Embedded clinical decision support system

By integrating a large language model into an OCT device and injecting OCT expertise, and using LoRA and AWQ technologies to compress the model, the problem of limited computing resources in OCT devices is solved, enabling rapid and accurate medical diagnostic support, improving the accuracy and consistency of diagnosis, and enhancing the accessibility and efficiency of medical resources.

CN118983081APending Publication Date: 2024-11-19SHANGHAI JIAOTONG UNIV
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411124235.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-15
Publication Date
2024-11-19

AI Technical Summary

Technical Problem

In existing technologies, OCT devices, due to limited computing resources, offline operation requirements, and lack of expertise, cannot effectively embed large language models to provide rapid and accurate medical diagnostic support.

Method used

Integrating a large language model into an OCT device, compressing it using low-rank adaptive (LoRA) and activation-aware weighted quantization (AWQ) techniques, injecting OCT expertise, and using a Q&A-based instruction fine-tuning method to optimize the clinical decision support system.

Benefits of technology

It enables rapid and accurate medical diagnostic support on OCT devices, reduces computing resource requirements, improves diagnostic accuracy and consistency, and enhances the accessibility and efficiency of medical resources.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118983081A_ABST
    Figure CN118983081A_ABST
Patent Text Reader

Abstract

The invention discloses an embedded clinical decision support system, and relates to the technical field of medical diagnosis and artificial intelligence, and the system comprises a storage module which comprises a model library and a knowledge library; an OCT information acquisition module; an OCT knowledge injection module; a clinical instruction fine tuning module; and a model compression module. OCT knowledge is injected through the large language model, provided clinical decision support is more accurate and more explanatory, variability of diagnosis results among different medical specialists is reduced, and clinical decision quality and patient nursing consistency are improved; according to the method, the fine-tuned large language model is compressed by activating sensing weight quantization, so that the demand on computing resources is remarkably reduced, the OCT equipment can be deployed in a wider clinical environment, and the accessibility of medical resources and the medical efficiency are improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of medical diagnosis and artificial intelligence technology, and in particular to an embedded clinical decision support system. Background Art

[0002] Since its introduction in 1991, Optical Coherence Tomography (OCT) has become the gold standard in the management of ophthalmic diseases and plays a key role in many clinical fields, including cardiovascular disease assessment, on-site diagnosis of the gastrointestinal tract and cervix, surgical robot navigation, and minimally invasive treatment monitoring. OCT uses near-infrared light to scan biological tissues, providing high-resolution images in two or three dimensions, greatly enriching doctors' diagnostic tools.

[0003] With the rapid development of deep learning algorithms, image processing and analysis technologies in OCT systems have been significantly improved. In particular, image segmentation and classification algorithms, these deep learning-based methods have significant advantages in accuracy and reasoning speed compared to traditional rule-based methods, effectively reducing the burden on medical personnel to convert image results into clinical decisions. For example, deep learning algorithms can accurately extract choroidal thickness from retinal OCT images, which is an important quantitative biomarker for evaluating eye diseases such as glaucoma and myopia.

[0004] However, existing deep learning algorithms usually run in a "black box" form, and interpreting their outputs requires a lot of domain knowledge, which may lead to variations in diagnostic results between different medical specialties. Therefore, it is necessary to embed medical expert systems in OCT devices to provide immediate and professional guidance to help clinicians make decisions.

[0005] The concept of medical expert systems has been around since the late 1960s. For example, the MYCIN system was designed to assist doctors in diagnosing bacterial infections and recommending appropriate antibiotic treatments. Other systems, such as INTERNIST-I and IBM Watson, have expanded the scope of medical expert systems to provide diagnostic support for multiple medical specialties. Although these systems are innovative, they are prone to errors when dealing with noisy data or uncertain inputs due to their reliance on well-defined rules, and scalability issues begin to emerge as the complexity of the systems increases.

[0006] The emergence of large language models (LLMs) has changed the paradigm of computer-aided diagnosis. Different from traditional rule-based systems, LLMs adopt learning-based models, especially Transformer architecture, which are trained on large amounts of text data through self-supervision to deeply understand the patterns and semantics of natural language. For example, fine-tuned LLMs such as ChatGPT are designed for generating human-style text responses in conversational settings. The natural language processing capabilities of these models make them suitable for tasks such as conversational agents, text summarization, and content generation.

[0007] Furthermore, although Large Language Models (LLMs) offer new possibilities in computer-aided diagnosis, it is not feasible to embed them directly into OCT devices. The main reasons include limited computing resources, offline operation requirements, and lack of expertise in OCT and its medical applications. Specifically, 1. Limited computing resources. Although OCT devices are usually equipped with GPUs with several to tens of GB of video memory for real-time processing and rendering, LLMs usually require larger memory during training and inference, which is beyond the computing power of OCT devices.

[0008] 2. Offline operation requirements. Due to the need to protect patient privacy and ensure data security, OCT devices are usually operated offline in clinical environments, which limits the usage model of LLMs that relies on remote servers or cloud platforms.

[0009] 3. Lack of expertise. Although LLMs perform well in processing and generating texts based on extensive knowledge, they lack expertise in OCT and its medical applications, which are essential for clinical decision support systems.

[0010] Therefore, technicians in this field are committed to developing an embedded clinical decision support system. Summary of the invention

[0011] In view of the above-mentioned defects of the prior art, the technical problem to be solved by the present invention is how to use an embedded clinical decision support system to enable OCT equipment to provide fast and accurate medical diagnosis support in the absence of professional knowledge of OCT and its medical applications, limited computing resources and offline conditions.

[0012] Integrating a large language model locally in the OCT device can eliminate dependence on network connectivity, provide clinical decision support without an Internet connection, protect patient privacy and ensure data security; using low-rank adaptation (LoRA) and activation-aware weight quantization (AWQ) technology to compress the large language model can reduce computing resource usage, speed up reasoning, and adapt to the hardware limitations of the OCT device; in addition, injecting a large amount of OCT expertise into the large language model and using a Q&A (generating questions and corresponding answers)-based instruction fine-tuning method can optimize the response to clinical problems and provide accurate and actionable clinical guidance.

[0013] In one embodiment of the present invention, an embedded clinical decision support system is provided, comprising: The storage module includes a model library and a knowledge library. The model library stores the base large language model, the knowledge-injected large language model, the fine-tuned large language model, and the compressed large language model. The knowledge library stores OCT knowledge data, including pre-processed OCT-related literature and articles. The OCT information acquisition module obtains OCT-related literature and articles from the biomedical paper database, performs preprocessing, and stores them in the knowledge base; The OCT knowledge injection module extracts preprocessed OCT-related literature and articles from the knowledge base, divides the training set and the validation set in proportion, connects the model library, trains the base large language model, injects OCT knowledge, and obtains the knowledge-injected large language model; The clinical instruction fine-tuning module connects to the knowledge base, automatically generates question-answer pairs based on pre-processed OCT-related literature and articles, fine-tunes the instructions of the large language model after knowledge injection, and saves the fine-tuned large language model to the model library; The model compression module is connected to the model library to compress the fine-tuned large language model to make it suitable for the computing resources of the OCT device. The compressed large language model is saved to the model library for the OCT device to call. The OCT information acquisition module, the knowledge base, the OCT knowledge injection module, the clinical instruction fine-tuning module and the model compression module are sequentially connected in communication, the storage library is respectively connected in communication with the OCT knowledge injection module, the clinical instruction fine-tuning module and the model compression module, and the knowledge base is connected in communication with the clinical instruction fine-tuning module; The OCT information acquisition module obtains OCT-related literature and articles, and stores them in the knowledge base of the storage module after preprocessing; the OCT knowledge injection module extracts the preprocessed OCT-related literature and articles from the knowledge base, and connects to the model base, trains the base large language model therein, injects OCT knowledge, obtains the knowledge-injected large language model and saves it to the model base; the clinical instruction fine-tuning module is connected to the knowledge base, automatically generates question-answer pairs based on the preprocessed OCT-related literature and articles, and connects to the model base, fine-tunes the instructions on the knowledge-injected large language model therein, and saves the fine-tuned large language model to the model base; the model compression module is connected to the model base, compresses the fine-tuned large language model therein, and saves the compressed large language model to the model base for OCT equipment to call.

[0014] Optionally, in the embedded clinical decision support system in the above embodiment, the OCT information acquisition module uses an API interface or a web crawler to acquire OCT-related documents and articles from a biomedical paper database.

[0015] Optionally, in the embedded clinical decision support system in any of the above embodiments, the biomedical paper database includes PubMed and OCT News.

[0016] Optionally, in the embedded clinical decision support system in any of the above embodiments, preprocessing includes removing irrelevant information and redundant content.

[0017] Optionally, in the embedded clinical decision support system in any of the above embodiments, the ratio of the training set to the validation set is 99:1.

[0018] Optionally, in the embedded clinical decision support system in any of the above embodiments, the training of the base large language model ends when the end condition is reached.

[0019] Furthermore, in the embedded clinical decision support system in the above embodiment, the termination condition is that the maximum number of training rounds is 3 or the training set loss function and the validation set loss function do not decrease for 3000 consecutive steps.

[0020] Optionally, in the embedded clinical decision support system in any of the above embodiments, the large language model includes Llama-2-7B.

[0021] Optionally, in the embedded clinical decision support system in any of the above embodiments, the instruction fine-tuning of the large language model after knowledge injection is performed using the formula: W' = W + BA Among them, W' is the updated model weight, W represents the original model weight, and BA is based on the low-rank adaptation (LoRA) technology to decompose the original parameters into two smaller matrices, called update matrices, which are used for parameter update.

[0022] Optionally, in the embedded clinical decision support system in any of the above embodiments, instruction fine-tuning also includes automated question-answer pair construction.

[0023] Furthermore, in the embedded clinical decision support system in the above embodiment, the automated question-answer pairs are constructed using a natural language processing tool, and the text paragraphs in the knowledge base are input into the natural language processing tool.

[0024] Furthermore, in the embedded clinical decision support system in the above embodiment, the natural language processing tool adopts ChatGPT.

[0025] Furthermore, in the embedded clinical decision support system in the above embodiment, ChatGPT uses a version no less than GPT-4.

[0026] Further, in the embedded clinical decision support system in the above embodiment, the following prompt words are used to set ChatGPT: "Assume you are an expert in optical coherence tomography. Read the following OCT excerpt. Identify key points and transform them into a comprehensive question answer format. Create a variety of questions includinginferential, comparative, critical, and detail-oriented ones. Formulateanswers that are precise, factually correct, and supported by the text. Ensure to include questions that encourage synthesis of information, evaluation of data, and understanding of broader implications in the field ofOCT." ChatGPT generates a question-answer pair, and the generated question is called question Q, and the generated answer is called answer A.

[0027] Furthermore, in the embedded clinical decision support system in the above embodiment, the automated question-answer pair adopts the Chain of thought (COT) reasoning method to generate a series of intermediate reasoning steps when processing complex reasoning tasks, and finally obtain the answer, requiring ChatGPT to re-answer question Q based on answer A to form an instruction training set for instruction fine-tuning.

[0028] Furthermore, in the embedded clinical decision support system in the above embodiment, COT guides the reasoning of the large language model by providing a gold standard of answers generated based on actual corpus, ensuring that the generated answers are accurate and well-founded, avoiding simple "hallucinations", and providing a comprehensive understanding of the OCT field related to the problem.

[0029] Furthermore, in the embedded clinical decision support system in the above embodiment, COT uses the following prompt words to set ChatGPT, "Provide an expert answer to the above question by considering the information given in the above OCT excerpt.".

[0030] Optionally, in the embedded clinical decision support system in any of the above embodiments, an activation-aware weight quantization (AWQ) method is used to compress the fine-tuned large language model.

[0031] The present invention injects OCT knowledge through a large language model, providing more accurate and interpretable clinical decision support, reducing the variability of diagnostic results between different medical specialties, and improving the quality of clinical decision-making and the consistency of patient care; the present invention uses activation-aware weight quantization to compress the fine-tuned large language model, significantly reducing the demand for computing resources, allowing OCT equipment to be deployed in a wider range of clinical environments, improving the accessibility of medical resources and medical efficiency.

[0032] The concept, specific structure and technical effects of the present invention will be further described below in conjunction with the accompanying drawings to fully understand the purpose, characteristics and effects of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 is a schematic diagram illustrating the structure of an embedded clinical decision support system according to an exemplary embodiment; Figure 2 is a schematic diagram illustrating a comparison of effects of using ChatGPT (GPT-4) on different clinical decision support systems according to an exemplary embodiment; Figure 3 is a schematic diagram illustrating a comparison of effects of answers of ChatOCT and PMC-LLaMA-13B and ChatDoctor according to an exemplary embodiment; Figure 4 FIG. 4 is a schematic diagram illustrating a comparison of effects of answers using ChatOCT and ChatOCT (large) according to an exemplary embodiment. DETAILED DESCRIPTION

[0034] The following describes several preferred embodiments of the present invention with reference to the drawings in the specification, so that the technical content is clearer and easier to understand. The present invention can be embodied in many different forms of embodiments, and the protection scope of the present invention is not limited to the embodiments mentioned in the text.

[0035] In the drawings, components with the same structure are indicated by the same numerical reference numerals, and components with similar structures or functions are indicated by similar numerical reference numerals. The size and thickness of each component shown in the drawings are arbitrarily shown, and the present invention does not limit the size and thickness of each component. In order to make the illustration clearer, the thickness of the components is schematically exaggerated in some places in the drawings.

[0036] The inventors have designed an embedded clinical decision support system, such as Figure 1 As shown, including: The storage module includes a model library and a knowledge base. The model library stores the base large language model, the knowledge-injected large language model, the fine-tuned large language model and the compressed large language model. The knowledge base stores OCT knowledge data, including pre-processed OCT-related literature and articles.

[0037] The OCT information acquisition module obtains OCT-related literature and articles from the biomedical paper database, performs preprocessing, and stores them in the knowledge base; uses the API interface or web crawler to obtain OCT-related literature and articles from the biomedical paper database, which includes PubMed and OCT News; preprocessing includes removing irrelevant information and redundant content.

[0038] The OCT knowledge injection module extracts preprocessed OCT-related literature and articles from the knowledge base, divides the training set and the validation set in proportion, the ratio of the training set to the validation set is 99:1, connects the model library, and trains the base large language model. The training ends when the base large language model reaches the end condition, which is that the maximum training rounds are 3 or the training set loss function and the validation set loss function do not decrease for 3000 consecutive steps; then OCT knowledge injection is performed to obtain a knowledge-injected large language model; the large language model includes Llama-2-7B.

[0039] The clinical instruction fine-tuning module connects to the knowledge base, automatically generates question-answer pairs based on pre-processed OCT-related literature and articles, fine-tunes the instructions of the large language model after knowledge injection, and saves the fine-tuned large language model to the model library; the formula for fine-tuning the instructions of the large language model after knowledge injection is: W' = W + BA Among them, W' is the updated model weight, W represents the original model weight, BA is based on the low rank adaptation (LoRA) technology to decompose the original parameters into two smaller matrices, called update matrices, which are used for parameter update; instruction fine-tuning also includes the construction of automated question-answer pairs, which are constructed using natural language processing tools. The text paragraphs in the knowledge base are input into the natural language processing tool. The natural language processing tool uses ChatGPT, and ChatGPT uses a version no less than GPT-4; Use the following prompt words to set ChatGPT: "Assume you are an expert in optical coherence tomography. Read the following OCT excerpt. Identify keypoints and transform them into a comprehensive question answer format. Create a variety of questions including inferential, comparative, critical, and detail-oriented ones. Formulate answers that are precise, factually correct, and supported by the text. Ensure to include questions that encourage synthesis of information, evaluation of data, and understanding of broader implications in the field of OCT.", ChatGPT generates a question-answer pair, calling the generated question Q and the generated answer A; the automated question-answer pair adopts the Chain of Thought (COT) reasoning method, and when dealing with complex reasoning tasks, generates a series of intermediate reasoning steps and finally obtains the answer, requiring ChatGPT to re-answer question Q based on answer A to form an instruction training set for instruction fine-tuning; COT guides the reasoning of the large language model by providing a gold standard of answers generated based on actual corpus, ensuring that the generated answers are accurate and well-founded, avoiding simple "illusions", and providing a comprehensive understanding of the OCT field related to the question; COT uses the following prompt words to set ChatGPT, "Provide an expert answer to the above question by considering the information given in the above OCT excerpt.".

[0040] The model compression module connects to the model library and compresses the fine-tuned large language model. The activation-aware weight quantization (AWQ) method is used to compress the fine-tuned large language model to make it suitable for the computing resources of the OCT device. The compressed large language model is saved to the model library for the OCT device to call.

[0041] The OCT information acquisition module obtains OCT-related literature and articles, and stores them in the knowledge base of the storage module after preprocessing; the OCT knowledge injection module extracts the preprocessed OCT-related literature and articles from the knowledge base, and connects to the model base, trains the base large language model therein, injects OCT knowledge, obtains the knowledge-injected large language model and saves it to the model base; the clinical instruction fine-tuning module is connected to the knowledge base, automatically generates question-answer pairs based on the preprocessed OCT-related literature and articles, and connects to the model base, fine-tunes the instructions on the knowledge-injected large language model therein, and saves the fine-tuned large language model to the model base; the model compression module is connected to the model base, compresses the fine-tuned large language model therein, and saves the compressed large language model to the model base for OCT equipment to call.

[0042] The above embodiment (named ChatOCT, i.e., compressed large language model) was tested and verified by comparison. 30 questions were randomly selected from 91,124 question-answer pairs as the test set for evaluation, and scored from 0 to 100. ChatGPT (GPT-4) was used and two ophthalmologists from Shanghai First People's Hospital were invited to conduct independent scoring evaluation. It was compared with the benchmark model (Llama-2-7B) and other LLM-based clinical decision support systems (PMC-LLaMA-13B, MedAlpaca-13B, ChatDoctor). During the evaluation process of ChatGPT (GPT-4) and ophthalmologists, a double-blind scoring method was used to mitigate potential bias. Specifically, the scorer did not know the source of each answer to ensure the objectivity and fairness of the scoring process.

[0043] The results of using ChatGPT (GPT-4) to score the responses of different clinical decision support systems are shown in Figure 2. Figure 2 As shown, in the dimension of comprehensibility, ChatOCT demonstrated a high degree of clarity, which is essential for effectively conveying complex medical information. ChatOCT was able to convey information clearly and without error, contributing to its practicality as a clinical decision support tool. Coherence is also another dimension in which ChatOCT performs well, as it consistently produces logically structured and coherent outputs. This feature is essential to ensure that the model's explanations are not only logically reasonable, but also reflect a reasonable decision-making process. Relevance is a key area of ​​evaluation, and ChatOCT's answers are considered highly relevant to the clinical context of OCT diagnosis. The model's ability to maintain focus on the core findings of OCT scans without introducing additional details demonstrates its professional training and fine-tuning in the medical field. Clinical practicality is the most important indicator, which reflects the actual applicability of the model in a real-world clinical setting. ChatOCT's performance on this dimension suggests that it has the potential to provide practical insights that can help the clinical decision-making process. It should be noted that the score on the dimension of simplicity is low, which is attributed to the inherent redundancy problem of LLM. It can be seen that ChatOCT and the fine-tuned large language model ChatOCT (large) performed well in all scoring dimensions.

[0044] In order to more intuitively compare the answers of different clinical decision support systems, three questions in the test set were randomly selected, and the answers of ChatOCT and two other high-scoring answer systems, PMC-LLaMA-13B and ChatDoctor, were listed, as follows: Figure 3Figure 2 shows a side-by-side comparison of the answers, highlighting key descriptive differences. Clearly, ChatOCT consistently provided answers that were more consistent with expectations in the clinical setting. Answers were not only technically accurate, but also provided rich explanations and context that aided clinician understanding and subsequent decision making. The answers provided by the other systems were more generic and lacked the specificity and detail of ChatOCT’s answers. The comparative analysis further confirmed ChatOCT’s robustness in generating medically plausible and context-rich answers. This comparative analysis highlights ChatOCT’s potential as a reliable clinical decision support tool.

[0045] Compare the answers provided by ChatOCT (large) and ChatOCT, such as Figure 4 As shown, despite the reduction in size and computational requirements of the ChatOCT model, it still maintains a high level of diagnostic accuracy comparable to the original model. There is a high level of agreement between the answers to the key descriptive sections of the two models, indicating that the compression process effectively retains the model's ability to provide reliable clinical decision support. This is particularly important in clinical settings where OCT devices may have limited computational resources. The ability to compress the model without sacrificing diagnostic accuracy paves the way for the ChatOCT system to be integrated into a wider range of OCT devices, potentially increasing the accessibility of advanced diagnostic tools in a variety of medical settings.

[0046] The computational efficiency of the ChatOCT model was further evaluated relative to the benchmark models Llama-2-7B and ChatOCT (large), as shown in the following table.

[0047] Evaluation metrics include time to first token (TTFT), initialization time latency (ITL), and GPU memory usage. The model compression process significantly reduced the memory requirements of the ChatOCT model from 25 GB to 4.74 GB. It not only supports smooth operation in existing OCT systems, but also enhances the deployment capabilities of the model in resource-limited environments. In addition, the improved memory efficiency translates into faster operation speeds, with significant improvements in both ITL and TTFT. It is particularly beneficial for real-time diagnostic applications, where fast responses are critical to improving patient flow and overall system performance. Faster response times help reduce waiting time for medical professionals and patients, thereby accelerating diagnostic procedures and treatment initiation. In clinical settings, this time saving has a significant impact on patient care outcomes and healthcare service efficiency.

[0048] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. An embedded clinical decision support system, characterized in that: include: A storage module, including a model library and a knowledge base, wherein the model library stores a base large language model, a knowledge-injected large language model, a fine-tuned large language model, and a compressed large language model, and the knowledge base stores OCT knowledge data, including pre-processed OCT-related literature and articles; An OCT information acquisition module obtains OCT-related literature and articles from a biomedical paper database, performs preprocessing, and stores them in the knowledge base; An OCT knowledge injection module extracts the preprocessed OCT-related literature and articles from the knowledge base, divides the training set and the validation set in proportion, connects the model library, trains the base large language model, performs OCT knowledge injection, and obtains a knowledge-injected large language model; A clinical instruction fine-tuning module is connected to the knowledge base, automatically generates question-answer pairs based on the preprocessed OCT-related literature and articles, performs instruction fine-tuning on the large language model after the knowledge injection, and saves the fine-tuned large language model to the model library; A model compression module, connected to the model library, compresses the fine-tuned large language model to make it suitable for the computing resources of the OCT device, and saves the compressed large language model to the model library for the OCT device to call; The OCT information acquisition module, the knowledge base, the OCT knowledge injection module, the clinical instruction fine-tuning module and the model compression module are communicatively connected in sequence, the storage library is communicatively connected with the OCT knowledge injection module, the clinical instruction fine-tuning module and the model compression module respectively, and the knowledge base is communicatively connected with the clinical instruction fine-tuning module.

2. The embedded clinical decision support system according to claim 1, characterized in that: The OCT information acquisition module uses an API interface or a web crawler to acquire the OCT-related documents and articles from the biomedical paper database.

3. The embedded clinical decision support system according to claim 1, characterized in that: The preprocessing includes removing irrelevant information and redundant content.

4. The embedded clinical decision support system as claimed in claim 3, characterized in that: The ratio of the training set to the validation set is 99:

1.

5. The embedded clinical decision support system according to claim 4, characterized in that: The training of the base large language model ends when the end condition is reached.

6. The embedded clinical decision support system according to claim 5, characterized in that: The termination condition is that the maximum number of training rounds is 3 or the training set loss function and the validation set loss function do not decrease for 3000 consecutive steps.

7. The embedded clinical decision support system according to claim 1, characterized in that: The formula used to fine-tune the instruction of the large language model after the knowledge injection is: W' = W + BA.

8. The embedded clinical decision support system according to claim 7, characterized in that: The instruction fine-tuning also includes automated question-answer pair construction.

9. The embedded clinical decision support system according to claim 8, characterized in that: The automated question-answer pairs are constructed using a natural language processing tool, and text paragraphs in the knowledge base are input into the natural language processing tool.

10. The embedded clinical decision support system according to claim 1, characterized in that: The compression of the fine-tuned large language model adopts an activation-aware weight quantization method.

Citation Information

Cited By

  • Organ transplantation clinical aid decision-making method, device and equipment based on AI large model, medium and product

    CN121096601A