Chronic disease management AI recommendation anti-illusion method and system based on knowledge graph

Abstract

The invention discloses a chronic disease management AI recommendation anti-illusion method and system based on a knowledge graph, and relates to the technical field of medical health artificial intelligence, and the method comprises the steps: obtaining chronic disease medical data of a patient, extracting a feature vector, and constructing a multi-modal health trajectory vector; constructing a chronic disease tag vector and a directional causal knowledge graph; executing map path search to obtain candidate treatment paths; constructing structured cue words, and inputting the structured cue words into the medical large language model to generate candidate schemes; and calculating a comprehensive illusion score to carry out anti-illusion judgment and correction, and outputting a credible personalized treatment recommendation. By constructing a directional causal knowledge graph, the causal reasonability and semantic consistency of treatment path retrieval are enhanced, and the matching precision of candidate paths and patient states is improved; by calculating the comprehensive illusion score and introducing the four-dimensional score item for anti-illusion judgment, the credibility of the output content of the large language model is improved, and the availability of AI recommendation in clinical aid decision making is guaranteed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical health artificial intelligence, in particular to a knowledge graph-based AI recommendation anti-hallucination method and system for chronic disease management. BACKGROUND

[0002] With the rapid development of artificial intelligence technology, large language models have gradually expanded their application in medical auxiliary diagnosis and treatment, becoming an important auxiliary tool in chronic disease management. At the same time, knowledge graphs are widely used in individualized diagnosis and treatment recommendation systems due to their ability to structure medical knowledge and support causal reasoning. At present, some studies have attempted to integrate multi-modal health data and medical knowledge graphs to generate personalized intervention paths, aiming to achieve fine management of chronic disease patients. In addition, related research has also made preliminary exploration on hallucination suppression based on retrieval enhancement and knowledge guidance to address the risk of hallucination phenomenon in large language models in medical generation tasks.

[0003] Existing chronic disease management recommendation methods generally have the following shortcomings, which limit their practical usability and recommendation credibility in clinical practice: first, most existing technologies do not construct a causal knowledge graph with clear directionality and evidence level, making it difficult to achieve precise treatment path retrieval based on causal paths. Without fully utilizing the causal relationship information and its credibility classification in medical records, path search often stays at the semantic association level, lacking structural modeling of patient state evolution mechanisms; second, existing technologies for hallucination recognition and correction of large language model output results are relatively rough, mostly using confidence filtering and simple rule constraints, failing to achieve a structured and multi-dimensional anti-hallucination evaluation mechanism. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides a knowledge graph-based AI recommendation anti-hallucination method for chronic disease management to address the lack of directionality in causal knowledge graphs and the weak identification and feedback of hallucination risks in large language model generated content.

[0006] To solve the above technical problems, the present application provides the following technical solutions:

[0007] In a first aspect, the present application provides a knowledge graph-based AI recommendation anti-hallucination method for chronic disease management, which includes obtaining chronic disease medical data of a patient and extracting feature vectors of each chronic disease medical data to form a multi-modal health trajectory vector by concatenation;

[0008] According to the patient's chronic disease demographic information, a chronic disease label vector is constructed;

[0009] Based on the patients' clinical data on chronic diseases, a directional causal knowledge graph is constructed.

[0010] The multimodal health trajectory vector is mapped to the node space of the causal knowledge graph. The set of nodes most relevant to the patient's state is selected as the starting point. A path search algorithm is executed in the causal knowledge graph to obtain a set of candidate treatment paths that match the patient's state.

[0011] Using multimodal health trajectory vectors and chronic disease label vectors as input, structured prompt words are constructed. These structured prompt words are then input into a large language model in the medical field to generate a set of candidate treatment plans.

[0012] For each candidate treatment plan in the candidate treatment plan set, a comprehensive hallucination score is calculated and an anti-hallucination judgment is made. Based on the candidate treatment plan that triggers the anti-hallucination judgment, an anti-hallucination strategy is executed to obtain a corrected candidate treatment plan.

[0013] The candidate treatment options that did not trigger the anti-hallucination judgment and the revised candidate treatment options were selected as the final recommended options.

[0014] As a preferred embodiment of the AI-recommended anti-hallucination method for chronic disease management based on knowledge graphs described in this invention, the chronic disease medical data includes electronic medical record text, medical images, and structured biometric data.

[0015] As a preferred embodiment of the AI-recommended anti-hallucination method for chronic disease management based on knowledge graphs described in this invention, the specific steps of constructing a directional causal knowledge graph based on the patient's clinical data on chronic diseases are as follows:

[0016] Obtain the patient's clinical data on chronic diseases, and use dependency syntax to extract the set of causal candidate statements from the patient's clinical data on chronic diseases.

[0017] Extract causal triples from the set of causal candidate statements to obtain a set of causal triples.

[0018] The causal triple set is assigned an evidence level score based on the literature source of the patient's chronic disease clinical data using an evidence level matching function.

[0019] By using graph construction functions, a directional causal knowledge graph is generated.

[0020] As a preferred embodiment of the AI-recommended anti-hallucination method for chronic disease management based on knowledge graphs described in this invention, the steps include: mapping multimodal health trajectory vectors to the node space of a causal knowledge graph, selecting the set of nodes most relevant to the patient's state as the starting point, executing a path search algorithm in the causal knowledge graph to obtain a set of candidate treatment paths matching the patient's state; and specifically, mapping the multimodal health trajectory vectors to the node space of a causal knowledge graph, selecting the set of nodes most relevant to the patient's state as the starting point, and executing a path search algorithm in the causal knowledge graph to obtain a set of candidate treatment paths matching the patient's state.

[0021] By using the embedding space projection matrix to map the multimodal health trajectory vectors, an embedding representation of the multimodal health trajectory vectors in the node space of the causal knowledge graph is generated.

[0022] The embedding representation of multimodal health trajectory vectors in the node space of causal knowledge graph is similar to that of patients by cosine similarity calculation. Nodes with similarity scores higher than the similarity screening threshold are selected to form the set of nodes most relevant to the patient's status.

[0023] The set of nodes most relevant to the patient's status is used as the starting node. A path search algorithm is used to search for the starting node to obtain the set of outgoing paths from the starting node.

[0024] The set of outgoing paths from the starting node is subjected to medical semantic filtering to filter out paths that are inconsistent with the medical context, and a set of candidate treatment paths that match the patient's status is obtained.

[0025] As a preferred embodiment of the knowledge graph-based AI recommendation method for preventing hallucinations in chronic disease management as described in this invention, the steps include: using multimodal health trajectory vectors and chronic disease tag vectors as inputs to construct structured prompt words; inputting the structured prompt words into a large language model in the medical field to generate a set of candidate treatment plans; and the specific steps are as follows.

[0026] Generate standardized vector input fragments, input the standardized vector input fragments into a semantic generation function, combine them with a medical task enhancement template, and construct structured prompt words;

[0027] The structured prompts are input into a large language model in the medical field to generate a set of candidate treatment options.

[0028] As a preferred embodiment of the knowledge graph-based AI-recommended anti-hallucination method for chronic disease management described in this invention, the steps include: calculating a comprehensive hallucination score for each candidate treatment plan in the candidate treatment plan set, performing an anti-hallucination judgment, executing an anti-hallucination strategy based on the candidate treatment plan that triggers the anti-hallucination judgment, and obtaining a corrected candidate treatment plan. The specific steps are as follows:

[0029] For each candidate treatment option in the set of candidate treatment options, a hallucination risk assessment is performed using a composite scoring function to obtain a comprehensive hallucination score for each candidate treatment option;

[0030] Candidate treatment options are assessed for anti-hallucination based on a comprehensive hallucination score.

[0031] As a preferred embodiment of the AI-recommended anti-hallucination method for chronic disease management based on knowledge graphs described in this invention, the composite scoring function includes a data deviation scoring item, a similar patient reference deviation scoring item, a causal knowledge graph rule conflict item, and a population bias scoring item.

[0032] As a preferred embodiment of the knowledge graph-based AI recommendation anti-hallucination method for chronic disease management described in this invention, the anti-hallucination judgment includes setting a hallucination trigger threshold based on the statistical distribution of comprehensive hallucination scores of hallucination samples in historical labeled data. If the comprehensive hallucination score of a candidate treatment plan is lower than the hallucination trigger threshold, an anti-hallucination strategy is executed to correct the candidate treatment plan and obtain the corrected candidate treatment plan.

[0033] If the overall hallucination score of a candidate treatment plan is higher than the hallucination trigger threshold, the candidate treatment plan will be the final recommended plan.

[0034] As a preferred embodiment of the knowledge graph-based AI recommendation anti-hallucination method for chronic disease management described in this invention, the anti-hallucination strategy includes normalizing the data deviation score item, the similar patient reference deviation score item, the causal knowledge graph rule conflict item, and the population bias score item, and calculating the negative contribution of each score item to the total score.

[0035] For rating items with negative contribution values, a prompt word field is generated using a prompt word field generation function to construct reinforcement prompt words;

[0036] The constructed enhanced prompts are input into a large language model in the medical field, and the comprehensive hallucination score is recalculated until it passes the anti-hallucination judgment. The revised candidate solution is then used as the final recommended solution.

[0037] Secondly, the present invention provides an AI recommendation anti-hallucination system for chronic disease management based on knowledge graphs, including a trajectory construction module for acquiring patients' chronic disease medical data, extracting feature vectors of each chronic disease medical data, and splicing them to form a multimodal health trajectory vector;

[0038] The tag generation module is used to construct chronic disease tag vectors based on the patient's chronic disease demographic information;

[0039] The graph construction module is used to construct a directional causal knowledge graph based on the patient's clinical data on chronic diseases.

[0040] The path search module is used to map multimodal health trajectory vectors to the node space of the causal knowledge graph, select the set of nodes most relevant to the patient's state as the starting point, execute the path search algorithm in the causal knowledge graph, and obtain a set of candidate treatment paths that match the patient's state.

[0041] The prompt generation module is used to construct structured prompt words by taking multimodal health trajectory vectors and chronic disease label vectors as inputs, and then inputting the structured prompt words into a large language model in the medical field to generate a set of candidate treatment plans.

[0042] The hallucination correction module is used to calculate a comprehensive hallucination score for each candidate treatment plan in the candidate treatment plan set, and to make an anti-hallucination judgment. Based on the candidate treatment plan that triggers the anti-hallucination judgment, the anti-hallucination strategy is executed to obtain the corrected candidate treatment plan.

[0043] The treatment plan recommendation module is used to select candidate treatment plans that have not triggered the anti-hallucination judgment and the revised candidate treatment plans as the final recommended plans.

[0044] The beneficial effects of this invention are as follows: by constructing a directional causal knowledge graph, the causal rationality and semantic consistency of treatment path retrieval are enhanced, and the matching accuracy between candidate paths and patient states is improved; by calculating a comprehensive hallucination score and introducing a four-dimensional scoring item for anti-hallucination judgment, the credibility of the output content of the large language model is improved, and the usability of AI recommendations in clinical auxiliary decision-making is guaranteed. Attached Figure Description

[0045] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Fig. 1 A flowchart for recommending anti-hallucination methods using knowledge graph-based AI for chronic disease management.

[0047] Fig. 2 A schematic diagram of an AI-based anti-hallucination system for chronic disease management based on knowledge graphs.

[0048] Fig. 3 This is a flowchart of the prompt word generation process.

[0049] Fig. 4 A flowchart for the process of rating and revising recommendations for hallucinations. Detailed Implementation

[0050] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0051] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0052] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0053] Reference Figs. 1-4 This is one embodiment of the present invention, which provides an AI-based anti-hallucination method for chronic disease management based on knowledge graphs, including the following steps:

[0054] S1: Obtain the patient's chronic disease medical data, extract the feature vectors of each chronic disease medical data, and concatenate them to form a multimodal health trajectory vector.

[0055] Furthermore, the patient's chronic disease medical data is retrieved from the hospital's EMR electronic medical record system. This chronic disease medical data includes electronic medical record text, medical images, and structured biomarker data. The electronic medical record text data includes the chief complaint, present medical history, diagnosis, and medical orders. The medical images include image sequences such as CT and MRI. The structured biomarker data includes continuous time-series test items, such as blood glucose, blood lipids, liver enzymes, and eGFR.

[0056] It should be noted that the above content has been agreed to by the user and is used for legitimate purposes.

[0057] Furthermore, the acquired patient chronic disease medical data will undergo data cleaning and standard terminology mapping to form a unified text format and indicator units.

[0058] The electronic medical record text data of patients with unified text format and indicator units are used to extract features through a pre-trained BERT model in the medical field. The sentence vector at the CLS position is extracted as the semantic feature vector of the medical record text.

[0059] It should be noted that the training process of the BERT model in the medical field is as follows: a large amount of medical corpus is collected, the text is segmented, desensitized and standardized, unsupervised training is carried out on the BERT architecture through the masked language modeling task, general medical semantics and contextual relationships are learned, and the language understanding and task adaptability of the BERT model in the chronic disease medical scenario are optimized through supervised learning.

[0060] Patient medical images with standardized text format and unit of measurement are sliced ​​to form slice sequences. The slice sequences are then processed by a pre-trained ResNet3D network to extract features and obtain medical image feature vectors.

[0061] It should be noted that the training process of the ResNet3D network is as follows: the patient's medical image is standardized, its size is normalized, and its channels are adjusted to construct a three-dimensional image tensor with a unified format as input. The predefined ResNet3D structure is loaded onto the pre-trained weights on the medical image dataset as initialization and based on the annotation task, cross-entropy is used as the loss function. The Adam optimizer is used to iteratively update the parameters on the training set, extracting features through forward propagation and optimizing parameters through backpropagation.

[0062] Patient structured biometrics with uniform text format and unit of measurement are constructed into a time series matrix. The time series matrix is ​​then used to extract time trend feature vectors through a pre-trained Transformer network. Each element in the matrix represents a structured biometric at a given time point.

[0063] It should be noted that the training process of the Transformer network is as follows: the structured time series data is time-aligned, normalized, and missing values ​​are filled in, and encoded into a fixed-length time series input matrix. A Transformer model based on the self-attention mechanism is constructed, the input sequence is positionally encoded, and global dependencies are extracted through the attention layer. According to the specific task, a regression loss function is set, and the Adam optimizer is used to perform multiple rounds of iterative training on the training set. The model parameters are continuously optimized through gradient backpropagation.

[0064] The semantic feature vectors of medical record text, medical image feature vectors, and time trend feature vectors are concatenated using a concatenation function to form the patient's multimodal health trajectory vector.

[0065] S2: Construct chronic disease label vectors based on the patient's chronic disease demographic information.

[0066] Patient demographic information for chronic diseases includes the patient's age, gender, body mass index, and a list of concomitant medical records.

[0067] Furthermore, chronic disease population information of patients is collected and statistically analyzed from the clinical system. Based on the statistical information of chronic disease population of patients, chronic disease label vectors are set and constructed. Specifically, according to the type of chronic disease diagnosed by the patient, combined with the corresponding standard disease code and clinical feature description, a disease label set is established, and the disease label set is mapped into a fixed-length vector using One-Hot encoding to form a structured chronic disease label vector.

[0068] Chronic disease label vectors include the patient's age normalized value, gender code, BMI level code, comorbidity label vector, disease stage code, lifestyle clustering label, and first-degree relative medical history label.

[0069] It should be noted that the above content has been agreed to by the user and is used for legitimate purposes.

[0070] S3: Construct a directional causal knowledge graph based on the patient's clinical data on chronic diseases;

[0071] Obtain the patient's clinical data on chronic diseases, and use dependency syntax to extract the set of causal candidate statements from the patient's clinical data on chronic diseases.

[0072] Furthermore, APIs or database query interfaces can be called from the hospital's electronic medical record system to obtain the patient's chronic disease clinical data, which includes inpatient medical records, outpatient follow-up records, discharge summaries, clinical medical orders, and summaries of examination and test reports.

[0073] Use a medical ontology library to build a vocabulary for chronic diseases, such as hypertension, diabetes, chronic obstructive pulmonary disease, and coronary heart disease.

[0074] The clinical data of patients with chronic diseases are filtered using chronic disease terms as query criteria to generate a structured clinical text dataset of patients. The specific filtering rule is whether the patient's chronic disease clinical data contains at least one chronic disease keyword. If it does, it is determined to be chronic disease-related text. The retrieved chronic disease-related text is recorded, and each record includes the patient number, record time, record type, and chronic disease text content.

[0075] We performed Chinese sentence segmentation on the chronic disease text content of each record in the patient's structured clinical text dataset, removed invalid symbols, filtered semantically invalid words using medical stop words, and used a Chinese dependency parsing tool to parse the dependency relationships between words, generating short sentences summarizing the patient's chronic disease with dependency relationships.

[0076] For example, “Patient A suffers from frequent headaches at night and blurred vision due to long-term uncontrolled hypertension”, where the chronic disease term is “hypertension”, the causal trigger term is “cause”, the causal entity is “long-term uncontrolled hypertension”, and the result entity is “frequent headaches at night and blurred vision”.

[0077] We collected causal verbs from short sentences summarizing patients' chronic diseases and constructed a dictionary of causal relationship trigger words.

[0078] Iterate through each generated patient chronic disease summary sentence, detect whether there is a causal relationship trigger word in the patient chronic disease summary sentence, and look for a matching structural combination in the dependency syntax structure. If the dependency relationship of the causal relationship trigger word is a predicate, and the cause entity and the result entity have medical entities, then the chronic disease summary sentence is identified as a causal candidate statement. Among them, medical entities include diseases, symptoms, drugs, indicators, etc.

[0079] It should be noted that each causal candidate statement includes the patient number, recording time, a short summary of the patient's chronic disease, causal trigger words, cause, effect, and dependency path.

[0080] For example, a short summary of a patient's chronic disease is "Patient A suffers from frequent headaches at night and blurred vision due to long-term uncontrolled hypertension". The trigger word for the causal relationship is "causes", because of "long-term uncontrolled hypertension", and the effect is "frequent headaches at night and blurred vision". The dependency path is "SBV (hypertension, causes) → HED (causes) → VOB (causes, headache) → COO (headache, blurred vision)".

[0081] After the traversal is complete, obtain the set of causal candidate statements.

[0082] Extract causal triples from the set of causal candidate statements to obtain a set of causal triples.

[0083] Furthermore, the semantic boundaries of cause and effect in the causal candidate statements are refined, potentially redundant modifications are removed, sentences containing multiple clauses are judged by parallel relationship COO or subordinate relationship SBV / POB, irrelevant or ambiguous words are removed, the smallest identifiable causal unit is retained, and the smallest structured statement is generated.

[0084] For each minimal structured statement, a causal triple is constructed, and each causal triple is standardized. The causal triple includes a cause entity, a causal trigger word, and a result entity. The standardization process is as follows: the cause entity and the result entity are uniformly normalized to form a unified terminology, the causal trigger word is extracted from the minimal structured statement, and the parallel result entities are decomposed into multiple causal triples.

[0085] For example, the phrase "frequent headaches at night and blurred vision due to long-term uncontrolled hypertension" can be broken down into two causal triplets: ["hypertension" "leads to" "frequent headaches at night"] and ["hypertension" "leads to" "blurred vision".]

[0086] Based on historical causal triple credibility score data, a causal triple screening threshold of α is defined, with a value range consistent with the credibility score of the causal triple. A credibility score is calculated for each causal triple after standardization. If S≥α, it means that the screening is passed. All causal triples that pass the screening are output as a structured causal triple set.

[0087] Furthermore, the formula for calculating the credibility score of causal triples is expressed as follows:

[0088]

[0089] Where S represents the confidence score of the causal triple, with a value range of [0,1], λ1 represents the weight coefficient of the causal trigger word score, f1 represents the confidence score of the causal trigger word, λ2 represents the weight coefficient of the dependency structure integrity score, f2 represents the dependency structure integrity score between cause and effect, λ3 represents the word distance score of the patient's chronic disease summary sentence, and f3 represents the weight coefficient of the word distance score of the patient's chronic disease summary sentence.

[0090] It should be noted that, in order to ensure that the three contributions in the causal triple credibility score are non-negative and have practical significance, the weight coefficients of the causal trigger word score, dependency structure integrity score and word distance score must satisfy 0<λ1+λ2+λ3≤1 and λ1, λ2, λ3∈[0,1].

[0091] It should be noted that the confidence score is obtained as follows: based on the usage habits of commonly used causal terms in the medical field, a confidence word list of causal trigger words in the medical field is constructed, and each word is assigned a score from 0 to 1. The higher the score, the more common and authoritative it is.

[0092] For example, strong causal trigger words such as "cause," "induce," and "induce" are assigned a causal trigger word score of 1.0; slightly weaker causal trigger words such as "related" are assigned a causal trigger word score of 0.7; and weak causal trigger words such as "it can be seen" and "sometimes occur" are assigned a causal trigger word score of 0.2.

[0093] The dependency structure integrity score is obtained as follows: Let the dependency path distance between the cause and effect words in the dependency syntax structure be a. The dependency path distance is normalized according to the dependency path pattern, and the normalized value is used as the dependency structure integrity score.

[0094] For example, if the dependency path SBV→HED→VOB is 3, then the dependency path “SBV (hypertension, leading to)→HED (leading to)→VOB (leading to, headache)→COO (headache, confusion)” is 3, and the normalized dependency structure integrity score of patient A is 1.

[0095] The word distance score for the patient's chronic disease summary sentence is obtained as follows: The word distance score is the normalized value of the difference between the starting position of the cause word and the starting position of the effect word in the patient's chronic disease summary sentence.

[0096] The causal triple set is assigned an evidence level score based on the literature source corresponding to the patient's chronic disease clinical data by using an evidence level matching function.

[0097] Furthermore, when extracting electronic medical record text data from patients' chronic disease clinical data, if the electronic medical record text data comes from a data source with structured citation information, such as literature citations or medical guideline abstracts attached to the medical record system, the literature information related to each causal triple is extracted simultaneously. The literature information includes the source of the literature, the type of the literature, and the level of the literature.

[0098] The sources of the literature include PubMed numbers, DOI numbers, and medical guideline numbers; the types of literature include systematic reviews, RCTs, observational studies, and medical record reports; and the levels of literature include the GRADE system and the Oxford EBM system.

[0099] If the source of the literature is not indicated, an approximate semantic matching method is used to compare the causal triple with the content of the open literature database.

[0100] Adopt an authoritative evidence grading system, such as GRADE, and set standards for grading evidence.

[0101] For example, the ordinal score is set to 1.0 for meta-analysis; 0.9 for RCT (randomized controlled trial); 0.8 for prospective cohort study; 0.7 for retrospective observational study; and 0.4 for case report; if the literature cannot be matched or the electronic medical record text data is not supported by literature, it is set to 0.1.

[0102] For each causal triple (C x ,R x E x The evidence level score is calculated by comparing the extracted literature content with the evidence level score using an evidence level matching function, and the evidence level score is assigned to each causal triple.

[0103] The evidence level matching function is expressed as:

[0104]

[0105] Where S2 represents the evidence level score assigned to the causal triplet. C represents the text representation vector after concatenating causal triples. x Let R represent the cause entity of the x-th causal triple. x E represents the trigger word for the causal relationship in the x-th causal triple. x Represents the result entity of the x-th causal triple. Let G be the text representation vector of the y-th document's main text segment. y Let represent the evidence level score of the y-th document, and let cos(,) represent the cosine similarity between the text representation vector after the concatenation of the causal triples and the text representation vector of the document's main text segment.

[0106] It should be noted that obtaining the text representation vector after concatenating causal triples involves concatenating each causal triple in a unified format of "symptom" → "cause" → "disease" to form a semantically complete sequence of natural language short sentences, and then extracting the text representation vector of each causal triple.

[0107] It should be noted that obtaining the text representation vector of a document's main text segment specifically involves: extracting the context representation vector of the entire text segment after word segmentation and cleaning of the selected document's main text segment.

[0108] If the causal triple matches multiple literature passages, the passage with the highest evidence level score is selected; if no matching literature is found, a default score, such as 0.1, is set.

[0109] By using a graph construction function, a directional causal knowledge graph is generated by combining the set of causal triples and the evidence level score of each causal triple.

[0110] Furthermore, the causal and consequent entities in the causal triples are standardized using medical concepts. Entity disambiguation and naming are achieved through medical terminology usage habits, and a standard entity dictionary is established, assigning each entity a unique entity identifier. Specifically, the standard entity dictionary is established by summarizing frequently occurring standard medical terms from chronic disease prevention and control guidelines, clinical pathway specifications, medical insurance drug catalogs, and electronic medical records. These terms are then standardized and uniformly processed based on medical terminology usage habits, extracting entities such as disease names, examination items, treatment methods, and drug names. Ambiguous and synonymous words are removed, constructing a structured entity entry set, and generating standardized entities based on entity categories and attribute relationships. A chronic disease management entity dictionary is constructed. Each standardized causal entity and result entity is used as a graph node. Each graph node contains an entity identifier, a standard entity name, an entity type, and a thesaurus. Each causal triple generates a directed edge from the causal entity to the result entity. The attributes associated with the directed edge include causal relationship trigger words, evidence level scores, causal triple credibility scores, and short sentences summarizing the patient's chronic disease. The graph nodes and directed edges of each causal triple are organized using a graph construction function. Each causal triple is mapped to a directional edge, with the starting point of the edge being the causal entity node and the ending point being the result entity node. The direction of the edge represents the causal logical order.

[0111] S4: Map the multimodal health trajectory vector to the node space of the causal knowledge graph, select the set of nodes most relevant to the patient's state as the starting point, execute the path search algorithm in the causal knowledge graph, and obtain the set of candidate treatment paths that match the patient's state.

[0112] By using the embedding space projection matrix to map the multimodal health trajectory vectors, an embedding representation of the multimodal health trajectory vectors in the node space of the causal knowledge graph is generated.

[0113] Furthermore, the multimodal health trajectory vector is denoted as... Where H represents the multimodal health trajectory vector, and d1 represents the dimension of the multimodal health trajectory vector; to achieve consistency with the embedding vector space of causal knowledge graph nodes, an embedding space projection matrix is ​​defined. Where d2 represents the dimension of the embedding space of the causal knowledge graph nodes; the linear projection method is used to... Mapping to the node space of the causal knowledge graph, we obtain the embedding representation of the multimodal health trajectory vector in the causal knowledge graph space. The training of the embedded spatial projection matrix can be achieved by minimizing the matching error.

[0114] It should be noted that the training process of the embedding spatial projection matrix is ​​as follows: Patient data with known treatment pathways are selected from historical medical data to form a set of corresponding samples. Each set of samples includes a multimodal health trajectory vector and a causal knowledge graph node representation vector. Each sample comes from the same patient to ensure semantic correspondence. Based on the training sample pairs, an embedding spatial projection matrix with dimensions of multimodal health trajectory space × causal knowledge graph node space is constructed. For each set of samples, the multimodal health trajectory vector and the embedding spatial projection matrix are linearly transformed to obtain the mapped representation vector. Then, the Euclidean distance between the multimodal health trajectory vector and the corresponding graph node vector is calculated as the basis for the training loss function. The weighted average of the distances of all sample pairs is used as the overall loss. The projection matrix parameters are optimized through backpropagation until the training iterations end or the minimum projection error is reached on the validation set, at which point training terminates.

[0115] The embedding representation of multimodal health trajectory vectors in the causal knowledge graph node space is used to calculate similarity using cosine similarity, thereby obtaining a list of similarity scores for causal knowledge graph nodes.

[0116] Furthermore, using cosine similarity as a vector similarity metric, a one-to-one comparison calculation is performed between the multimodal health trajectory vector and the embedding vector of each node in the causal knowledge graph to obtain the similarity score of the causal knowledge graph node. Each node and its corresponding similarity score are recorded as a set of matching results during the calculation process, and finally a list of similarity scores of causal knowledge graph nodes is formed.

[0117] A similarity screening threshold τ is calculated using a dynamic threshold strategy. Nodes with similarity scores higher than the similarity screening threshold are selected to form a set of nodes most relevant to the patient's condition.

[0118] Furthermore, a similarity screening threshold is defined using a dynamic threshold strategy, and the calculation formula is expressed as follows:

[0119] τ=μ+α·σ;

[0120] Where τ represents the similarity screening threshold, with a value range consistent with the similarity score ([0,1]), μ represents the mean of all similarity scores in the similarity score list, σ represents the standard deviation of the similarity scores, and α represents an adjustment parameter used to ensure that the similarity screening threshold is within a meaningful range, with a value range of [0,1].

[0121] The set of nodes most relevant to the patient's status is used as the starting node. A path search algorithm is used to search for the starting node and obtain the set of exit paths from the starting node.

[0122] Furthermore, to avoid path combination explosion and semantic drift, a suitable maximum path search depth is set, such as a maximum of 3 layers. Path constraints are set, such as prohibiting cyclic paths. For each node in the set of nodes most relevant to the patient's state, the following operations are performed: starting from the current node, find the target node in the causal knowledge graph that has an outgoing edge pointing to it, and record it as a first-level neighbor; continue to perform the same outgoing edge expansion on the first-level neighbors to obtain second-level neighbors; continue to perform the expansion until the set maximum path depth is reached; each node sequence that starts from the starting node and passes through directional causal edges in sequence constitutes a path, and the path is saved to the outgoing path set to obtain the outgoing path selection set of the starting node.

[0123] The set of outgoing paths from the starting node is subjected to medical semantic filtering to filter out paths that are inconsistent with the medical context, and a set of candidate treatment paths that match the patient's status is obtained.

[0124] Furthermore, a medical semantic filtering rule base is constructed using patients' chronic disease clinical data. For example, "symptoms" cannot be used as a downstream node of "treatment"; "examination items" should not be directly connected to the "treatment method" node, but should be transferred through the "diagnosis" node.

[0125] Furthermore, the entity type sequence of each node in the outgoing path selection set of the starting node is extracted, and the entity type sequence is matched with the medical semantic filtering rule base. If there is a case in the outgoing path that violates the invalid structure defined in the rule base, it is judged as semantic inconsistency and is removed; if all node type combinations in the outgoing path meet the semantic rules, it is judged as a legal path and the path is retained.

[0126] Finally, a set of candidate treatment paths matching the patient's condition is obtained.

[0127] S5: Using multimodal health trajectory vectors and chronic disease label vectors as input, construct structured prompt words, input the structured prompt words into a large language model in the medical field, and generate a set of candidate treatment plans.

[0128] Based on the set of candidate treatment paths matched with the patient's status, the multimodal health trajectory vector and chronic disease label vector are subjected to field classification, dimension flattening and semantic annotation operations to generate standardized vector input fragments.

[0129] Furthermore, the elements in the multimodal health trajectory vector are divided into multimodal health trajectory vector field groups according to their source and semantic category. The multimodal health trajectory vector field groups include physiological signal field groups, examination result field groups, symptom manifestation field groups, drug use field groups, diagnosis field groups, and treatment response field groups.

[0130] For example, the physiological signal field group includes heart rate variability, systolic blood pressure, diastolic blood pressure, body temperature, etc.; the symptom manifestation field group includes fatigue, dizziness, chest tightness, etc.; and the medication use field group includes medication name, dosage, and cycle, etc.

[0131] The elements in the chronic disease label vector are divided into chronic disease label vector field groups according to their source and semantic category. The chronic disease label vector field groups include patient basic attribute field groups, chronic disease label field groups, complication field groups, genetic risk field groups, etc.

[0132] For example, the patient basic attribute field group includes age, gender, BMI, etc., the chronic disease label field group includes type II diabetes, hypertension classification, coronary heart disease course, etc., the complication field group includes retinopathy, kidney damage, etc., and the genetic risk field group includes family history, genotype, etc.

[0133] Furthermore, a flattening operation is performed on each data dimension within each field group. Specifically, for numerical fields, such as systolic blood pressure values, the quantified values ​​are retained; for textual fields, such as symptom names, they are converted into unified medical terminology codes.

[0134] Furthermore, semantic tags and structural descriptions are added to each field group and its corresponding subordinate dimensions. The semantic tags and structural descriptions include the semantic category, time attribute, unit and measurement standard, and semantic unique identifier.

[0135] Furthermore, all fields and content that have been categorized, flattened in dimensions, and semantically annotated are concatenated into a unified format input fragment according to structured rules, forming a standardized vector input fragment for the final interaction of the large language model.

[0136] It should be noted that the structured rules are as follows: assign uniform field names to each type of input data; use fixed sentence structures or structural formats to textualize numerical information; complete the medical units for numerical information; and arrange all fields in a fixed order in the prompt words.

[0137] The standardized vector input fragment is fed into the semantic generation function, combined with the medical task enhancement template, to construct structured prompt words and output the structured prompt words.

[0138] Furthermore, the standardized vector input fragments, which have undergone field classification, dimension flattening, and semantic annotation, are used as input to the semantic generation function. Based on the type of medical task being performed, the corresponding medical task prompt template is selected from the pre-built medical task enhancement template library. The standardized vector input fragments are then filled into the "Insert Standardized Vector Input Fragment" placeholder of the medical task prompt template to construct a structured prompt word with domain semantics and a clear task objective.

[0139] For example, the structured prompt might be: Based on the patient's health status and diagnostic information, and in conjunction with current medical consensus and evidence-based treatment standards, recommend a personalized and clinically reasonable treatment plan. The output should include the following three items: plan description, key medications, and expected efficacy; Patient basic attributes: Male, 62 years old; BMI: 27.8; Chronic disease symptoms: Fatigue (lasting approximately 6 days); Chronic disease diagnosis information: Stage II hypertension (initial diagnosis: [date]); Physiological information: Systolic blood pressure 142 mmHg (average of the last 7 days); Medication used: Valsartan 80 mg, once daily; Complications: Retinopathy (positive screening).

[0140] Furthermore, a medical task enhancement template library is constructed. Specifically, based on the core objectives of chronic disease management, a standardized set of medical task types is pre-defined as the functional scope for constructing medical task prompt templates. This standardized set of medical task types includes tasks such as recommending treatment plans and extracting key diagnostic and treatment points. Typical medical dialogue data and clinical communication examples are collected from question-and-answer transcriptions in real medical scenarios and recommended wording structures in medical guidelines and evidence-based medicine. The data is then segmented, subject-verb structure analyzed, and task intent categorized. General prompt language and task objective expression patterns are extracted. Based on the summarized task objective expression patterns and medical task semantics, structured input data is guided to fill in prompt words, thus constructing prompt templates that can be used in different patient scenarios.

[0141] For example, in the "Recommend Treatment Plan" task, the medical task enhancement template skeleton is: "Based on the following patients' health status and diagnostic information, and in conjunction with current medical consensus and evidence-based treatment standards, recommend a personalized and clinically reasonable treatment plan. The output should include the following three items: plan description, key drugs, and expected efficacy; [Insert standardized vector input fragment]".

[0142] The structured prompts are input into a large language model in the medical field to generate a set of candidate treatment options.

[0143] Furthermore, the medical domain large language model is an open language generation model that has been fine-tuned with medical corpus, such as Med-PaLM. An existing medical domain large language model is called and loaded, and the temperature, Top-K, maximum generation length, and number of generation rounds are set. The structured prompt words are used as complete input and input into the medical domain large language model, which triggers the contextual generation mechanism of the medical domain large language model to generate a set of candidate treatment plans containing medical semantic structure. The result generated by the medical domain large language model in each round is a candidate treatment plan, and each candidate treatment plan includes the treatment drug and dosage.

[0144] Furthermore, in a preferred embodiment, the training process of the large language model in the medical field includes two parts: a basic training stage and an instruction fine-tuning stage. Specifically, in the basic training stage, multi-source heterogeneous medical text data covering the chronic disease field is used. This multi-source heterogeneous medical text data includes clinical treatment guidelines, PubMed summaries, structured records and chief complaints of patients' electronic medical records for chronic diseases, and medical clinical communication corpora. The multi-source heterogeneous medical text data undergoes unified format standardization and sensitive information desensitization processing, low-quality corpora are removed, and unified terminology mapping is performed to ensure that the training corpus has high-quality medical semantic support. Based on the cleaned medical corpus, a standard pre-trained language model architecture, such as BERT, Med-PaLM, or GPT variants, is selected as the basic framework for language modeling. Causal language modeling is set, and the training is applied to the large language model in the medical field. The language model undergoes large-scale unsupervised pre-training, involving billions of data points, to build the basic language construction capabilities and semantic relationship perception capabilities of the medical domain's large language model in medical scenarios. Building upon this, to enhance the model's adaptability to real-world clinical question-and-answer tasks, a set of question-and-answer instruction sample pairs for chronic disease management scenarios is constructed, covering tasks such as treatment suggestion generation, medication regimen analysis, and complication intervention suggestions. Structured information from real patient cases or simulated patient status descriptions are used as input, with reference to medical guidelines and real medical orders as output labels. Supervised learning is used to fine-tune the large language model's instructions. During the fine-tuning phase, a learning rate adjustment mechanism and gradient freezing strategy are employed to control the stability of the medical domain's large language model. After training, the medical domain's large language model is evaluated based on a validation set constructed from real clinical tasks, completing the training process.

[0145] S6: Calculate the comprehensive hallucination score for each candidate treatment plan in the candidate treatment plan set, and make an anti-hallucination judgment. Execute the anti-hallucination strategy based on the candidate treatment plan that triggers the anti-hallucination judgment to obtain the corrected candidate treatment plan.

[0146] For each candidate treatment option in the candidate treatment option set, a hallucination risk assessment is performed using a composite scoring function to obtain a comprehensive hallucination score for each candidate treatment option. The composite scoring function includes a data deviation score, a similar patient reference deviation score, a causal knowledge graph rule conflict score, and a population bias score.

[0147] Furthermore, the data deviation scoring item is represented as follows:

[0148]

[0149] Wherein, D1(t) i ) represents candidate treatment plan t i The data support score, where m represents the total number of feature vectors involved in the comparison, and KL(|) represents the Kullback-Leibler divergence. Indicating in candidate treatment options t i The probability distribution generated by the k-th eigenvector in the regression prediction model. This represents the empirical probability distribution of the k-th feature vector in the patient's real chronic disease data.

[0150] It should be noted that obtaining Specifically, a pre-trained multi-task regression model is selected as the base model for the regression prediction model. The drug type, dosage, course of treatment, related diagnosis, and patient health characteristics in the candidate treatment plan are uniformly encoded into a fixed-dimensional input vector. The multi-task output layer of the regression prediction model simultaneously predicts the probability distribution of the candidate treatment plan. During the training process, a dataset containing the historical electronic medical records of patients with chronic diseases is used to extract the true feature distribution of each patient as a supervision signal. The KL divergence between the predicted distribution and the true distribution is constructed as a loss function, and the KL divergence is minimized in each round of training to approximate the true statistical distribution.

[0151] It obtains the empirical probability distribution of each feature vector from historical real-world chronic disease patient data through statistical modeling methods.

[0152] Furthermore, the similar patient reference deviation scoring item is represented as follows:

[0153]

[0154] Wherein, D2(t) i) represents candidate treatment plan t i The document similarity bias score, where K represents the number of subsets in the reference set used for comparison. v(t) represents the set of treatment options contained in the j-th reference subset. i ) represents candidate treatment plan t i The vectorized representation of v(t) j ) indicates the reference treatment plan t j The vectorized representation of , where |||2 represents the Euclidean distance.

[0155] It should be noted that the reference collection is a collection of review articles on guidelines related to the target chronic disease, complications and population characteristics, which are retrieved from authoritative medical databases such as PubMed and NICE. The structured treatment path information is extracted and vectorized using medical named entity recognition technology. Finally, it is a collection containing multiple population feature subsets that are constructed according to different patient population characteristics.

[0156] Furthermore, conflicting items in the medical rules of the knowledge graph are represented as follows:

[0157]

[0158] Among them, O3(t) i ) represents candidate treatment plan t i Rule consistency score, This represents the set of medical rules extracted from the medical rule knowledge graph. This indicates that the rule does not satisfy the indicator function when the candidate treatment plan t i The value is 1 when the medical rule r is not satisfied, and 0 otherwise. This indicates that the logical condition is not met.

[0159] It should be noted that the construction of the medical rule knowledge graph specifically involves extracting ruleable knowledge in the field of chronic diseases from authoritative medical knowledge resources, including clinical treatment guidelines, pharmacopoeias, and drug instructions.

[0160] For example, the clinical treatment guidelines extracted the statement that "thiazolidinediones are contraindicated in patients with type II diabetes and heart failure"; and the pharmacopoeia and drug instructions extracted the statement that "aspirin should not be used in combination with warfarin for a long period of time, as there is a risk of bleeding."

[0161] The extracted ruleable knowledge in the chronic disease domain is transformed into triples or conditional logic rules with constraints. A medical rule knowledge graph structure is constructed using a data structure consistent with the causal knowledge graph. The triple form includes entities, rule relations, and entities.

[0162] For example, in the form of a triplet, “insulin pump therapy” - “contraindicated” - “renal insufficiency”; and a conditional logic rule with constraints, “IF ‘patient age’ > 75 and glomerular filtration rate < 60, THEN ‘avoid using SGLT2 inhibitors’”.

[0163] Furthermore, the population bias scoring item is represented as follows:

[0164]

[0165] Among them, D4(t) i ) represents candidate treatment plan t i Compared to the population characteristic deviation score of the reference plan for similar patient groups, P represents the current patient's chronic disease demographic feature vector. Indicates candidate treatment plan t i The matched cluster center vector of the population.

[0166] It should be noted that obtaining the cluster center vector of the population specifically involves selecting candidate treatment plan t. i Historical patient records were retrieved from the clinical system, showing all patients who received candidate treatments in real-world medical settings. i For each patient in the patient population, demographic features are extracted and concatenated into a fixed-dimensional demographic feature vector to form a patient population sample set. A K-means clustering algorithm is then performed on this sample set, with a set K-means clustering value to calculate the cluster centers. These cluster centers are then used as candidate treatment plans. i For feature center vectors in different populations, the cluster center vector closest to the current patient demographic feature vector P is selected as the cluster center vector. The selection formula is expressed as follows:

[0167]

[0168] Among them, E K Let e ​​represent the Kth cluster center vector. k This represents the minimum value among all cluster center vectors.

[0169] Furthermore, the composite scoring function is expressed as:

[0170] A(t i )=ω1D1(t i )+ω2D2(t i )+ω3O3(t i )+ω4D4(t i );

[0171] Where A(t) i ) represents candidate treatment plan ti The comprehensive hallucination score is calculated as follows: ω1 represents the weight coefficient of the data deviation score item, ω2 represents the weight coefficient of the similar patient reference deviation score item, ω3 represents the weight coefficient of the knowledge graph medical rule conflict item, and ω4 represents the weight coefficient of the population bias score item.

[0172] It should be noted that the range of values ​​for each weight coefficient in the composite scoring function is ω. q ∈[0,1] and satisfy Where q = 1, 2, 3, 4, these correspond to the weight factors of the four dimensions.

[0173] Candidate treatment options are assessed for anti-hallucination based on a comprehensive hallucination score.

[0174] The anti-hallucination judgment includes setting a hallucination trigger threshold based on the statistical distribution of comprehensive hallucination scores of hallucination samples in historical labeled data. If the comprehensive hallucination score of a candidate treatment plan is lower than the hallucination trigger threshold, an anti-hallucination strategy is implemented to modify the candidate treatment plan and obtain the modified candidate treatment plan. If the comprehensive hallucination score of a candidate treatment plan is higher than the hallucination trigger threshold, the candidate treatment plan is taken as the final recommended plan.

[0175] Furthermore, based on the statistical distribution of comprehensive hallucination scores from hallucination samples in historical labeled data, a hallucination trigger threshold θ is set, with a value range of real numbers; each candidate treatment plan t is... i Calculated comprehensive hallucination score A(t) i Compare A(t) with the hallucination trigger threshold θ. i If t < θ, then the candidate treatment plan t is considered. i The presence of hallucination risk triggers anti-hallucination strategies; conversely, the absence of hallucination risk will trigger anti-hallucination strategies. i Candidate treatment options are output as recommended options.

[0176] Furthermore, the anti-hallucination strategy specifically involves normalizing four scoring items: data deviation score, similar patient reference deviation score, causal knowledge graph rule conflict score, and population bias score. This yields the scores for candidate treatment plans across each scoring item, and the negative contribution of each score to the total score is calculated. The formula for calculating the negative contribution is as follows:

[0177]

[0178] Among them, C q (t i ) represents candidate treatment plan t i The negative contribution value ω in the q-th rating item q This represents the weight coefficient of the q-th rating item. Indicates candidate treatment plan t iThe standardized score in the q-th rating item.

[0179] If C q (t i If ) is negative, it means that the qth scoring item is a characteristic segment source leading to the risk of hallucination.

[0180] For each C q (t i The feature vectors of rating items with negative values ​​are used to generate corresponding prompt word fields through a prompt word field generation function. Get the set of prompt words for negative rating items

[0181] It should be noted that a prompt word field generation function is defined, which is a calling function. Specifically, the prompt word field generation function calls different prompt word field parsing logics to extract prompt word fields based on the scoring item type. The scoring item types include data deviation type and graph rule conflict type. The extracted prompt word fields are uniformly output as key-value pair structures, with each type of scoring item corresponding to a set of fields. The key of the prompt word field is a standardized name, and the value is structured content or the corresponding numerical value. Specifically, the prompt word field parsing logic is as follows: If the scoring item type is data deviation type, the output value of the feature in the prediction model is obtained from the candidate treatment plan based on the feature name in the scoring item. At the same time, the historical mean or individual measured value of the corresponding feature is read from the patient status information, and the deviation value is obtained by subtraction. The "feature name", "predicted value", "actual value" and "deviation value" are output as prompt word fields. If the scoring item type is graph rule conflict type, the conflicting drug name, triggering rule text and corresponding patient label provided in the scoring item are directly read, and the fields are directly mapped to "conflicting drug", "rule content" and "patient label".

[0182] Based on the obtained negative score item prompt word field, and through semantic generation function, combined with medical task improvement template, reinforced prompt words are constructed.

[0183] For example, a strong prompt might say, "Please regenerate a chronic disease management plan suitable for the following patients, and avoid the following unreasonable recommendations:" ".

[0184] The constructed enhanced prompts are input into a large language model in the medical field to obtain the corrected candidate solutions. The comprehensive hallucination score is then recalculated until the anti-hallucination judgment is passed, and the corrected candidate solutions are output as recommended solutions.

[0185] S7: The candidate treatment options that do not trigger the anti-hallucination judgment and the revised candidate treatment options will be the final recommended options.

[0186] This embodiment also provides a knowledge graph-based AI recommendation system for chronic disease management to prevent hallucinations, including:

[0187] The trajectory construction module is used to acquire patients' chronic disease medical data, extract feature vectors from each chronic disease medical data, and concatenate them to form a multimodal health trajectory vector;

[0188] The tag generation module is used to construct chronic disease tag vectors based on the patient's chronic disease demographic information;

[0189] The graph construction module is used to construct a directional causal knowledge graph based on the patient's clinical data on chronic diseases.

[0190] The path search module is used to map multimodal health trajectory vectors to the node space of the causal knowledge graph, select the set of nodes most relevant to the patient's state as the starting point, execute the path search algorithm in the causal knowledge graph, and obtain a set of candidate treatment paths that match the patient's state.

[0191] The prompt generation module is used to construct structured prompt words by taking multimodal health trajectory vectors and chronic disease label vectors as inputs, and then inputting the structured prompt words into a large language model in the medical field to generate a set of candidate treatment plans.

[0192] The hallucination correction module is used to calculate a comprehensive hallucination score for each candidate treatment plan in the candidate treatment plan set, and to make an anti-hallucination judgment. Based on the candidate treatment plan that triggers the anti-hallucination judgment, the anti-hallucination strategy is executed to obtain the corrected candidate treatment plan.

[0193] The treatment plan recommendation module is used to select candidate treatment plans that have not triggered the anti-hallucination judgment and the revised candidate treatment plans as the final recommended plans.

[0194] This embodiment also provides a computer device applicable to the use of knowledge graph-based AI-recommended anti-hallucination methods for chronic disease management, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the knowledge graph-based AI-recommended anti-hallucination method for chronic disease management as proposed in the above embodiment.

[0195] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0196] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the AI-based anti-hallucination method for chronic disease management based on knowledge graphs as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0197] In summary, this invention enhances the causal rationality and semantic consistency of treatment path retrieval by constructing a directional causal knowledge graph, thereby improving the matching accuracy between candidate paths and patient states. Furthermore, by calculating a comprehensive hallucination score and introducing a four-dimensional scoring item for anti-hallucination judgment, the credibility of the output content of the large language model is improved, ensuring the usability of AI recommendations in clinical decision support.

[0198] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A knowledge graph-based AI recommendation method for preventing hallucinations in chronic disease management, characterized by: include, Acquire patients' chronic disease medical data, extract feature vectors from each chronic disease medical data, and concatenate them to form a multimodal health trajectory vector; Based on the demographic information of patients with chronic diseases, construct chronic disease label vectors; Based on the patients' clinical data on chronic diseases, a directional causal knowledge graph is constructed. The multimodal health trajectory vector is mapped to the node space of the causal knowledge graph. The set of nodes most relevant to the patient's state is selected as the starting point. A path search algorithm is executed in the causal knowledge graph to obtain a set of candidate treatment paths that match the patient's state. Using multimodal health trajectory vectors and chronic disease label vectors as input, structured prompt words are constructed. These structured prompt words are then input into a large language model in the medical field to generate a set of candidate treatment plans. For each candidate treatment plan in the candidate treatment plan set, a comprehensive hallucination score is calculated and an anti-hallucination judgment is made. Based on the candidate treatment plan that triggers the anti-hallucination judgment, an anti-hallucination strategy is executed to obtain a corrected candidate treatment plan. The candidate treatment options that did not trigger the anti-hallucination judgment and the revised candidate treatment options were selected as the final recommended options.

2. The knowledge graph-based AI recommendation method for preventing hallucinations in chronic disease management as described in claim 1, characterized in that: The chronic disease medical data includes electronic medical record texts, medical images, and structured biometric data.

3. The knowledge graph-based AI recommendation method for preventing hallucinations in chronic disease management as described in claim 2, characterized in that: The specific steps for constructing a directional causal knowledge graph based on the patient's clinical data on chronic diseases are as follows: Obtain the patient's clinical data on chronic diseases, and use dependency syntax to extract the set of causal candidate statements from the patient's clinical data on chronic diseases. Extract causal triples from the set of causal candidate statements to obtain a set of causal triples. The causal triple set is assigned an evidence level score based on the literature source of the patient's chronic disease clinical data using an evidence level matching function. By using graph construction functions, a directional causal knowledge graph is generated.

4. The knowledge graph-based AI recommendation method for preventing hallucinations in chronic disease management as described in claim 3, characterized in that: The process involves mapping multimodal health trajectory vectors to the node space of a causal knowledge graph, selecting the set of nodes most relevant to the patient's state as the starting point, and executing a path search algorithm within the causal knowledge graph to obtain a set of candidate treatment paths matching the patient's state. The specific steps are as follows: By using the embedding space projection matrix to map the multimodal health trajectory vectors, an embedding representation of the multimodal health trajectory vectors in the node space of the causal knowledge graph is generated. The embedding representation of multimodal health trajectory vectors in the node space of causal knowledge graph is similar to that of patients by cosine similarity calculation. Nodes with similarity scores higher than the similarity screening threshold are selected to form the set of nodes most relevant to the patient's status. The set of nodes most relevant to the patient's status is used as the starting node. A path search algorithm is used to search for the starting node to obtain the set of outgoing paths from the starting node. The set of outgoing paths from the starting node is subjected to medical semantic filtering to filter out paths that are inconsistent with the medical context, and a set of candidate treatment paths that match the patient's status is obtained.

5. The knowledge graph-based AI recommendation method for preventing hallucinations in chronic disease management as described in claim 4, characterized in that: The process involves using multimodal health trajectory vectors and chronic disease label vectors as input to construct structured cue words. These structured cue words are then input into a large language model in the medical field to generate a set of candidate treatment plans. The specific steps are as follows: Generate standardized vector input fragments, input the standardized vector input fragments into a semantic generation function, combine them with a medical task enhancement template, and construct structured prompt words; The structured prompts are input into a large language model in the medical field to generate a set of candidate treatment options.

6. The knowledge graph-based AI recommendation method for preventing hallucinations in chronic disease management as described in claim 5, characterized in that: The steps involve calculating a comprehensive hallucination score for each candidate treatment plan in the candidate treatment plan set, performing an anti-hallucination judgment, executing an anti-hallucination strategy based on the candidate treatment plan that triggers the anti-hallucination judgment, and obtaining a corrected candidate treatment plan. For each candidate treatment option in the set of candidate treatment options, a hallucination risk assessment is performed using a composite scoring function to obtain a comprehensive hallucination score for each candidate treatment option; Candidate treatment options are assessed for anti-hallucination based on a comprehensive hallucination score.

7. The knowledge graph-based AI recommendation method for preventing hallucinations in chronic disease management as described in claim 6, characterized in that: The composite scoring function includes a data deviation scoring item, a similar patient reference deviation scoring item, a causal knowledge graph rule conflict item, and a population bias scoring item.

8. The knowledge graph-based AI recommendation method for preventing hallucinations in chronic disease management as described in claim 7, characterized in that: The anti-hallucination judgment includes setting a hallucination trigger threshold based on the statistical distribution of comprehensive hallucination scores of hallucination samples in historical labeled data. If the comprehensive hallucination score of a candidate treatment plan is lower than the hallucination trigger threshold, an anti-hallucination strategy is executed to modify the candidate treatment plan and obtain the modified candidate treatment plan. If the overall hallucination score of a candidate treatment plan is higher than the hallucination trigger threshold, the candidate treatment plan will be the final recommended plan.

9. The knowledge graph-based AI recommendation method for preventing hallucinations in chronic disease management as described in claim 8, characterized in that: The anti-hallucination strategy includes normalizing the data deviation score item, the similar patient reference deviation score item, the causal knowledge graph rule conflict item, and the population bias score item, and calculating the negative contribution of each score item to the total score. For rating items with negative contribution values, a prompt word field is generated using a prompt word field generation function to construct reinforcement prompt words; The constructed enhanced prompts are input into a large language model in the medical field, and the comprehensive hallucination score is recalculated until it passes the anti-hallucination judgment. The revised candidate solution is then used as the final recommended solution.

10. A knowledge graph-based AI recommendation anti-hallucination system for chronic disease management, based on the knowledge graph-based AI recommendation anti-hallucination method for chronic disease management according to any one of claims 1 to 9, characterized in that: include, The trajectory construction module is used to acquire patients' chronic disease medical data, extract feature vectors from each chronic disease medical data, and concatenate them to form a multimodal health trajectory vector; The tag generation module is used to construct chronic disease tag vectors based on the patient's chronic disease demographic information; The graph construction module is used to construct a directional causal knowledge graph based on the patient's clinical data on chronic diseases. The path search module is used to map multimodal health trajectory vectors to the node space of the causal knowledge graph, select the set of nodes most relevant to the patient's state as the starting point, execute the path search algorithm in the causal knowledge graph, and obtain a set of candidate treatment paths that match the patient's state. The prompt generation module is used to construct structured prompt words by taking multimodal health trajectory vectors and chronic disease label vectors as inputs, and then inputting the structured prompt words into a large language model in the medical field to generate a set of candidate treatment plans. The hallucination correction module is used to calculate a comprehensive hallucination score for each candidate treatment plan in the candidate treatment plan set, and to make an anti-hallucination judgment. Based on the candidate treatment plan that triggers the anti-hallucination judgment, the anti-hallucination strategy is executed to obtain the corrected candidate treatment plan. The treatment plan recommendation module is used to select candidate treatment plans that have not triggered the anti-hallucination judgment and the revised candidate treatment plans as the final recommended plans.

Images