Ophthalmology Triage Knowledge Graph Construction System Based on Improved Gradient Boosting

Abstract

This invention belongs to the field of smart healthcare technology. It discloses an ophthalmology triage knowledge graph construction system based on improved gradient boosting. The system includes: collecting multi-source raw ophthalmology data; performing word segmentation on the multi-source raw ophthalmology data to obtain ophthalmology medical vocabulary, and dividing the ophthalmology medical vocabulary into symptom vocabulary and disease vocabulary; identifying professional terms within the symptom vocabulary, and constructing coarse inference algorithms and fine inference algorithms; constructing an ophthalmology entity relationship network based on disease vocabulary, professional terms, and a predefined symptom weight set, and integrating the coarse inference algorithm and fine inference algorithm to construct an ophthalmology triage knowledge graph. This invention can dynamically analyze complex symptoms, improve diagnostic accuracy, optimize medical resource allocation, and meet patients' self-assessment needs, thereby achieving rapid response and accurate recommendations for user symptoms, improving patient access to medical care efficiency and the quality of medical services.

Description

Technical Field

[0001] This invention relates to the field of smart healthcare technology, and more specifically, to an ophthalmic triage knowledge graph construction system based on improved gradient enhancement. Background Technology

[0002] With the aging population and the widespread use of electronic devices, the incidence of ophthalmic diseases is rising, and the uneven distribution of ophthalmic medical resources is becoming increasingly prominent. Patients often lack professional self-assessment capabilities before seeking medical treatment, leading to inefficient allocation of medical resources. Traditional ophthalmic triage systems mainly rely on simple rule matching or keyword retrieval, which cannot handle complex symptom descriptions and ambiguous expressions, resulting in limited accuracy and practicality. To assist doctors and patients in making scientific and efficient diagnoses, the construction of intelligent ophthalmic triage systems has become a research hotspot.

[0003] Knowledge graphs, as an effective tool for expressing domain knowledge and semantic relationships, can integrate and represent diseases, symptoms, and their inherent connections in the medical field, providing a solid data foundation and reasoning support for intelligent triage. However, existing medical knowledge graphs mostly focus on static knowledge representation and lack effective reasoning mechanisms, making them difficult to adapt to scenarios like ophthalmology triage that require dynamic analysis and precise reasoning. At the same time, the ophthalmology field has a large number of professional terms, and the same symptom may have multiple ways of being expressed, which increases the complexity and challenges of knowledge graph construction.

[0004] In view of this, the present invention proposes an ophthalmology triage knowledge graph construction system based on improved gradient boosting to solve the above problems. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art and to achieve the above objectives, the present invention provides the following technical solution: an ophthalmology triage knowledge graph construction system based on improved gradient boosting, comprising:

[0006] The data acquisition module is used to collect raw ophthalmic data from multiple sources.

[0007] The word segmentation module is used to segment multi-source ophthalmology raw data, obtain ophthalmology medical vocabulary, and divide ophthalmology medical vocabulary into symptom vocabulary and disease vocabulary.

[0008] The thesaurus building module is used to identify specialized terms in disease-related vocabulary and to build a thesaurus.

[0009] The coarse reasoning module is used to build a coarse reasoning algorithm. The coarse reasoning algorithm performs preliminary reasoning on the collected user-input symptoms based on a predefined set of symptom weights and a thesaurus, and obtains coarse reasoning results.

[0010] The fine-grained inference module is used to build fine-grained inference algorithms. The fine-grained inference algorithms use pre-built improved gradient models to perform deep inference on the coarse inference results to obtain fine-grained inference results.

[0011] The graph construction module is used to build an ophthalmology entity relationship network based on disease vocabulary, professional vocabulary and predefined disease weight sets, and to integrate coarse reasoning algorithm and fine reasoning algorithm to build an ophthalmology triage knowledge graph.

[0012] Furthermore, methods for classifying ophthalmological terminology into symptom terms and disease terms include:

[0013] Obtain disease seeds and symptom seeds from the ophthalmology medical vocabulary, and mark all words in the ophthalmology medical vocabulary that are not disease seeds or symptom seeds as words to be segmented; convert each word in the ophthalmology medical vocabulary into a corresponding word vector, and calculate the average disease similarity between each word to be segmented and all disease seeds, and the average symptom similarity between each word to be segmented and all symptom seeds according to the word vector. The average symptom similarity is the mean of the symptom similarity between the word to be segmented and each symptom seed.

[0014] Based on the comparison between the average disease similarity and the corresponding average symptom similarity of each word to be classified, each word to be classified is divided into symptom words or disease words.

[0015] Furthermore, methods for constructing a thesaurus include:

[0016] All disease seeds are treated as professional terms, and disease terms that are not professional terms are treated as terms to be analyzed. Based on the word vectors, the comprehensive similarity between each term to be analyzed and each professional term is calculated in turn. The comprehensive similarity with the highest value for each term to be analyzed is marked as the maximum similarity. The term to be analyzed with the maximum similarity greater than the preset similarity threshold is grouped with the corresponding professional term as a synonym group. A thesaurus is constructed based on all synonym groups.

[0017] Furthermore, methods for calculating comprehensive similarity include:

[0018] Assign character weights to each character position in the professional terminology and determine the corresponding editing operation for each character position; compare the professional terminology with the terminology to be analyzed to determine the editing operation corresponding to each character position in the terminology to be analyzed; set corresponding operation weights for each editing operation, and calculate the weighted edit distance and the maximum possible distance based on the character weight of each character position and the operation weight of the corresponding editing operation; calculate the literal similarity based on the weighted edit distance and the maximum possible distance.

[0019] A preset ratio set is used, which includes the ratio coefficients corresponding to literal similarity and disease similarity. Based on the ratio set, the disease similarity and literal similarity between the words to be analyzed and the professional words are weighted and summed to obtain the comprehensive similarity.

[0020] Furthermore, methods for obtaining rough inference results include:

[0021] Collect ophthalmology-specific vocabulary and build an expanded dictionary; select a word segmentation tool, and based on the default and expanded dictionaries of the word segmentation tool, perform word segmentation on the user's input symptoms to extract the user's symptoms vocabulary; according to the thesaurus, map each user's symptoms vocabulary to the corresponding professional vocabulary and mark it as the current professional vocabulary; the symptoms weighting includes the vocabulary weight between each professional vocabulary and each disease vocabulary;

[0022] The word weight of each current professional term is obtained from the disease weight set. The word weights of the same disease terms are added together in turn to obtain the credibility of each disease term. All credibility and corresponding disease terms are used as a rough inference result.

[0023] Furthermore, methods for obtaining detailed reasoning results include:

[0024] Disease-related terms in the rough inference results are marked as candidate terms. Professional terms corresponding to each candidate term are obtained according to the symptom weight set. Professional terms corresponding to the same candidate term are used as a set of standard symptoms. All current professional terms are used as a set of user symptoms. The set of standard symptoms and the set of user symptoms are collectively referred to as the symptom set. The term frequency-inverse document frequency of each term in each symptom set is calculated in turn. The vector space of each symptom set is modeled based on the term frequency-inverse document frequency to obtain the symptom vector of each symptom set.

[0025] Based on the symptom vector, the symptom similarity between the user's symptom set and each standard symptom set is calculated, and used as the confidence score of the candidate words corresponding to the standard symptom set. The coarse inference results and all confidence scores are input into the improved gradient model to predict the reliability set. The candidate words with a reliability score greater than the preset reliability threshold in the reliability set are used as diagnostic words, and all diagnostic words are used as the fine inference results.

[0026] Furthermore, the method for obtaining symptom vectors is as follows: construct a vocabulary list based on professional terms in the thesaurus; arrange the term frequency-inverse document frequency of each term in each symptom set in the order of professional terms in the vocabulary list to form the symptom vector of each symptom set; if there are terms in the vocabulary list that do not appear in the symptom set, then set the term frequency-inverse document frequency of the corresponding term in the corresponding symptom set to 0.

[0027] Furthermore, the steps for improving the gradient model construction include:

[0028] Step S1: Construct the initial feature space;

[0029] Step S2: Construct a disease relationship network based on the initial feature space;

[0030] Step S3: Extract the kink feature vector from the disease relationship network and merge it with the initial feature space to form an enhanced feature space;

[0031] Step S4: Iteratively train the gradient boosting model based on the enhanced feature space, and introduce a kink entropy regularization term. After the iterative training is completed, the improved gradient model is obtained.

[0032] Furthermore, the knot feature vector includes node centrality and knot invariants;

[0033] The methods for calculating the knot entropy regularization term include:

[0034] Based on the knot invariant, all nodes in the disease relationship network are divided into... There are several knot types, and the number of nodes in each knot type is counted. The relative frequency of each knot type is obtained based on the ratio between the number of nodes in each knot type and the total number of nodes. The knot entropy is calculated based on the relative frequency of each knot type. The knot entropy regularization term is obtained by multiplying the knot entropy by a preset regularization strength parameter.

[0035] Furthermore, the method for introducing the knot entropy regularization term is to add the knot entropy regularization term to the loss function during the iterative training process of the gradient boosting model.

[0036] The technical effects and advantages of this invention based on an improved gradient boosting-based ophthalmology triage knowledge graph construction system are as follows:

[0037] By collecting and segmenting multi-source ophthalmological data, ophthalmological medical vocabulary is accurately divided into symptom-related terms and disease-related terms, laying a solid lexical foundation for subsequent knowledge graph construction. A comprehensive similarity calculation method, combining literal and semantic similarity, effectively identifies specialized terms within symptom-related terms, constructing a specialized thesaurus covering synonyms and improving the understanding of user input descriptions. Based on a predefined symptom weight set and thesaurus, a coarse inference algorithm is designed to quickly identify potential disease candidates from user symptoms, providing a foundation for subsequent refined inference. An improved gradient boosting model is employed, integrating similarity scores based on symptom vectors. By analyzing and modeling relationships based on kink features, the diagnostic accuracy and reliability of inference algorithms are significantly enhanced, and the accuracy and reliability of fine-grained inference results are improved. By fusing coarse and fine-grained inference results, an ophthalmological knowledge graph containing disease entities, symptom entities, and their relationships is constructed, effectively supporting the knowledge representation and inference functions of intelligent triage systems. The entire process from raw data to knowledge representation is automated, enabling dynamic analysis of complex symptoms, improving diagnostic accuracy, optimizing medical resource allocation, and meeting patients' self-assessment needs. This leads to rapid response and accurate recommendations for user symptoms, improving patient access efficiency and the quality of medical services, and promoting the development of smart healthcare. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of the ophthalmology triage knowledge graph construction system based on improved gradient boosting, according to Embodiment 1 of the present invention. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] Example 1

[0041] Please see Figure 1 As shown in this embodiment, the ophthalmology triage knowledge graph construction system based on improved gradient enhancement includes a data acquisition module, a word segmentation processing module, a lexicon building module, a coarse reasoning module, a fine reasoning module, and a graph construction module; each module is connected via wired and / or wireless means to realize data transmission between modules.

[0042] The data acquisition module is used to collect raw ophthalmic data from multiple sources.

[0043] Multi-source ophthalmic original data includes but is not limited to medical institution data, open medical data, online platform data, etc.; among them, medical institution data is real clinical data from medical institutions such as hospitals and ophthalmic centers, including but not limited to electronic medical record data, ophthalmic examination reports, etc.; open medical data is data from authoritative public resources, standard terminology libraries, and academic databases, with the characteristics of open access or semi-open authorization, including but not limited to medical terminology libraries, medical literature abstracts, etc.; online platform data is user interaction content from Internet medical platforms, Q&A communities, and medical guidance APPs, including but not limited to online consultation dialogues, health Q&A community content, etc.;

[0044] Electronic medical record data includes but is not limited to chief complaints, current medical history, diagnosis records, etc.; ophthalmic examinations include but are not limited to optometry results, slit lamp examination results, etc.; medical terminology libraries such as ICD-11 (disease classification), SNOMED CT (clinical terminology library), MeSH (medical subject headings), etc.; medical literature abstracts such as literature abstracts from platforms such as PubMed, CNKI, Wanfang, Google Scholar, etc., containing the latest clinical research conclusions; online consultation dialogues such as structured dialogue texts where patients describe symptoms and doctors give preliminary suggestions, etc.; health Q&A community content such as community users' self-posted questions and doctors' or other users' replies, etc., and health Q&A communities such as Baidu Health, Dingxiangyuan Forum, Zhihu Health Topics, etc.

[0045] The word segmentation processing module is used to perform word segmentation processing on multi-source ophthalmic original data, obtain ophthalmic medical vocabulary, and divide the ophthalmic medical vocabulary into disease symptom vocabulary and disease vocabulary.

[0046] The methods for obtaining ophthalmic medical vocabulary include:

[0047] Using a general Chinese word segmentation tool (such as Jieba, IK Analyzer, NLPIR, etc.) to perform word segmentation processing on multi-source ophthalmic original data to extract candidate medical vocabulary; sequentially performing word property filtering and rule filtering on the candidate medical vocabulary to remove non-entity vocabulary and invalid vocabulary in the candidate medical vocabulary; those skilled in the art review the filtered candidate medical vocabulary to obtain ophthalmic medical vocabulary; word property filtering is to remove vocabulary of specific word properties, such as removing verbs, adverbs, conjunctions, prepositions, etc.; rule filtering is to remove invalid or irrelevant vocabulary according to preset rules, such as removing stop words (such as "de", "le", "shi", etc.), specific characters (such as numbers, special symbols, etc.), domain-irrelevant words (that is, common words in non-medical fields, such as "feeling", "yesterday", "I", etc.); the specific content of word property filtering and rule filtering is all pre-determined by those skilled in the art according to actual needs.

[0048] Symptom terms refer to words that describe a patient's subjective symptoms or abnormal manifestations found in objective examinations, such as blurred vision, eye pain, corneal edema, etc.; disease terms refer to words that have clear diagnostic significance and represent the physical manifestation of a disease, such as glaucoma, cataract, macular degeneration, etc.

[0049] Methods for categorizing ophthalmological terminology into symptom-specific and disease-specific terms include:

[0050] In the ophthalmological medical vocabulary, words collected from ICD-11 were labeled as disease seeds, and words collected from SNOMED CT were labeled as symptom seeds. Words in the ophthalmological medical vocabulary not labeled as disease seeds or symptom seeds were labeled as words to be segmented. Using pre-trained word vector models (such as Word2Vec, FastText, etc.), each word in the ophthalmological medical vocabulary was converted into a corresponding word vector. Based on the word vectors, the average disease similarity between each word to be segmented and all disease seeds, and the average symptom similarity between each word to be segmented and all symptom seeds were calculated sequentially. The average disease similarity and average symptom similarity of each word to be segmented were compared. If the average disease similarity was greater than the average symptom similarity, the corresponding word to be segmented was classified as a disease word; if the average disease similarity was less than the average symptom similarity, the corresponding word to be segmented was classified as a symptom word; if the average disease similarity was equal to the average symptom similarity, the corresponding word to be segmented was manually segmented by a person skilled in the art. Disease seeds were classified as disease words, and symptom seeds were classified as symptom words.

[0051] Among them, the average disease similarity is the mean of the disease similarity between the words to be segmented and each disease seed, and the average symptom similarity is the mean of the symptom similarity between the words to be segmented and each symptom seed. Both the disease similarity and the symptom similarity are cosine similarity.

[0052] The thesaurus building module is used to identify specialized terms in disease-related vocabulary and build a thesaurus.

[0053] Methods for building a thesaurus include:

[0054] All disease seeds are treated as professional terms, while disease terms that are not professional terms are treated as terms to be analyzed. Based on the word vectors, the comprehensive similarity between each term to be analyzed and each professional term is calculated sequentially. The comprehensive similarity with the highest value for each term to be analyzed is marked as the maximum similarity. Each maximum similarity is compared with a preset similarity threshold. Terms to be analyzed with a maximum similarity greater than the similarity threshold are grouped with their corresponding professional terms as synonym pairs; terms to be analyzed with a maximum similarity less than or equal to the similarity threshold are not grouped with their corresponding professional terms as synonym pairs. A thesaurus is constructed based on all synonym pairs. The similarity threshold is preset by those skilled in the art based on actual conditions.

[0055] Methods for calculating overall similarity include:

[0056] Each character position in the specialized vocabulary is assigned a character weight, and the corresponding editing operation is determined for each character position. The character weight increases with the character position, meaning the character position closer to the end of the vocabulary has a higher weight, reflecting the significant contribution of the characters at the end of the vocabulary to the meaning. The specialized vocabulary is compared with the vocabulary to be analyzed to determine the editing operation corresponding to each character position in the vocabulary to be analyzed. A corresponding operation weight is set for each editing operation. Based on the character weight of each character position and the operation weight of the corresponding editing operation, the weighted editing distance and the maximum possible distance are calculated respectively. The editing operations include insertion, replacement, and deletion. The specific values ​​of the character weight and operation weight are determined by those skilled in the art based on actual application scenarios and practical experience.

[0057] The literal similarity is calculated based on the weighted edit distance and the maximum possible distance. A preset ratio set is used, which includes the ratio coefficients corresponding to the literal similarity and the symptom similarity. This set is preset by those skilled in the art based on the actual situation. Based on the ratio set, the symptom similarity and literal similarity between the words to be analyzed and the professional words are weighted and summed to obtain the comprehensive similarity.

[0058] The weighted edit distance is calculated as follows: the product of the character weights corresponding to the same character position and the operation weight is used as the weight product of the corresponding character positions; the weight products of all character positions are summed sequentially to obtain the weighted edit distance; it should be noted that when the editing operation is insertion, since there is no original character at the insertion position, i.e., there is no corresponding character position, the character weight of the insertion position is defined as the average of the character weights of the two character positions before and after the insertion position; if the insertion position is the beginning or end of a word, i.e., there are only two adjacent character positions, then the character weights of the adjacent character positions are used as the character weight of the insertion position;

[0059] The maximum possible distance is calculated as follows: all character positions in the word to be analyzed are deleted and inserted sequentially, and the corresponding weighted edit distance is calculated as the maximum possible distance; the maximum possible distance is the maximum value of the weighted edit distance among all character edit paths; for example, the word to be analyzed is gastritis, and the professional term is chronic gastritis, so gastritis is deleted sequentially in the word to be segmented, and then chronic gastritis is inserted in the professional term.

[0060] The method for calculating literal similarity is as follows: the ratio of the weighted edit distance to the maximum possible distance is used as the normalized edit distance, and the difference between the weighted edit distance and the normalized edit distance is used as the literal similarity.

[0061] The coarse reasoning module is used to build a coarse reasoning algorithm. The coarse reasoning algorithm performs preliminary reasoning on the collected user-input symptoms based on a predefined set of symptom weights and a thesaurus, and obtains coarse reasoning results.

[0062] Methods for obtaining rough inference results include:

[0063] We collect ophthalmology-specific vocabulary and construct an expanded dictionary. Using a word segmentation tool, based on its default and expanded dictionaries, we segment user-input symptoms to extract relevant vocabulary. According to a thesaurus, each user-input symptom vocabulary is mapped to a corresponding professional term and marked as the current professional term. A symptom weight set includes the vocabulary weights between each professional term and various disease terms, used to measure the degree to which symptoms support disease diagnosis. We obtain the vocabulary weight of each current professional term from the symptom weight set, and sequentially add the vocabulary weights of terms with the same disease to obtain the credibility of each disease term. All credibility scores and their corresponding disease terms are used as a rough inference result; the credibility scores in the rough inference result are all greater than 0.

[0064] Among them, ophthalmology terminology refers to the professional symptom terms used in ophthalmological diagnosis, description, and communication, which are acquired through expert interviews, treatment guidelines, textbooks, and other means;

[0065] This embodiment uses IKAnalyzer as the word segmentation tool. IKAnalyzer is a lightweight open-source Chinese word segmentation tool based on the Java language. It mainly includes components such as IKAnalyzer.cfg.xml (the word segmenter extension configuration file), IKAnalyzer2012_u6.jar (the main JAR file), stopword.dic (the stop dictionary), LICENSE.txt, and NOTICE.txt (Apache copyright notice). The default dictionary is the basic dictionary that comes with the word segmentation tool when it is installed or deployed. It contains commonly used words, general terms, and common phrases to help the word segmentation tool perform basic segmentation of ordinary text. The extended dictionary adopts a configuration-based dictionary expansion method. Each ophthalmology professional term in the extended dictionary is placed line by line in the diseasename.txt text file. Notepad++ is used to convert it into a UTF-8 encoded document without BOM, diseasename.dic. The user-defined extended dictionary path in the IKAnalyzer.cfg.xml document is modified to realize the configuration of the extended dictionary.

[0066] The disease weight set is predefined by those skilled in the art based on a combination of factors, including clinical knowledge, standard treatment pathways, professional terminology systems, and expert consensus.

[0067] The fine-grained inference module is used to build fine-grained inference algorithms. These algorithms use pre-built improved gradient models to perform deep inference on the coarse inference results to obtain fine-grained inference results.

[0068] Methods for obtaining detailed inference results include:

[0069] Disease-related terms in the rough inference results are marked as candidate terms. Professional terms corresponding to each candidate term are obtained based on the symptom weight set, i.e., professional terms whose weights are not zero between the symptom weight set and the candidate term. Professional terms corresponding to the same candidate term are used as a set of standard symptoms, and all current professional terms are used as a set of user symptoms. The set of standard symptoms and the set of user symptoms are collectively referred to as the symptom set. The term frequency-inverse document frequency of each term in each symptom set is calculated in turn, and the vector space of each symptom set is modeled based on the term frequency-inverse document frequency to obtain the symptom vector of each symptom set.

[0070] Based on the symptom vector, the symptom similarity between the user's symptom set and each standard symptom set is calculated, and used as the confidence score of the candidate words corresponding to the standard symptom sets; where the symptom similarity is cosine similarity; the coarse inference results and all confidence scores are input into the improved gradient model to predict the reliability set; where the reliability set includes the reliability of each candidate word; the candidate words in the reliability set whose reliability is greater than a preset reliability threshold are taken as diagnostic words, and all diagnostic words are used as the fine inference results.

[0071] The term frequency-inverse document frequency (IF-IVF) measures the importance of a word within a set of symptoms, specifically as the product of term frequency and IVF. Term frequency reflects the frequency of a word's occurrence within a set of symptoms, specifically as the ratio of the number of times a word appears to the most frequent word in that set. The number of times a word appears is the total number of times it appears in the set of symptoms, and the most frequent word is the total number of times the word appears most frequently in the corresponding symptom set. IVF is used to reduce the weight of common words, and its expression is: In the formula, Inverse document frequency, The total number of symptoms. This represents the number of symptom sets containing the corresponding words when calculating word frequencies.

[0072] The method for obtaining symptom vectors is as follows: A thesaurus is constructed based on specialized vocabulary from a thesaurus, serving as the dimensional basis for the symptom vectors; the term frequency-inverse document frequency (TNF) of each word in each symptom set is arranged sequentially according to the order of specialized vocabulary in the thesaurus, forming the symptom vector for each symptom set; if a word in the thesaurus does not appear in the symptom set, its TNF in the corresponding symptom set is set to 0; that is, the number of TNFs in each symptom vector is consistent with the number of specialized vocabulary words in the thesaurus, which helps to quantify, compare, and analyze the relationship between user symptom sets and standard symptom sets.

[0073] The steps for improving the gradient model construction include:

[0074] Step S1: Construct the initial feature space;

[0075] Step S2: Construct a disease relationship network based on the initial feature space;

[0076] Step S3: Extract the kink feature vector from the disease relationship network and merge it with the initial feature space to form an enhanced feature space;

[0077] Step S4: Iteratively train the gradient boosting model based on the enhanced feature space, and introduce a kink entropy regularization term. After the iterative training is completed, the improved gradient model is obtained.

[0078] In step S1 above, the initial feature space includes the credibility and confidence levels of multiple different disease terms collected by those skilled in the art, as well as the reliability set according to actual experience; wherein, the credibility and confidence levels of the same disease terms are used as a set of initial feature vectors, and the corresponding reliability is used as the target value.

[0079] In step S2 above, the method for constructing the disease relationship network includes:

[0080] A pre-trained word vector model is used to convert each disease word in the initial feature space into a corresponding disease vector. Based on the disease vectors, the disease similarity between every two disease words is calculated and labeled as the relation similarity. Based on the relation similarity, an adjacency matrix is ​​constructed. Based on all disease words, disease similarities, and adjacency matrices in the initial feature space, a disease relation network is constructed.

[0081] The size of the adjacency matrix is , To construct the adjacency matrix based on the number of disease terms in the initial feature space, the method is as follows: when the similarity between two disease terms is greater than a preset threshold, the corresponding position in the adjacency matrix is ​​assigned the corresponding relation similarity value; based on the relation similarity, a similarity set is obtained for each disease term, including the one with the highest relation similarity to the disease term. A disease term, When one of the two disease terms has a similarity set that includes the other, the corresponding position in the adjacency matrix is ​​assigned the corresponding relation similarity. When the relation similarity between two disease terms is less than or equal to a preset relation threshold and neither is in the other's similarity set, the corresponding position in the adjacency matrix is ​​assigned zero. The relation threshold is preset by those skilled in the art based on the actual situation.

[0082] All disease terms in the initial feature space are treated as nodes in the disease relationship network. An edge is constructed between two nodes corresponding to non-zero elements in the adjacency matrix, and the weight of the edge is the value of the corresponding non-zero element.

[0083] Step S3 above: The method for extracting kink feature vectors from the disease relationship network includes:

[0084] The kink eigenvectors include node centrality and kink invariants;

[0085] Methods for extracting node centrality from disease relationship networks include:

[0086] Initialize the center vector, which includes... There are several numerical values, each initially set to 1. An iterative process is defined as follows: the product of the adjacency matrix and the center vector is normalized to obtain the modulation vector; the center vector is updated based on the modulation vector; the iterative process is executed repeatedly, and the modulation difference is calculated after each iteration until the modulation difference is less than a preset difference threshold. At this point, the iterative process stops, and the center vector is obtained, which includes the node centrality of each node. The modulation difference is the Euclidean distance between the updated center vector and the original center vector. The difference threshold is preset by those skilled in the art based on actual conditions.

[0087] Methods for extracting kink invariants from disease relationship networks include:

[0088] Calculate the clustering coefficient of each node in turn, and subtract each clustering coefficient from the result to obtain the number of kinks in each node. Use the number of kinks in all nodes as the kink invariant.

[0089] The expression for the clustering coefficient is: In the formula, For the first Clustering coefficients of nodes, For the first The number of edges between the neighboring nodes of a node. For the first The number of neighboring nodes of the i-th node, where the neighboring nodes are those of the i-th node. There are nodes that are connected by edges.

[0090] The method for forming the enhanced feature space is as follows: merge the initial feature vector of each node with the knot feature vector to form the enhanced feature vector; replace the initial feature vector of each node in the initial feature space with the corresponding enhanced feature vector to form the enhanced feature space.

[0091] In step S4 above, the calculation method for the knot entropy regularization term includes:

[0092] Based on the knot invariant, all nodes are divided into There are several knot types, and the number of nodes in each knot type is counted. The relative frequency of each knot type is obtained based on the ratio between the number of nodes in each knot type and the total number of nodes. Based on the relative frequency of each knot type, the knot entropy is calculated. The knot entropy regularization term is obtained by multiplying the knot entropy by a preset regularization strength parameter. Here, the knot entropy is the information entropy, and the regularization strength parameter is preset by those skilled in the art according to the actual situation.

[0093] The method for introducing the knot entropy regularization term is as follows: the knot entropy regularization term is added to the loss function during the iterative training process of the gradient boosting model; the gradient boosting model is an existing technology, and the specific training process will not be elaborated on here.

[0094] The graph construction module is used to build an ophthalmology entity relationship network based on disease vocabulary, professional vocabulary and predefined disease weight sets, and to integrate coarse reasoning algorithm and fine reasoning algorithm to build an ophthalmology triage knowledge graph.

[0095] The method for constructing an ophthalmology entity relationship network is as follows: disease terms are treated as disease entities, and professional terms are treated as symptom entities; a connection edge is constructed between each disease entity and each symptom entity; based on the symptom weight set, the word weights between each disease entity and symptom entity are obtained and used as the connection weights of the corresponding connection edges; based on the disease entities, symptom entities, and the connection weights of the connection edges between disease entities and symptom entities, an ophthalmology entity relationship network is constructed.

[0096] This embodiment accurately divides ophthalmological medical vocabulary into symptom-related terms and disease-related terms through the collection and word segmentation of multi-source ophthalmological data, laying a solid lexical foundation for subsequent knowledge graph construction. A comprehensive similarity calculation method, combining literal and semantic similarity, effectively identifies specialized terms within symptom-related terms, constructing a specialized thesaurus covering synonyms and improving the understanding of user input descriptions. Based on a predefined symptom weight set and thesaurus, a coarse inference algorithm is designed to quickly identify potential disease candidates from user symptoms, providing a foundation for subsequent refined inference. An improved gradient boosting model is employed, integrating similarity based on symptom vectors. Degree analysis and relationship modeling based on kink features significantly enhance the diagnostic accuracy and reliability of inference algorithms, and improve the accuracy and reliability of fine-grained inference results. By fusing coarse and fine-grained inference results, an ophthalmological knowledge graph containing disease entities, symptom entities, and their relationships is constructed, which can effectively support the knowledge representation and inference functions of intelligent triage systems. The entire process from raw data to knowledge representation is automated, enabling dynamic analysis of complex symptoms, improving diagnostic accuracy, optimizing the allocation of medical resources, and meeting patients' self-assessment needs. This allows for rapid response and accurate recommendations for user symptoms, improving patient access efficiency and the quality of medical services, and promoting the development of smart healthcare.

[0097] Example 2

[0098] This application also provides an electronic device. The electronic device may include one or more processors and one or more memories. The memories store computer-readable code that, when executed by the one or more processors, can perform the ophthalmology triage knowledge graph construction system based on improved gradient boosting as described above.

[0099] The method or system according to the embodiments of this application can also be implemented using the architecture of the electronic device shown in this application. The electronic device may include a bus, one or more CPUs, ROM, RAM, a communication port connected to a network, input / output, a hard disk, etc. The storage device in the electronic device, such as a ROM or hard disk, may store the ophthalmic triage knowledge graph construction system based on improved gradient boosting provided in this application. Furthermore, the electronic device may also include a user interface. Of course, the architecture shown in this application is merely exemplary; when implementing different devices, one or more components in the electronic device shown in this application may be omitted according to actual needs.

[0100] Example 3

[0101] One embodiment of this application discloses a computer-readable storage medium. The computer-readable storage medium stores computer-readable instructions. When the computer-readable instructions are executed by a processor, the ophthalmic triage knowledge graph construction system based on improved gradient boosting according to the embodiment of this application, as described with reference to the above figures, can be executed. The storage medium includes, but is not limited to, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and cache memory. Non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc.

[0102] Furthermore, according to embodiments of this application, the processes described in the above-referenced flowcharts can be implemented as computer software programs. For example, this application provides a non-transitory machine-readable storage medium storing machine-readable instructions that can be executed by a processor to perform instructions corresponding to the method steps provided in this application, such as an ophthalmic triage knowledge graph construction system based on improved gradient boosting. When this computer program is executed by a central processing unit (CPU), it performs the functions defined in the method of this application.

[0103] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0104] It should be noted that, in this document, 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. Unless otherwise specified, 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.

[0105] In the description of this invention, it should be understood that the terms "first," "second," etc., are used only for distinguishing descriptions and should not be construed as indicating or implying relative importance.

[0106] In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0107] In the description of this invention, "several" means one or more, and "a large number" means two or more.

[0108] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0109] All formulas in this manual are dimensionless and calculated numerically. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0110] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. An ophthalmic triage knowledge graph construction system based on improved gradient boosting, characterized in that, include: The data acquisition module is used to collect raw ophthalmic data from multiple sources. The word segmentation module is used to segment multi-source ophthalmology raw data, obtain ophthalmology medical vocabulary, and divide ophthalmology medical vocabulary into symptom vocabulary and disease vocabulary. Methods for categorizing ophthalmological terminology into symptom-specific and disease-specific terms include: Obtain disease seeds and symptom seeds from the ophthalmology medical vocabulary, and mark all words in the ophthalmology medical vocabulary that are not disease seeds or symptom seeds as words to be segmented; convert each word in the ophthalmology medical vocabulary into a corresponding word vector, and calculate the average disease similarity between each word to be segmented and all disease seeds, and the average symptom similarity between each word to be segmented and all symptom seeds according to the word vector. The average symptom similarity is the mean of the symptom similarity between the word to be segmented and each symptom seed. Based on the comparison between the average disease similarity of each word to be classified and the corresponding average symptom similarity, each word to be classified is divided into symptom words or disease words. The thesaurus building module is used to identify specialized terms in disease-related vocabulary and to build a thesaurus. Methods for building a thesaurus include: All disease seeds are treated as professional terms, and disease terms that are not professional terms are treated as terms to be analyzed. Based on the word vectors, the comprehensive similarity between each term to be analyzed and each professional term is calculated in turn. The comprehensive similarity with the highest value for each term to be analyzed is marked as the maximum similarity. The term to be analyzed with the maximum similarity greater than the preset similarity threshold is grouped with the corresponding professional term as a synonym group. A thesaurus is constructed based on all synonym groups. Methods for calculating overall similarity include: Assign character weights to each character position in the professional terminology and determine the corresponding editing operation for each character position; compare the professional terminology with the terminology to be analyzed to determine the editing operation corresponding to each character position in the terminology to be analyzed; set corresponding operation weights for each editing operation, and calculate the weighted edit distance and the maximum possible distance based on the character weight of each character position and the operation weight of the corresponding editing operation; calculate the literal similarity based on the weighted edit distance and the maximum possible distance. A preset ratio set is used, which includes the ratio coefficients corresponding to literal similarity and disease similarity. Based on the ratio set, the disease similarity and literal similarity between the words to be analyzed and the professional words are weighted and summed to obtain the comprehensive similarity. The coarse reasoning module is used to build a coarse reasoning algorithm. The coarse reasoning algorithm performs preliminary reasoning on the collected user-input symptoms based on a predefined set of symptom weights and a thesaurus, and obtains coarse reasoning results. The fine-grained inference module is used to build fine-grained inference algorithms. The fine-grained inference algorithms use pre-built improved gradient models to perform deep inference on the coarse inference results to obtain fine-grained inference results. The graph construction module is used to build an ophthalmology entity relationship network based on disease vocabulary, professional vocabulary and predefined disease weight sets, and to integrate coarse reasoning algorithm and fine reasoning algorithm to build an ophthalmology triage knowledge graph.

2. The ophthalmic triage knowledge graph construction system based on improved gradient boosting according to claim 1, characterized in that, Methods for obtaining rough inference results include: Collect ophthalmology-specific vocabulary and build an expanded dictionary; select a word segmentation tool, and based on the default and expanded dictionaries of the word segmentation tool, perform word segmentation on the user's input symptoms to extract the user's symptoms vocabulary; according to the thesaurus, map each user's symptoms vocabulary to the corresponding professional vocabulary and mark it as the current professional vocabulary; the symptoms weighting includes the vocabulary weight between each professional vocabulary and each disease vocabulary; The word weight of each current professional term is obtained from the disease weight set. The word weights of the same disease terms are added together in turn to obtain the credibility of each disease term. All credibility and corresponding disease terms are used as a rough inference result.

3. The ophthalmic triage knowledge graph construction system based on improved gradient boosting according to claim 2, characterized in that, Methods for obtaining detailed inference results include: Disease-related terms in the rough inference results are marked as candidate terms. Professional terms corresponding to each candidate term are obtained according to the symptom weight set. Professional terms corresponding to the same candidate term are used as a set of standard symptoms. All current professional terms are used as a set of user symptoms. The set of standard symptoms and the set of user symptoms are collectively referred to as the symptom set. The term frequency-inverse document frequency of each term in each symptom set is calculated in turn. The vector space of each symptom set is modeled based on the term frequency-inverse document frequency to obtain the symptom vector of each symptom set. Based on the symptom vector, the symptom similarity between the user's symptom set and each standard symptom set is calculated, and used as the confidence score of the candidate words corresponding to the standard symptom set. The coarse inference results and all confidence scores are input into the improved gradient model to predict the reliability set. The candidate words with a reliability score greater than the preset reliability threshold in the reliability set are used as diagnostic words, and all diagnostic words are used as the fine inference results.

4. The ophthalmic triage knowledge graph construction system based on improved gradient boosting according to claim 3, characterized in that, The method for obtaining symptom vectors is as follows: construct a vocabulary list based on professional terms in the thesaurus; arrange the term frequency-inverse document frequency of each word in each symptom set in the order of professional terms in the vocabulary list to form the symptom vector of each symptom set; if there are words in the vocabulary list that do not appear in the symptom set, then set the term frequency-inverse document frequency of the corresponding word in the corresponding symptom set to 0.

5. The ophthalmic triage knowledge graph construction system based on improved gradient boosting according to claim 4, characterized in that, The steps for improving the gradient model construction include: Step S1: Construct the initial feature space; Step S2: Construct a disease relationship network based on the initial feature space; Step S3: Extract the kink feature vector from the disease relationship network and merge it with the initial feature space to form an enhanced feature space; Step S4: Iteratively train the gradient boosting model based on the enhanced feature space, and introduce a kink entropy regularization term. After the iterative training is completed, the improved gradient model is obtained.

6. The ophthalmic triage knowledge graph construction system based on improved gradient boosting according to claim 5, characterized in that, The kink eigenvectors include node centrality and kink invariants; The methods for calculating the knot entropy regularization term include: Based on the knot invariant, all nodes in the disease relationship network are divided into... There are several knot types, and the number of nodes in each knot type is counted. The relative frequency of each knot type is obtained based on the ratio between the number of nodes in each knot type and the total number of nodes. Based on the relative frequency of each knot type, the knot entropy is calculated; the knot entropy regularization term is obtained by multiplying the knot entropy by a preset regularization strength parameter.

7. The ophthalmic triage knowledge graph construction system based on improved gradient boosting according to claim 6, characterized in that, The method for introducing the knot entropy regularization term is to add the knot entropy regularization term to the loss function during the iterative training process of the gradient boosting model.

Images