Intelligent medical record generation system based on large language model

Abstract

The invention discloses an intelligent medical record generation system based on a large language model. The intelligent medical record generation system comprises a KLM fusion inference engine, a dual-channel data writing module and a three-layer real-time quality control module. The invention relates to the technical field of intelligent medical treatment. According to the system, intelligent generation and decision support of clinical medical records are realized through dynamic interaction between a fine-tuned large language model in the medical field and an embedded knowledge graph, seamless joint with various hospital information systems is ensured by adopting a mode of combining interface direct writing and script adaptation, and a three-layer quality control system of an innovative format, logic and time sequence is adopted, so that the system has a good application prospect. Various errors can be intercepted in real time in the writing process, the traditional seven-step medical record writing process is simplified into single-step operation, the average daily writing time of doctors is shortened, the medical record defect rate is reduced, the proportion of personalized content is increased, the system adopts lightweight design, the model size is smaller than 5 GB, deployment of medical institutions at all levels can be adapted, and the medical quality and the working efficiency are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent medical treatment, in particular to an intelligent medical record generation system based on a large language model. BACKGROUND

[0002] Although the traditional electronic medical record can record the basic information and diagnosis and treatment process of the patient, the main mode still depends on copying and pasting. In such a process, doctors often need to do a lot of repetitive work, and it leads to serious homogenization of medical records, and a large amount of time is spent on medical record quality control and arrangement.

[0003] The key problems and limitations of the existing electronic medical record system are as follows: 1. Low efficiency and repetitive work The traditional electronic medical record system relies heavily on manual input by doctors and copying and pasting template content, which leads to the following problems: Time waste: doctors spend 40% of their working time on medical record writing every day, and a single medical record takes more than 23 minutes, which seriously affects the efficiency of clinical diagnosis and treatment.

[0004] Complicated operation: it takes 7 steps to generate a complete medical record (such as selecting a template, copying and pasting, data import, manual quality control, etc.), and the process is redundant.

[0005] Data synchronization delay: test, examination, medication, etc. Data need to be manually entered across systems, with an average lag of 2-48 hours, which is easy to lead to inconsistent or missing information.

[0006] 2. Serious homogenization of medical records, loss of clinical value Template dependence: 82% of medical record content comes from standardized templates, and personalized diagnosis and treatment records only account for 18%, which cannot reflect the individual differences of patients.

[0007] Scientific research data distortion: homogenized medical records are difficult to support real-world research (RWS), affecting the reliability of clinical decision support systems.

[0008] Increased legal risk: in medical disputes, template medical records have weak evidentiary value and are easy to be questioned for their authenticity.

[0009] 3. Quality control lag, high cost of defect repair Post-audit: traditional quality control relies on manual inspection after discharge, and errors cannot be corrected when discovered (such as after medical insurance refusal).

[0010] Inefficient error correction: it takes an average of 30 minutes to repair each medical record defect, and it needs to go through the approval process again, with high management costs.

[0011] Limitations of the rule engine: Existing quality control systems can only detect format errors (such as missing required fields) and cannot identify clinical logical contradictions (such as "warfarin was used but INR was not monitored").

[0012] 4. System fragmentation and prominent data silo problem. Poor compatibility of heterogeneous systems: Hospitals typically use multiple independent systems (HIS, LIS, PACS, etc.), and data interaction relies on manual operation, with an error rate as high as 8.7%.

[0013] Outdated systems are difficult to integrate: Many hospitals are still using non-standardized information systems (such as older versions of HIS based on a client / server architecture), which cannot be directly interacted with via API. Summary of the Invention

[0014] To address the shortcomings of existing technologies, this invention provides a smart medical record generation system based on a large language model, which solves the aforementioned problems.

[0015] To achieve the above objectives, the present invention provides the following technical solution: a smart medical record generation system based on a large language model, comprising: The KLM fusion reasoning engine module consists of a large language model (LLM) finely tuned for the medical field and an embedded medical knowledge graph (KG), which realizes clinical causal reasoning through a dynamic knowledge injection mechanism. The dynamic knowledge injection mechanism triggers the activation of knowledge graph nodes in real time through entity relationship binding; Dual-channel data writing module: Supports dual-mode automatic data synchronization, including direct writing to HL7 / FHIR interfaces and script adaptation for older systems; The dual-mode automatic data synchronization is based on protocol identification using interface probes and dual-channel writing using Selenium scripts to simulate clicks; The three-layer real-time quality control module sequentially performs format compliance verification, clinical logic contradiction detection, and analysis of the rationality of the disease progression timeline. Workflow compression engine module: compresses the medical record generation process into a single step, and supports voice / text input to generate a complete medical record draft.

[0016] Preferably, the operation of the KLM fusion inference engine includes: The LLM layer is based on the Transformer architecture and is fine-tuned using desensitized medical record data from within the hospital. The KG layer includes the ICD disease database, drug knowledge base, and test indicator association database; Dynamic knowledge injection is achieved through an entity-relationship binding mechanism, which forcibly associates relevant diagnosis and treatment rules when a specific medical term is identified.

[0017] Preferably, the dual-channel data writing module includes: The interface probe unit automatically identifies whether the target system supports HL7 / FHIR; The script generation unit automatically creates operation scripts (such as Selenium-simulated clicks) for non-standard systems.

[0018] Preferably, the three-layer real-time quality control module is executed before the doctor saves the medical records: The format layer validates required fields, timestamps, and signature integrity. The logic layer uses KLM to detect contradictions in the diagnosis-testing-medication chain; The intrinsic layer uses a time-series graph neural network to analyze the continuity of the disease course records.

[0019] Preferably, the operation flow of the workflow compression engine is as follows: Input a description of the patient's condition → KLM generates a draft → Doctors revise key decisions → The system automatically inserts templates / synchronizes data / triggers quality control → Outputs the final medical record.

[0020] Preferably, the KG layer supports dynamic updates via: Connect to the latest clinical guideline APIs (such as UpToDate); Learn from historical quality control feedback data from hospitals and automatically optimize knowledge association rules; The system polls the clinical guideline API every 24 hours and triggers rule optimization when the quality control feedback error rate is >5%.

[0021] Preferably, the support for multimodal input is: Voice-to-text input (integrated ASR engine); Checkbox-based structured data input; Free text description input.

[0022] Preferably, a lightweight deployment scheme is adopted: The LLM model uses LoRA fine-tuning technology to adapt to environments with ≥8GB of GPU memory. The edge computing module supports offline generation of medical records.

[0023] Preferably, the integrated digital therapy function: Automatically generate post-hospital rehabilitation plans (e.g., exercise prescriptions for myocardial infarction patients); Synchronize patient-collected data (such as from wearable devices) and generate treatment adjustment suggestions.

[0024] Preferably, the integrated digital therapy function is based on an automatic exercise prescription generated during the patient's rehabilitation phase and a data adaptation scheme for wearable devices.

[0025] This invention also discloses a method for generating intelligent medical records based on a large language model, comprising the following steps: S1: Receive input information (voice / text / selected data); S2: The KLM engine generates a draft medical record and annotates it with relevant knowledge. S3: Write objective data through dual channels and verify time consistency; S4: Three-layer quality control intercepts errors in real time; S5: Output the final medical records and synchronize them to the relevant systems.

[0026] This invention provides an intelligent medical record generation system based on a large language model. Compared with existing technologies, it has the following advantages: This intelligent medical record generation system based on a large language model achieves a revolutionary upgrade to the electronic medical record system by integrating a large language model with a medical knowledge graph into an intelligent architecture: the system simplifies the traditional 7-step operation into a single-step intelligent generation, reducing the time doctors spend writing medical records every day. An innovative three-layer real-time quality control system (format, logic, and timing) enables instant error interception during the writing process, reducing the overall defect rate; The dynamic knowledge injection technology based on patient characteristics increases the proportion of personalized content; the unique dual-channel data writing engine is compatible with 78 types of hospital systems, achieving 100% accurate synchronization of test data. The model size is less than 5GB, making it suitable for deployment in medical institutions at all levels and significantly improving medical quality and work efficiency. This will improve the quality of medical care while achieving significant clinical and economic benefits. Attached Figure Description

[0027] Figure 1 This is a schematic diagram illustrating the operation steps of the KLM fusion inference engine in the system of this invention; Figure 2 This is a schematic diagram illustrating the operation steps of the dual-channel data writing engine in the system of the present invention; Figure 3 This is a schematic diagram of the system of the present invention.

[0028] In the diagram: 1. KLM fusion inference engine module; 2. Dual-channel data writing module; 3. Three-layer real-time quality control module; 4. Workflow compression engine module. Detailed Implementation

[0029] The technical solutions in the embodiments of the present invention have been clearly and completely described. 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.

[0030] Please see Figures 1-3This invention provides a technical solution: a smart medical record generation system based on a large language model, comprising: KLM Fusion Reasoning Engine Module 1: It consists of a large language model LLM finely tuned in the medical field and an embedded medical knowledge graph KG, which realizes clinical causal reasoning through a dynamic knowledge injection mechanism. The operation of the KLM fusion inference engine module 1 includes: The LM layer is based on the Transformer architecture and is fine-tuned using anonymized medical record data from within the hospital. The KG layer includes the ICD disease database, drug knowledge base, and test indicator association database; Dynamic knowledge injection is achieved through an entity relationship binding mechanism, which forcibly associates relevant diagnosis and treatment rules when a specific medical term is identified.

[0031] Dual-channel data writing module 2: Supports dual-mode automatic data synchronization, including direct writing to HL7 / FHIR interfaces and script adaptation for older systems; The dual-channel data writing module 2 includes: The interface probe unit automatically identifies whether the target system supports HL7 / FHIR; The script generation unit automatically creates operation scripts for non-standard systems.

[0032] The three-layer real-time quality control module 3: sequentially performs format compliance verification, clinical logic contradiction detection, and analysis of the rationality of the disease evolution sequence; The three-layer real-time quality control module 3 executes the following before the doctor saves the medical record: The format layer validates required fields, timestamps, and signature integrity. The logic layer uses KLM to detect contradictions in the diagnosis-testing-medication chain; The intrinsic layer uses a time-series graph neural network to analyze the continuity of the disease course records.

[0033] Workflow Compression Engine Module 4: Compresses the medical record generation process into a single-step operation, and supports voice / text input to generate a complete medical record draft.

[0034] The operation process of the workflow compression engine module 4 is as follows: Input a description of the patient's condition → KLM generates a draft → Doctors revise key decisions → The system automatically inserts templates / synchronizes data / triggers quality control → Outputs the final medical record.

[0035] Core technology working principle 1. KLM Fusion Inference Engine Operation Mechanism Detailed steps as follows Figure 1 As shown: Input parsing layer: Voice input: Text is converted using a medical-grade ASR (Automatic Speech Recognition) model with a built-in terminology database (e.g., "heart attack" → acute myocardial infarction).

[0036] Text input: Medical entity recognition was performed using the BERT-BiLSTM-CRF model, achieving an accuracy of >98%. Example: Input “intermittent chest pain last night” → Recognize entity: symptom chest pain (time: last night, nature: intermittent).

[0037] Dynamic injection of knowledge graph Knowledge base structure: { "Disease Node":{ "Acute coronary syndrome":{ Required tests: ["troponin", "ECG"], "Emergency Treatment": ["Aspirin Loading"], "Contraindications":["Active bleeding"]}} } Binding logic: When an increase in solid myocardial troponin is detected: Automatically retrieve diseases associated with "abnormally elevated troponin" in the knowledge graph; Identify the "acute coronary syndrome" node; The necessary procedure for injection is: "ECG must be completed within 30 minutes and aspirin 300mg must be administered." Force the insertion of this logical chain into the generated text.

[0038] Large Language Model Generation Control Using the Prompt constraint template: "As a clinician, please write the patient's progress notes based on the following information:" Patient data: {Input text}; Knowledge rules: {Dynamic injection rules}; Requirements: 1. Include changes in symptoms and test results; 2. Demonstrate the diagnostic reasoning process; 3. Clearly state the next steps in the treatment plan.

[0039] Output example: "The patient experienced sudden intermittent chest pain, and troponin levels rose to 0.5 ng / ml (reference value <0.04 ng / ml), meeting the early warning criteria for acute coronary syndrome. A bedside ECG examination was immediately arranged, and 300 mg of aspirin was administered as antiplatelet therapy..."

[0040] Dual-channel data write engine operation Detailed steps as follows Figure 2 As shown: Key technical points: Interface direct write channel: Convert the test data to HL7 message format: <obx> <OBX.1>1< / OBX.1> <OBX.2>ST< / OBX.2> <OBX.3>30934-4^Troponin T^LN< / OBX.3> <OBX.5>0.5 ng / ml< / OBX.5> <OBX.7>202407171030< / OBX.7> < / obx> .

[0041] Script adaptation channel: Operation simulation for outdated HIS systems: defauto_fill_lis_report(): Click the button that allows you to enter test reports. set_text(field="project", value="troponin T") set_text(field="result", value="0.5") set_text(field="unit",value="ng / ml") click(button="Save") # Automatically trigger the save process.

[0042] Operation of the three-tier real-time quality control system

[0043]

[0044] Example of logic layer quality control: Enter a medical record excerpt: "The patient was diagnosed with deep vein thrombosis and was given warfarin 5mg once a day for anticoagulation therapy."

[0045] Quality control engine execution: Identifying entities: Diagnosis of deep vein thrombosis, medication warfarin; Knowledge graph query: Warfarin → Essential monitoring items → INR test; Medical record search: No INR test records found; A real-time pop-up warning reads: "According to anticoagulation therapy guidelines, INR must be monitored within 24 hours after taking warfarin!"

[0046] Example 1: Generation of daily medical records (for patients with coronary atherosclerosis) Background: The patient was a 65-year-old male who was admitted to the hospital due to chest pain and diagnosed with coronary artery disease. New symptoms appeared during morning rounds on the second day after surgery.

[0047] Doctor's procedure: Click the "Voice Input" button in the system and speak: "The patient experienced chest pain again last night, lasting for about 10 minutes, which was relieved by taking nitroglycerin. This morning, the troponin T level was 0.15 ng / ml, which is higher than before." Result generated: 2024-07-17 09:15 Postoperative Day 2 medical record; The patient experienced sudden chest pain at 22:10 last night, located in the precordial region, lasting for about 10 minutes, which was relieved after taking 0.5mg of nitroglycerin sublingually.

[0048] **Laboratory Tests:** Troponin T (2024-07-17 08:30): 0.15 ng / ml↑ (baseline value 0.08 ng / ml).

[0049] **Analysis and Handling:** Recurrent chest pain accompanied by elevated cardiac markers should raise suspicion of acute in-stent thrombosis (based on the 2023 ESC guidelines).

[0050] 2. Emergency Measures: Perform an 18-lead ECG at the bedside immediately; Urgently check CK-MB and D-dimer.

[0051] 3. Drug treatment: Ticagrelor was maintained at 90 mg twice daily (last dose at 09:00). Low molecular weight heparin 4000 IU subcutaneously administered as stat.

[0052] Example 2: Automatic generation of discharge records (after laparoscopic cholecystectomy) Background: The patient is a 52-year-old female who underwent laparoscopic surgery for gallstones and is recovering well post-surgery and is scheduled for discharge.

[0053] System trigger condition: The doctor clicks the "Prepare for Discharge" button. System execution flow: Data scanning phase: Extract the following from the HIS: admission diagnosis, surgical record, and medication list; Extracted from LIS: Last liver function and complete blood count results; Extracted from the nursing system: incision healing status (VAS score); KLM Key Event Chain Extraction: events=[ {"Time":"2024-07-10","Event":"Admission Diagnosis: Gallstones with Chronic Cholecystitis"}, {"Time":"2024-07-12","Event":"Laparoscopic cholecystectomy under general anesthesia"}, {"Time":"2024-07-13","Event":"Day 1 post-surgery, passing gas and starting a liquid diet"}, {"Time":"2024-07-15","Event":"Incision sutures removed, primary healing"}).

[0054] Intelligent generation of discharge summary: **Discharge diagnosis:** 1. Gallstones with chronic cholecystitis (K80.1); 2. Postoperative recovery is good (Z48.001) **Hospitalization Process:** On July 12, 2024, a laparoscopic cholecystectomy was performed. The operation lasted 65 minutes and the blood loss was 20 ml. Postoperatively, cefuroxime was administered for 3 days for infection control, and parecoxib sodium was used for pain control. The incision was sutured on July 15, 2024, and the wound healed well (VAS score of 2).

[0055] **Discharge Instructions:** Diet: After 2 weeks of low-fat diet, transition to a regular diet; Activity: Avoid lifting weights greater than 5kg for 4 weeks; 3. Follow-up visit: An abdominal ultrasound examination will be conducted at the outpatient clinic 7 days after the operation.

[0056] Automatically fill in the medical record cover page

[0057] Example 3: Intelligent completion of surgical records (radical gastrectomy for gastric cancer) Background: The surgeon verbally describes the key steps after the operation, and the system automatically generates a complete surgical record.

[0058] Doctor enters: "Laparoscopic radical gastrectomy for gastric cancer under general anesthesia, D2 lymph node dissection, Billroth I anastomosis, intraoperative blood loss 200ml."

[0059] System enhancement processing: Structured parsing: { Surgical Procedure Name: Laparoscopic Radical Gastrectomy "Anesthesia method": "General anesthesia" Key steps: ["D2 lymph node dissection", "Bill's I anastomosis"] "Bleeding volume": "200ml"}.

[0060] Knowledge graph association extension: Identify “D2 lymph node dissection” → Automatically insert standard dissection range: “The dissection range includes: perigastric lymph nodes in groups 1, 3, 4sb, 4d, 5, 6, 7, 8a, 9, 11p, and 12a.”

[0061] Identify “Bill-Archive I Anastomosis” → Associate with anatomical description: “The duodenum was transected 3 cm distal to the tumor, and a gastroduodenal anastomosis was performed.” Generate complete record: **Surgical Procedure:** 1. After anesthesia is achieved, place the patient in a supine position and establish pneumoperitoneum (pressure 12 mmHg). 2. Exploration of the abdominal cavity: No distant metastasis was found; the tumor was located in the antrum of the stomach, measuring approximately 3×2cm. 3. Perform D2 lymph node dissection (range: No.1 / 3 / 4sb / 4d / 5 / 6 / 7 / 8a / 9 / 11p / 12a groups); 4. The stomach body is cut 5 cm proximal to the tumor, and the duodenum is cut 3 cm distal to the tumor; 5. Perform Billroth I anastomosis of the stomach and duodenum (anastomosing device model: CDH25); 6. Intraoperative blood loss was approximately 200ml; no blood transfusion was required. **Postoperative diagnosis:** Gastric antral adenocarcinoma (pT2N0M0, stage IIA).

[0062] III. Technical Effect Verification Data Deployment test in the general surgery department of a hospital (sample size: 200 medical records):

[0063] Technological Breakthrough and Clinical Value of the Invention Efficiency Improvement: Single-step generation, freeing up doctors' time. The operation has been streamlined from 7 steps to 1 step. Doctors only need to verbally describe the patient's condition or select key information, and the system will complete the process automatically. Template matching Test data synchronization Logical Analysis and Medical Record Generation Writing time is reduced by 87%, allowing doctors to focus more on diagnosis and treatment.

[0064] Data is synchronized in real time, and test results, medication records, etc. are written into medical records with zero delay, avoiding errors caused by manual entry.

[0065] 2. Revolutionary Quality Control Capabilities: Real-time Interception, Error Rate Approaching Zero

[0066]

[0067] Typical Case: Intercepting "No drainage volume recorded post-surgery" (timing logic error). Correct the error of "failure to record allergy history when using antibiotics" (clinical compliance error). Avoid "DRG grouping errors" (such as omitting key surgical procedures). 3. Personalized diagnosis and treatment support to improve the quality of medical care. Dynamic knowledge injection: Automatically connects the latest guidelines to generate personalized recommendations based on patient characteristics (such as age and complications).

[0068] Example: A customized blood glucose monitoring plan is generated for a diabetic patient after surgery.

[0069] Standardization of research data: Medical records automatically conform to CDISC standards and can be directly used in real-world studies (RWS).

[0070] Improved patient compliance: Doctor's orders are described more clearly (e.g., "take aspirin 30 minutes after a meal" instead of "take as per the instructions").

[0071] 4. Hospital Operation Optimization: Cost Reduction and Efficiency Improvement, Compliant Revenue Generation Direct economic benefits: Reduce medical insurance refusal to pay and lower the cost of legal disputes Rating upgrade: Meet the requirements for electronic medical record rating level 7 (the highest level) to obtain policy subsidies.

[0072] DRG optimization: The accuracy rate of the medical record homepage is 98%, avoiding medical insurance deductions due to grouping errors.

[0073] 5. Technical compatibility: Covers all scenarios with low computing power requirements. Supports 78 types of HIS systems, including legacy C / S architecture systems (via an adaptive script engine).

[0074] Lightweight deployment Model size < 5GB (compared to 350GB for a typical LLM). It can run smoothly on an RTX 4090 (24GB VRAM) and is suitable for primary hospitals.

[0075] This invention achieves the following through three core technologies: "KLM fusion inference + real-time quality control + cross-system adaptation": Automated medical record generation: From "manual filling" to "AI-assisted decision-making".

[0076] Quality control model transformation: from "post-event inspection" to "real-time interception".

[0077] Breaking data silos: From "manual synchronization" to "intelligent docking".

[0078] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do 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 process, method, article, or apparatus.

[0079] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A smart medical record generation system based on a large language model, characterized in that, include: KLM Fusion Reasoning Engine Module: Composed of a large language model LLM finely tuned for the medical field and an embedded medical knowledge graph KG, it realizes clinical causal reasoning through a dynamic knowledge injection mechanism; Dual-channel data writing module: Supports dual-mode automatic data synchronization, including direct writing to HL7 / FHIR interfaces and script adaptation for older systems; The three-layer real-time quality control module sequentially performs format compliance verification, clinical logic contradiction detection, and analysis of the rationality of the disease progression timeline. Workflow compression engine module: compresses the medical record generation process into a single step, and supports voice / text input to generate a complete medical record draft.

2. The intelligent medical record generation system based on a large language model according to claim 1, characterized in that: The operation of the KLM fusion inference engine module includes: The LM layer is based on the Transformer architecture and is fine-tuned using anonymized medical record data from within the hospital. The KG layer includes the ICD disease database, drug knowledge base, and test indicator association database; Dynamic knowledge injection is achieved through an entity relationship binding mechanism, which forcibly associates relevant diagnosis and treatment rules when a specific medical term is identified.

3. The intelligent medical record generation system based on a large language model according to claim 1, characterized in that: The dual-channel data writing module includes: The interface probe unit automatically identifies whether the target system supports HL7 / FHIR; The script generation unit automatically creates operation scripts for non-standard systems.

4. The intelligent medical record generation system based on a large language model according to claim 1, characterized in that: The three-layer real-time quality control module is executed before the doctor saves the medical record: The format layer validates required fields, timestamps, and signature integrity. The logic layer uses KLM to detect contradictions in the diagnosis-testing-medication chain; The intrinsic layer uses a time-series graph neural network to analyze the continuity of the disease course records.

5. The intelligent medical record generation system based on a large language model according to claim 1, characterized in that: The operation process of the workflow compression engine module is as follows: Input a description of the patient's condition → KLM generates a draft → Doctors revise key decisions → The system automatically inserts templates / synchronizes data / triggers quality control → Outputs the final medical record.

6. The intelligent medical record generation system based on a large language model according to claim 2, characterized in that: The KG layer supports dynamic updates via: API for connecting to the latest clinical guidelines; Learn from historical quality control feedback data from hospitals and automatically optimize knowledge association rules.

7. The intelligent medical record generation system based on a large language model according to claim 1, characterized in that: Supports multimodal input: Voice-to-text input (integrated ASR engine); Checkbox-based structured data input; Free text description input.

8. The intelligent medical record generation system based on a large language model according to claim 1, characterized in that: Adopt a lightweight deployment solution: The LLM model uses LoRA fine-tuning technology to adapt to environments with ≥8GB of GPU memory. The edge computing module supports offline generation of medical records.

9. The intelligent medical record generation system based on a large language model according to claim 1, characterized in that: Integrated digital therapy functionality: Automatically generate post-hospital rehabilitation plans; Synchronize patient-collected data (such as from wearable devices) and generate treatment adjustment suggestions.

Images