turning post-event rectification into pre-event prevention, ensuring medical record quality and avoiding the waste of medical resources due to error retrospection.

[0016] 2. This invention addresses specific needs by integrating model routing with specialized data, making auxiliary suggestions more aligned with clinical practice. Its network adaptation dimension and three-level dynamic anchor point adjustment strategy ensure that only core anchor points and Delta are transmitted in weak network conditions, significantly increasing transmission success rates. This allows for smooth operation even in network-scarce scenarios such as township hospitals. Multi-level caching and automatic expansion mechanisms guarantee uninterrupted collaborative editing across multiple terminals in teaching hospitals. This allows the system to be both deeply rooted in grassroots medical networks and capable of supporting complex operations in large hospitals, becoming a cornerstone of digital medical records covering all medical scenarios. Attached Figure Description

[0017] Further details, features, and advantages of this application are disclosed in the following description of exemplary embodiments in conjunction with the accompanying drawings, in which: Figure 1 This is a system structure diagram of the present invention; Figure 2 This is a flowchart of the method of the present invention; Figure 3 This describes the entire process of differential transmission. Detailed Implementation

[0018] Several embodiments of this application will now be described in more detail with reference to the accompanying drawings to enable those skilled in the art to implement this application. This application may be embodied in many different forms and for various purposes and should not be limited to the embodiments set forth herein. These embodiments are provided to make this application thorough and complete, and to fully convey the scope of this application to those skilled in the art. The embodiments described do not limit this application.

[0019] Unless otherwise defined, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It will be further understood that terms such as those defined in commonly used dictionaries shall be interpreted as having a meaning consistent with their meaning in the relevant field and / or the context of this specification, and shall not be interpreted in an idealized or overly formal sense unless expressly defined herein.

[0020] Example 1 Its specific implementation method is combined with the appendix Figure 1 Appendix Figure 2 and attached Figure 3 Please provide a detailed explanation.

[0021] Appendix Figure 1 This invention provides a structural block diagram of an electronic medical record real-time intelligent assistance and generative quality control system, which shows the connection relationship between the client interaction and snapshot management module and the generative quality control and feedback module, and marks the main functional interaction flow of each module.

[0022] Appendix Figure 2 The flowchart of a real-time intelligent assistance and generative quality control method for electronic medical records provided in this embodiment of the invention illustrates the complete steps from client interaction and snapshot construction to generative quality control and feedback closed loop.

[0023] In this embodiment, it includes: The client interaction and snapshot management module is configured to achieve smooth editing and context capture through a rich text editing engine, multimodal input adaptation and operation behavior capture, and to ensure efficient snapshot management by relying on the fine-grained division, incremental update mechanism and local caching of the snapshot engine. Specifically, it includes: It is based on open-source frameworks (such as TinyMCE Medical Edition) for secondary development, supports medical-specific formats (such as table entry of vital signs, nested structured fields), and integrates a rich text editing engine with shortcut keys (such as Ctrl+S to trigger snapshot updates). Multimodal input adaptation includes support for speech-to-text (using a medical-specific speech model to recognize professional terms such as heart rate of 90 beats / min), handwriting trajectory recognition (adapting to doctors' handwriting habits), and real-time conversion of input content into text snapshots; Record all editing actions via JavaScript event listeners, including: Text operations: Input / delete characters, cut / paste (distinguish between internal and external paste; external content requires additional quality control checks); Structural operations: chapter folding / expanding, template insertion (such as calling the "hypertension medical history template"), field linkage (such as automatically associating the "blood glucose value" field when "diabetes" is entered); Metadata changes: cursor position (accurate to line number and character index), selected area (such as highlighted "chest pain" text); Medical records are split into four levels: document, chapter, text block, and sentence. Each text block (such as a paragraph) is assigned a unique ID. Snapshots only record the changed text block IDs and content hashes. When the content of a text block changes, only the hash value of that block is updated (using the Blake3 algorithm, which is 3 times faster than MD5), and the operation type of "add / modify / delete" is marked. A baseline snapshot is generated after each transmission, and subsequent Delta calculations are based on the latest baseline to avoid accumulated errors. The client caches the most recent preset number of snapshots (approximately 100KB / time) locally for use during offline editing (e.g., if editing continues when the network is disconnected, the total Delta will be calculated after the network is connected).

[0024] The intelligent differential transmission module is configured as a data bridge between the client and the server. By minimizing Delta and simplifying identifiers, it achieves a balance between real-time performance and efficiency through dynamic priority of transmission scheduling strategy and network adaptive anti-interference. Specifically, it includes: Extracting precise differences based on an optimized version of the Myers algorithm: For example: Text modification scenario: When the doctor changes "patient has had a fever for 3 days" to "patient has had a fever for 3 days, with a maximum body temperature of 39℃", only "Chapter ID=2, text block ID=5, starting from the 3rd character, delete 'burn' and replace it with 'heat', and add ', with a maximum body temperature of 39℃' at the end of the sentence" is transmitted; If a sub-section "Past Medical History" is inserted under the "Present Illness" section, the transmitted content is "Operation Type = Insert, Parent Chapter ID = 2, New Chapter ID = 6, Position = 3rd item", without needing to carry the chapter template content (the server automatically associates the template library through the chapter ID); For nested data such as tables and examination items (e.g., "complete blood count: white blood cells") Red blood cells This allows you to locate and modify cells by using the table ID, row number, and column number, transmitting only the changed values ​​and units. Identifiers are encapsulated using the TLV (Type-Length-Value) encoding format, for example: Patient ID: Type=0x01 (1 byte), Length=6 (bytes), Value=“P123456” (6 bytes after binary compression); Session ID: Type=0x02 (1 byte), Length=4 (bytes), Value=32-bit integer (e.g., 0x1A3B5C7D); Hash value: A truncated 16-byte Blake3 hash (the full hash is 32 bytes) is used to reduce the amount of data transmitted while ensuring uniqueness; The total length of all identifiers is strictly controlled within a preset number of bytes, which reduces the amount of data compared to plaintext transmission; Priority is determined based on the degree of impact of the operation on subsequent AI assistance (e.g., diagnosis modification directly affects recommended medication), the amount of data (e.g., single-word modification vs. template insertion), and the frequency of doctor operations (e.g., cursor movement can reach 10 times per second). Priority execution details: High priority (delay ≤100ms): Transmitted via independent TCP connection, using "timestamp and sequence number" to ensure order, suitable for chapter switching, diagnostic code modification, and key value entry (such as blood pressure, blood sugar). Medium priority (delay ≤ 500ms): A batch merging strategy is adopted, and the queue to be transmitted is checked every 200ms to merge consecutive modifications of the same text block (such as consecutive input of "cough, cough phlegm" is merged into one transmission). Low priority (delay ≤ 1s): After accumulating 5 operations, merge and transmit, and only send when the network is idle (such as cursor movement, font color adjustment) to avoid consuming bandwidth.

[0025] The AES-128-GCM algorithm is used to encrypt the transmitted content. The key is dynamically allocated by the server (once per session), balancing security and computational efficiency (encryption / decryption time < 1ms). Perform corresponding layered compression based on different data types: Text-based Delta: Compression using the LZ77 algorithm (for repetitive terms such as "hypertension" and "diabetes"), achieving a compression rate of 40% to 60%; Identifiers and metadata: Since binary encoding has been used, only simple Huffman coding compression is performed (reducing the amount of data by an additional 10%~15%). Compression selection strategy: When the data size of a single packet is less than 100 bytes, compression is skipped (to avoid compression overhead exceeding benefits); when it is greater than 100 bytes, compression is automatically enabled. Module Collaboration Interface With the client snapshot module: obtain Delta and identifier through callback functions to trigger transport checks (such as evaluating whether transport is needed immediately after a snapshot update); Reconstruction module with server: Defines a unified data packet format (including version number, checksum, and data length), and supports the server to be compatible with different client versions of Delta format through version number; With the caching module: query historical transmission status through session ID to ensure consistency between server-side cache and client-side snapshot chain.

[0026] Network adaptive interference immunity includes: By periodically sending a preset number of byte probe packets (once every 500ms), the round-trip time (RTT) and packet loss rate are calculated to dynamically classify network levels. For example: High-quality network (RTT < 100ms, packet loss rate < 1%): Enables instant transmission and supports real-time synchronization of all operation types; General network (RTT 100~500ms, packet loss rate 1%~5%): Enable redundant verification for low and medium priority operations (add CRC32 checksum to each packet). Weak network (RTT > 500ms, packet loss rate > 5%): Automatically switch to critical data priority mode, only transmit patient ID, session ID, and core text Delta, suspend the transmission of format and metadata, and resend them after the network is restored; The transmission status table is maintained based on the session ID (recording the Delta sequence number that has been successfully received). If the transmission is interrupted, the client will only send the unacknowledged Delta fragments after reconnecting, avoiding full retransmission.

[0027] This also includes enhancing data robustness through multi-dimensional anchor information to combat network failures during data transmission: Each data packet sent carries three levels of anchor information, including primary anchor, secondary anchor, and semantic anchor; Main anchor point acquisition process: Extract the chapter ID (such as "CH02_Present Illness" or "CH03_Physical Examination") from the editing status of the electronic medical record. This is the basic identifier for the structured division of the medical record, marking which major module the modification occurred in. Using text blocks as the smallest unit (such as a continuous text or an input paragraph), the hash value of the currently modified text block is calculated using a hash algorithm (such as Blake3 or SHA-256). The first few bytes (such as 8 bytes) of the hash result are taken as the starting block hash for quickly verifying the integrity of the text block. By concatenating the chapter ID with the starting block hash to form a unique identifier, the server can immediately identify which chapter the modification occurred in and which text block it is associated with, thus locking in the basic scope for context reconstruction.

[0028] Secondary anchor point acquisition process: The hash of the text block that was successfully transmitted (or cached locally) before the current text block is recorded as the predecessor block hash, which represents the state of the text block before the modification occurred; Based on the locally maintained edit snapshot chain (a sequence of text blocks that record historical modifications), predict the text block that may be modified next (or the text block that will be associated with the current modification operation), calculate its hash value, and use it as the hash of the successor block; By concatenating the hash of the predecessor block with the hash of the successor block, a hash association of predecessor, current, and successor is formed, allowing the server to determine the modification order through the hash chain. Even if network transmission fails, the time sequence can be restored according to the hash chain.

[0029] Semantic anchor point acquisition process: The medical named entity recognition (NER) model is invoked to scan the currently modified text content and extract the key entities with the most medical value (such as the disease name "myocardial infarction", the examination indicator "ST segment elevation", and the test result "elevated troponin"). Entities that are strongly related to the diagnosis and treatment of the medical record are given priority. Using a standardized medical terminology coding system, the identified medical entities are converted into compact codes (such as using 4-byte binary codes to represent "ST segment elevation"), thus compressing the volume of semantic information. From the encoded medical entities, select the three nearest and most semantically related entities (for example, related to the symptom "chest pain", "ST segment elevation", "elevated troponin" and "myocardial infarction") and combine them into semantic anchors. This allows the server to understand the text modification intent through entity association and assist in repairing or verifying data.

[0030] Based on the network level classification returned by the byte probe packet, the relevant parameters of the anchor point are dynamically adjusted. When the network bandwidth is low, the amount of anchor point information transmitted is reduced, such as reducing the number of semantic anchor points selected or adopting a more concise encoding method. When the network packet loss rate is high, the redundancy of the anchor point is increased, such as repeatedly sending key anchor point information or using error correction coding to protect the anchor point data, so as to improve the transmission reliability of anchor point information in harsh network environments. Create an anchor mapping table to define the relationships between primary anchors, secondary anchors, and semantic anchors. For example, the primary anchor can quickly locate the related secondary anchor hash chain and the scope of medical entities represented by the semantic anchor. This allows the server to more efficiently utilize this anchor information for data repair and context reconstruction in the event of network failures, incomplete or out-of-order data. Each anchor point is assigned a context-based weight; for example, anchor points located in key parts of the medical record (such as diagnosis conclusions, treatment plans, etc.) are given higher weights, while anchor points in less important text areas are given lower weights. During data repair and recovery, operations are prioritized based on the information from high-weight anchor points to improve the accuracy and reliability of the recovery results.

[0031] Based on network conditions and the weight values ​​of anchor points, a phased approach to anchor point transmission is adopted. When network conditions are good, the main anchor point, secondary anchor point, and semantic anchor point are transmitted in their entirety. However, when the network is congested or unstable, anchor points with higher weight values ​​are transmitted first to ensure that the server can perform basic context positioning and obtain key information. Other semantic anchor points are then transmitted as needed, thereby maximizing the effectiveness of data transmission even under network constraints.

[0032] The server-side context reconstruction and management module is configured as the central hub connecting the front-end transmission and the back-end AI. The data integration engine merges multi-source data and handles conflicts. Through multi-level caching and elastic scaling, it achieves efficient resource management and quickly and accurately restores the editing scene. Specifically, it includes: By querying patient basic data (such as age, gender, and allergy history) through the HL7FHIR interface, the returned fields are limited to only 10 necessary fields to reduce the amount of data. For specific departments (such as cardiology), supplement specialized data (such as electrocardiogram report summaries); Retrieve previously edited content (such as modification records within 30 minutes) from the Redis cache via session ID; If multiple devices are editing simultaneously (e.g., a doctor is editing on both a computer and a tablet), resolve conflicts based on timestamps and version numbers: keep the latest modification (after the timestamp), mark the conflicting part, and notify the client (e.g., "The tablet client modified this part at 10:05, should we merge it?"). Multi-level caching architecture: L1 cache (memory): Stores the complete context of active sessions (≤1000), with a response time ≤10ms; L2 cache (SSD): Stores compressed snapshots of inactive sessions (1000-10000), with a response time ≤100ms; Cold data archiving: Sessions older than 24 hours are compressed and stored in the database, and retrieved when needed.

[0033] Real-time monitoring of CPU / memory usage; when the number of sessions exceeds a threshold (e.g., 2000), automatic scaling of cache nodes (based on K8s container orchestration) to avoid performance bottlenecks.

[0034] The real-time intelligent auxiliary generation module is configured as the core engine to improve doctors' efficiency. It uses AI to realize model routing and optimization, generate continuation suggestions and other content, and push them in a non-intrusive way to avoid invalid suggestions. Specifically, it includes: Model routing automatically matches models based on chapter type, including Present illness / progress record: Utilizes the "chronological narrative model" (excels at organizing symptom descriptions chronologically); Physical examination: Call the "Structured Generative Model" (outputs standardized formats such as "body temperature: 38.5℃, pulse: 90 beats / min"); Diagnostic conclusion: Use the "clinical reasoning model" (combining symptoms to recommend possible diagnoses, such as "upper respiratory tract infection? pneumonia to be ruled out"); Model optimization employs a strategy of pre-filtering with small models and deep generation with large models, including: First, use a lightweight model (such as a tweaked version of BERT for healthcare) to determine whether suggestions need to be generated (e.g., simple data entry does not require suggestions); When necessary, call up a large model and only input the current chapter context (instead of the entire medical record) to reduce computational load; The auxiliary content generation generates subsequent content based on the current input (e.g., inputting "patient cough" generates "accompanied by coughing up sputum, which is white mucous sputum, without hemoptysis"). Complete the information prompts: Identify missing information (such as "The current medical history does not record the cause of the illness") and generate guiding questions (such as "Is it necessary to add: Did the patient catch a cold before the onset of the illness?"). Template Recommendation: Recommend specialist templates based on patient characteristics (such as children, the elderly) (e.g., "Patient Illness History Template for Children with Fever Clinic"). The push notification method is non-intrusive and sent to the client. Lightweight suggestions (such as terminology replacement) are displayed as underlined prompts below the text; suggestions for long texts (such as paragraph continuation) are displayed in a floating window on the right and can be inserted with one click.

[0035] The generative quality control and feedback module is configured as a gatekeeper to ensure the quality of medical records. The multi-dimensional quality control engine achieves multi-dimensional quality control through a combination of rule engine and AI. The feedback and iteration mechanism corrects errors in real time during the editing process rather than requiring rework afterward through hierarchical feedback and closed-loop optimization. Specifically, it includes: Rigid rules are executed based on the rule engine (such as "admission records must include 'chief complaint' and 'present illness' sections" and "surgical records must include the amount of blood loss"). AI quality control is based on medical knowledge graphs, clinical guidelines, and large-scale model training to deeply understand the clinical relevance and logic of medical record texts (such as "whether the symptom description conforms to clinical logic" and "whether the medication dosage matches the patient's weight"). Quality control dimensions include completeness, accuracy, and consistency; Completeness: Check required fields (e.g., "Allergy History" is not filled in) and missing sections (e.g., "Discharge Summary" is missing "Discharge Instructions"). Accuracy: Identify incorrect terms (e.g., "lung cancer" should be written as "lung cancer") and abnormal values ​​(e.g., "blood pressure 12 / 80 mmHg" has an incorrect unit); Consistency: Verify whether the "present illness history" matches the "diagnosis" (e.g., diagnosed as "diabetes" but blood glucose value not recorded), and whether there is a contradiction in the time (e.g., "admission time 2024-11-01, present illness history says 'onset on 2024-12-02'"). Tiered feedback is adapted to different problem severity levels: the prompt level (blue) provides optimization suggestions (such as changing "palpitation" to "heart palpitation"), respecting the doctor's judgment; the warning level (yellow) focuses on safety risks (such as failure to record drug allergy history), requiring confirmation before proceeding; the error level (red) addresses core contradictions (such as conflict between diagnostic codes and text), forcibly correcting to ensure the bottom line. Closed-loop optimization involves recording the doctor adoption rate, allowing low-value suggestions to undergo manual review, and updating the rule base and AI model monthly (such as adding specialized disease quality control logic), so that the quality control capability can dynamically evolve with clinical needs and continuously improve the accuracy of medical record quality control. The system records doctors' adoption of quality control suggestions. For suggestions with low adoption rates (such as overly sensitive "insufficient symptom description length" alerts), a manual review process is automatically triggered. Clinical experts and quality control physicians assess whether the suggestion is a "false alarm" or "unreasonable rule" and decide whether to adjust the rule threshold or delete the suggestion. We integrate the latest industry guidelines (such as the newly added "Monkeypox Diagnosis and Treatment Quality Control Rules") and hospital-specific needs (such as adjustments to specialist medical record templates) every month, and update rigid verification conditions. Based on training with massive amounts of medical record data, the AI ​​optimizes clinical logic judgment (such as improving the accuracy of identifying the correlation between "tumor targeted therapy and gene testing"), making AI quality control more aligned with actual diagnosis and treatment scenarios.

[0036] Example 2 Please see Figure 2 A real-time intelligent assistance and generative quality control method for electronic medical records includes the following parts: Client-side interaction and snapshot creation: Doctors edit medical records using a rich text editing engine that supports medical-specific format input and multimodal input, while capturing editing behavior. Medical records are split into four levels: document, chapter, text block, and sentence. A unique ID is assigned to each changing text block, a hash value is calculated using the Blake3 algorithm, the operation type is marked, and a fine-grained snapshot is generated and cached locally to record the editing status, laying the foundation for subsequent differential transmission. Offline editing is possible when the network is disconnected, and the total difference data (Delta) is calculated when the network is connected. Intelligent Differential Transmission and Anchor Point Enhancement: Based on an optimized version of the Myers algorithm, the differential value (Delta) of medical record text is extracted. Identifiers such as patient ID and session ID are encapsulated using TLV encoding and compressed for transmission. Transmission priorities are divided according to operation impact, data volume, and frequency, and scheduling is carried out through strategies such as independent connection, batch merging, and idle transmission. AES-128-GCM encryption is used in combination with a layered compression algorithm to ensure secure and efficient transmission. At the same time, primary anchor points, secondary anchor points, and semantic anchor points are added to data packets to enhance the ability to resist network failures and dynamically adapt the anchor point transmission strategy to network conditions. Server-side context reconstruction and management: After receiving Delta and anchor point information, the server queries patient basic and specialty data via the HL7FHIR interface, combines previous edited content cached in Redis, and integrates multi-source data; in case of multi-terminal editing conflicts, it resolves them based on timestamps and version numbers; it manages the context through multi-level caching (active sessions in memory, compressed snapshots of inactive sessions stored on SSD, and cold data archived in the database), monitors resources, automatically expands capacity when sessions exceed thresholds, and efficiently reconstructs and maintains the medical record editing context; Real-time intelligent auxiliary generation: Automatically matches chronological narrative, structured generation, clinical reasoning and other models according to the type of medical record chapter (present illness, physical examination, etc.); after pre-filtering by small models, large models are called when necessary, and auxiliary content is generated by inputting only the current chapter context, including continuation suggestions, missing information completion prompts, and specialty template recommendations, etc., and pushed to the client in a non-invasive manner (underline prompts, floating window display) to assist doctors in efficient editing; Generative quality control and feedback loop: Relying on a rule engine to execute rigid quality control rules, AI quality control processes the flexible logic of clinical practice, scanning medical records from the dimensions of completeness (checking required fields and chapters), accuracy (identifying terminology and numerical errors), and consistency (verifying the matching of diagnosis and medical history, and timeline); providing graded feedback on problems (prompt-level optimization suggestions, warning-level safety confirmation, and error-level mandatory correction), recording the adoption rate of doctor suggestions, suggesting manual review for low adoption rates, and updating the rule base and AI model monthly to achieve dynamic evolution of quality control capabilities and ensure the quality of medical records.

[0037] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0038] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

[0039] It should be noted that, in this document, the use of relational terms such as "first" and "second" is merely for distinguishing one entity or operation from another, and does not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0040] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0041] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0042] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0043] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0044] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0045] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0046] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. An electronic medical record real-time intelligent assistant and generation type quality control system, characterized in that, Comprise: Client interaction and snapshot management module is configured to edit and context capture, relying on the fine-grained division of snapshot engine, incremental update mechanism and local cache to guarantee snapshot management; Intelligent differential transmission module is configured to minimize and identifier simplify through Delta, with dynamic priority of transmission scheduling strategy and network adaptive anti-interference, to achieve balance between real-time and efficiency; Server context reconstruction and management module is configured to data integration engine to fuse multi-source data and handle conflicts, through multi-level cache and elastic expansion to realize efficient management of resources; Real-time intelligent auxiliary generation module is configured to realize model routing and optimization based on AI, generate continuation suggestion content and push in a non-intrusive way; Quality control and feedback module is configured, multi-dimensional quality control engine realizes multi-dimensional quality control through rule engine and AI dual driving, feedback and iteration mechanism realizes real-time error correction in the editing process through hierarchical feedback and closed-loop optimization.

2. The electronic medical record real-time intelligent auxiliary and generation type quality control system according to claim 1, characterized in that, Client interaction and snapshot management module, specifically comprising: Based on the secondary development of open source framework, support medical special format, and integrate shortcut keys; Multi-modal input adaptation includes supporting speech to text, handwriting trajectory recognition, and real-time conversion of input content into text snapshots; Listen to all editing actions, including text operations, structure operations, and metadata changes; Split the medical record into four levels, assign a unique ID to each text block, and record only the changed text block ID and content hash in the snapshot; When the text block content changes, only update the hash value of the block and mark the operation type; generate a baseline snapshot after each transmission, and calculate the subsequent Delta based on the latest baseline; The client locally caches the last preset number of snapshots for offline editing.

3. The electronic medical record real-time intelligent assisted and generated quality control system according to claim 1, characterized in that, Intelligent differential transmission module, specifically comprising: Based on the optimized version of Myers algorithm to extract accurate differences: For nested data, locate and modify the cell through table ID, row number, and column number, and only transmit the changed value and unit; Use TLV encoding format to package identifiers, and strictly control the total length of all identifiers within the preset byte number; Based on the influence degree of operation on subsequent AI assistance, data size, and doctor operation frequency as the basis for priority division; Encrypt the transmission content, and the key is dynamically allocated by the server; perform corresponding hierarchical compression based on different types of data.

4. The electronic medical record real-time intelligent assisted and generated quality control system according to claim 3, characterized in that, Network adaptive anti-interference, including: By periodically sending a preset number of byte probe packets, calculate the round trip time and packet loss rate, and dynamically divide the network level; Based on the session ID to maintain the transmission state table, if the transmission is interrupted, the client reconnects and only sends the unconfirmed Delta fragment.

5. The electronic medical record real-time intelligent assisted and generated quality control system according to claim 4, characterized in that, Also includes enhancing data robustness through multi-dimensional anchor information: Each data packet sent carries three levels of anchor information, including primary anchor, secondary anchor, and semantic anchor; Primary anchor: chapter ID and start block hash are spliced; Secondary anchor: the predecessor block hash and the successor block hash are spliced; Semantic anchor: from the encoded medical entities, select the most closely related entity combination; The data returned by the byte probe packet divides the network level, dynamically adjusts the related parameters of the anchor point; An anchor mapping table is created to explicitly indicate the association between the primary anchor, the secondary anchor and the semantic anchor. A context-based weight value is assigned to each anchor. According to the network condition and the weight value of the anchor, the anchor is transmitted in stages. When the network is congested or unstable, the anchor with a larger weight value is preferentially transmitted.

6. The electronic medical record real-time intelligent assisted and generated quality control system according to claim 1, characterized in that, The service end context reconstruction and management module specifically comprises: Querying patient basic data, limiting returned fields, and reducing data volume; For a specific department, supplementing specialized data; calling the pre-edited content of the Redis cache through the session ID; If multiple terminals are edited at the same time, conflicts are solved based on timestamps and version numbers: the latest modification is retained, the conflicting part is marked and the client is notified; real-time monitoring of CPU / memory usage, automatic expansion of cache nodes when the number of sessions exceeds the threshold.

7. The electronic medical record real-time intelligent assisted and generated quality control system according to claim 1, characterized in that, The real-time intelligent auxiliary generation module specifically comprises: Model routing automatically matches models according to chapter types, including present illness history / medical history record, diagnosis conclusion; Model optimization adopts a small model pre-filtering and a large model deep generation strategy, including: using a lightweight model to determine whether a suggestion needs to be generated; calling a large model when necessary, and only inputting the current chapter context; auxiliary content generation generates subsequent content based on the current input; The push method is non-intrusive to the client, and lightweight suggestions are displayed as underlined prompts below the text; long text suggestions are displayed in a floating window on the right side.

8. The electronic medical record real-time intelligent assisted and generated quality control system according to claim 1, characterized in that, The generative quality control and feedback module specifically comprises: Based on the rule engine, rigid rules are executed, and AI quality control is based on medical knowledge graph, clinical guidelines and large model training; quality control dimensions include integrity, accuracy and consistency; Graded feedback adapts to different problem severity, with prompt level for optimization suggestions, respecting doctor's judgment; warning level focuses on safety risks that need to be confirmed before proceeding; error level targets core contradictions, and forced correction ensures the bottom line; Closed-loop optimization records the doctor's adoption rate, allowing low-value suggestions to enter manual review, and updates the rule library and AI model monthly.

9. The method of real-time intelligent assistance and generative quality control of electronic medical record, according to the real-time intelligent assistance and generative quality control system of electronic medical record of any one of claims 1-8, characterized in that, It includes: Client interaction and snapshot construction: according to the four levels of document, chapter, text block and sentence, the medical record is divided, a unique ID is assigned to the changed text block, a hash value is calculated, the operation type is marked, and a fine-grained snapshot is generated and cached locally; Intelligent differential transmission and anchor enhancement: based on the optimized version of Myers algorithm, the medical record text difference is extracted, an identifier is packaged and compressed, and the main anchor, secondary anchor and semantic anchor are attached to the data packet; Service end context reconstruction and management: after receiving the Delta and anchor information, the service end queries the patient's basic and specialized data, combines the pre-edited content of the Redis cache, and integrates multi-source data; the context is managed through multi-level caching; Real-time intelligent auxiliary generation: automatically match models according to medical record chapter types; generate auxiliary content and push it to the client in a non-intrusive manner; Generative quality control and feedback loop: relying on the rule engine to execute rigid quality control rules, using AI quality control to handle clinical flexibility, scanning medical records from the integrity, accuracy and consistency dimensions; graded feedback to problems.

Images