Method and system for constructing asymptomatic knowledge graph

Abstract

The invention relates to the field of artificial intelligence, in particular to an asymptomatic knowledge graph construction method and system. The method comprises the following steps: acquiring single-symptom text data of the asymptomatic and discriminating diseases; constructing an information extraction model based on a long short-term memory network and a deep hidden Markov model; performing information extraction on the single symptom text data by using an information extraction model to obtain a plurality of disease triples; constructing an identification triple by using the disease triple; and constructing the asymptomatic knowledge graph based on the identification triad. The problem that in the prior art, an asymptomatic knowledge graph is not accurate enough is solved. According to the method, the disease triad is extracted from the special single symptom text data of the asymptomatic disease and the identified disease through the constructed information extraction model, then the identification triad containing the quantitative support degree is constructed, and finally the knowledge graph for asymptomatic disease identification is established. The deep fusion of multi-source medical knowledge and the accurate quantification of symptom-disease association are realized, and the accuracy of the asymptomatic knowledge graph is improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence, specifically to a method and system for constructing a knowledge graph for ALS (Amyotrophic Lateral Sclerosis). Background Technology

[0002] Amyotrophic Lateral Sclerosis (ALS), a progressive neurodegenerative disease, is characterized by damage to upper and lower motor neurons. Clinical manifestations include muscle weakness, atrophy, and fasciculations. The disease progresses rapidly and has a very poor prognosis, making early and accurate diagnosis crucial for slowing its progression. However, the clinical differential diagnosis of ALS faces significant challenges: its early symptoms lack specificity and highly overlap with those of various diseases such as cervical spondylosis, Hirayama disease, myasthenia gravis, and syringomyelia. Common manifestations such as limb weakness and sensory abnormalities can easily lead to clinical misdiagnosis.

[0003] Clinical statistics show that the early misdiagnosis rate of ALS is as high as 40%, mainly because current diagnoses rely on doctors' experience and lack structured diagnostic knowledge. Current medical knowledge is scattered across multiple sources such as textbooks, guidelines, and research literature, exhibiting heterogeneous forms and loose connections, making it difficult to quickly retrieve quantitative correlation data between different diseases and common symptoms.

[0004] Existing medical knowledge graphs are mostly general practice-based, with a rough depiction of specific knowledge about ALS and disease differentiation, and have not formed a targeted symptom-disease differentiation system, resulting in an inaccurate knowledge graph for ALS.

[0005] Therefore, a more accurate knowledge graph of ALS is urgently needed. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method and system for constructing a knowledge graph for ALS (Amyotrophic Lateral Sclerosis), which solves the problem that existing knowledge graphs for ALS are not accurate enough.

[0007] To achieve the above objectives, one aspect of the present invention provides a method for constructing a knowledge graph for ALS (Amyotrophic Lateral Sclerosis), the method comprising: acquiring single-symptom text data of ALS and differential diagnosis diseases; constructing an information extraction model based on a long short-term memory network and a deep hidden Markov model; using the information extraction model to extract information from the single-symptom text data to obtain multiple disease triples; constructing differential diagnosis triples using the disease triples; and constructing an ALS knowledge graph based on the differential diagnosis triples.

[0008] This invention acquires single-symptom text data of ALS and its differential diagnosis, constructs an information extraction model using long short-term memory networks and deep hidden Markov models, accurately extracts disease triples and further constructs differential diagnosis triples, and finally builds an ALS knowledge graph based on the differential diagnosis triples. This effectively integrates multi-source medical knowledge, quantifies the association between symptoms and diseases, and improves the accuracy of the ALS knowledge graph.

[0009] Optionally, the step of obtaining single-symptom text data for ALS and disease identification includes: obtaining raw text data for ALS and disease identification; standardizing the raw text data to obtain standardized text data; performing dependency parsing on the standardized text data to obtain disease entities, symptom entities, associations, and semantic boundaries; dividing and semantically completing the standardized text data based on the disease entities, symptom entities, and associations using the semantic boundaries to obtain initial single-symptom text data; obtaining the initial source coefficients of the initial single-symptom text data, and concatenating the initial source coefficients with the initial single-symptom text data to obtain single-symptom text data.

[0010] This invention ensures comprehensive information by acquiring multi-source original text data on ALS and disease identification. Standardization is then applied to unify terminology, resolve ambiguities, and eliminate text heterogeneity and ambiguity. Next, dependency parsing is used to accurately extract core elements and semantic boundaries. These elements are then combined to segment the text and complete its semantics, ensuring structural integrity and semantic clarity. Finally, source coefficients are assigned to the text to quantify data credibility, constructing structured single-symptom text data and improving its accuracy and reliability.

[0011] Optionally, the standardization process of the original text data to obtain standardized text data includes: constructing a standardized terminology dictionary, and identifying and replacing synonymous expressions in the original text data based on the standardized terminology dictionary to obtain optimized text data; performing dependency parsing on the optimized text data to obtain pronouns and their corresponding antecedents; and replacing the pronouns with the antecedents to obtain standardized text data.

[0012] This invention improves the data quality of the original text by constructing a standardized terminology dictionary specifically for the identification of ALS, identifying and replacing synonymous expressions of disease and symptoms in the original text, unifying entity representations to reduce ambiguity, and then performing dependency parsing on the optimized text to accurately identify pronouns and their corresponding antecedents, replacing pronouns with antecedents to resolve referential ambiguity.

[0013] Optionally, the information extraction model constructed based on the Long Short-Term Memory Network and the Deep Hidden Markov Model includes: setting a loss function based on Conditional Random Field loss and Binary Cross-Entropy loss; constructing a joint extraction sub-model based on the Long Short-Term Memory Network using the loss function; constructing a confidence calculation model based on the Deep Hidden Markov Model; and constructing an information extraction model based on the joint extraction sub-model and the confidence calculation model.

[0014] This invention optimizes the sequence labeling task for entity recognition and the binary classification task for associated attributes by setting a joint loss function that combines conditional random field loss and binary cross-entropy loss. Then, a joint extraction sub-model with a shared coding layer is constructed based on a long short-term memory network to simultaneously realize entity recognition and attribute determination to improve extraction efficiency. Subsequently, a confidence calculation model is constructed based on a deep hidden Markov model to perform sequence calibration and confidence quantification on the preliminary results. Finally, the two models are integrated to form an information extraction model, which improves the accuracy of medical information extraction.

[0015] Optionally, constructing a differential diagnosis triplet using the disease triplet includes: obtaining multiple ALS triplets and multiple differential diagnosis triplets based on the disease triplets; and obtaining a differential diagnosis triplet based on the ALS triplets and the differential diagnosis triplets.

[0016] This invention first classifies and screens multiple ALS triples and differential disease triples from disease triples, clearly distinguishing the knowledge units of the core disease and the control disease to avoid category confusion. Then, based on the two types of triples, it extracts symptom entities, compares the same symptoms, constructs association pairs and quantifies the support, and finally integrates them to form differential triples. This accurately establishes a quantitative association between the same symptoms and the two diseases, improving the scientificity and accuracy of differential triples.

[0017] Optionally, obtaining multiple ALS triples and multiple differential disease triples based on the disease triples includes: classifying the disease triples into a first category to obtain multiple sets of similar triples; fusing the sets of similar triples to obtain multiple fused similar triples; and classifying the fused similar triples into a second category to obtain multiple ALS triples and multiple differential disease triples.

[0018] This invention first classifies disease triples by disease entity, associated attribute, and symptom entity, merging similar triples into sets to avoid duplication and redundancy. Then, it merges sets of similar triples to resolve conflicting knowledge and integrate multi-source evidence to improve knowledge consistency and credibility. Finally, it performs a second classification by disease type to accurately distinguish between ALS triples and disease identification triples, thereby improving the accuracy of ALS triples and disease identification triples.

[0019] Optionally, fusing the sets of similar triples to obtain multiple fused triples includes: obtaining a first conflicting triple set and a second conflicting triple set through comparisons between the sets of similar triples; calculating a first comprehensive confidence set for the first conflicting triple set and a second comprehensive confidence set for the second conflicting triple set; calculating confidence reference values ​​for the first comprehensive confidence set and the second comprehensive confidence set based on linear interpolation and a pre-set quantile; and filtering from the first comprehensive confidence set and the second comprehensive confidence set based on the confidence reference values. The system obtains a first preferred set and a second preferred set; calculates a first fusion weight and a second fusion weight based on the first preferred set and the second preferred set; performs weighted fusion on the first preferred set using the first fusion weight to obtain a first confidence level, and performs weighted fusion on the second preferred set using the second fusion weight to obtain a second confidence level; compares the first confidence level and the second confidence level, and performs conflict resolution on the set of similar triples based on the comparison result to obtain an optimized set of similar triples; calculates the fusion confidence level of the optimized set of similar triples, and constructs similar fusion triples using the fusion confidence level.

[0020] This invention first compares sets of similar triples to filter out conflicting sets with mutually exclusive attributes, accurately locating contradictory knowledge; then, it calculates the comprehensive confidence metric to quantify the confidence of each triple, and uses linear interpolation and quantiles to determine reference values, effectively filtering low-confidence data; subsequently, it selects the optimal set based on the reference values, calculates the fusion weight based on the set size, and performs weighted fusion to obtain the overall confidence of the two sets; conflict resolution is achieved through objective numerical comparison; finally, it calculates the fusion confidence of the optimized set and constructs similar fused triples, completely solving the knowledge conflict and redundancy problem of multi-source data and improving the accuracy of similar fused triples.

[0021] Optionally, obtaining the identification triplet based on the ALS triplet and the identification disease triplet includes: extracting ALS symptom entities from each of the ALS triplets; extracting identification disease symptom entities from each of the identification disease triplets; comparing each ALS symptom entity and each identification disease symptom entity one by one to obtain identical symptom entities; constructing triplet association pairs based on the identical symptom entities; and constructing identification triplets based on the triplet association pairs.

[0022] This invention first extracts symptom entities from the ALS triplet and the disease identification triplet respectively, accurately separating the core symptom dimensions of the two types of diseases. Then, it compares and filters out the same symptom entities one by one, locking in the key entry point for disease identification. Next, it constructs triplet association pairs based on the same symptom entities, establishing the corresponding association between symptoms and the two types of diseases. Finally, it integrates the core elements based on the association pairs to construct the identification triplet, realizing the structured presentation of the symptom-disease identification relationship and improving the scientificity and accuracy of the identification triplet.

[0023] Optionally, the triplet association pair includes a first disease triplet and a second disease triplet. Constructing a discriminative triplet based on the triplet association pair includes: extracting a first association attribute, a first overall confidence level, and a first disease entity from the first disease triplet; extracting a second association attribute, a second overall confidence level, and a second disease entity from the second disease triplet; calculating a first support level of the same symptom entity to the first disease entity based on the first association attribute and the first overall confidence level; calculating a second support level of the same symptom entity to the second disease entity based on the second association attribute and the second overall confidence level; and constructing a discriminative triplet based on the first disease entity, the first support level, the same symptom entity, the second support level, and the second disease entity.

[0024] This invention first extracts association attributes, overall confidence scores, and disease entities from the first and second disease triplet pairs of the triplet association pair to accurately capture core identification elements. Then, based on the "support" or "not support" category of the association attributes, the overall confidence score is directly used or negatively calculated to obtain the corresponding support score, quantifying the direction and intensity of the influence of symptoms on the two diseases. Finally, the first disease entity, first support score, same symptom entity, second support score, and second disease entity are integrated to construct the identification triplet, realizing the structured integration of identification information and further improving the scientificity and accuracy of the identification triplet.

[0025] Another aspect of the present invention provides a knowledge graph construction system for ALS (Amyotrophic Lateral Sclerosis), comprising: a processor, an input device, an output device, and a memory, wherein the processor, the input device, the output device, and the memory are interconnected, wherein the memory is used to store a computer program, the computer program including program instructions, and the processor is configured to invoke the program instructions to execute the ALS knowledge graph construction method according to any one of the preceding aspects of the present invention.

[0026] The present invention provides a knowledge graph construction system for ALS, which is compact, stable, highly integrated and simple in structure. It can stably execute the ALS knowledge graph construction method provided in the preceding aspect of the present invention, further improving the overall applicability and practical application capability of the present invention. Attached Figure Description

[0027] Figure 1 This is a flowchart of a method for constructing a knowledge graph of ALS according to an embodiment of the present invention; Figure 2 This is a schematic diagram of a knowledge graph construction system for ALS (Amyotrophic Lateral Sclerosis) according to an embodiment of the present invention. Detailed Implementation

[0028] Specific embodiments of the present invention will now be described in detail. It should be noted that the embodiments described herein are for illustrative purposes only and are not intended to limit the invention. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that these specific details are not necessary to practice the invention. In other instances, well-known circuits, software, or methods have not been specifically described to avoid obscuring the invention.

[0029] Throughout this specification, references to "an embodiment," "an embodiment," "an example," or "an example" mean that a particular feature, structure, or characteristic described in connection with that embodiment or example is included in at least one embodiment of the invention. Therefore, the phrases "in an embodiment," "in an embodiment," "an example," or "an example" appearing in various places throughout the specification do not necessarily refer to the same embodiment or example. Furthermore, specific features, structures, or characteristics can be combined in one or more embodiments or examples in any suitable combination and / or sub-combination. Moreover, those skilled in the art will understand that the illustrations provided herein are for illustrative purposes and are not necessarily drawn to scale.

[0030] Please see Figure 1 In order to solve the problems in the prior art, in an alternative embodiment, such as Figure 1 The method for constructing a knowledge graph for ALS (Amyotrophic Lateral Sclerosis) shown includes the following steps: Step S1: Obtain text data of single symptoms of ALS and disease identification.

[0031] The specific steps involved in obtaining textual data on single symptoms of ALS and identifying the disease include the following: Step S101: Obtain the original text data of ALS and disease identification.

[0032] In this embodiment, multi-source, heterogeneous raw text data is acquired. Data sources include authoritative medical works and treatment guidelines, Chinese and English research literature, structured medical knowledge bases and encyclopedias, and verified authoritative medical and health websites. Authoritative medical works and treatment guidelines, such as textbooks on neurology and the "Chinese Guidelines for the Diagnosis and Treatment of Amyotrophic Lateral Sclerosis," provide standardized and normative knowledge. Chinese and English research literature is retrieved from academic databases such as CNKI and PubMed, using journal articles and dissertations with core keywords such as "ALS," "cervical spondylosis," and "Hirayama disease" to obtain cutting-edge research results and clinical evidence. Structured medical knowledge bases and encyclopedias utilize relevant entries from open graphs such as CMeKG and OMAHA. Verified authoritative medical and health websites publish disease introductions, patient education materials, and case discussions. All data are recorded at the time of collection, including their source and type, and are initially classified and labeled according to the recognized authority of their source (e.g., guidelines, core journals, ordinary web pages). The data scope explicitly covers amyotrophic lateral sclerosis (ALS) and at least eight predefined clinically easily confused differential diseases (such as cervical spondylosis, Hirayama disease, myasthenia gravis, syringomyelia, etc.), where differential diseases are those that are easily confused with ALS.

[0033] Step S102: Standardize the original text data to obtain standardized text data.

[0034] The standardization process for obtaining standardized text data from the original text data includes the following sub-steps: Step S10201: Construct a standardized terminology dictionary, and based on the standardized terminology dictionary, identify and replace synonyms in the original text data to obtain optimized text data.

[0035] In this embodiment, in order to unify the description of symptoms and disease entities in the text, a standardized terminology dictionary specifically for the identification of ALS is first constructed. The dictionary's core coverage is clearly defined, limited to two main categories: disease entities requiring identification and core symptom entities. Terminology from authoritative sources is collected. Disease entity information originates from clinical practice guidelines and authoritative medical textbooks, while symptom entity information comes from standardized symptom terminology sets. Synonyms, abbreviations, common names, and frequently used clinical expressions for both types of entities are simultaneously compiled. Structured dictionary entry is completed, establishing a one-to-one mapping relationship between "standardized terminology" and "synonymous terminology set." For example, in the symptom category, "muscle weakness," "muscle weakness," and "limb weakness" are grouped into the same synonymous terminology set, with "muscle weakness" as the standardized term. Similarly, "muscle twitching," "muscle fasciculations," and "muscle fasciculations" are grouped into the same synonymous terminology set, with "muscle twitching" as the standardized term. In the disease category, "Amyotrophic Lateral Sclerosis (ALS)," "Amyotrophic Lateral Sclerosis," and "ALS" are grouped into the same synonymous terminology set, with "ALS" as the standardized term. Likewise, "cervical spondylosis," "cervical disc herniation," and "cervical radiculopathy" are grouped into the same synonymous terminology set, with "cervical spondylosis" as the standardized term. After the standardized terminology dictionary is fully constructed, a dual-mode matching and replacement engine is built using the Python programming language to perform layered scanning and replacement on the original text data. First, a string exact matching method is used to accurately identify and replace synonyms in the dictionary that have no expression variations. Then, a regular expression matching method is used to perform fuzzy matching and replacement on synonyms that have expression variations or are easily confused. At the same time, the execution order of "exact matching first, then regular expression matching" is strictly followed to avoid the situation where short term matching covers the wrong replacement of long terms. Finally, optimized text data with completely consistent terminology and significantly reduced entity ambiguity is output.

[0036] Step S10202: Perform dependency parsing on the optimized text data to obtain the pronouns and their corresponding antecedents.

[0037] In this embodiment, to resolve referential ambiguities in medical texts and ensure the accuracy of entity extraction, natural language processing tools are used to perform dependency parsing on the optimized text data. First, each sentence is parsed to construct a dependency parsing tree, identifying typical pronouns such as "its," "the disease," "this symptom," and "the latter," as well as nominal pronouns such as "this disease" and "the aforementioned manifestations." Then, by analyzing dependency relationship arcs such as pronouns and noun modifiers in the parsing tree, and combining the relative positions and grammatical roles of pronouns and candidate antecedents (i.e., standardized disease or symptom terminology) in the sentence, based on the principle of nearest neighbor priority and consistency of grammatical roles, the specific disease or symptom entity referred to by each pronoun is determined as its antecedent, thereby completing the accurate parsing of referential relationships.

[0038] Step S10203: Replace the pronoun with the antecedent to obtain standardized text data.

[0039] In this embodiment, each sentence in the optimized text is traversed to accurately locate the identified pronouns (such as "its", "the disease", "this symptom") and their corresponding antecedents (i.e., standardized disease or symptom entity terms). Following the principle of one-to-one precise mapping, each pronoun in the sentence is directly replaced with its corresponding antecedent. For scenarios where the same antecedent corresponds to multiple pronouns, the consistency of the replacement rules is maintained, and finally, standardized text data with clear referential relationships and unambiguous entity descriptions is output.

[0040] Step S103: Perform dependency parsing on the standardized text data to obtain disease entities, symptom entities, associations, and semantic boundaries.

[0041] In this embodiment, natural language processing tools are invoked to perform fine-grained dependency parsing analysis on the standardized text data after referential resolution. By parsing the grammatical structure of the sentences, the core disease entities (such as "ALS") and symptom entities (such as "muscle weakness") defined by the standardized dictionary are first identified and labeled. Secondly, by analyzing the dependency relationship paths connecting the disease and symptoms, predicates or prepositional phrases expressing the clinical association between the two are identified and classified as association relationships (e.g., "manifests as" corresponds to "symptom" relationship, "often accompanied by" corresponds to "accompanied by" relationship, "exclude" corresponds to "exclude" relationship). At the same time, based on the subject-predicate structure, clause conjunctions, and punctuation marks, the internal hierarchical structure of the sentences is analyzed, and the start and end ranges of each phrase or clause expressing a relatively independent medical fact are automatically delineated, which are the semantic boundaries.

[0042] Step S104: Based on the disease entity, the symptom entity, and the relationship, the standardized text data is divided and semantically completed using the semantic boundaries to obtain initial single-symptom text data.

[0043] In this embodiment, complex long sentences are first segmented into several independent short text segments according to semantic boundaries. Then, based on explicit rules, core element verification and semantic completion are performed on each segment: First, core element verification is performed by loading a standardized terminology dictionary and a predefined relational word list. For each semantic segment, string matching and semantic similarity matching based on medical pre-trained word vectors are used to identify whether it contains at least one standardized disease entity and at least one standardized symptom entity. At the same time, verb phrase matching is used to detect whether there are relational words connecting the above two. If all three are present, the core elements of the segment are considered complete; if the verification finds that the segment is missing a disease entity, symptom entity, or relational word, then... Semantic completion is performed according to the following priority rules: The source range for completion prioritizes candidates within the original sentence to which the current segment belongs; if none are found, it backtracks to the adjacent previous sentence. For entity completion, when multiple candidate entities of the same type exist within the range, the priority is given to the candidate entity with the closest contextual distance; if the distances are the same, the candidate entity corresponding to the original text with the higher source coefficient is selected; if the source coefficients are the same, the order of standardized terms defined in the standardized terminology dictionary is prioritized. For relation completion, if a relation is missing, it is completed based on the core predicate connecting entity pairs in dependency parsing; if analysis is not possible, the default relation "related to" is used, which will be distinguished in subsequent confidence calculations. Ultimately, each segmented and completed text segment forms a clear and structurally consistent initial single-symptom text data, with a standardized format of "disease entity - relation - symptom entity".

[0044] Step S105: Obtain the initial source coefficient of the initial single-symptom text data, and concatenate the initial source coefficient with the initial single-symptom text data to obtain single-symptom text data.

[0045] In this embodiment, an authority level mapping table is predefined. International / domestic treatment guidelines are assigned a value of 1, research papers in core journals (SCI / EI / Chinese Core Journals) are assigned a value of 0.8, ordinary journals or dissertations are assigned a value of 0.6, popular science articles from authoritative medical institution websites are assigned a value of 0.4, and other online information is assigned a value of 0.2. By tracing the source identifiers carried in the data pipeline, the corresponding numerical coefficients, i.e., the initial source coefficients, are automatically retrieved from this mapping table by querying the data. Finally, a structured splicing method is used to concatenate the initial source coefficients with the initial single-symptom text data to form a complete single-symptom text data containing evidence strength labels.

[0046] Step S2: Construct an information extraction model based on long short-term memory network and deep hidden Markov model.

[0047] The construction of the information extraction model based on Long Short-Term Memory Networks and Deep Hidden Markov Models specifically includes the following sub-steps: Step S201: Set the loss function based on conditional random field loss and binary cross-entropy loss.

[0048] In this embodiment, to achieve synergistic optimization of entity recognition and attribute classification tasks in medical texts using a joint extraction sub-model, a multi-task joint loss function is constructed, with Conditional Random Field (CRF) loss and Binary Cross-Entropy (BCE) loss as its core. The CRF loss constrains the sequence labeling process for entity recognition, accurately capturing the boundary features and category associations of disease and symptom entities by modeling the transition probabilities between adjacent labels, thus addressing the entity fragmentation and category confusion issues that easily occur in single-label prediction. The BCE loss targets the "supported / unsupported" attribute classification task. Considering the potential imbalance between the two classes of samples in medical texts, a higher loss weight is assigned to "unsupported" samples to improve the model's ability to discriminate minority attributes. Finally, the two losses are integrated through weighted fusion to form a final loss function that balances optimization for both tasks.

[0049] Step S202: Construct a joint extraction sub-model based on the loss function using the Long Short-Term Memory network.

[0050] In this embodiment, in order to realize the identification of disease entities and symptom entities in medical text and the simultaneous determination of their association attributes (support / non-support), a joint extraction sub-model based on Long Short-Term Memory Network (LSTM) is constructed.

[0051] The joint extraction sub-model employs a multi-task learning architecture with a shared encoding layer and dual-task output heads: A word embedding layer is set at the bottom layer to map words in standardized text data into low-dimensional dense vectors; subsequently, a bidirectional LSTM encoding layer is connected to capture the contextual semantic features and long-distance dependencies of the text sequence (including the strength of tone features of the association relationship). This encoding layer is shared for both tasks to reduce parameter redundancy and improve feature reuse efficiency. Two output heads are set in parallel above the encoding layer: the first is an entity recognition output head, which connects to a CRF layer based on the encoded features, and outputs the sequence labeling results for disease entities and symptom entities by modeling the transition probabilities between labels; the second is an attribute classification output head.

[0052] The classification output head performs the following operations: First, it locates and extracts the contextual semantic feature vectors corresponding to the disease entities and symptom entities identified by the entity recognition head from the feature sequence output by the encoding layer (for example, taking the average of all labeled vectors within each entity span as the entity's representation); next, it fuses the extracted disease entity feature vectors and symptom entity feature vectors (e.g., concatenation, addition, or interactive attention calculation) to form a joint representation vector, which comprehensively encodes the semantics of the disease and symptom and their association patterns in the current context; finally, it inputs this joint representation vector into a fully connected layer, which outputs binary classification probability values ​​of "support" and "not support" after passing through the Softmax activation function. The model learns the mapping rules from entity semantics and contextual association patterns to "support / not support" attributes by training on massive labeled data (labeled disease, symptom entities, and their associated attributes).

[0053] Step S203: Construct a confidence calculation model based on the deep hidden Markov model.

[0054] In this embodiment, the Deep Hidden Markov Model (DHMM) is a probabilistic graphical model that integrates deep learning and Hidden Markov Models. It introduces a multi-layered neural network as the generative model for observation states, building upon the traditional Hidden Markov Model. This model automatically extracts high-order abstract features from raw input data (such as text and sequential signals) using deep neural networks and inputs these features as observation states into the Hidden Markov chain for temporal modeling. This significantly improves the modeling ability for complex, high-dimensional, and nonlinear observation data while preserving the state transition patterns of the sequence. DHMM not only captures the hidden state transition patterns behind the observation sequence but also enhances the model's fit to the observation distribution through the representation learning capabilities of neural networks. It is suitable for scenarios with temporal dependencies and hidden structural features, such as speech recognition, natural language processing, and biological sequence analysis, and demonstrates strong expressive power in tasks that require simultaneously modeling the dynamics of sequences and the deep semantics of data.

[0055] To quantify the reliability of the extraction results from the joint extraction sub-model, a confidence calculation model is constructed based on a deep hidden Markov model. This model employs a multi-layered hidden state design to evaluate the reliability of the extraction results at different granularities: the bottom-level hidden states model the matching quality of local semantic fragments (such as entity words themselves); the middle-level hidden states model the logical consistency between entities and attributes; and the top-level hidden states model the global compatibility of the extraction results with external prior medical knowledge. The state evolution between each layer of hidden states is characterized by a transition probability matrix.

[0056] The observation layer of the model takes as input multidimensional features extracted from the sub-models, including but not limited to: label encoding of entity sequences, probability distribution of attribute classification, and contextual semantic vector representation of text. These features are standardized and concatenated to form a unified observation sequence.

[0057] The model is trained using a self-supervised learning approach, eliminating the need for manual confidence labeling. Using the output features of the aforementioned joint extraction sub-model on the training set as the observation sequence, the transition probability matrix and emission probability matrix of the model are iteratively optimized through a forward-backward algorithm (Baum-Welch algorithm) to maximize the model's likelihood function, thereby learning the mapping pattern from observed features to implicit reliability states.

[0058] During the inference phase, the initial output features of the joint extraction sub-model on the single-symptom text are input into the trained confidence calculation model. The optimal hidden state sequence is inferred using the Viterbi algorithm, and the generation probability of this sequence is calculated. This generation probability is normalized to obtain a value between 0 and 1, which serves as the initial confidence score for this extraction. This confidence score comprehensively reflects the basic credibility of the extraction results based on sequence consistency, semantic rationality, and compatibility with medical knowledge.

[0059] Step S204: Construct an information extraction model based on the joint extraction sub-model and the confidence calculation model.

[0060] In this embodiment, the joint extraction sub-model and the confidence calculation model are integrated in series to form a complete information extraction model, realizing the integrated output of entity and attribute extraction and extraction confidence quantification. The specific integration logic is as follows: First, the joint extraction sub-model receives single symptom text data and outputs structured intermediate results, including disease entities, symptom entities, and binary classification probability values ​​of "support / not support" (e.g., P(support) = 0.95, P(not support) = 0.05). Then, the intermediate results are used as observation features and input into the confidence calculation model to output the initial confidence. Finally, the information extraction model simultaneously outputs four core results: disease entity, symptom entity, related attribute determination (taking the larger of the binary classification probability values, i.e., if P(support) ≥ 0.5, it is determined as "support", otherwise it is "not support"), and confidence (initial confidence multiplied by the binary classification probability value).

[0061] Step S3: Use the information extraction model to extract information from the single symptom text data to obtain multiple disease triples.

[0062] In this embodiment, to extract standardized medical knowledge units from structured single-symptom text data, each single-symptom text data is input into the information extraction model one by one: The model first outputs the disease entity, symptom entity, "support / not support" binary classification probability value and association attribute judgment result through the joint extraction sub-model, and then outputs the initial confidence score through the confidence score calculation model. Subsequently, the final confidence score is calculated according to the rules: the binary classification probability value corresponding to the association attribute judgment result is taken (e.g., if the judgment is "support", then P(support) is taken), and multiplied with the initial confidence score to obtain the confidence score of the comprehensive representation attribute judgment confidence and extraction confidence (e.g., initial confidence score 0.98 × P(support) 0.95 = 0.931). Finally, the disease entity, the association attribute with "confidence score + source coefficient", and the symptom entity are integrated to form a standardized disease triplet, with a unified format of (disease entity, association attribute (support / not support) {confidence score: X, source coefficient: Y}, symptom entity). For example: (ALS, associated attributes (support) {confidence: 0.931, source coefficient: 1.0}, muscle twitching).

[0063] The training set (covering different diseases, symptoms, and original association types) is formed by stratified sampling from single-symptom text data. The training set is manually annotated by medical experts. The annotation content includes three parts: first, the text start and end index of the disease entity and the "disease entity" type label; second, the text start and end index of the symptom entity and the "symptom entity" type label; and third, the corresponding association attribute label ("support" or "do not support") of the disease-symptom entity pair. The annotation is based on the semantic tendency of the original association in the sample (e.g., "manifests as" or "often accompanied by" is labeled "support", "no" or "excluded" is labeled "do not support").

[0064] Disease triples satisfy the following formula: in, For confidence level, As the initial confidence level, This represents the binary classification probability.

[0065] Step S4: Construct a discrimination triplet using the disease triplet.

[0066] The construction of a differential diagnosis triple using the disease triple specifically includes the following sub-steps: Step S401: Based on the disease triplet, obtain multiple ALS triplets and multiple disease identification triplets.

[0067] Specifically, obtaining multiple ALS triads and multiple differential disease triads based on the aforementioned disease triads includes the following sub-steps: Step S40101: Perform a first classification on the disease triplet to obtain multiple sets of triplets of the same type.

[0068] In this embodiment, the disease entity, associated attribute (support / not support), and symptom entity in each disease triple are used as classification keys. Through exact matching, all disease triples that are identical in all three key values ​​are grouped into the same set, which is the set of similar triples. After classification, each set represents all evidence obtained from different data sources regarding whether a specific disease supports or excludes a specific symptom. The triples within a set differ only in the confidence level, which characterizes the strength of the evidence, and the source coefficient, which characterizes the authority of the source.

[0069] Step S40102: Merge the set of similar triples to obtain multiple similar fused triples.

[0070] The specific steps for fusing the set of similar triples to obtain multiple fused triples include the following: Step S4010201: The first conflicting triplet set and the second conflicting triplet set are obtained by comparing the sets of triplets of the same type.

[0071] In this embodiment, firstly, disease entities and symptom entities are used as unique matching identifiers to group all sets of similar triples. That is, all sets of similar triples within the same group have completely identical disease entities and symptom entities, differing only in the association attribute (support / not support) or parameters related to the strength of evidence. Then, each group is traversed, and the association attribute type of the sets of similar triples within the group is checked: if there are sets of similar triples with the same association attribute of "support" and sets of similar triples with the same association attribute of "not support" within the same group, then two sets of mutually exclusive opposing evidence are formed. Based on the set, the set with the association attribute "support" is marked as the first conflicting triple set, and the set with the association attribute "not support" is marked as the second conflicting triple set. If all triple sets of the same type in the same group have completely identical association attributes (both are "support" or both are "not support"), then it is determined that there is no knowledge conflict in the group, and there is no need to generate a conflict set. The process directly enters the subsequent fusion process. By first grouping by disease-symptom label and then judging the mutual exclusivity of attributes within the group, the conflicting sets with truly opposing evidence are accurately identified, avoiding misjudging the attribute differences of different disease-symptom pairs as conflicts.

[0072] Step S4010202: Calculate the first comprehensive confidence set of the first conflicting triple set and the second comprehensive confidence set of the second conflicting triple set, respectively.

[0073] In this embodiment, to quantify and compare the evidentiary credibility of the two conflicting triple sets, a comprehensive confidence calculation operation is performed on the first conflicting triple set and the second conflicting triple set, respectively. That is, the product of the confidence and the source coefficient of each triple in the first comprehensive confidence set of the first conflicting triple set is calculated to obtain the comprehensive confidence (i.e., individual comprehensive confidence) of each triple. Finally, the first comprehensive confidence set (from the first conflicting triple set) and the second comprehensive confidence set (from the second conflicting triple set) are formed.

[0074] Step S4010203: Calculate the confidence reference values ​​of the first comprehensive confidence set and the second comprehensive confidence set based on linear interpolation and pre-set quantiles, respectively.

[0075] In this embodiment, to objectively measure the overall credibility level of the first and second conflicting triplet sets and avoid interference from the composite confidence values ​​of individual extreme individuals within the sets on subsequent screening decisions, credibility reference values ​​are calculated for both sets of composite confidence. First, the composite confidence values ​​of all individuals within the first composite confidence set are sorted in ascending order. Using a pre-set quantile parameter (preferably the 50th quantile, i.e., the median, which has strong resistance to extreme values), a value reflecting the overall central tendency of the set's confidence is calculated using linear interpolation and defined as the first credibility reference value. The second composite confidence set is then processed using the same sorting method, quantile parameter, and calculation method to obtain the second credibility reference value. This credibility reference value is not the confidence value of a single individual triplet, but rather a quantification of the overall credibility of the entire set of conflicting evidence. It provides a unified, objective, and robust judgment benchmark for subsequent screening of high-credibility triplets, avoiding deviations from the quality of true evidence due to individual outliers.

[0076] Linear interpolation is a simple numerical estimation method. When the quantile position i calculated by the formula i=n×p (where n is the number of confidence levels built into the set and p is the preset quantile) is not an integer, the confidence values ​​corresponding to the two integer positions adjacent to i are taken. Then, based on the proportion of the integer part and the decimal part of i, the two confidence values ​​are weighted and calculated to obtain a continuous and accurate confidence reference value, thereby avoiding the calculation deviation caused by forced rounding.

[0077] Step S4010204: Based on the confidence reference value, filter the first comprehensive confidence set and the second comprehensive confidence set to obtain a first preferred set and a second preferred set.

[0078] In this embodiment, it is first clarified that the comprehensive confidence level of each individual in the first comprehensive confidence level set corresponds one-to-one with a triple in the first conflicting triple set. A screening rule is set that the comprehensive confidence level of each individual in the first and second comprehensive confidence levels sets is greater than or equal to the corresponding confidence reference value (first confidence reference value and second confidence reference value). The first comprehensive confidence level set is traversed, and the comprehensive confidence levels of individuals that meet the rule are retained, and their corresponding triples are retained simultaneously. These triples are then aggregated to obtain the first preferred set. Using the same screening rule and corresponding logic, the second comprehensive confidence level set is traversed, and the confidence levels that meet the rule and their corresponding triples in the second conflicting triple set are retained simultaneously. Finally, the second preferred set is obtained, which effectively filters out triples with low confidence levels in each set and retains high-quality knowledge units with greater evidentiary value.

[0079] Step S4010205: Calculate the first fusion weight and the second fusion weight based on the first preferred set and the second preferred set.

[0080] In this embodiment, the number of triples contained in the first preferred set and the second preferred set are counted respectively, and then the two counts are added together to obtain the total number of triples in the two preferred sets. Then, the number of triples in the first preferred set is divided by the total number to obtain the first fusion weight, and the number of triples in the second preferred set is divided by the total number to obtain the second fusion weight. The sum of the two fusion weights is equal to 1. The fusion weight is directly determined by the number of valid triples in the two preferred sets. The more triples there are, the more sufficient the corresponding conflict evidence is, and the higher its proportion in the final fusion.

[0081] Step S4010206: Use the first fusion weight to perform weighted fusion on the first preferred set to obtain a first confidence level, and use the second fusion weight to perform weighted fusion on the second preferred set to obtain a second confidence level.

[0082] In this embodiment, the individual comprehensive confidence score of each triple in the first preferred set is extracted, and the first fusion weight is used to calculate the weighted average of these individual comprehensive confidence scores. The result is the first confidence score. Then, the same weighted average method is used to extract the individual comprehensive confidence scores of all triples in the second preferred set, and the second fusion weight is used to complete the calculation to obtain the second confidence score.

[0083] Step S4010207: Compare the first confidence level and the second confidence level, and resolve the conflict of the set of similar triples according to the comparison result to obtain an optimized set of similar triples.

[0084] In this embodiment, the numerical values ​​of the first and second credibility are directly compared. If the first credibility is greater than the second credibility, the evidence of the associated attribute corresponding to the first preferred set is determined to be more reliable. The triples corresponding to this attribute are retained, and the triples with opposing attributes in the second preferred set are removed. If the second credibility is greater than the first credibility, the reverse retention and removal operation is performed. In extreme scenarios, if the two values ​​are equal, the average source coefficient (the average source coefficient of all triples in the set) of the two preferred sets can be used for auxiliary determination. Triples corresponding to the set with the higher source coefficient are retained first. Finally, all the triples retained after determination are integrated to obtain an optimized set of similar triples with unified attributes and qualified credibility. Conflict resolution is achieved through objective comparison of quantified credibility, completely avoiding the bias caused by subjective decision-making, and ensuring that the final output set of similar triples has consistent and reliable medical knowledge attributes. Conflict resolution refers to deleting conflicting triples with lower credibility.

[0085] Step S4010208: Calculate the fusion confidence of the optimized set of similar triples, and construct similar fusion triples using the fusion confidence.

[0086] In this embodiment, each triplet in the optimized set of similar triplets is first traversed, and its individual confidence score is multiplied by the corresponding source coefficient to obtain the individual comprehensive confidence score of each triplet. Then, the mean of all individual comprehensive confidence scores is calculated to obtain the overall fusion confidence score of the set. Next, the unified disease entity, associated attribute, and symptom entity in the optimized set of similar triplets are extracted, and the calculated overall fusion confidence score is set as the attribute value of the associated attribute. Finally, multiple similar triplets are integrated into a unique similar fusion triplet to achieve deduplication and fusion.

[0087] Step S40103: Perform a second classification on the aforementioned fusion triplets to obtain multiple ALS triplets and multiple differential disease triplets.

[0088] In this embodiment, to clearly distinguish between knowledge units related to ALS and control diseases, a second classification (by disease type) operation is performed on all similar fusion triples. First, all similar fusion triples are traversed, and the disease entity information contained in each triple is extracted. Using whether the disease entity is ALS as the core classification criterion, similar fusion triples whose disease entities explicitly point to ALS are grouped into multiple ALS triples. Similar fusion triples whose disease entities are other diseases requiring differentiation from ALS are grouped into multiple disease differentiation triples.

[0089] Step S402: Obtain the identification triplet based on the ALS triplet and the identification disease triplet.

[0090] The specific steps for obtaining the differential diagnosis triad based on the ALS triad and the differential diagnosis triad include the following: Step S40201: Extract ALS symptom entities from each of the ALS triples.

[0091] In this embodiment, each ALS triple is traversed sequentially, and the symptom entity content contained therein is extracted to obtain the ALS symptom entity. After the symptom entity extraction of all ALS triples is completed, the corresponding ALS symptom entity set can be obtained.

[0092] Step S40202: Extract the disease symptom entities from each of the disease identification triples.

[0093] In this embodiment, each disease identification triple is traversed sequentially, and the symptom entity content contained therein is extracted to obtain the disease identification symptom entity. After the symptom entity extraction of all disease identification triples is completed, the disease identification symptom entity set corresponding to the disease identification can be obtained.

[0094] Step S40203: Compare each of the ALS symptom entities with each of the identified disease symptom entities one by one to obtain the same symptom entities.

[0095] In this embodiment, all extracted ALS symptom entities are first summarized to form an ALS symptom entity set. Then, all disease identification symptom entities extracted from disease identification triples are summarized to form a disease identification symptom entity set. Subsequently, each symptom entity in the ALS symptom entity set is matched and compared one by one with each symptom entity in the disease identification symptom entity set. Finally, symptom entities that are completely identical in the two sets are selected and summarized to obtain a set of identical symptom entities.

[0096] Step S40204: Construct triplet association pairs based on the entities with the same symptoms.

[0097] In this embodiment, all identical symptom entities (including recurring symptom entities) are traversed. For each identical symptom entity, its two corresponding source triples are traced. The triple related to ALS and containing the symptom entity is defined as the first disease triple, and the other triple related to the identified disease and containing the symptom entity is defined as the second disease triple. These two triples are then paired and combined to form a set of triple association pairs corresponding to the identical symptom entity. This process is performed for each identical symptom entity, ultimately resulting in multiple sets of triple association pairs.

[0098] Step S40205: Construct a discrimination triplet based on the triplet association pair.

[0099] The construction of the identification triplet based on the triplet association pairs specifically includes the following sub-steps: Step S4020501: Extract the first associated attribute, the first comprehensive confidence level, and the first disease entity from the first disease triplet.

[0100] In this embodiment, the first disease triplet in each triplet association pair is traversed sequentially, and the three types of core information—first association attribute, first comprehensive confidence, and first disease entity—are extracted from it. After extracting the elements of all first disease triplets, the first core information set corresponding to each triplet association pair can be obtained.

[0101] Step S4020502: Extract the second association attribute, the second comprehensive confidence level, and the second disease entity from the second disease triplet.

[0102] In this embodiment, the second disease triples in each triplet association pair are traversed sequentially, and the three types of core information—second association attribute, second comprehensive confidence level, and second disease entity—are extracted from them respectively. After the element extraction of all second disease triples is completed, the set of second core information corresponding to each triplet association pair can be obtained.

[0103] Step S4020503: Calculate the first support of the same symptom entity to the first disease entity based on the first association attribute and the first comprehensive confidence level.

[0104] In this embodiment, for each pair of triplet associations, a classification calculation is performed on the corresponding first association attribute and first comprehensive confidence score. If the first association attribute is "supportive," the corresponding first comprehensive confidence score is directly used as the first support score; if the first association attribute is "unsupportive," the corresponding first comprehensive confidence score is negative, and the resulting value is the first support score. The sign and magnitude of the first support score directly represent the degree of numerical support or weakening of the same symptom entity for the first disease entity.

[0105] Step S4020504: Calculate the second support of the same symptom entity to the second disease entity based on the second association attribute and the second comprehensive confidence level.

[0106] In this embodiment, for each pair of triplet associations, a classification calculation is performed on the corresponding second association attribute and second comprehensive confidence level. If the second association attribute is "supportive," the corresponding second comprehensive confidence level is directly used as the second support level; if the second association attribute is "unsupportive," the corresponding second comprehensive confidence level is negative, and the resulting value is the second support level. The sign and magnitude of the second support level directly represent the degree of numerical support or weakening of the same symptom entity for the second disease entity.

[0107] Step S4020505: Construct a discrimination triplet based on the first disease entity, the first support, the same symptom entity, the second support, and the second disease entity.

[0108] In this embodiment, the first disease entity, first support, same symptom entity, second support and second disease entity corresponding to each group are traversed. These five core elements are structurally integrated, and the same symptom entity is used as the link to connect the two disease entities and their corresponding support indicators, and finally construct a discrimination triple with clear discrimination value.

[0109] Step S5: Construct an ALS knowledge graph based on the identification triplet.

[0110] In this embodiment, the core structure of the ALS knowledge graph is first defined based on Neo4j's graph data model. The first disease entity (ALS) and the second disease entities (various differential diagnoses) are abstracted as disease type nodes, and basic attributes such as disease name and disease classification are configured for the nodes. Entities with the same symptom are abstracted as symptom type nodes, and core attributes such as symptom name and symptom category are configured. At the same time, the relationship between nodes is defined. The relationship type between symptom nodes and disease nodes is symptom association, and a support value (first support / second support) is added to this relationship. Then, all differential triples are converted into node data and relationship data formats that Neo4j can recognize. Through Neo4j's Cypher query statement, disease nodes, symptom nodes, and attributed relationships are imported into the Neo4j database in batches. Finally, node creation, relationship association, and attribute mounting are completed in Neo4j, forming an ALS knowledge graph with ALS and differential diagnoses as the core, symptoms as the link, and support as the quantitative feature.

[0111] The query uses the standardized disease name as the core query keyword, matches the corresponding "symptom" node, and simultaneously associates the "symptom association" between the symptom node and all "disease" nodes in the atlas (including ALS and various differential diagnoses). Then, it extracts the support value associated with each association. For each disease entity, it summarizes the support values ​​of all associations between the symptom node and the disease node (where "no support" is quantified as negative) and calculates the algebraic sum to obtain the net support of the standardized disease for that disease entity. This algebraic sum integrates the net effect of both positive and negative evidence, objectively reflecting the overall tendency and strength of the symptom's support or weakening of the disease. Finally, the query results interface outputs the standardized disease name, all corresponding disease entity names, and the comprehensive support value for each disease entity, clearly presenting the quantitative support or weakening degree of a single standardized disease for different diseases, providing direct quantitative reference for clinical differentiation.

[0112] like Figure 2 As shown, in another aspect, the present invention also provides a knowledge graph construction system for ALS, including: a processor, an input device, an output device, and a memory, wherein the processor, the input device, the output device, and the memory are interconnected, wherein the memory is used to store a computer program, the computer program includes program instructions, and the processor is configured to call the program instructions to execute the relevant steps of a relevant embodiment of the ALS knowledge graph construction method of the present invention.

[0113] This invention provides a knowledge graph construction system for ALS (Amyotrophic Lateral Sclerosis). The functional components can be integrated into a single processing unit, or each component can exist independently, or two or more components can be integrated into one unit. The integrated components can be implemented in hardware or as software functions.

[0114] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A method for constructing a knowledge graph for ALS (Amyotrophic Lateral Sclerosis), characterized in that, The method includes: Acquire textual data of single symptoms of ALS and its identification; An information extraction model is constructed based on long short-term memory networks and deep hidden Markov models; The information extraction model is used to extract information from the single-symptom text data to obtain multiple disease triples; Construct a differential diagnosis triplet using the disease triplet; A knowledge graph of ALS is constructed based on the aforementioned identification triplet.

2. The method for constructing a knowledge graph for ALS according to claim 1, characterized in that, The single-symptom text data for obtaining ALS and identifying the disease includes: Obtain raw text data on ALS and disease identification; The original text data is standardized to obtain standardized text data; Dependency parsing is performed on the standardized text data to obtain disease entities, symptom entities, relationships, and semantic boundaries. Based on the disease entity, the symptom entity, and the relationship, the standardized text data is divided and semantically completed using the semantic boundaries to obtain initial single-symptom text data. Obtain the initial source coefficient of the initial single-symptom text data, and concatenate the initial source coefficient with the initial single-symptom text data to obtain the single-symptom text data.

3. The method for constructing a knowledge graph for ALS according to claim 2, characterized in that, The standardization process for obtaining standardized text data from the original text data includes: A standardized terminology dictionary is constructed, and based on the standardized terminology dictionary, synonyms are identified and replaced in the original text data to obtain optimized text data; Dependency parsing is performed on the optimized text data to obtain pronouns and their corresponding antecedents. Standardized text data is obtained by replacing the pronoun with the antecedent.

4. The method for constructing a knowledge graph for ALS according to claim 1, characterized in that, The information extraction model constructed based on long short-term memory networks and deep hidden Markov models includes: The loss function is set based on conditional random field loss and binary cross-entropy loss; A joint extraction sub-model is constructed based on the loss function of the Long Short-Term Memory network. Construct a confidence calculation model based on a deep hidden Markov model; An information extraction model is constructed based on the joint extraction sub-model and the confidence calculation model.

5. The method for constructing a knowledge graph for ALS according to claim 1, characterized in that, The construction of a differential triad using the disease triads includes: Based on the disease triplet, multiple ALS triplets and multiple differential disease triplets were obtained; The differential diagnosis triad was obtained based on the ALS triad and the differential diagnosis triad.

6. The method for constructing a knowledge graph for ALS according to claim 5, characterized in that, The process of obtaining multiple ALS triads and multiple differential disease triads based on the disease triads includes: The disease triplet is classified into multiple sets of similar triplets by performing a first classification. The set of triples of the same type is fused to obtain multiple fused triples of the same type; The aforementioned fusion triplets were further classified into multiple ALS triplets and multiple differential disease triplets.

7. The method for constructing a knowledge graph for ALS according to claim 6, characterized in that, The process of fusing the set of similar triples to obtain multiple fused triples of the same type includes: The first conflicting triplet set and the second conflicting triplet set are obtained by comparing the sets of triplets of the same kind. Calculate the first comprehensive confidence set of the first conflicting triple set and the second comprehensive confidence set of the second conflicting triple set, respectively; The confidence reference values ​​for the first comprehensive confidence set and the second comprehensive confidence set are calculated based on linear interpolation and pre-set quantiles, respectively. Based on the confidence reference value, a first preferred set and a second preferred set are obtained by filtering from the first comprehensive confidence set and the second comprehensive confidence set, respectively; Calculate the first fusion weight and the second fusion weight based on the first preferred set and the second preferred set; The first set of preferred products is weighted and fused using the first fusion weight to obtain a first confidence level, and the second set of preferred products is weighted and fused using the second fusion weight to obtain a second confidence level. The first confidence level and the second confidence level are compared, and the conflict resolution of the set of similar triples is performed based on the comparison result to obtain an optimized set of similar triples. Calculate the fusion confidence of the optimized set of similar triples, and use the fusion confidence to construct similar fusion triples.

8. The method for constructing a knowledge graph for ALS according to claim 5, characterized in that, The differential diagnosis triad obtained based on the ALS triad and the differential disease triad includes: Extract ALS symptom entities from each of the aforementioned ALS triples; Extract disease symptom entities from each of the disease identification triples; Each of the ALS symptom entities and each of the differential disease symptom entities are compared one by one to obtain the same symptom entities; Construct triplet association pairs based on the entities with the same symptoms; Based on the triplet association pairs, construct the discriminative triplet.

9. The method for constructing a knowledge graph for ALS according to claim 8, characterized in that, The triplet association pairs include a first disease triplet and a second disease triplet, and the construction of discriminative triplets based on the triplet association pairs includes: Extract the first associated attribute, the first comprehensive confidence level, and the first disease entity from the first disease triple; Extract the second association attribute, the second comprehensive confidence level, and the second disease entity from the second disease triple; Based on the first association attribute, the first support degree of the same symptom entity to the first disease entity is calculated using the first comprehensive confidence degree; Calculate the second support of the same symptom entity to the second disease entity based on the second association attribute and the second comprehensive confidence score. Construct a discrimination triple based on the first disease entity, the first support, the same symptom entity, the second support, and the second disease entity.

10. A knowledge graph construction system for ALS (Amyotrophic Lateral Sclerosis), characterized in that, include: The system includes a processor, an input device, an output device, and a memory, all interconnected, wherein the memory stores a computer program comprising program instructions, and the processor is configured to invoke the program instructions to execute a method for constructing an ALS knowledge graph as described in any one of claims 1 to 9.

Images