VTE risk intelligent dynamic assessment method and device based on large language model

Abstract

The invention discloses a VTE risk intelligent dynamic assessment method and device based on a large language model, which can process medical record writing habits of different medical institutions, consider semantic understanding and logic inference, and support dynamic risk prediction in combination with low-cost conventional indexes, so as to simply and conveniently integrate into clinical work, reduce medical care burden and improve the risk assessment efficiency. The problems of VTE risk factor definition standardization, information extraction accuracy, user interface interaction friendliness and VTE event dynamic monitoring ability in the prior art are solved, effective help and support are practically provided for patient management, the patient is helped to recover early, and bad outcome is reduced. The method comprises the following steps: (1) calling and preprocessing data; (2) identifying and judging risk factors; (3) performing dynamic risk prediction and joint modeling; (4) triggering and early warning; (5) performing doctor checking and error correction feedback; and (6) performing individualized intervention suggestion and disposal recommendation.

Description

Technical Field

[0001] This invention relates to the field of medical information processing technology, and in particular to a VTE risk intelligent dynamic assessment method based on a large language model, and a VTE risk intelligent dynamic assessment device based on a large language model. Background Technology

[0002] Venous thromboembolism (VTE) includes deep vein thrombosis (DVT) and pulmonary embolism (PE). As a common and serious complication of hospitalized patients worldwide, VTE can lead to pulmonary circulatory obstruction, acute respiratory failure, right ventricular failure, and even sudden death. In Europe and the United States, its incidence ranks third among acute cardiovascular events, after acute coronary syndrome and stroke. VTE is one of the leading causes of preventable death in hospitals, and early identification and intervention can significantly reduce its morbidity and mortality. Therefore, hospitalized medical and surgical patients with one or more major risk factors should undergo systematic risk assessment and stratified management. In recent years, automated risk identification, information extraction, and dynamic management based on electronic medical records have become important directions in medical informatics, clinical research, and public health. Current research has explored and applied VTE risk assessment systems based on technologies such as natural language processing and artificial intelligence, mainly including: 1. Information extraction methods based on keyword matching or traditional natural language processing (NLP): Keywords or diagnostic descriptions related to VTE risk factors are extracted from medical record texts according to preset rules through medical dictionary matching or regular expressions, and mapped to scoring scale items.

[0003] 2. Application exploration of large language models (LLM): Existing studies have combined general large language models with medical knowledge bases and risk assessment rules, and retrained them on specific medical corpora through fine-tuning methods to improve their ability to identify and calculate professional terms, clinical expressions and multi-rating scales; in the risk assessment of specific medical records, this type of method has shown high accuracy and has certain application potential.

[0004] 3. Predictive models based on machine learning (ML) or deep learning (DL): Current research suggests that after structuring the information in electronic medical records, classification models (such as logistic regression, random forest) or deep learning models (such as convolutional neural networks, dual-branch networks, etc.) can be used to predict VTE risk.

[0005] Overall, although existing VTE risk assessment systems have made some progress in natural language processing, machine learning / deep learning, and large language models, the following problems still need to be addressed in practical clinical applications: 1. The assessment of VTE risk factors relies heavily on physician experience and knowledge, making keyword matching or NLP inefficient. Furthermore, vertical LLM (Limited Learning Model) implementation remains to be effectively implemented: Currently, some items in VTE risk assessment scales are highly subjective in clinical evaluation, relying on physician experience and knowledge. They require comprehensive judgment based on multi-source data such as past medical history, hospitalization duration, and laboratory tests, impacting the accuracy and repeatability of the assessment. Existing Clinical Decision Support Systems (CDSS) heavily rely on keyword matching or traditional NLP models for extracting medical record information, resulting in low accuracy, especially with large text medical records (a common form in electronic medical record systems). Keyword matching is prone to errors, while NLP methods rely heavily on manual feature engineering, processing only local context (such as phrases within a window). This results in low information processing capacity and a lack of medical information understanding and reasoning capabilities, particularly regarding subjective assessment items, leading to frequent missed and incorrect judgments. Moreover, differences in document formats and recording habits across different medical systems further reduce the stability, adaptability, and deployment costs of NLP recognition. Therefore, in practical applications, healthcare professionals have low willingness to use NLP, failing to effectively reduce their workload. Industry-specific LLM employing ultra-long text understanding, semantic and logic enhancement is one solution. However, due to a lack of sufficient disease-specific training data and the professional technical and hardware support required for fine-tuning, and doubts about the extent to which fine-tuning improves capabilities within specific disease areas, the cost-effectiveness of LLM deployment in the industry is questionable, with some exhibiting a phenomenon where form outweighs substance. Clinical application of LLM is still in its early stages. Limited by factors such as context length constraints, high fine-tuning costs, high technical barriers, high hardware and energy consumption requirements, a scarcity of high-quality disease-specific training data, attention basin deficiencies, and frequent hallucinations, the application of LLM for medical information extraction accuracy and decision support is relatively limited. The vast majority remain in the conceptual and experimental stages, lacking large-scale adoption and clinical implementation, awaiting the discovery of effective solutions.

[0006] 2. Interaction methods rely on manual triggering, resulting in a poor user experience: Most CDSS actively pop up windows during operation or require doctors to actively operate on a specific interface to complete the risk assessment. This not only increases the workload but also leads to a poor user experience, low willingness to use, and even a sense of rejection, which is not conducive to the continuous monitoring of VTE risk.

[0007] 3. Lack of dynamic risk prediction capability for thrombotic events: Existing systems or ML / DL methods mostly remain at a single time point in static risk stratification, or rely on complex indicators such as genomics and proteomics, lacking the ability to analyze longitudinal risk trends and predict dynamic outcomes, especially failing to fully explore and utilize existing simple and objective longitudinal indicators. Time-related DL methods such as LSTM (Long Short-Term Memory Network) require time series data to meet the conditions of equal intervals and uniform sampling (data balance requirements). However, data collection in clinical practice is affected by both changes in patient condition and physician decisions, resulting in an irregular, unbalanced distribution (or random missing data), making it difficult to successfully build the model. Alternatively, simplified modeling methods such as merging time points or extending time intervals may be used, failing to fully utilize longitudinal dynamic information. Furthermore, such models have "black box" characteristics, insufficient clinical interpretability, and are prone to questioning of prediction results by medical staff, thus affecting their approval for application in the medical system and their adoption and promotion in actual diagnosis and treatment. New methods need to be explored and applied. Summary of the Invention

[0008] To overcome the shortcomings of existing technologies, the technical problem this invention aims to solve is to provide a VTE risk intelligent dynamic assessment method based on a large language model. This method can handle the different medical record writing habits of various medical institutions, take into account semantic understanding and logical inference, and support dynamic risk prediction by combining low-cost routine indicators. It can then be easily and conveniently integrated into clinical work, reducing the burden on medical staff. This addresses the challenges of existing technologies in terms of standardization of VTE risk factor definition, accuracy of information extraction, user interface interactivity, and dynamic monitoring capabilities of VTE events. It effectively helps and supports patient management, helps patients recover as soon as possible, and reduces adverse outcomes.

[0009] The technical solution of this invention is: a VTE risk intelligent dynamic assessment method based on a large language model, which includes the following steps: (1) Data retrieval and preprocessing: Taking the variables required for VTE risk assessment as the core, relevant data are retrieved from multiple sources such as the Hospital Information System (HIS), Laboratory Information System (LIS), Medical Order System (CPOE), and Electronic Medical Record System (EMR). During the retrieval and preprocessing stage, key-value pairs are generated and irrelevant information is excluded to improve LLM understanding and save context length. Attributes are marked for each variable to control subsequent dynamic capture and evaluation result updates. Variable marking includes static variables and dynamic variables. (2) Risk factor identification and judgment, implemented based on LLM cue word engineering and chain reasoning technology; (3) Dynamic risk prediction and joint modeling: real-time monitoring of patients' test data. When the longitudinal data of a patient's specific test indicator reaches the preset minimum threshold, the mode based solely on risk assessment variables is automatically or manually switched to the joint mode. (4) Triggering and warning: Execute model calculations and update results, and provide prompts based on the updated results; (5) Doctors verify and correct the prediction results manually; (6) Individualized intervention and treatment recommendations: After obtaining individualized risk prediction results, based on the preset clinical guidelines and drug contraindication database, combined with bleeding risk, renal function and medical order status, corresponding prevention and treatment recommendations are output, and consultation reminders are triggered simultaneously for clinical team reference and decision-making.

[0010] Compared with the prior art, the present invention has the following beneficial technical effects: 1. High accuracy and low deployment cost This invention is based on a general-purpose LLM platform and introduces cue word engineering technology. By constructing cue words that integrate expert knowledge and continuously optimizing them based on practical applications, the model can automatically identify and process multiple variables required for VTE risk assessment without fine-tuning or retraining, while maintaining a high data extraction accuracy (over 99%). The technical difficulty is low. Simultaneously, the system retains complete contextual information during data preprocessing, avoiding block processing to prevent semantic fragmentation, and further eliminates irrelevant variables, thereby effectively reducing token consumption and computational overhead. Compared to existing solutions that rely on large-scale fine-tuning and block calls, this invention significantly reduces computational resources and deployment costs, and demonstrates strong applicability and stability in VTE risk assessment tasks.

[0011] 2. It can monitor dynamically in real time and has high practicality. This invention introduces a dynamic, multi-time-point assessment mechanism into VTE risk prediction. Assessment tasks can be triggered automatically or manually to ensure the generation of the latest risk prediction results at points in time when the patient's condition changes. The system continuously monitors and integrates the patient's longitudinal laboratory indicators, employing parametric mathematical modeling methods to analyze and model the constantly updated longitudinal data, reflecting the real-time trend of VTE risk changes. This method uses objective, simple, and clinically readily available indicators, eliminating the need for high-cost genomic testing and demonstrating strong universality. It effectively overcomes the shortcomings of traditional machine learning methods, such as insufficient interpretability and inability to utilize dynamic or imbalanced clinical data, thereby significantly improving the real-time nature and clinical applicability of risk assessment.

[0012] 3. Explanatory power and traceability The system introduces a closed-loop mechanism of "automatic extraction—chain reasoning—manual verification—prompt word optimization," combined with parametric modeling methods, to extract, reason, and judge multi-source medical data, while retaining the reasoning basis to achieve full traceability. Compared with the "black box" problem of traditional ML / DL models, this invention not only significantly improves prediction accuracy but also enhances the interpretability of results and decision-making transparency, facilitating clinical review and reference by medical staff.

[0013] A VTE risk intelligent dynamic assessment device based on a large language model is also provided, which includes: The data retrieval and preprocessing module focuses on the variables required for VTE risk assessment. It retrieves relevant data from multiple sources, including the Hospital Information System (HIS), Laboratory Information System (LIS), Medical Order System (CPOE), and Electronic Medical Record System (EMR). During the retrieval and preprocessing stages, it generates key-value pairs and removes irrelevant information to improve LLM comprehension and save context length. It also labels each variable with attributes to control subsequent dynamic data retrieval and evaluation result updates. Variable labeling includes static and dynamic variables. The risk factor identification and judgment module is implemented based on LLM's prompt word engineering and chain reasoning technology; The dynamic risk prediction and joint modeling module monitors patients' test data in real time. When the longitudinal data of a patient's specific test indicator reaches a preset minimum threshold, it automatically or manually switches from the mode based solely on risk assessment variables to the joint mode. The triggering and early warning module performs model calculations and updates results, and provides prompts based on the updated results; The doctor's verification and error correction feedback module allows for manual review and correction of the prediction results; The individualized intervention and treatment recommendation module, after obtaining individualized risk prediction results, outputs prevention and treatment recommendations based on preset clinical guidelines and medication contraindication databases, combined with bleeding risk, renal function and medical order status, and simultaneously triggers consultation reminders for clinical team reference and decision-making. Attached Figure Description

[0014] Figure 1 A flowchart of the VTE risk intelligent dynamic assessment method based on a large language model according to the present invention is shown. Detailed Implementation

[0015] like Figure 1 As shown, this VTE risk intelligent dynamic assessment method based on a large language model includes the following steps: (1) Data retrieval and preprocessing: Taking the variables required for VTE risk assessment as the core, relevant data are retrieved from multiple sources such as the Hospital Information System (HIS), Laboratory Information System (LIS), Medical Order System (CPOE), and Electronic Medical Record System (EMR). During the retrieval and preprocessing stage, key-value pairs are generated and irrelevant information is excluded to improve LLM understanding and save context length. Attributes are marked for each variable to control subsequent dynamic capture and evaluation result updates. Variable marking includes static variables and dynamic variables. (2) Risk factor identification and judgment, implemented based on LLM cue word engineering and chain reasoning technology; (3) Dynamic risk prediction and joint modeling: real-time monitoring of patients' test data. When the longitudinal data of a patient's specific test indicator reaches the preset minimum threshold, the mode based solely on risk assessment variables is automatically or manually switched to the joint mode. (4) Triggering and warning: Execute model calculations and update results, and provide prompts based on the updated results; (5) Doctors verify and correct the prediction results manually; (6) Individualized intervention and treatment recommendations: After obtaining individualized risk prediction results, based on the preset clinical guidelines and drug contraindication database, combined with bleeding risk, renal function and medical order status, corresponding prevention and treatment recommendations are output, and consultation reminders are triggered simultaneously for clinical team reference and decision-making.

[0016] Compared with the prior art, the present invention has the following beneficial technical effects: 1. High accuracy and low deployment cost This invention is based on a general-purpose LLM platform and introduces cue word engineering technology. By constructing cue words that integrate expert knowledge and continuously optimizing them based on practical applications, the model can automatically identify and process multiple variables required for VTE risk assessment without fine-tuning or retraining, while maintaining a high data extraction accuracy (over 99%). The technical difficulty is low. Simultaneously, the system retains complete contextual information during data preprocessing, avoiding block processing to prevent semantic fragmentation, and further eliminates irrelevant variables, thereby effectively reducing token consumption and computational overhead. Compared to existing solutions that rely on large-scale fine-tuning and block calls, this invention significantly reduces computational resources and deployment costs, and demonstrates strong applicability and stability in VTE risk assessment tasks.

[0017] 2. It can monitor dynamically in real time and has high practicality. This invention introduces a dynamic, multi-time-point assessment mechanism into VTE risk prediction. Assessment tasks can be triggered automatically or manually to ensure the generation of the latest risk prediction results at points in time when the patient's condition changes. The system continuously monitors and integrates the patient's longitudinal laboratory indicators, employing parametric mathematical modeling methods to analyze and model the constantly updated longitudinal data, reflecting the real-time trend of VTE risk changes. This method uses objective, simple, and clinically readily available indicators, eliminating the need for high-cost genomic testing and demonstrating strong universality. It effectively overcomes the shortcomings of traditional machine learning methods, such as insufficient interpretability and inability to utilize dynamic or imbalanced clinical data, thereby significantly improving the real-time nature and clinical applicability of risk assessment.

[0018] 3. Explanatory power and traceability The system introduces a closed-loop mechanism of "automatic extraction—chain reasoning—manual verification—prompt word optimization," combined with parametric modeling methods, to extract, reason, and judge multi-source medical data, while retaining the reasoning basis to achieve full traceability. Compared with the "black box" problem of traditional ML / DL models, this invention not only significantly improves prediction accuracy but also enhances the interpretability of results and decision-making transparency, facilitating clinical review and reference by medical staff.

[0019] Preferably, in step (1), static variables are not changed after the first retrieval; dynamic variables adopt an incremental update mechanism, and only newly added or changed records are retained each time they are retrieved; all structured and unstructured records are unified to a relative timeline starting from the admission time when they are merged, and a traceable identifier is retained for each record: source system, source text name, time, and variable key-value pairs; with the help of the long text reception and strong reasoning capabilities of the general-purpose high-parameter LLM, the input is controlled by variable list + irrelevant variable exclusion + time window limit while retaining the complete context and source label, and the LLM call is triggered only when new evidence is detected or evaluation is required; the final output is a structured dataset with state attributes, data source and time information.

[0020] Preferably, in step (2), a prompt word system for VTE risk factors is constructed. For each risk variable, at least two senior clinical experts independently provide medical definitions, judgment rules, sources of judgment basis, possible medical descriptions, evidence priority levels, and mutual exclusion condition judgment mechanisms, and a fixed structured output pattern is specified. Chain reasoning is used to decompose the complex judgment process into logical steps, and reasoning records are generated for each variable, thereby ensuring the accuracy and interpretability of the results.

[0021] Preferably, in step (3), a joint model that has been pre-fitted based on a historical queue is called, and the aforementioned risk assessment variables at the current time point are used as covariates to construct a joint model together with the repeated measurements of the test indicators. When the model receives new test results, it performs incremental updates and generates an individualized risk prediction map and the probability of VTE occurrence. If the condition of ≥2 tests is not met, the risk assessment-based mode is maintained, and the joint modeling process is automatically switched when the condition is met.

[0022] Preferably, in step (4), the triggering includes two modes: automatic triggering and manual triggering. Automatic triggering involves the system capturing the latest medical information at a fixed time each day, re-performing the risk assessment, and completing the dynamic update. Manual triggering involves the doctor clicking on the corresponding interface at any time to perform the risk assessment and prediction calculation. After each update, the system judges the individualized prediction results based on the preset risk level or dynamic assessment threshold, and generates a warning message containing the risk level and prompt text. This message is then pushed to medical staff through the medical information system or mobile terminal. At the same time, the triggering time, the source of evidence, and the handling status are recorded for traceability.

[0023] Preferably, in step (5), the system generates a list to be reviewed after each update, displaying the final judgment result of the variable, the time point, the basis fragment and the source label; the doctor confirms or corrects, and the system automatically records the operator, time, differences before and after modification and the reasons. At the same time, this part of the content also serves as the basis for subsequent prompt word updates and optimizations, realizing a closed-loop mechanism of automatic extraction, chain reasoning, manual verification and prompt word optimization.

[0024] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium. When executed, the program includes the steps of the methods of the above embodiments. The storage medium can be ROM / RAM, magnetic disk, optical disk, memory card, etc. Therefore, corresponding to the method of the present invention, the present invention also includes a VTE risk intelligent dynamic assessment device based on a large language model. This device is typically represented in the form of functional modules corresponding to the steps of the method. The device includes: The data retrieval and preprocessing module focuses on the variables required for VTE risk assessment. It retrieves relevant data from multiple sources, including the Hospital Information System (HIS), Laboratory Information System (LIS), Medical Order System (CPOE), and Electronic Medical Record System (EMR). During the retrieval and preprocessing stages, it generates key-value pairs and removes irrelevant information to improve LLM comprehension and save context length. It also labels each variable with attributes to control subsequent dynamic data retrieval and evaluation result updates. Variable labeling includes static and dynamic variables. The risk factor identification and judgment module is implemented based on LLM's prompt word engineering and chain reasoning technology; The dynamic risk prediction and joint modeling module monitors patients' test data in real time. When the longitudinal data of a patient's specific test indicator reaches a preset minimum threshold, it automatically or manually switches from the mode based solely on risk assessment variables to the joint mode. The triggering and early warning module performs model calculations and updates results, and provides prompts based on the updated results; The doctor's verification and error correction feedback module allows for manual review and correction of the prediction results; The individualized intervention and treatment recommendation module, after obtaining individualized risk prediction results, outputs prevention and treatment recommendations based on preset clinical guidelines and medication contraindication databases, combined with bleeding risk, renal function and medical order status, and simultaneously triggers consultation reminders for clinical team reference and decision-making.

[0025] Preferably, in the data retrieval and preprocessing module, static variables are not changed after the first retrieval; dynamic variables adopt an incremental update mechanism, retaining only newly added or changed records each time they are retrieved; all structured and unstructured records are unified to a relative timeline starting from the admission time when they are merged, and a traceable identifier is retained for each record: source system, source text name, time, and variable key-value pairs; leveraging the long text reception and strong reasoning capabilities of the general-purpose high-parameter LLM, the data retrieval and preprocessing module controls the input by using a variable list + irrelevant variable exclusion + time window limitation while retaining complete context and source labels, and triggers the LLM call only when new evidence is detected or when evaluation is required; the final output is a structured dataset with state attributes, data source, and time information; In the risk factor identification and judgment module, a prompt word system for VTE risk factors is constructed. For each risk variable, at least two senior clinical experts independently provide medical definitions, judgment rules, sources of judgment basis, possible medical descriptions, evidence priority levels, and mutual exclusion condition judgment mechanisms, and a fixed structured output pattern is specified. Chain reasoning is used to decompose the complex judgment process into logical steps, and a reasoning record is generated for each variable, thereby ensuring the accuracy and interpretability of the results.

[0026] Preferably, in the dynamic risk prediction and joint modeling module, a joint model pre-fitted based on a historical queue is invoked, and the aforementioned risk assessment variables at the current time point are used as covariates to construct a joint model together with the repeated measurements of the test indicators. When the model receives new test results, it performs incremental updates and generates an individualized risk prediction map and the probability of VTE occurrence. If the condition of ≥2 tests is not met, the risk assessment-based mode is maintained, and the joint modeling process is automatically switched when the condition is met. In the triggering and early warning module, the system supports two modes: automatic triggering and manual triggering. Automatic triggering involves the system capturing the latest medical information at a fixed time each day, re-performing the risk assessment, and completing the dynamic update. Manual triggering allows doctors to click to perform risk assessment and prediction calculations at any time on the corresponding interface. After each update, the system judges the individualized prediction results based on preset risk levels or dynamic assessment thresholds and generates early warning information containing risk levels and prompts. This information is then pushed to medical staff through the medical information system or mobile terminal. At the same time, the triggering time, source of evidence, and handling status are recorded for traceability.

[0027] Preferably, in the doctor verification and error correction feedback module, the system generates a list to be reviewed after each update, displaying the final judgment result of the variable, the time point, the basis fragment, and the source tag; when the doctor confirms or corrects, the system automatically records the operator, time, differences before and after the modification, and the reason. At the same time, this part of the content also serves as the basis for subsequent prompt word updates and optimizations, realizing a closed-loop mechanism of automatic extraction, chain reasoning, manual verification, and prompt word optimization.

[0028] To better illustrate the present invention, the following two embodiments demonstrate the specific workflow of the system in the auxiliary diagnosis task of VTE risk populations, and further illustrate the collaboration process between the various modules of the system.

[0029] Example 1: VTE-assisted risk assessment system in clinical decision-making scenarios 1. After admitting a new patient, medical staff complete nursing records, initial assessments, and prescribe medications. At this time, they extract multi-source patient data through interfaces of hospital systems such as HIS / EMR / LIS. Using a Python program on the hospital server, they remove irrelevant fields and anonymize privacy information from the data, aligning the timeline according to a variable list and time window. This unifies structured and unstructured information into a structured dataset acceptable to a large language model, and retains traceable tags for each record, including the source system, source text name, time, and variable name-value pair.

[0030] 2. After the data preprocessing module outputs the structured dataset, the Python program first establishes a connection with the LLM through the pre-configured API interface: BASE_URL = os.getenv("OPENAI_BASE_URL", "https: / / <gateway-host> / v1") API_KEY = os.getenv("LLM_API_KEY", "sk-xxxx") It actively loads a prompt template that matches the current task, and then submits the dataset and prompts simultaneously. The model uses chain reasoning technology to perform risk factor analysis step by step, generates intermediate reasoning nodes based on medical definitions and judgment rules, and finally forms a result that includes risk classification results and judgment basis reviewed by doctors.

[0031] resp = c.chat.completions.create( model=os.getenv("LLM_MODEL","qwen-max"), messages=[{"role":"system","content":"You are a medical expert in the field of thrombosis. Please provide a structured conclusion step by step using chain reasoning."}, {"role":"user","content": tpl.format(**data)}], temperature=0.2 ) print(resp.choices[0].message.content) # => JSON with steps, explicitly showing each step of the reasoning. For example: 1- Active tumor: The medical record does not mention any malignant tumors, cancer-related or treatment-related conditions, nor does it mention any use of tumor-related drugs, therefore it is judged as 0.

[0032] 2- Past history of VTE: The past history did not mention deep vein thrombosis, pulmonary embolism, pulmonary artery thrombosis or other venous thrombotic events, so it was judged as 1.

[0033] 3 - Age ≥ 70: Born in 1962, aged 61, not yet 70, therefore judged as 0.

[0034] 4-Heart failure or respiratory failure: The admission diagnosis field contains "Type I respiratory failure", so it is judged as 1.

[0035] 5- Acute infection and / or rheumatic disease: The medical record mentions "the patient has had a fever for 10 days, accompanied by cough and yellow sputum", and multiple antibiotics were prescribed (such as cefoperazone sodium, tigecycline, acyclovir, etc.). The lung CT scan upon admission showed bilateral inflammatory bowel disease, so it meets the criteria for acute infection and is classified as 1.

[0036] 6-Obesity: BMI is 24.5, which is below 30, and there is no description of "obesity", so it is judged as 0.

[0037] 7- Currently undergoing hormone therapy: The medical history and prescriptions do not mention the use of estrogen and progesterone drugs, nor are there any records of "drug contraception" in the marital and reproductive history, therefore the result is 0.

[0038] 8- Recent (≤1 month) trauma or surgery: The medical history mentions "cholecystectomy", the surgery was performed before admission (2024-07-22) and within 1 month, therefore it is classified as 1.

[0039] 9-Immobilization: The doctor's orders include "intensive care", "special care", "endotracheal intubation", "invasive mechanical ventilation", "tracheotomy", etc., and the admission department is the intensive care unit, which meets the criteria for bed rest. Therefore, it is judged as 1.

[0040] 10- Acute myocardial infarction and / or ischemic stroke: The admission diagnosis does not include the relevant disease, and the relevant symptoms are not mentioned in the initial course of treatment. Therefore, the evidence is insufficient and the result is judged as 0.

[0041] 11-Known thrombotic tendency: No thrombotic tendency-related diseases were mentioned in the initial test, such as protein C deficiency, antiphospholipid antibody syndrome, etc., therefore it was judged as 0.

[0042] The patient's final risk assessment score was 7, indicating high risk.

[0043] 3. Doctors can confirm or correct the results of this assessment. Afterwards, the system will automatically output corresponding prevention and treatment suggestions based on the confirmed assessment scores and sub-item results, combined with international guidelines (such as ACCP, ASH, ESC) and Chinese prevention and treatment guidelines.

[0044] For example: Low molecular weight heparin is preferred; if eGFR <30 ml / min, unfractionated heparin is recommended. If there are contraindications to bleeding (active bleeding, platelet count <50×10⁻⁶) 9 For conditions such as / L, mechanical prevention (IPC / GCS) is recommended. It is recommended to reassess the risk within 48 hours and conduct imaging diagnostic tests if necessary.

[0045] Note: The above information is for reference only. The final treatment decision should be made by the attending physician based on the patient's condition. If necessary, consultation with relevant departments should be sought for joint decision-making.

[0046] Example 2: VTE Risk Prediction and Intervention Recommendations Based on a Dynamic Assessment System During a patient's hospitalization, the system collects dynamic data through the LIS system interface, including laboratory indicators such as D-dimer and lymphocytes, and records them according to the testing time. When a specific test indicator is detected to have "≥2" longitudinal data points, the system automatically (or manually) switches from the mode based solely on risk assessment variables to the combined mode.

[0047] The dynamic monitoring module consists of the following parts: Longitudinal submodel (LME): Fitting a random-effects linear mixture model to dynamic laboratory indicators (such as D-dimer). , in, For the current patient in time Longitudinal observations, and These are column vectors of covariates for fixed effects and random effects, respectively. These are the fixed-effects regression coefficients. For individual-specific random effects, This represents the independent and identically distributed measurement error term.

[0048] Survival sub-model: , in, For the baseline risk function, The risk assessment variable at the current point in time is used as a covariate, and γ is its regression coefficient.

[0049] Joint model: , in, Indicates time The "current value" Indicates time The first derivative (trend of change). and For the joint model used to characterize the current value of the longitudinal process With rate of change The regression coefficient for the impact of the endpoint event risk, also known as the "association parameter". For the current value, each unit increase increases the risk of the event occurring. For each unit increase in the rate of change, the risk changes by a factor of [number]. times.

[0050] After completing parameter estimation, the joint modeling module will automatically generate risk change curves: The horizontal axis represents time (current time and future predicted time points 12 hours / 24 hours / 48 hours / 72 hours), and the vertical axis represents the probability of VTE events occurring. The system automatically plots the risk change trajectory curve over time and marks the upward, downward, or stable trends. Key nodes (such as turning points where the risk rises rapidly) are automatically marked in red for doctors to focus on.

[0051] The doctor confirms or corrects the assessment results, and then the system will automatically output corresponding prevention and treatment suggestions based on the confirmed assessment results, combined with international guidelines (such as ACCP, ASH, ESC) and Chinese prevention and treatment guidelines.

[0052] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A VTE risk intelligent dynamic assessment method based on a large language model, characterized by: It includes the following steps: (1) Data retrieval and preprocessing: Taking the variables required for VTE risk assessment as the core, relevant data are retrieved from multiple sources such as the hospital information system (HIS), laboratory system (LIS), medical order system (CPOE), and electronic medical record system (EMR). Key-value pairs are generated and irrelevant information is excluded during the retrieval and preprocessing stages to improve LLM understanding and save context length. Each variable is labeled with attributes to control the subsequent dynamic fetching and evaluation results updates. Variable labeling includes static variables and dynamic variables. (2) Risk factor identification and judgment, implemented based on LLM cue word engineering and chain reasoning technology; (3) Dynamic risk prediction and joint modeling: real-time monitoring of patients' test data. When the longitudinal data of a patient's specific test indicator reaches the preset minimum threshold, the mode based solely on risk assessment variables is automatically or manually switched to the joint mode. (4) Triggering and warning: Execute model calculations and update results, and provide prompts based on the updated results; (5) Doctors verify and correct the prediction results manually; (6) Individualized intervention and treatment recommendations: After obtaining individualized risk prediction results, based on the preset clinical guidelines and drug contraindication database, combined with bleeding risk, renal function and medical order status, corresponding prevention and treatment recommendations are output, and consultation reminders are triggered simultaneously for clinical team reference and decision-making.

2. The intelligent dynamic assessment method for VTE risk based on a large language model according to claim 1, characterized in that: In step (1), static variables are not changed after the first retrieval; dynamic variables adopt an incremental update mechanism, and only newly added or changed records are retained each time they are retrieved; all structured and unstructured records are unified to a relative timeline starting from the admission time when they are merged, and a traceable identifier is retained for each record: source system, source text name, time, and variable key-value pairs; with the help of the long text reception and strong reasoning capabilities of the general-purpose high-parameter LLM, the input is controlled by variable list + irrelevant variable exclusion + time window limit while retaining the complete context and source label, and the LLM call is triggered only when new evidence is detected or evaluation is required; the final output is a structured dataset with state attributes, data source and time information.

3. The intelligent dynamic assessment method for VTE risk based on a large language model according to claim 2, characterized in that: In step (2), a prompt word system for VTE risk factors is constructed. For each risk variable, at least two senior clinical experts independently provide medical definitions, judgment rules, sources of judgment basis, possible medical descriptions, evidence priority levels, and mutual exclusion condition judgment mechanisms, and a fixed structured output pattern is specified. Chain reasoning is used to decompose the complex judgment process into logical steps, and reasoning records are generated for each variable, thereby ensuring the accuracy and interpretability of the results.

4. The intelligent dynamic assessment method for VTE risk based on a large language model according to claim 3, characterized in that: In step (3), the joint model that has been fitted based on the historical queue is called, and the aforementioned risk assessment variables at the current time point are used as covariates to construct the joint model together with the repeated measurements of the test indicators. When the model receives new test results, it performs incremental updates and generates individualized risk prediction maps and VTE occurrence probabilities. If the condition of ≥2 tests is not met, the risk assessment-based mode is maintained, and the joint modeling process is automatically switched when the condition is met.

5. The intelligent dynamic assessment method for VTE risk based on a large language model according to claim 4, characterized in that: In step (4), the triggering includes two modes: automatic triggering and manual triggering. Automatic triggering involves the system capturing the latest medical information at a fixed time each day, re-performing the risk assessment, and completing the dynamic update. Manual triggering involves the doctor clicking on the corresponding interface at any time to perform the risk assessment and prediction calculation. After each update, the system judges the individualized prediction results based on the preset risk level or dynamic assessment threshold, and generates a warning message containing the risk level and prompt text. This message is then pushed to medical staff through the medical information system or mobile terminal. At the same time, the triggering time, the source of evidence, and the handling status are recorded for traceability.

6. The intelligent dynamic assessment method for VTE risk based on a large language model according to claim 5, characterized in that: In step (5), the system generates a list of items to be reviewed after each update, displaying the final judgment result of the variable, the time point, the basis fragment and the source label; the doctor confirms or corrects, and the system automatically records the operator, time, differences before and after the modification and the reason. At the same time, this part of the content also serves as the basis for subsequent prompt word updates and optimizations, realizing a closed-loop mechanism of automatic extraction, chain reasoning, manual verification and prompt word optimization.

7. A VTE risk intelligent dynamic assessment device based on a large language model, characterized in that: It includes: The data retrieval and preprocessing module focuses on the variables required for VTE risk assessment. It retrieves relevant data from multiple sources, including the Hospital Information System (HIS), Laboratory Information System (LIS), Medical Order System (CPOE), and Electronic Medical Record System (EMR). During the retrieval and preprocessing stages, it generates key-value pairs and excludes irrelevant information to improve LLM comprehension and save context length. Each variable is labeled with attributes to control the subsequent dynamic fetching and evaluation results updates. Variable labeling includes static variables and dynamic variables. The risk factor identification and judgment module is implemented based on LLM's prompt word engineering and chain reasoning technology; The dynamic risk prediction and joint modeling module monitors patients' test data in real time. When the longitudinal data of a patient's specific test indicator reaches a preset minimum threshold, it automatically or manually switches from the mode based solely on risk assessment variables to the joint mode. The triggering and early warning module performs model calculations and updates results, and provides prompts based on the updated results; The doctor's verification and error correction feedback module allows for manual review and correction of the prediction results; The individualized intervention and treatment recommendation module, after obtaining individualized risk prediction results, outputs prevention and treatment recommendations based on preset clinical guidelines and medication contraindication databases, combined with bleeding risk, renal function and medical order status, and simultaneously triggers consultation reminders for clinical team reference and decision-making.

8. The VTE risk intelligent dynamic assessment device based on a large language model according to claim 7, characterized in that: In the data retrieval and preprocessing module, static variables remain unchanged after the first retrieval; dynamic variables employ an incremental update mechanism, retaining only newly added or modified records each time they are retrieved; all structured and unstructured records are unified to a relative timeline starting from the admission time during merging, and each record retains a traceable identifier: source system, source text name, time, and variable key-value pairs; leveraging the long text reception and strong reasoning capabilities of the general-purpose high-parameter LLM, the data retrieval and preprocessing module controls the input by using a variable list, irrelevant variable exclusion, and time window restrictions, while preserving complete context and source labels, and triggers LLM calls only when new evidence is detected or when evaluation is required; the final output is a structured dataset with state attributes, data source, and time information; In the risk factor identification and judgment module, a prompt word system for VTE risk factors is constructed. For each risk variable, at least two senior clinical experts independently provide medical definitions, judgment rules, sources of judgment basis, possible medical descriptions, evidence priority levels, and mutual exclusion condition judgment mechanisms, and a fixed structured output pattern is specified. Chain reasoning is used to decompose the complex judgment process into logical steps, and a reasoning record is generated for each variable, thereby ensuring the accuracy and interpretability of the results.

9. The VTE risk intelligent dynamic assessment device based on a large language model according to claim 8, characterized in that: In the dynamic risk prediction and joint modeling module, a joint model that has been pre-fitted based on a historical queue is called. The aforementioned risk assessment variables at the current time point are used as covariates and the repeated measurements of the test indicators are used to construct the joint model. When the model receives new test results, it performs incremental updates and generates individualized risk prediction maps and VTE occurrence probabilities. If the condition of ≥2 tests is not met, the risk assessment-based mode is maintained, and the joint modeling process is automatically switched when the condition is met. In the triggering and early warning module, the system supports two modes: automatic triggering and manual triggering. Automatic triggering involves the system capturing the latest medical information at a fixed time each day, re-performing the risk assessment, and completing the dynamic update. Manual triggering allows doctors to click to perform risk assessment and prediction calculations at any time on the corresponding interface. After each update, the system judges the individualized prediction results based on preset risk levels or dynamic assessment thresholds and generates early warning information containing risk levels and prompts. This information is then pushed to medical staff through the medical information system or mobile terminal. At the same time, the triggering time, source of evidence, and handling status are recorded for traceability.

10. The VTE risk intelligent dynamic assessment device based on a large language model according to claim 9, characterized in that: In the doctor verification and error correction feedback module, the system generates a list of items to be reviewed after each update, displaying the final judgment result of the variable, the time point, the basis fragment, and the source tag; when the doctor confirms or corrects, the system automatically records the operator, time, differences before and after the modification, and the reason. At the same time, this part of the content also serves as the basis for subsequent prompt word updates and optimizations, realizing a closed-loop mechanism of automatic extraction, chain reasoning, manual verification, and prompt word optimization.

Images