Optimization method and device for auxiliary diagnosis and treatment of large medical model and computer program product

Abstract

The invention discloses an optimization method and device for medical large model auxiliary diagnosis and treatment and a computer program product, and the method comprises the steps: firstly carrying out the term standardization of target case information of a target patient, and keywords in a preliminary diagnosis and treatment scheme outputted by an LLM; after TF-IDF vectors and dense semantic vectors of a target keyword set composed of the keywords and the standard medical terms of the keywords are generated, fusion processing is carried out on the TF-IDF vectors and the dense semantic vectors, and mixed vectors are obtained; a clinical path knowledge graph conforming to personalized features of a target patient is dynamically constructed by using more accurate candidate knowledge texts retrieved from a medical knowledge base, and a preliminary diagnosis and treatment scheme output by a medical LLM is continuously and intelligently corrected and detected medical contradictory points are adjusted in combination with a portrait of the target patient, so that the clinical path knowledge graph is obtained. And finally, obtaining a final optimized diagnosis and treatment scheme. Therefore, closed-loop correction optimization of medical LLM auxiliary diagnosis and treatment results is realized, and an ideal optimization effect is achieved.

Description

Technical Field

[0001] This application relates to the field of medical technology, and in particular to an optimization method, device and computer program product for medical large-scale model-assisted diagnosis and treatment. Background Technology

[0002] With the continuous development of artificial intelligence and information technology, various large language models (LLMs) have emerged in people's lives and work, providing various conveniences. For example, with their powerful natural language understanding and generation capabilities, large models have been used for clinical diagnosis and treatment plan generation. Their advantage lies in their ability to process unstructured medical record texts (such as chief complaints and present medical history) and generate recommendation results that closely resemble clinical expressions.

[0003] However, LLM suffers from at least two problems in its application to assisted diagnosis and treatment: First, "illusion" occurs frequently, and erroneous results are difficult to detect by the model itself. Second, the timeliness of knowledge is insufficient, because LLM relies on pre-trained data and struggles to integrate the latest medical evidence in real time, causing recommended solutions to lag behind clinical practice. To address this, those skilled in the art have attempted to combine LLM generation with knowledge graph retrieval to reduce model illusions, such as generating preliminary solutions using LLM and then verifying them through literature retrieval. However, because such solutions are essentially a loose coupling of "generation-retrieval," they do not form a closed loop and still struggle to achieve the ideal optimization effect for LLM-assisted diagnosis and treatment in medicine. Summary of the Invention

[0004] The main objective of this application is to provide an optimization method, device, and computer program product for medical large model-assisted diagnosis and treatment. By integrating standard medical terminology and dynamically constructing a personalized clinical pathway knowledge graph, it can achieve closed-loop correction and optimization of medical LLM-assisted diagnosis and treatment results. This can solve problems such as non-standard terminology, illusions, outdated information, internal contradictions, and clinical thinking biases that occur in medical LLM in assisted diagnosis and treatment recommendations, thereby achieving ideal optimization results.

[0005] This application provides an optimization method for medical large-scale model-assisted diagnosis and treatment, including:

[0006] Obtain the target patient's case information and use the case information to generate a target patient profile.

[0007] Based on the target patient profile, diagnostic and treatment instructions are generated and input into the medical big data model for auxiliary diagnostic and treatment processing to obtain a preliminary diagnosis and treatment plan. After generating a TF-IDF vector and a dense semantic vector of the target keyword set composed of the target case information, keywords in the preliminary diagnosis and treatment plan and their standard medical terms, the TF-IDF vector and the dense semantic vector are fused to obtain a hybrid vector.

[0008] After retrieving candidate knowledge texts that meet preset conditions from the medical knowledge base based on the hybrid vector, an evidence set is generated using the candidate knowledge texts; and a clinical pathway knowledge graph is constructed using the evidence set.

[0009] Using the clinical pathway knowledge graph, target patient profile, and evidence set, the preliminary diagnosis and treatment plan are optimized through closed-loop correction until the final optimized diagnosis and treatment plan is obtained.

[0010] In one possible implementation, generating a target patient profile using target case information includes:

[0011] The target case information is subjected to named entity recognition and relation extraction using a language model pre-trained on a medical corpus, resulting in entity recognition and relation extraction results.

[0012] Based on a standard medical terminology database, the unstructured text of the target case information is mapped to standard medical terms and associated with standard classification codes. Based on the entity recognition results, relation extraction results, mapped standard medical terms, and associated standard classification codes, a target patient profile corresponding to the target patient is generated.

[0013] In one possible implementation, after generating treatment instructions based on the target patient profile and inputting them into a large medical model for assisted diagnosis and treatment processing to obtain a preliminary diagnosis and treatment plan, the method further includes:

[0014] Extract N target keywords from the target case information, preliminary diagnosis, and treatment plan, and determine the standard medical terms corresponding to the N target keywords; where N is a positive integer greater than 0.

[0015] The N target keywords are fused to obtain a fused target keyword set, and a TF-IDF vector and a dense semantic vector corresponding to the fused target keyword set are generated.

[0016] In one possible implementation, retrieving candidate knowledge texts that meet preset conditions from a medical knowledge base based on a hybrid vector includes:

[0017] Based on the target keyword set and its corresponding hybrid vector, knowledge texts with a comprehensive score higher than the preset score threshold are selected from the real-time updated medical knowledge base as candidate knowledge texts.

[0018] The comprehensive score is calculated by weighting and summing the semantic similarity between medical knowledge texts in the medical knowledge base and the target keyword set, the evidence level of the medical knowledge texts, the time decay of the medical knowledge texts, and the authority of the medical knowledge texts.

[0019] After retrieving candidate knowledge texts that meet preset conditions from the medical knowledge base based on the hybrid vector, an evidence set is generated using the candidate knowledge texts; and a clinical pathway knowledge graph is constructed using the evidence set, including:

[0020] The top-K candidate knowledge texts with the highest comprehensive scores are selected from all candidate knowledge texts to form an evidence set. Based on the evidence set, a clinical pathway knowledge graph is constructed using the core classes and relationships of the clinical pathway defined by the OWL ontology language. The clinical pathway knowledge graph includes standardized processes and key decision nodes from initial diagnosis, examination, differential diagnosis to confirmed diagnosis and treatment.

[0021] In one possible implementation, the optimization process of using the clinical pathway knowledge graph, target patient profile, and evidence set to perform closed-loop correction on the preliminary diagnosis and treatment plan until the final optimized diagnosis and treatment plan is obtained includes:

[0022] The preliminary diagnosis and treatment plan are compared with the diagnosis and treatment pathways in the clinical pathway knowledge graph. Based on the first comparison result, the preliminary diagnosis and treatment plan are revised to obtain the revised diagnosis and treatment plan.

[0023] The revised diagnosis and treatment plan is compared with the target patient profile, and the revised diagnosis and treatment plan is adjusted based on the second comparison result to obtain the adjusted diagnosis and treatment plan.

[0024] Using the evidence set, the adjusted diagnosis and treatment plan is optimized to obtain the final optimized diagnosis and treatment plan.

[0025] In one possible implementation, the step of comparing the preliminary diagnosis and treatment plan with the treatment pathway in the clinical pathway knowledge graph, and revising the preliminary diagnosis and treatment plan based on the first comparison result to obtain the revised diagnosis and treatment plan, includes:

[0026] Starting from three dimensions—node alignment, path consistency detection, and constraint verification—the preliminary diagnosis and treatment plan are compared with the diagnosis and treatment path in the clinical pathway knowledge graph. Jumps, omissions, or deviations in the medical big model's thinking chain are identified as the first comparison result. Based on the first comparison result, the preliminary diagnosis and treatment plan output by the medical big model are corrected to obtain the corrected diagnosis and treatment plan.

[0027] In one possible implementation, the step of comparing the revised diagnosis and treatment plan with the target patient profile, and adjusting the revised diagnosis and treatment plan based on the obtained second comparison result to obtain the adjusted diagnosis and treatment plan, includes:

[0028] The revised diagnosis and treatment plan is compared with the target patient profile to determine whether there are any inherent contradictions in medical knowledge. This is used as the second comparison result. Based on the second comparison result, the revised diagnosis and treatment plan is adjusted to obtain the adjusted diagnosis and treatment plan.

[0029] In one possible implementation, the step of using an evidence set to optimize the adjusted diagnosis and treatment plan to obtain a final optimized diagnosis and treatment plan includes:

[0030] When optimizing the adjusted diagnosis and treatment plan using the evidence set, if there is a correction to the diagnosis or a new treatment plan, the process of generating the TF-IDF vector and dense semantic vector corresponding to the target keyword set consisting of the target case information, the corrected diagnosis or the new treatment plan is repeated. Then, the TF-IDF vector and dense semantic vector are fused and processed, and subsequent steps are performed until the final optimized diagnosis and treatment plan is obtained.

[0031] This application embodiment also provides an optimized device for medical large-scale model-assisted diagnosis and treatment, including: a processor, a memory, and a system bus;

[0032] The processor and the memory are connected via the system bus;

[0033] The memory is used to store one or more programs, the one or more programs including instructions, which, when executed by the processor, cause the processor to perform any of the above-described optimization methods for medical large-scale model-assisted diagnosis and treatment.

[0034] This application also provides a computer program product, which, when run on a terminal device, causes the terminal device to execute any of the above-described optimization methods for medical big data-assisted diagnosis and treatment.

[0035] This application provides an optimization method, device, and computer program product for medical big data-assisted diagnosis and treatment. First, it acquires target patient case information and generates a target patient profile based on this information. Then, based on the target patient profile, it generates treatment instructions and inputs them into the medical big data model for assisted diagnosis and treatment, resulting in a preliminary diagnosis and treatment plan. After generating a TF-IDF vector and a dense semantic vector representing a target keyword set composed of keywords from the target patient information, the preliminary diagnosis, and the treatment plan, along with their standard medical terminology, the TF-IDF vector and the dense semantic vector are fused to obtain a hybrid vector. Next, based on the hybrid vector, candidate knowledge texts that meet preset conditions are retrieved from a medical knowledge base, and an evidence set is generated using these candidate knowledge texts. This evidence set is then used to construct a clinical pathway knowledge graph. Finally, the clinical pathway knowledge graph, the target patient profile, and the evidence set are used to perform closed-loop correction optimization of the preliminary diagnosis and treatment plan until a final optimized diagnosis and treatment plan is obtained.

[0036] As can be seen, when optimizing the results of medical LLM-assisted diagnosis and treatment, this application first standardizes the keywords in the target case information, the preliminary diagnosis and treatment plan output by LLM, and then, after generating the TF-IDF vector and dense semantic vector of the target keyword set composed of each keyword and its standard medical terminology (e.g., deduplication and merging), the two are fused (e.g., weighted summation) to obtain a hybrid vector. This reduces the dispersion of the weight of medical terminology synonyms and improves the accuracy of the medical knowledge text subsequently selected from the real-time updated medical knowledge base.

[0037] Building upon this foundation, this application further utilizes more accurate candidate knowledge texts retrieved from medical knowledge bases to dynamically construct a clinical pathway knowledge graph that aligns with the personalized characteristics of the target patient. Combined with the target patient profile, this allows for continuous intelligent correction of the initial diagnosis and treatment plan output by the medical LLM, as well as adjustments to detected medical inconsistencies, until a final optimized diagnosis and treatment plan is obtained. This achieves closed-loop correction and optimization of the medical LLM-assisted diagnosis and treatment results, resolving issues such as non-standard terminology, illusions, outdated information, internal contradictions, and clinical reasoning biases that arise in medical LLM's assisted diagnosis and treatment recommendations, thereby achieving the desired optimization effect. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 A flowchart illustrating an optimization method for medical large-scale model-assisted diagnosis and treatment provided in an embodiment of this application;

[0040] Figure 2 Example diagram of the target patient profile provided in the embodiments of this application;

[0041] Figure 3 Example diagram of clinical pathway knowledge graph provided in the embodiments of this application;

[0042] Figure 4 This is a schematic diagram of the structure of an optimized device for medical large-scale model-assisted diagnosis and treatment provided in an embodiment of this application. Detailed Implementation

[0043] With the widespread application of large medical models (LLM) in clinical diagnosis and treatment planning tasks, two potential core shortcomings have become increasingly apparent. Specifically, first, the problem of "illusion" occurs frequently, and erroneous results are difficult to detect by the model itself. Second, the timeliness of knowledge is insufficient. This is because LLM relies on pre-trained data and struggles to integrate the latest medical evidence in real time, causing recommended solutions to lag behind clinical practice.

[0044] In response, those skilled in the art have attempted to combine LLM generation with knowledge graph retrieval to reduce model illusion, such as by generating a preliminary scheme through LLM and then verifying it through literature retrieval. However, since such schemes are essentially loosely coupled "generation-retrieval" and do not form a closed loop, it is still difficult to achieve the ideal optimization effect for medical LLM-assisted diagnosis and treatment.

[0045] To address the aforementioned shortcomings, this application provides an optimization method for medical big data-assisted diagnosis and treatment. First, target patient case information is acquired, and a target patient profile is generated using this information. Then, based on the target patient profile, diagnostic instructions are generated and input into the medical big data model for assisted diagnosis and treatment, resulting in a preliminary diagnosis and treatment plan. After generating a TF-IDF vector and a dense semantic vector representing a target keyword set composed of keywords from the target patient information, preliminary diagnosis, and treatment plan, along with their standard medical terminology, the TF-IDF vector and dense semantic vector are fused to obtain a hybrid vector. Next, candidate knowledge texts meeting preset conditions are retrieved from a medical knowledge base based on the hybrid vector, and an evidence set is generated using these candidate knowledge texts. This evidence set is then used to construct a clinical pathway knowledge graph. Finally, this clinical pathway knowledge graph, the target patient profile, and the evidence set are used to perform closed-loop correction optimization of the preliminary diagnosis and treatment plan until the final optimized diagnosis and treatment plan are obtained.

[0046] As can be seen, when optimizing the results of medical LLM-assisted diagnosis and treatment, this application first standardizes the keywords in the target case information, the preliminary diagnosis and treatment plan output by LLM, and then, after generating the TF-IDF vector and dense semantic vector of the target keyword set composed of each keyword and its standard medical terminology (e.g., deduplication and merging), the two are fused (e.g., weighted summation) to obtain a hybrid vector. This reduces the dispersion of the weight of medical terminology synonyms and improves the accuracy of the medical knowledge text subsequently selected from the real-time updated medical knowledge base.

[0047] Building upon this foundation, this application further utilizes more accurate candidate knowledge texts retrieved from medical knowledge bases to dynamically construct a clinical pathway knowledge graph that aligns with the personalized characteristics of the target patient. Combined with the target patient profile, this allows for continuous intelligent correction of the initial diagnosis and treatment plan output by the medical LLM, as well as adjustments to detected medical inconsistencies, until a final optimized diagnosis and treatment plan is obtained. This achieves closed-loop correction and optimization of the medical LLM-assisted diagnosis and treatment results, resolving issues such as non-standard terminology, illusions, outdated information, internal contradictions, and clinical reasoning biases that arise in medical LLM's assisted diagnosis and treatment recommendations, thereby achieving the desired optimization effect.

[0048] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0049] See Figure 1 This is a flowchart illustrating an optimization method for medical large-scale model-assisted diagnosis provided in this embodiment. The method includes the following steps:

[0050] S101: Obtain the target patient's case information and use the case information to generate a target patient profile corresponding to the target patient.

[0051] In this embodiment, any patient used to generate an optimized treatment plan using the medical large-scale model-assisted diagnosis and treatment optimization method provided in this application is defined as the target patient, and the medical record information related to the target patient's visit is defined as the target case information. It should be noted that this embodiment does not limit the language type of the target case information; for example, the target case information can be Chinese medical record information or English medical record information, etc. This embodiment also does not limit the length of the target case information; for example, the target medical record can be sentence text or paragraph text, etc. This embodiment also does not limit the length of the target case information; for example, the target case information can come from data from a Hospital Information System (HIS), Laboratory Information System (LIS), or Picture Archiving and Communication System (PACS), etc. Furthermore, this embodiment does not limit the content of the target case information. For example, the target case information may include: medical records (such as outpatient medical records, admission records, progress notes, etc.), examination and test reports (such as blood routine, biochemistry, CT, MRI reports, etc.), and past medical history information. For example, the target malicious query text may be "how to make dangerous items", etc.

[0052] Furthermore, after obtaining the target case information related to the target patient's visit, firstly, existing or future named entity recognition and relation extraction methods can be utilized. For example, a language model pre-trained on a medical corpus (such as BioBERT) can be used to perform named entity recognition and relation extraction on the target case information, obtaining entity recognition results and relation extraction results. Then, standard medical terminology databases such as the Unified Medical Language System (UMLS) can be combined to accurately map the unstructured text in the target case information into standard medical terms and associate it with standard classification codes such as SNOMED CT (Systematic Medical Nomenclature—Clinical Terminology) or ICD-10 (International Classification of Diseases, 10th Revision). Finally, a structured "patient profile" containing named entities such as the target patient's symptoms, signs, medical history, medications, adverse events, and their time, severity, and causal relationships is generated and defined as the target patient profile for subsequent steps S102 and S104.

[0053] For example: Suppose the target patient's case information obtained from the HIS system, LIS system, PACS system, and other sources is as follows:

[0054] "The patient, a 69-year-old male, presented with a history of recurrent abdominal distension for over a month. He developed abdominal distension without any obvious cause over a month prior, accompanied by poor appetite and fatigue, but no fever, abdominal pain, or other discomfort. Physical examination revealed: T 36.8℃, P 60 bpm, R 18 bpm, BP 109 / 74 mmHg. He was alert but lethargic, with jaundice of the skin and sclera, no palmar erythema or spider angiomas, and no significantly enlarged superficial lymph nodes. The abdomen was distended, without significant tenderness or rebound tenderness. The liver and spleen were not palpable below the costal margin, and there was no tenderness upon percussion of the liver area. The gallbladder was not palpable, and Murphy's sign was negative; shifting dullness was positive. There was no edema in the lower extremities, and the neurological examination was unremarkable. He had a history of long-term alcohol consumption."

[0055] Complete blood count + CRP: WBC 5.9*10^9 / L, HGB 176g / L, PLT 70*10^9 / L, CRP 11.3mg / L; PT 19.2S; Biochemistry: ALB 29.3g / L, GLB 26.8g / L, ALT 100U / L, AST 170U / L, TB 67.8μmol / L, DB 32.7μmol / L, Cr 69.4μmol / L; Viral hepatitis test, tumor markers, ANA, TNI+BNP and other tests showed no significant abnormalities.

[0056] Abdominal ultrasound showed ascites.

[0057] Other data...

[0058] The target patient profile generated using the above target case information can be as follows: Figure 2 As shown.

[0059] S102: Based on the target patient profile, generate treatment instructions and input them into the medical big model for auxiliary treatment processing to obtain a preliminary diagnosis and treatment plan; after generating a TF-IDF vector and a dense semantic vector consisting of the target keyword set composed of the target case information, keywords in the preliminary diagnosis and treatment plan and their standard medical terms, fuse the TF-IDF vector and the dense semantic vector to obtain a hybrid vector.

[0060] In this embodiment, after generating the target patient profile corresponding to the target patient in step S101, the content contained in the target patient profile can be further incorporated into prompts to generate treatment instructions. The generated treatment instructions are then input into a medical big model (the specific composition structure is not limited and can be selected according to the actual situation and experience values, such as iFlytek medical big model, etc.) for auxiliary treatment processing to obtain a preliminary diagnosis and treatment plan for the target patient.

[0061] For example, the following is an example of how to incorporate the content of the target patient profile into the generated treatment instructions based on prompts:

[0062] "Based on the target patient profile, please provide a specific diagnosis and next treatment plan. The output should use standardized medical terminology."

[0063] ###

[0064] The target patient profile includes the following:

[0065] ###".

[0066] Furthermore, after inputting the generated diagnostic and treatment instructions into the Medical Large Model (LLM), the model outputs an example of a preliminary diagnosis and treatment plan for the target patient, as shown below:

[0067] ###

[0068] 1. Primary diagnosis: Alcoholic cirrhosis (decompensated stage) (ICD-10: K70.3).

[0069] 2. Further inspection recommendations: ......

[0070] 3. Treatment plan: ...... ......

[0072] ###".

[0073] Furthermore, we can first utilize existing or future keyword extraction methods, such as the BioBERT model, to extract all keywords from the target case information, the preliminary diagnosis output by the model, and the treatment plan (here defined as N target keywords, where N is not limited and can be any positive integer greater than 0). We can then determine the standard medical terms corresponding to these target keywords (to eliminate synonym ambiguity, such as mapping "myocardial infarction" and "AMI" to CUI:C0027051). Then, we can perform a fusion process on the N target keywords (and their corresponding standard medical terms) (such as deduplication and merging) to obtain the fused target keyword set.

[0074] For example, based on the target case information in the above example, the target keywords and their corresponding standard medical terms extracted could be: "Male (SNOMED:248153007), 69 years old (SNOMED:416070008), abdominal distension (R14.0), poor appetite (R63.0), ascites (R18.8), elevated alanine aminotransferase (R74.0), elevated aspartate aminotransferase (R74.0)...". Meanwhile, based on the preliminary diagnosis and treatment plan in the above example, the target keywords and their corresponding standard medical terms extracted could be: "Alcoholic cirrhosis (K70.3), ascites (R18.8), spontaneous bacterial peritonitis (K65.2)...".

[0075] Based on this, after deduplication and merging all target keywords and their standard medical terms, the resulting set of target keywords can be: "Male (SNOMED:248153007), 69 years old (SNOMED:416070008), abdominal distension (R14.0), poor appetite (R63.0), ascites (R18.8), elevated alanine aminotransferase (R74.0), elevated aspartate aminotransferase (R74.0), alcoholic cirrhosis (K70.3), ascites (R18.8), spontaneous bacterial peritonitis (K65.2)...".

[0076] Furthermore, existing or future vector generation methods can be used to generate TF-IDF vectors (which can be represented as V) for the target keyword set. tfidf ) and dense semantic vectors (which can be represented as V) semantic Then, the two are fused (e.g., by weighted summation) to obtain a hybrid vector (which can be represented as V). hybrid This reduces the weight dispersion of medical terminology synonyms and improves the accuracy of subsequent steps in selecting medical knowledge texts (i.e., candidate knowledge texts mentioned later) from the real-time updated medical knowledge base.

[0077] TF (Term Frequency) measures the frequency with which a word appears in a single document, reflecting the word's representativeness within that document. For example, if "C0027051 (myocardial infarction)" appears 10 times in a medical article, it indicates that the article primarily discusses myocardial infarction. A high TF value indicates high relevance.

[0078] Inverse Document Frequency (IDF) measures the rarity of a term within a collection of documents, reflecting its uniqueness. The rarer the term (the fewer documents containing it), the higher its IDF value and the stronger its discriminative power. For example, if "C0027051 (myocardial infarction)" appears in only 3 out of 1000 articles, these 3 articles are considered valuable and highly specific. Conversely, a term like "Cxxxxx (patient)" appears in every article, resulting in a low IDF that cannot be used to distinguish between articles.

[0079] Based on this, if the TF-IDF = TF×IDF of a word is higher, it means that the word appears more often (high TF) and is also unique in the entire document collection (high IDF), then the word is the core essence of this article.

[0080] Similarly, the TF corresponding to each standard medical term in the target keyword set mentioned above in this application refers to the number of times that standard medical term appears in a medical document (e.g., assuming C0027051 appears once in each of medical documents 1, 2, and 3, then TF=1); the IDF corresponding to each standard medical term refers to the proportion of medical documents containing that standard medical term to the total number of medical documents (e.g., assuming there are 1000 total medical documents, of which 3 contain C0027051, then IDF=log (1000 / 3)≈2.52). Furthermore, the TF-IDF of C0027051 is TF×IDF=1×2.52≈2.52.

[0081] Based on this, a similar calculation method can be used to calculate the TF-IDF vector corresponding to each standard medical term in the target keyword set. Then, concatenating these vectors yields the TF-IDF vector (V) corresponding to the entire target keyword set. tfidf ).

[0082] Simultaneously, the BioBERT model can be used to transform the text corresponding to the target keyword set into a 768-dimensional (other dimensions are acceptable; 768 dimensions here is merely an example and not a limitation) dense vector (used to capture the semantic relationships between various standard medical terms). For example, it can be transformed into a 768-dimensional dense vector: "[0.12, 0.35, -0.08, ..., 0.21]", which contains 768 floating-point numbers, serving as the dense semantic vector (V) corresponding to the entire target keyword set. semantic ).

[0083] Furthermore, the TF-IDF vector (V) corresponding to the entire target keyword set is... tfidf ) and dense semantic vectors (Vsemantic By performing a weighted summation, the mixture vector (V) can be obtained. hybrid ), used to perform subsequent step S103, wherein the blending vector (V hybrid The specific formula is as follows:

[0084] V hybrid = 0.3 * V semantic + 0.7 * V tfidf

[0085] It should be noted that 0.3 and 0.7 represent the TF-IDF vector (Vi) and Vi, respectively. tfidf ) and dense semantic vectors (V semantic The weight of ) is only an example value and is not specifically limited. It can be set to other values ​​according to the actual situation and experience.

[0086] S103: After retrieving candidate knowledge texts that meet preset conditions from the medical knowledge base based on the hybrid vector, generate an evidence set using the candidate knowledge texts; and use the evidence set to construct a clinical pathway knowledge graph.

[0087] In this embodiment, the target keyword set and its corresponding hybrid vector (V) are obtained through step S102. hybrid Furthermore, firstly, keywords from the target keyword set can be used as search keywords to filter highly relevant knowledge texts with a comprehensive score higher than a preset score threshold (the specific value is not limited and can be set according to actual situation and experience) from real-time updated medical knowledge bases (such as UpToDate, NCCN guidelines, etc.) as candidate knowledge texts.

[0088] It should be noted that this application does not limit the calculation method for the comprehensive score of each medical knowledge text in the medical knowledge base. One optional implementation method is that the comprehensive score can be obtained by weighted summation of the semantic similarity between the medical knowledge text and the target keyword set, the evidence level (GRADE) of the medical knowledge text, the time decay of the medical knowledge text (which can be reflected in the publication year of the literature), and the authority of the medical knowledge text (which can be reflected in the journal impact factor, etc.). The specific calculation formula is as follows:

[0089]

[0090] Wherein, Score represents the overall score of each medical knowledge text in the medical knowledge base; sem E represents the semantic similarity between the corresponding medical knowledge text and the target keyword set; grade Indicates the level of evidence (GRADE) for the corresponding medical knowledge text; T decay Indicates the degree of time decay of the corresponding medical knowledge text; Ijournal This represents the authority of the corresponding medical knowledge text (such as journal impact factor); α, β, γ, and δ are weighting coefficients, and their specific values ​​are not limited. They can be set according to the actual situation and experience, but it is necessary to ensure that α+β+γ+δ=1. For example, considering that clinical decision-making requires extremely high accuracy and evidentiary strength, the four can be set as follows: α=0.4; β=0.3; γ=0.2; δ=0.1.

[0091] Specifically, S sem The determination method can be as follows: First, use a Sentence-BERT model (such as sentence-transformers / all-MiniLM-L6-v2) to encode medical knowledge text fragments, and then calculate the resulting encoded vector and the mixture vector (V). hybrid The cosine similarity between S and S is used as the basis for determining the cosine similarity between S and S. sem In practical applications, key segments of medical knowledge text (such as abstracts and conclusions) can first be segmented into windows of 50-100 characters; then, Sentence-BERT can be used to generate segment vectors (represented as V). segment Then, calculate its interaction with the mixture vector (V). hybrid The cosine similarity between S and S is used as the basis for determining the cosine similarity between S and S. sem The specific calculation formula is as follows:

[0092]

[0093] And E grade The value can be determined based on the pre-defined correspondence between various medical knowledge texts and their levels of evidence and their quantitative scores (specifically, any integer score between 1 and 5). An example of the correspondence between various medical knowledge texts (types) and their levels of evidence and their quantitative scores (specifically, any integer score between 1 and 5) is shown in Table 1 below:

[0094] Table 1

[0095] Level of evidence Medical knowledge text types / research design Quantitative score Level 1 Guidelines, systematic reviews / meta-analyses, and well-homogenized randomized controlled trials (RCTs). 5 Level 2 Randomized controlled trials (RCTs) 4 Level 3 Cohort studies, case-control studies 3 Level 4 Case series reports, uncontrolled cohort studies 2 Level 5 Expert opinions, medical record reports 1

[0096] Understandably, if the type of medical knowledge text cannot be identified, that is, if the corresponding quantitative score cannot be found from the correspondence shown in Table 1 as E, then... grade Then any integer score between 1 and 5 (such as 2) can be randomly selected as E. grade The default value.

[0097] And, T decay The specific calculation formula is as follows:

[0098]

[0099] Among them, t now Indicates the current year; t publish This indicates the year of publication of the corresponding medical knowledge text; k represents the attenuation coefficient, the specific value of which is not limited and can be set according to the actual situation and experience. For example, it can be set to 0.3, i.e., k=0.3 (adjustable).

[0100] Furthermore, with I journal Taking journal impact factor as an example, the specific calculation formula is as follows:

[0101] I journal =Actual impact factor / Highest impact factor in the field

[0102] For example, the highest impact factor in the cardiovascular field is 20.0, and the impact factor of a certain journal (a medical knowledge text) is 10.5, then I journal =10.5 / 20.0=0.525.

[0103] Based on this, after selecting candidate knowledge texts according to the comprehensive scores of each medical knowledge text, the top-K candidate knowledge texts with the highest comprehensive scores can be selected from all candidate knowledge texts to form an evidence set. Based on this evidence set, a clinical pathway knowledge graph is constructed using the core classes and relationships of the clinical pathway defined by the OWL (Web Ontology Language) ontology language, which is used to execute the subsequent step S104. The clinical pathway knowledge graph may include, but is not limited to, the standardized process and key decision nodes from initial diagnosis, examination, differential diagnosis to confirmed diagnosis and treatment.

[0104] For example: Based on the above example, suppose that based on the comprehensive score of each medical knowledge text, three candidate knowledge texts are selected as follows:

[0105] 1. Candidate knowledge text A (Doc A): "20xx Clinical Practice Guidelines: Management of Decompensated Cirrhosis".

[0106] Content: This article provides a detailed introduction to the diagnosis and treatment procedures for ascites, SBP, and hepatic encephalopathy.

[0107] Type: Guide (Level 1).

[0108] Year: xx.

[0109] Journal: xxx (IF≈25.0, highest in the field is about 30.0).

[0110] 2. Candidate knowledge text B (Doc B): "Meta-analysis of xxx combined with diuretics for the treatment of ascites in cirrhosis".

[0111] Content: The efficacy of spironolactone and furosemide combined with albumin was analyzed.

[0112] Type: Meta-analysis (Level 1).

[0113] Year: xx.

[0114] Journal: xxx (IF≈17.0).

[0115] 3. Candidate knowledge text C (Doc C): "Diagnosis and treatment report of xxx alcoholic cirrhosis xxx".

[0116] Content: Describes the treatment process of a single patient.

[0117] Type: Case report (Level 5).

[0118] Year: xx.

[0119] Journal: xxx (IF≈2.0).

[0120] Then set the current year to 2024, i.e., t. now =2024; and setting the attenuation coefficient to 0.3, i.e., k=0.3, then using the above formulas, the comprehensive scores of the above three candidate knowledge texts (i.e., Doc A, Doc B, and Doc C) can be calculated as follows:

[0121] Score A=0.4(0.95)+0.3(1.0)+0.2(0.74)+0.1(0.83)=0.38+0.3+0.148+0.083=0.911.

[0122] Score B=0.4(0.80)+0.3(1.0)+0.2(0.16)+0.1(0.57)=0.32+0.3+0.032+0.057=0.709.

[0123] Score C=0.4(0.75)+0.3(0.2)+0.2(1.0)+0.1(0.07)=0.30+0.06+0.2+0.007=0.567.

[0124] It is evident that the first candidate knowledge text (Doc A) will be included in the evidence set as the top-1 high-quality evidence for constructing standard clinical pathways; the second candidate knowledge text (Doc B) can be used as supplementary evidence; and the third candidate knowledge text (Doc C) has the lowest weight, forming a rich evidence set.

[0125] Based on this, and using the aforementioned evidence set (especially the high-quality first candidate knowledge text (i.e., Doc A)), OWL can be used to define an ontology, allowing entities in the target keyword set (such as alcoholic cirrhosis, spironolactone, and paracentesis) to be organized into a structured clinical pathway knowledge graph that conforms to the personalized characteristics of the target patient. Figure 3 As shown.

[0126] S104: Utilize clinical pathway knowledge graphs, target patient profiles, and evidence sets to perform closed-loop optimization of preliminary diagnoses and treatment plans until the final optimized diagnoses and treatment plans are obtained.

[0127] In this embodiment, after generating a target patient profile corresponding to the target patient in step S101, obtaining a preliminary diagnosis and treatment plan output by the medical big data model in step S102, and generating an evidence set and a clinical pathway knowledge graph that conforms to the personalized characteristics of the target patient in step S103, the preliminary diagnosis and treatment plan can be further optimized by using the clinical pathway knowledge graph, the target patient profile, and the evidence set to perform closed-loop correction until the final optimized diagnosis and treatment plan is obtained. This solves the problems that may exist in the preliminary diagnosis and treatment plan, such as non-standard terminology, illusions, outdated information, internal contradictions, and clinical thinking biases, thereby achieving the ideal optimization effect.

[0128] Specifically, one optional implementation is that the implementation process of step S104 may include the following steps S1041-S1043:

[0129] S1041: Compare the preliminary diagnosis and treatment plan with the diagnosis and treatment pathway in the clinical pathway knowledge graph, and revise the preliminary diagnosis and treatment plan based on the first comparison result to obtain the revised diagnosis and treatment plan.

[0130] In this implementation, the preliminary diagnosis and treatment plan can be compared with the diagnosis and treatment path in the clinical pathway knowledge graph from three dimensions: node alignment, path consistency detection, and constraint verification. Jumps, missing or deviated content in the medical big model thinking chain can be identified as the first comparison result. Based on the first comparison result, the preliminary diagnosis and treatment plan output by the medical big model can be corrected to obtain the corrected diagnosis and treatment plan, which is then used to execute the subsequent step S1042.

[0131] Specifically, firstly, the preliminary diagnosis and treatment plan output by the medical big data model can be transformed into structured "thought chain triplets" and anchored to standard nodes in the clinical pathway knowledge graph. Then, by comparing the "thought chain triplets" with the standard paths in the clinical pathway knowledge graph, common errors such as jumps, missing content, or deviations in the medical big data model's thought chain are identified, serving as the first comparison result. Next, based on the obtained comparison results, the preliminary diagnosis and treatment plan output by the medical big data model are corrected (which can be understood as rewriting) according to preset correction principles (the specific content is not limited, such as supplementing jump nodes, adding safety constraints, adjusting temporal priorities, etc.), resulting in the corrected diagnosis and treatment plan.

[0132] For example: Based on the above example, suppose that by comparing the preliminary diagnosis and treatment plan output by the medical big data model with the standard path in the clinical pathway knowledge graph, the jump error in the medical big data model's thought chain is identified as "LLM directly jumping from 'symptoms' to 'treatment,' skipping the necessary 'examination / decision' node." That is, in the medical big data model's thought chain, it is "Ascites → Direct use of diuretics," while the standard path in the clinical pathway knowledge graph is "Ascites → Decision node: Exclude SBP? → Examination: Abdominal paracentesis → If negative → Diuretics." Therefore, based on this first comparison result, it can be determined that there is a risk, as the sequential logic of "first puncture to rule out infection, then large-volume diuresis" is not clearly emphasized. If SBP (i.e., spontaneous bacterial peritonitis) is present, blind diuresis may induce kidney damage or shock. Consequently, the preliminary diagnosis and treatment plan can be adaptively modified to obtain a revised diagnosis and treatment plan.

[0133] S1042: Compare the revised diagnosis and treatment plan with the target patient profile to determine whether there are any inherent contradictions in medical knowledge, and use this as the second comparison result. Based on the second comparison result, adjust the revised diagnosis and treatment plan to obtain the adjusted diagnosis and treatment plan.

[0134] In this implementation, after obtaining the revised diagnosis and treatment plan through step S1041, the results of the path correction can be further compared with the target patient profile (i.e., the actual information of the original medical records related to the target patient) to detect whether there are any inherent contradictions in medical knowledge, such as the recommended diagnosis not matching the target patient's symptoms, or the recommended drugs conflicting with the target patient's comorbidities / allergy history. This is used as the second comparison result. Once a contradiction is triggered, the revised diagnosis and treatment plan can be adjusted according to the second comparison result to obtain the adjusted diagnosis and treatment plan, which is then used to execute the subsequent step S1043.

[0135] Specifically, testing for inherent contradictions in medical knowledge can protect the following two aspects:

[0136] Firstly, there are contradictions between diagnosis and symptoms, medical history, and examination results.

[0137] For scenarios such as "diagnosis failing to explain patient symptoms, contradictions between diagnosis and test results, and conflicts between diagnosis and past medical history," the core approach can revolve around "verification of the correlation between disease, symptoms, and tests." Based on authoritative clinical guidelines (such as the NCCN guidelines), medical textbooks, and high-quality literature, a knowledge chain of "disease-necessary symptoms-typical symptoms-exclusionary symptoms-key tests-test thresholds" can be reconstructed. Then, through forward verification, reverse exclusion, and relational reasoning, diagnostic inconsistencies can be detected.

[0138] Among them, positive validation refers to checking whether the "necessary symptoms / key examinations for the recommended diagnosis" exist in the target patient profile (e.g., if the diagnosis is "acute myocardial infarction" but the target patient has no "chest pain" and "normal troponin levels", then a contradiction is triggered).

[0139] Reverse exclusion refers to checking whether there are "exclusionary symptoms / examination results for the recommended diagnosis" in the target patient profile (e.g., if the diagnosis is "common cold" but the target patient has "high fever lasting 10 days + lung CT consolidation shadows", then a contradiction is triggered).

[0140] Relational reasoning refers to using the taboo relationship of "disease-comorbidity" in the graph (e.g., if the diagnosis is "gout" but the target patient has no history of "hyperuricemia" and no symptoms of "redness, swelling, heat and pain in the joints", it does not conform to the pathogenesis logic of gout, thus triggering a contradiction).

[0141] For example: Based on the above example, suppose the target patient profile includes the following:

[0142] 1. Symptoms / signs: Abdominal distension, ascites (ultrasound positive), jaundice (TB 67.8), no fever, no abdominal pain.

[0143] 2. Test results: PLT 70 (low), PT 19.2s (prolonged), AST 170 / ALT 100, WBC 5.9 (normal), CRP 11.3 (slightly elevated).

[0144] 3. Past medical history: Long-term alcohol consumption.

[0145] 4. History of drug allergies: None (default).

[0146] 5. Complications: Hypersplenism (presumed), coagulation dysfunction.

[0147] Furthermore, let's assume the revised diagnostic and treatment plan includes the following:

[0148] 1. Diagnosis: Decompensated alcoholic cirrhosis, suspected SBP (spontaneous bacterial peritonitis).

[0149] 2. Medications: Spironolactone, furosemide, human serum albumin, cephalosporin antibiotics (recommended in the hypothetical model due to suspected SBP), propranolol (recommended in the hypothetical model for the prevention of bleeding).

[0150] 3. Paracentesis and laboratory tests of ascites (including routine ascites fluid analysis, biochemistry, and bacterial culture): To determine the nature of the ascites (differentiating between transudate and exudate) and rule out spontaneous bacterial peritonitis (SBP). Although the patient has no fever or abdominal pain, a slight increase in white blood cell count (WBC) and C-reactive protein (CRP) should raise suspicion of occult infection.

[0151] The two are then compared to determine whether there are any inherent contradictions between the diagnosis and symptoms, medical history, examinations, and other medical knowledge mentioned in the first aspect above. Specific tests can include:

[0152] Test point A: Diagnosis of "spontaneous bacterial peritonitis (SBP)".

[0153] Positive verification: The typical features of SBP are fever, abdominal pain, and peritoneal irritation signs.

[0154] Comparison of target patient profiles: The patient had no fever, no abdominal pain, and no tenderness or rebound tenderness.

[0155] Key diagnostic tests: The gold standard for SBP diagnosis is ascites PMN >250 / mm³. Currently, only CRP is slightly elevated (11.3), and WBC is normal (5.9).

[0156] Judgment: [Weak contradiction / Insufficient evidence].

[0157] Test point B: Diagnosis of "alcoholic cirrhosis".

[0158] Positive verification: long history of alcohol consumption + AST>ALT (typical characteristics of alcoholism) + signs of portal hypertension (ascites / hypersplenism).

[0159] Target patient profile comparison: Completely matches.

[0160] Judgment result: [No contradiction].

[0161] Secondly, the detection of drug-related conflicts.

[0162] By integrating standard drug instructions, drug databases, pharmacovigilance information, clinical pharmacy guidelines, and the latest guidelines, a knowledge chain of "drug-allergenic component-contraindications-contraindications-cautious population" can be reconstructed (e.g., "bronchial asthma" is a contraindication for "drug xx," and "severe bradycardia" is a contraindication for "drug xx"). The "recommended drugs" (e.g., drug xx) output from the diagnosis are located in the atlas, their "contraindication list" is extracted, and matched one by one with the "allergy history (e.g., none), comorbidities (e.g., asthma), and symptoms (e.g., heart rate 50 bpm)" in the target patient profile. If there is an overlap (e.g., "asthma" is listed as a contraindication for drug xx), a "drug-comorbidity conflict" is triggered.

[0163] For example: Based on the examples above regarding the content of the target patient profile and the content of the revised diagnosis and treatment plan, comparing the two to determine whether the drug-related contradictions mentioned in the second aspect exist can be specifically tested as follows:

[0164] Detection point C: Drug "xx" (non-selective beta-blocker, NSBB).

[0165] Drug atlas knowledge: Drug xx is used to prevent varicose vein bleeding.

[0166] Contraindications / Cautions: Severe hypotension (SBP < 90 mmHg). Use with caution in patients with refractory ascites or at risk of acute kidney injury (may reduce renal perfusion and induce hepatorenal syndrome).

[0167] Comparison of target patient profile: The target patient's BP was 109 / 74 mmHg (acceptable), but he had a large amount of ascites and was in the active phase of jaundice (TB 67.8).

[0168] Judgment result: [Potential risk contradiction].

[0169] S1043: Utilize the evidence set to optimize the adjusted diagnosis and treatment plan, resulting in the final optimized diagnosis and treatment plan.

[0170] In this implementation, after obtaining the adjusted diagnosis and treatment plan through step S1042, the adjusted diagnosis and treatment plan can be further optimized using the evidence set. If a correction to the diagnosis strategy or a new treatment plan is found during the optimization process, the process of generating the TF-IDF vector and dense semantic vector corresponding to the target keyword set composed of the target case information, the corrected diagnosis, or the new treatment plan can be re-executed. Then, the TF-IDF vector and dense semantic vector are fused and processed, and subsequent steps (i.e., steps S102-S104 above) are performed until the final optimized diagnosis and treatment plan is obtained.

[0171] Specifically, the evidence set (including Doc A, etc.) constructed through operations such as retrieval and sorting mentioned in step S103 can be used to perform final optimization of the adjusted diagnosis and treatment plan. This includes reordering drugs, replacing them with treatment plans ranked higher, or supplementing with the latest treatment recommendations. If there are corrected diagnoses or newly added treatment plans, the process can re-enter S102 and subsequent steps for iterative verification until all steps pass verification, resulting in the final optimized diagnosis and treatment plan. Furthermore, to prevent getting stuck in an infinite loop, the upper limit of the number of iterations can be preset on the client side, and a prompt can be given for the last output result.

[0172] For example, based on the above examples, examples of optimizing the adjusted diagnosis and treatment plan using the evidence set can include the following three points:

[0173] 1. Revision of Nutritional Support Strategy (based on [Doc A] Guideline): Previous guidelines recommended "moderate protein intake," but the latest guidelines clearly state that patients with cirrhosis often have sarcopenia, which severely affects prognosis. Therefore, the dietary recommendation has been revised to "high-protein diet (1.2-1.5 g / kg / day)," and protein should not be restricted unless severe hepatic encephalopathy occurs.

[0174] 2. Micronutrient supplementation (based on [Doc C] A guidance): In view of the unique metabolic characteristics of patients with alcoholic liver disease, the "Zinc" and "Vitamin D" level testing and supplementation recommendations have been added, because zinc deficiency can aggravate hyperammonemia.

[0175] 3. Refined indications for albumin use (based on [Doc C] A & [Doc C] B): It is clarified that albumin is not only for volume expansion, but also for preventing diuretic-induced circulatory dysfunction, especially when the amount of fluid drained by paracentesis is large or the dose of diuretics is high.

[0176] Based on this, an example of the final optimized diagnosis and treatment plan can be shown below:

[0177] I. Clinical Diagnosis

[0178] Primary diagnosis: Alcoholic cirrhosis (decompensated, Child-Pugh C);

[0179] Other diagnoses: Ascites (moderate to large amount); hypersplenism; hypoalbuminemia; hyperbilirubinemia;

[0180] Differential diagnosis: Spontaneous bacterial peritonitis (SBP); Latent hepatic encephalopathy.

[0181] II. Further inspection

[0182] 1. [Urgent] Diagnostic paracentesis: xxx [Doc A].

[0183] 2. [New] Nutrition and micronutrient assessment: xxx [Doc C].

[0184] 3. Upper gastrointestinal endoscopy: xxx.

[0185] III. Comprehensive Treatment Plan

[0186] 1. Basic and Nutritional Therapy

[0187] Absolute abstinence from alcohol: Prevention and treatment of alcohol withdrawal syndrome.

[0188] Dietary management: xxx [Doc A].

[0189] Trace elements: xxx [Doc C].

[0190] 2. Management of ascites and edema

[0191] Combined diuresis: xxx. Adjustment principle: xxx [Doc A].

[0192] Human serum albumin: xxx. Supporting evidence: xxx [Doc B].

[0193] 3. Liver protection and anti-inflammation

[0194] ......

[0195] 4. Portal hypertension and hemorrhage prevention

[0196] Drug Xx: Not to be used at this time.

[0197] Reason: xxx.

[0198] IV. Monitoring Indicators

[0199] Daily: weight, waist circumference...

[0200] The next day: Kidney function...

[0201] Weekly: Liver function...

[0202] V. Supporting Evidence

[0203] [Doc A] .......

[0204] [Doc B] .......

[0205] [Doc C] .......

[0206] VI. Iteration State

[0207] Status: Validation passed - Output final result.

[0208] Note: Although "zinc" and "vitamin D" supplementation recommendations have been added, this is a minor adjustment related to nutritional support and does not involve core contraindications or change the main diagnostic pathway, nor does it trigger new contradictory testing loops. This output is the final optimized diagnosis and treatment plan.

[0209] In summary, this embodiment provides an optimization method for medical big data-assisted diagnosis and treatment. First, it acquires the target patient's case information and generates a target patient profile based on this information. Then, based on the target patient profile, it generates treatment instructions and inputs them into the medical big data model for assisted diagnosis and treatment, resulting in a preliminary diagnosis and treatment plan. After generating a TF-IDF vector and a dense semantic vector representing a target keyword set composed of keywords from the target case information, the preliminary diagnosis, and the treatment plan, along with their standard medical terminology, the TF-IDF vector and the dense semantic vector are fused to obtain a hybrid vector. Next, based on the hybrid vector, candidate knowledge texts that meet preset conditions are retrieved from a medical knowledge base, and an evidence set is generated using these candidate knowledge texts. This evidence set is then used to construct a clinical pathway knowledge graph. Finally, the clinical pathway knowledge graph, the target patient profile, and the evidence set are used to perform closed-loop correction optimization of the preliminary diagnosis and treatment plan until the final optimized diagnosis and treatment plan are obtained.

[0210] As can be seen, when optimizing the results of medical LLM-assisted diagnosis and treatment, this application first standardizes the keywords in the target case information, the preliminary diagnosis and treatment plan output by LLM, and then, after generating the TF-IDF vector and dense semantic vector of the target keyword set composed of each keyword and its standard medical terminology (e.g., deduplication and merging), the two are fused (e.g., weighted summation) to obtain a hybrid vector. This reduces the dispersion of the weight of medical terminology synonyms and improves the accuracy of the medical knowledge text subsequently selected from the real-time updated medical knowledge base.

[0211] Building upon this foundation, this application further utilizes more accurate candidate knowledge texts retrieved from medical knowledge bases to dynamically construct a clinical pathway knowledge graph that aligns with the personalized characteristics of the target patient. Combined with the target patient profile, this allows for continuous intelligent correction of the initial diagnosis and treatment plan output by the medical LLM, as well as adjustments to detected medical inconsistencies, until a final optimized diagnosis and treatment plan is obtained. This achieves closed-loop correction and optimization of the medical LLM-assisted diagnosis and treatment results, resolving issues such as non-standard terminology, illusions, outdated information, internal contradictions, and clinical reasoning biases that arise in medical LLM's assisted diagnosis and treatment recommendations, thereby achieving the desired optimization effect.

[0212] Furthermore, embodiments of this application also provide an optimized device for medical large-scale model-assisted diagnosis and treatment, including: a processor, a memory, and a system bus;

[0213] The processor and the memory are connected via the system bus;

[0214] The memory is used to store one or more programs, the one or more programs including instructions, which, when executed by the processor, cause the processor to perform any of the above-described implementations of the optimized method for medical large-scale model-assisted diagnosis and treatment.

[0215] refer to Figure 4 The diagram illustrates a structural schematic of an optimized device suitable for implementing the medical large-scale model-assisted diagnosis and treatment embodiments of this application. The optimized device for medical large-scale model-assisted diagnosis and treatment embodiments of this application may include, but is not limited to, terminals and servers equipped with medical large-scale models. Figure 4 The medical large-scale model-assisted diagnosis and treatment optimization device shown is merely an example of an electronic device and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0216] like Figure 4 As shown, the optimized device for medical model-assisted diagnosis and treatment may include a processor (e.g., a central processing unit, graphics processing unit, etc.) 401, which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 402 or a program loaded from storage device 408 into random access memory (RAM) 403. When the optimized device for medical model-assisted diagnosis and treatment is powered on, RAM 403 also stores various programs and data required for device operation. The processor 401, ROM 402, and RAM 403 are interconnected via bus 404. Input / output (I / O) interface 405 is also connected to bus 404.

[0217] Typically, the following devices can be connected to I / O interface 405: input devices 406 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 407 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 408 including, for example, memory cards, hard drives, etc.; and communication devices 409. Communication device 409 allows the electronic device to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 4 An optimized device for assisting diagnosis and treatment with a large medical model is shown; however, it should be understood that implementation or possession of all the devices shown is not required. More or fewer devices may be implemented alternatively.

[0218] Furthermore, embodiments of this application also provide a computer-readable storage medium storing instructions that, when executed on a terminal device, cause the terminal device to perform any of the above-described optimization methods for medical big data-assisted diagnosis and treatment.

[0219] Furthermore, this application embodiment also provides a computer program product, which, when run on a terminal device, causes the terminal device to execute any of the above-described optimization methods for medical big data-assisted diagnosis and treatment.

[0220] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that all or part of the steps in the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network communication device such as a media gateway, etc.) to execute the methods described in various embodiments or some parts of the embodiments of this application.

[0221] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0222] It should also 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 a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0223] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An optimization method for medical large-scale model-assisted diagnosis and treatment, characterized in that, include: Obtain the target patient's case information and use the case information to generate a target patient profile. Based on the target patient profile, diagnostic and treatment instructions are generated and input into the medical big data model for auxiliary diagnostic and treatment processing to obtain a preliminary diagnosis and treatment plan. After generating a TF-IDF vector and a dense semantic vector of the target keyword set composed of the target case information, keywords in the preliminary diagnosis and treatment plan and their standard medical terms, the TF-IDF vector and the dense semantic vector are fused to obtain a hybrid vector. After retrieving candidate knowledge texts that meet preset conditions from the medical knowledge base based on the hybrid vector, an evidence set is generated using the candidate knowledge texts; and a clinical pathway knowledge graph is constructed using the evidence set. Using the clinical pathway knowledge graph, target patient profile, and evidence set, the preliminary diagnosis and treatment plan are optimized through closed-loop correction until the final optimized diagnosis and treatment plan is obtained.

2. The method according to claim 1, characterized in that, The process of generating a target patient profile using target case information includes: The target case information is subjected to named entity recognition and relation extraction using a language model pre-trained on a medical corpus, resulting in entity recognition and relation extraction results. Based on a standard medical terminology database, the unstructured text of the target case information is mapped to standard medical terms and associated with standard classification codes. Based on the entity recognition results, relation extraction results, mapped standard medical terms, and associated standard classification codes, a target patient profile corresponding to the target patient is generated.

3. The method according to claim 1, characterized in that, After generating treatment instructions based on the target patient profile and inputting them into a large medical model for assisted diagnosis and treatment to obtain a preliminary diagnosis and treatment plan, the method further includes: Extract N target keywords from the target case information, preliminary diagnosis, and treatment plan, and determine the standard medical terms corresponding to the N target keywords; where N is a positive integer greater than 0. The N target keywords are fused to obtain a fused target keyword set, and a TF-IDF vector and a dense semantic vector corresponding to the fused target keyword set are generated.

4. The method according to claim 1, characterized in that, The step of retrieving candidate knowledge texts that meet preset conditions from the medical knowledge base based on the hybrid vector includes: Based on the target keyword set and its corresponding hybrid vector, knowledge texts with a comprehensive score higher than the preset score threshold are selected from the real-time updated medical knowledge base as candidate knowledge texts. The comprehensive score is calculated by weighting and summing the semantic similarity between medical knowledge texts in the medical knowledge base and the target keyword set, the evidence level of the medical knowledge texts, the time decay of the medical knowledge texts, and the authority of the medical knowledge texts. After retrieving candidate knowledge texts that meet preset conditions from the medical knowledge base based on the hybrid vector, an evidence set is generated using the candidate knowledge texts; and a clinical pathway knowledge graph is constructed using the evidence set, including: The top-K candidate knowledge texts with the highest comprehensive scores are selected from all candidate knowledge texts to form an evidence set. Based on the evidence set, a clinical pathway knowledge graph is constructed using the core classes and relationships of the clinical pathway defined by the OWL ontology language. The clinical pathway knowledge graph includes standardized processes and key decision nodes from initial diagnosis, examination, differential diagnosis to confirmed diagnosis and treatment.

5. The method according to any one of claims 1-4, characterized in that, The optimization process, which utilizes the clinical pathway knowledge graph, target patient profile, and evidence set to perform closed-loop correction on the preliminary diagnosis and treatment plan until a final optimized diagnosis and treatment plan is obtained, includes: The preliminary diagnosis and treatment plan are compared with the diagnosis and treatment pathways in the clinical pathway knowledge graph. Based on the first comparison result, the preliminary diagnosis and treatment plan are revised to obtain the revised diagnosis and treatment plan. The revised diagnosis and treatment plan is compared with the target patient profile, and the revised diagnosis and treatment plan is adjusted based on the second comparison result to obtain the adjusted diagnosis and treatment plan. Using the evidence set, the adjusted diagnosis and treatment plan is optimized to obtain the final optimized diagnosis and treatment plan.

6. The method according to claim 5, characterized in that, The process of comparing the preliminary diagnosis and treatment plan with the treatment pathways in the clinical pathway knowledge graph, and revising the preliminary diagnosis and treatment plan based on the first comparison result, to obtain the revised diagnosis and treatment plan, includes: Starting from three dimensions—node alignment, path consistency detection, and constraint verification—the preliminary diagnosis and treatment plan are compared with the diagnosis and treatment path in the clinical pathway knowledge graph. Jumps, omissions, or deviations in the medical big model's thinking chain are identified as the first comparison result. Based on the first comparison result, the preliminary diagnosis and treatment plan output by the medical big model are corrected to obtain the corrected diagnosis and treatment plan.

7. The method according to claim 5, characterized in that, The process involves comparing the revised diagnostic and treatment plan with the target patient profile, and adjusting the revised diagnostic and treatment plan based on the second comparison result to obtain the adjusted diagnostic and treatment plan, including: The revised diagnosis and treatment plan is compared with the target patient profile to determine whether there are any inherent contradictions in medical knowledge. This is used as the second comparison result. Based on the second comparison result, the revised diagnosis and treatment plan is adjusted to obtain the adjusted diagnosis and treatment plan.

8. The method according to claim 5, characterized in that, The process of optimizing the adjusted diagnosis and treatment plan using the evidence set to obtain the final optimized diagnosis and treatment plan includes: When optimizing the adjusted diagnosis and treatment plan using the evidence set, if there is a correction to the diagnosis or a new treatment plan, the process of generating the TF-IDF vector and dense semantic vector corresponding to the target keyword set consisting of the target case information, the corrected diagnosis or the new treatment plan is repeated. Then, the TF-IDF vector and dense semantic vector are fused and processed, and subsequent steps are performed until the final optimized diagnosis and treatment plan is obtained.

9. An optimized device for medical large-scale model-assisted diagnosis and treatment, characterized in that, include: Processor, memory, system bus; The processor and the memory are connected via the system bus; The memory is used to store one or more programs, the one or more programs including instructions that, when executed by the processor, cause the processor to perform the method according to any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed on a terminal device, cause the terminal device to perform the method described in any one of claims 1-8.

11. A computer program product, characterized in that, The program product includes a computer program that, when executed by an electronic product, can implement the method described in any one of claims 1-8.

Images