Anchors matching-based medical data analysis and processing system

Abstract

The application discloses a kind of medical data analysis and processing systems based on anchor point matching, it is related to big data acquisition technical field, system includes: medical database, patient data entry unit, big data keyword extraction unit and retrieval analysis unit and screening unit;The medical database is established by big data acquisition, determines the hospital center anchor point of the hospital corresponding to each data;The patient data entry unit is used to provide patient to enter medical demand description text;The big data keyword extraction unit is used to extract medical demand big data keyword from medical demand description text;The retrieval analysis unit is used to obtain the anchor point position of the object vector;The screening unit is used to calculate the hospital center anchor point closest to the center anchor point position, and the hospital corresponding to the hospital center anchor point is used as the best hospital of the patient.The application significantly improves the automation, intelligentization and practicality level of medical service acquisition.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of big data collection, and particularly relates to a medical data analysis and processing system based on anchor point matching. BACKGROUND

[0002] In the current medical service system, the intelligentization and self-service of the patient's medical path gradually become an important direction of the development of medical information technology. With the uneven distribution of medical resources between cities and regions, the diversification of medical needs and the rising demand of patients for personalized services, the traditional manual guidance, manual registration and static information retrieval method cannot meet the actual needs of the majority of patients for medical service efficiency, accuracy and accessibility. Especially in the context of multi-modal information mixing and strong heterogeneity of patient medical background, how to realize the user-friendly medical self-service system, accurately understand the natural language demand input by the user, and complete the intelligent matching and recommendation based on structured medical big data, has become a key problem to be solved.

[0003] Currently, there are many researches and systems trying to explore in medical service recommendation, semantic analysis and hospital navigation. The existing technical solutions can be roughly divided into the following categories: the first category is a medical information retrieval system based on keyword matching. This kind of system usually adopts simple string matching, Boolean logic operation or TF-IDF model, and realizes preliminary screening recommendation by mapping the keywords input by the patient with the static tags or service items in the hospital database. For example, some registration platforms provide the function of querying according to the screening conditions such as department, disease, hospital level, etc. Representative platforms such as 'Weiyi' and 'Good Doctor Online' also integrate the condition screening module, but are basically limited to menu selection or keyword search, and lack the ability to understand the implicit semantics in natural language. When facing fuzzy input, non-standard expression or cross-category demand description, this kind of system often cannot establish the correct service matching path, resulting in a lack of personalization or serious deviation in the recommended results. The second category is a medical semantic recognition system introducing natural language processing (NLP) model. With the wide application of deep learning in language models, some systems have begun to try to use pre-trained language models such as BERT, ERNIE, BioBERT, etc. to perform semantic representation on the medical text input by the user, so as to extract disease, examination, treatment, etc. for structured expression. Some researches such as 'Medical Consultation Dialogue System Based on BERT' have been able to convert the consultation content into a structured knowledge graph entity. But most of this kind of system stays at the experimental research level, lacks the whole process processing ability of heterogeneous medical institutions and nationwide medical data, and especially relies on artificial rules in the multi-dimensional matching strategy between entities and hospital resources, which is difficult to adapt to the matching complexity of high-dimensional semantic space. The third category is a hospital recommendation platform based on recommendation system technology. These systems mostly draw lessons from collaborative filtering, matrix decomposition and other algorithms in e-commerce or content recommendation, model the patient's historical access behavior and hospital service records, infer the medical path of similar groups of people, and realize the type recommendation of 'people who have seen this disease have also seen this hospital'. Representative researches include medical service recommendation based on graph neural network and collaborative recommendation based on knowledge graph. But this kind of system generally relies on large-scale behavior data and user preference history, and for the first-time users of the platform or the first-time medical users, the recommendation model cold start problem is serious. In addition, its model explanation is poor, the structure matching precision is low, and especially lacks real-time understanding and structured ability of the current input semantic content of the user, which is not suitable for high-risk and high-precision scenarios such as medical guidance. SUMMARY

[0004] To solve the above technical problems, a medical data analysis and processing system based on anchor point matching is provided, which comprises a medical database, a patient data input unit, a big data keyword extraction unit, a retrieval analysis unit and a screening unit.

[0005] To achieve the above purposes, the technical scheme adopted by the present application is as follows:

[0006] A medical data analysis and processing system based on anchor point matching, the system comprises: a medical database, a patient data input unit, a big data keyword extraction unit, a retrieval analysis unit and a screening unit; the medical database is established by big data collection, wherein each data corresponds to a hospital; each data includes a plurality of big data keywords with different categories, and the medical demand big data keywords are stored in the form of sparse inner product approximation table; according to the sparse inner product approximation table corresponding to all medical demand big data keywords of each data, the hospital center anchor point of the hospital corresponding to each data is determined; the patient data input unit is used for providing the patient to input the medical demand description text; the big data keyword extraction unit is used for extracting the medical demand big data keywords from the medical demand description text; the retrieval analysis unit is used for grouping all medical demand big data keywords into a retrieval entity, extracting features from the retrieval entity, representing the features as vectors according to the vector space model, obtaining the object vector set; determining the number of sparse inner product approximation tables, and constructing the sparse inner product approximation table according to the sparse inner product approximation function family; mapping each object vector in the object vector set through the sparse inner product approximation table to obtain the anchor point position of the object vector; the screening unit is used for calculating the center anchor point position common to all object vectors in the retrieval entity, and then calculating the hospital center anchor point closest to the center anchor point position, and taking the hospital corresponding to the hospital center anchor point as the best hospital for the patient.

[0007] Further, the big data keyword categories at least include: location, department name, hospital level, treatment period, examination item, symptom description, treatment method, expert qualification and medical insurance category; the primary key in each data in the medical database is the hospital number of the hospital.

[0008] Further, after the big data keyword extraction unit extracts the medical demand big data keywords from the medical demand description text, it calculates the confidence of the medical demand big data keywords to determine whether the medical demand big data keywords can be classified into a specific big data keyword category.

[0009] Further, the number of sparse inner product approximation tables M satisfies the following constraint relationship:

[0010]

[0011] Wherein, δ is the tolerance of the missing matching rate; P is the number of object vectors in the object vector set; P1 is the threshold of the number of big data keywords.

[0012] Further, the confidence is:

[0013]

[0014] Wherein, Γ c (w i ) is the confidence score of the i-th keyword w i belonging to the big data keyword category c; is the j-th standard reference keyword in the big data keyword category c; is the embedding space semantic similarity between the keyword w i and is the document frequency of the standard reference keyword in the medical database; n is the number of standard reference keywords in the big data keyword category.

[0015] Further, the sparse inner product approximation function family is:

[0016]

[0017] Wherein, is the anchor point in the m-th sparse inner product approximation table for the i-th object vector f i ; T is the number of hash bits of the sparse inner product approximation table; f i,k is the value of the i-th object vector in the k-th dimension; is the random projection weight of the t-th hash function corresponding to the k-th dimension in the m-th sparse inner product approximation table; ρ i,k is the normalized confidence deviation of the i-th object vector f i in the k-th dimension; is the index value of the hash anchor point generated by bit splicing; d is the dimension of the object vector.

[0018] Further, the center anchor position is:

[0019]

[0020] Wherein, z={z1,z2,…,z M ​} is the central anchor candidate vector, z1 is the first central anchor candidate, z2 is the second central anchor candidate, z M is the Mth central anchor candidate; is the hash density frequency of the anchor in the mth sparse inner product approximation table; is the discrete integer vector space composed of all anchor points.

[0021] Further, the best hospital is:

[0022]

[0023] where H * is the number of the best hospital; is the coordinate of the hospital central anchor in the mth sparse inner product approximation table; is the anchor position of the mth sparse inner product approximation table of hospital h; δ(·) is the anchor matching score function, which is the inverse of the Hamming distance; is the query frequency of hospital h in the mth sparse inner product approximation table near the hospital central anchor; GeoD(h, u) is the geographical distance between hospital h and the location u of the patient; is the hospital set.

[0024] Further, each medical demand big data keyword corresponds to a big data keyword category; the number of extracted medical demand big data keywords exceeds the set big data keyword quantity threshold; the number of medical demand big data keywords under each big data keyword category is at most one.

[0025] Compared with the prior art, the application has the beneficial effects that: it has highly intelligent, structured and fine medical demand understanding and service matching capability, which significantly improves the accuracy and efficiency of user medical navigation. By constructing a big data keyword category system covering multiple semantic dimensions such as location, symptoms, examination, hospital level, expert qualification, etc., the system can automatically analyze the natural language medical demand input by the user into a structured semantic unit, avoiding the ambiguity and ambiguity in the traditional manual selection or keyword retrieval mode. The system introduces a confidence evaluation mechanism to classify and attribute each keyword, ensuring the accuracy and uniqueness of the structure expression. At the structure matching level, the system innovatively uses sparse anchor point modeling and multi-table mapping strategy to realize efficient discrete expression of semantic vectors, which not only guarantees semantic fidelity, but also has good computing efficiency. By constructing a center needle point model, the system realizes the aggregation reasoning of multiple keyword structure information, making the matching process more stable and having overall semantic consistency. In the hospital selection stage, the system integrates multiple factor scoring mechanisms such as structure matching degree, historical access heat and geographic location distance, effectively balancing medical service capability and medical convenience, and significantly improving the personalization and practicality of the recommendation results. Overall, the system has systematic technical advantages in medical semantic understanding, structure modeling and intelligent recommendation, and can provide high-precision and high-efficiency medical service self-help support in the multi-source heterogeneous medical data scene, and has significant practical application value. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 The system structure diagram of the medical data analysis and processing system based on anchor point matching proposed by the application;

[0027] Figure 2 The performance comparison experimental results of the medical data analysis and processing system based on anchor point matching and the prior art;

[0028] Figure 3 The experimental results of the influence of the number of sparse inner product approximation tables on the accuracy of hospital center anchor point matching;

[0029] Figure 4 The experimental results of the influence of geographic distance on the best hospital recommendation effect. DETAILED DESCRIPTION

[0030] The following description is used to disclose the application so that those skilled in the art can implement the application. The preferred embodiments in the following description are only as examples, and other obvious modifications can be made by those skilled in the art.

[0031] REFERENCE Figure 1As shown, a medical data analysis and processing system based on anchor point matching, the system comprises: a medical database, a patient data entry unit, a big data keyword extraction unit and a retrieval analysis unit and a screening unit; the medical database is established by big data collection, wherein each data corresponds to a hospital; each data includes a plurality of big data keywords with different semantic categories, and the medical demand big data keywords are stored in the form of sparse inner product approximation table; according to the sparse inner product approximation table corresponding to all medical demand big data keywords of each data, the hospital center anchor point of the hospital corresponding to each data is determined; the patient data entry unit is used for providing the patient to enter the medical demand description text; the big data keyword extraction unit is used for extracting the medical demand big data keywords from the medical demand description text; the retrieval analysis unit is used for forming a retrieval entity by all medical demand big data keywords, after feature extraction of the retrieval entity, according to the vector space model, the feature is expressed as a vector form, and an object vector set is obtained; the number of sparse inner product approximation tables is determined, and the sparse inner product approximation table is constructed according to the sparse inner product approximation function family; each object vector in the object vector set is mapped through the sparse inner product approximation table respectively, and the anchor point position of the object vector is obtained; the screening unit is used for calculating the center anchor point position common to all object vectors in the retrieval entity, and then calculating the hospital center anchor point closest to the center anchor point position, and taking the hospital corresponding to the hospital center anchor point as the best hospital of the patient.

[0032] The medical database part is the data basis of the system, which is constructed by real-time big data collection mechanism. The data sources cover online platforms of medical institutions at all levels, government archives, medical insurance designated directories, diagnosis and treatment capacity recording systems and third-party medical evaluation platforms, etc. The structure design of collected data is based on multi-dimensional medical service characteristics. Each hospital is modeled as a data record, which contains multiple service capability fields with different semantic categories. For example, the scope of hospital specialist diagnosis and treatment, the category of diseases treated, the type of diagnosis and treatment technology, the configuration of auxiliary diagnosis and treatment equipment, the level of medical insurance reimbursement ratio, the region and traffic accessibility, the level of medical staff qualification, etc. These contents are abstracted as medical big data keywords and classified according to semantic categories, forming a multi-level keyword hierarchy system. The keywords are stored in a sparse structure and mapped to a sparse inner product approximation table through a hash function family as a discrete structural index of hospital features. This design allows the system to maintain high processing efficiency and good semantic retrieval resolution when facing large-scale hospital data sets.

[0033] The patient data entry unit is mainly responsible for obtaining the medical demand description input by the user. The input content is usually expressed in natural language form, including patient condition description, desired treatment plan, medical cost preference, geographical location constraint, hospital level requirement, etc. This part provides a human-computer interaction interface, and performs language normalization processing on the input content, such as removing noise words, synonym normalization, medical terminology standardization, and syntax structure parsing, to ensure that the subsequent processing module can obtain a clear structure and complete semantic original data carrier. The processed text will be transmitted to the keyword extraction module for semantic analysis.

[0034] The task of the keyword extraction module is to identify the key medical service intention words in the medical demand description text and determine the category to which each keyword belongs. In this system, each keyword must be uniquely assigned to a category, such as disease category, treatment method category, diagnostic equipment category, payment method category, medical level category, etc. This classification is a prerequisite for system semantic modeling, ensuring the orthogonality of vector modeling and the independence of anchor reasoning. The keyword extraction module relies on the entity dictionary constructed based on medical ontology and combines the field language model for context semantic analysis to extract key word groups with semantic direction. The system limits each category to retain at most one keyword to avoid information redundancy or conflict within the category. The extraction result forms a keyword set, which serves as the "search entity" for subsequent processing of the system.

[0035] The main task of the retrieval analysis unit is to convert the above-mentioned keyword set into a spatial representation suitable for structural semantic search, and prepare for hospital data pre-matching through sparse inner product approximation mechanism. First, the system will vectorize each keyword through a medical semantic embedding model to obtain its expression in a high-dimensional semantic space. This vector not only represents the literal meaning of the keyword, but also integrates the context relationship between the keyword and other related words in the medical semantic system. For example, "diabetes" and "endocrinology department", "insulin therapy", "blood glucose meter" will show close geometric distribution in this space, while "orthopedic surgery" and "tumor targeting" have obvious distinction. The semantic vector model used by the system can be a BERT-based medical special pre-training model, or a semantic vector dictionary constructed by traditional Word2Vec training medical data corpus. The keyword vector set constitutes the feature expression of the user input demand in the semantic space.

[0036] The vectorized keyword set also needs to be mapped into the structural index system of the system. To this end, the system introduces a sparse inner product approximation table modeling mechanism, which projects each vector into a predefined anchor position in the discrete space by constructing multiple approximation hash tables. These hash tables are generated using a specific family of approximation hash functions, which have a high probability of preserving the proximity relationship between similar vectors in the original semantic space, i.e., two semantically similar keywords will have their mapping result anchor points fall into adjacent or the same bucket in the hash table. Each keyword will be mapped to multiple hash tables to obtain multiple anchor positions, which provides a redundant fault tolerance mechanism in the spatial dimension and improves the robustness of the system to natural language variants. For example, if the user describes the requirement as "liver function abnormality", the system can identify that it has a cross position with "liver disease" and "cirrhosis" in the anchor point space, and thus retrieve the hospital related to liver disease treatment from the hospital data.

[0037] The acquisition of anchor points is only the first step, and the system also needs to integrate the anchor point set from multiple keywords to solve the center anchor point position of the retrieval entity in the sparse space. This process is based on the anchor point aggregation reasoning algorithm, i.e., the system performs weighted clustering calculation on the position of all keyword anchors in each hash table to determine the "semantic center point" in the multi-dimensional anchor point space. This center point represents the position of the user's overall demand in the structural semantic space and is the reference for subsequent similarity matching with hospital feature anchors. Since different keywords have different semantic weights when expressing user intent, the system also adjusts the weights during the anchor point aggregation process, for example, the anchor point contribution of disease keywords will be higher than that of geographic keywords or cost keywords, to ensure that the final matching is more focused on the core medical demand.

[0038] After completing the construction of the center anchor point of the retrieval entity, the system enters the screening stage. At this time, the system will traverse the center anchor point position of each hospital in the medical database and measure the sparse space distance with the user retrieval center anchor point. This distance is not a simple numerical difference, but an approximate similarity measure based on anchor point structure, which may use anchor point position Boolean overlap, hash collision rate, mapping sequence edit distance, etc. Finally, the system selects the hospital center anchor point closest to the retrieval center anchor point as the optimal matching result. The hospital in the matching result is the best treatment institution recommended by the system to the patient, which has the best fit between medical ability, service characteristics and user expectations. This process does not require manual selection by the user, and the system automatically completes it based on structural reasoning, truly realizing the automated service process from natural language input to high semantic matching.

[0039] During the operation of the entire system, the most core technical bottleneck and innovation is concentrated on how to convert unstructured natural language medical needs into structured indexable information units, and to realize efficient and accurate matching in the vast medical service space. Keyword extraction is the basis of semantic understanding, semantic vectorization is the key to semantic mapping, sparse anchor point modeling is the core of structure indexing, and central anchor point reasoning and approximate distance matching determine the accuracy and rationality of the final recommendation. By organically combining the above technical modules, the system breaks through the traditional rough matching mechanism based on keywords or tags, and establishes a medical service matching system with semantic understanding as the core, spatial indexing as the basis, and structure reasoning as the engine.

[0040] The system is versatile and highly adaptable. In addition to being suitable for single-user self-service recommendations, it can also be embedded in government medical systems, medical insurance management systems, and telemedicine platforms to achieve intelligent allocation of medical resources in a wider range of scenarios. Because the system is based on sparse approximate table structures, its data updating, expansion, and maintenance costs are much lower than those of traditional dense index systems, making it suitable for real-time matching requirements in large-scale medical data access and dynamic change scenarios. In addition, the system has good explainability, and each matching process can trace the keyword semantic path, anchor point space projection, and final distance calculation logic, providing data support and technical guarantee for later user behavior analysis, recommendation strategy optimization, and medical behavior supervision.

[0041] Further, the big data keywords at least include: location, department name, hospital level, treatment period, examination item, symptom description, treatment method, expert qualification, and medical insurance category; and the primary key in each piece of data in the medical database is the hospital number of the hospital.

[0042] Further, the big data keyword extraction unit extracts the medical demand big data keywords from the medical demand description text, and then calculates the confidence of the medical demand big data keywords to determine whether the medical demand big data keywords can be classified into a specific big data keyword category.

[0043] The position category is used to identify the geographical range where the user expects to visit, such as a city, a county, a subway station, etc.; the department name category involves the clinical department that the user wants to visit, such as internal medicine, orthopedics, obstetrics and gynecology, pediatrics, etc.; the hospital level category corresponds to the user's expectation of the level of the medical institution, such as a three-A hospital, a two-level hospital, etc.; the appointment time category includes the user's expected time, such as weekdays, holidays, weekend night clinics, etc.; the examination item category is used to identify the examination needs mentioned by the user, such as CT, MRI, blood routine, gastroscopy, etc.; the symptom description category is used to extract subjective narratives related to physical conditions or pathological reactions, such as cough, abdominal pain, fever, mass, etc.; the treatment method category covers the intervention methods expected by the patient, such as surgery, medication, physiotherapy, traditional Chinese medicine, etc.; the expert qualification category identifies whether the user has the demand for high-end medical personnel, such as chief physician, postdoctor, overseas expert, etc.; the medical insurance category identifies the payment mode of the patient, such as urban employee medical insurance, new rural cooperative medical care, self-payment, public medical care, etc. Each category in the system serves as a semantic container, responsible for receiving keywords from natural language input that belong to that category.

[0044] The design of the medical database takes the hospital as the smallest data unit, and each record corresponds to a specific hospital. The record uses the hospital number as the unique primary key, ensuring that each entity in the database has certainty. The database fields are structured according to the above keyword categories, and the fields under each category correspond to the service characteristics of the hospital. For example, under the "examination item" category, record all the examination capabilities of the hospital, and under the "expert qualification" category, record the titles and expertise of the resident experts of the hospital; in the "medical insurance category" field, record the medical insurance settlement methods supported by the hospital. This design ensures that the system can directly compare the user's demand with the service capability fields of the hospital during the matching process, forming a precise correspondence in the semantic dimension. After the user enters the medical demand description text through the interface, the text is first transmitted to the big data keyword extraction unit for natural language analysis. The system uses a set of language models optimized for the medical context to perform word segmentation, part-of-speech tagging, named entity recognition, and syntax analysis on the input text, extracting words or phrases that may have medical semantic orientation, and initially building a set of original keyword sets. These original keywords often have semantic ambiguity, i.e., the same word may have multiple candidate categories, such as "director" which may refer to expert qualifications or as a department title. Therefore, after the initial keyword extraction, the system needs to further determine whether each keyword can be uniquely classified into a specific category.

[0045] To solve the above classification problem, the system integrates a keyword attribution judgment mechanism based on confidence evaluation within the keyword extraction module. The system calculates the attribution confidence score Γ i of each extracted medical demand keyword w c for each category c.c (w i ). The confidence calculation is based on the semantic similarity of the keyword and all reference words in the standard category library, and the frequency of the reference words in the medical database is introduced as a weighting factor, thereby forming an information fusion judgment model that takes into account semantic proximity and commonality in the medical field. Finally, each keyword will select a category with the highest confidence value as its candidate category, and compare the score with the preset confidence threshold. Only when the confidence exceeds the threshold value, the system considers that the keyword can be effectively classified into the category, and it is included in the structured keyword set to participate in subsequent vector modeling and sparse anchor mapping. If the confidence is insufficient, the system will exclude the keyword or mark it as an unclassified word to avoid interfering with the construction of the system structure vector space. By introducing the confidence evaluation mechanism, the system solves the problem of uncertain word meaning in natural language input. Users may use ambiguous vocabulary, acronyms, colloquial expressions or omitted structures for description, and traditional keyword matching models often cannot accurately identify their semantic attribution. However, the system quantitatively models the semantic similarity of words in the category context, and combines the word frequency distribution in real medical corpus to build an interpretable and adjustable keyword attribution judgment mechanism, thereby realizing high-fidelity conversion from natural language input to structured semantic output. The conversion is a prerequisite for the subsequent object vector modeling, sparse anchor mapping, central anchor reasoning and hospital recommendation of the system. In addition, the system also supports dynamic expansion of keyword categories and standard reference word library. When high-frequency words in the user's input text cannot be effectively classified, the system will perform manual review and semantic expansion through the background log recording module, classify them into existing categories or create new categories, and update the standard reference word library and document frequency statistics. This mechanism ensures that the system's semantic recognition ability can be continuously optimized and improved with the accumulation of user data, forming a continuous learning and dynamically enhanced semantic perception ability, thereby continuously improving the system's practicality and adaptability in actual medical service scenarios.

[0046] Reference Figure 2 , further, the number of sparse inner product approximation tables M satisfies the following constraint relationship:

[0047]

[0048] where δ is the tolerance for the missing matching rate; P is the number of object vectors in the object vector set; P1 is the threshold value of the number of big data keywords.

[0049] In the medical data analysis and processing system based on anchor point matching, the performance of the entire system is highly dependent on its approximate matching capability on large-scale medical data sets. The core technical basis of this capability is to achieve fast and high-fidelity object vector retrieval through the construction of a sparse inner product approximation table. The retrieval analysis unit of the system is based on the sparse inner product approximation table as the basic structure, which is used to map the semantic vector of the patient's demand to the discrete anchor point space, and to realize the efficient comparison of high-dimensional semantic features in the sparse space. In order to ensure the accuracy and stability of the matching, the number of sparse inner product approximation tables, denoted as M, must satisfy certain constraint conditions. This constraint relationship reflects the mathematical balance mechanism between system fault tolerance and recognition rate in the retrieval scene of multiple objects, multiple categories, and multiple anchor points.

[0050] The design of this formula is based on the comprehensive consideration of the system's query error tolerance, keyword quantity threshold, and object vector size. First, the parameter δ represents the acceptable probability of missing matches, also known as the tolerance rate of missing matches. In the medical self-service recommendation scenario, missing matches mean that the hospital results retrieved by the user do not cover their actual semantic intent, which will directly affect the effectiveness of medical services. Therefore, the value of this parameter should be controlled at a very small level, usually between 10 -2 and 10 -6 . The specific value can be set according to the application environment. When δ is smaller, it means that the system needs higher anchor point space density coverage to reduce the matching omission caused by approximate hash mapping, and thus requires an increase in the number of sparse tables M.

[0051] The parameter P is the number of object vectors in the object vector set. In this system, this set comes from the semantic vector mapped by each keyword in the keyword set obtained after the system extracts and classifies the keywords from the patient's input medical demand text. Therefore, P is actually equivalent to the number of keywords involved in semantic modeling in the user's demand. Each keyword occupies a class position in the anchor point space as an independent semantic dimension, and the anchor point mapped by it will participate in the aggregation operation of the final central anchor point. The more the number of objects, the wider the anchor point space it covers. If we want to maintain a certain matching accuracy and recall rate, the system needs to build more sparse inner product tables to meet the needs of all objects simultaneously hitting in the hash space.

[0052] P1 is the system preset threshold of the number of keywords. In the medical self-service system, to prevent semantic space from being excessively dispersed, the system limits the extraction of one keyword under each category to form a high-quality and low-redundancy semantic input structure. P1 represents the lower limit of the number of keywords in an effective query allowed by the system, that is, if the number of keywords input by the user does not exceed the threshold, the system will consider that the semantic information is insufficient and will not start the query process. The parameter and the number of object vectors P are mutually constrained to form the basic control of the system on the input size and information density. When P is much higher than P1, it indicates that the user input information is sufficient, and in order to cover higher-dimensional anchor point space, the system will increase the number of sparse tables to prevent the anchor point from deviating from the aggregation center due to sparse high-dimensional space mapping.

[0053] The fractional term on the right side of the formula is the minimum coverage estimation expression in the sparse inner product space. The numerator represents the minimum space coverage strength required by the system to ensure that the probability of missing matching does not exceed δ. The denominator part log(P)-log(P1) represents the information density ratio between the number of input objects and the system keyword threshold. When P is much greater than P1, the denominator becomes larger, the entire fraction becomes smaller, and the number of sparse tables required is reduced; on the contrary, if P is close to P1, the system considers that the input information dimension is low or the category distribution is narrow, and in order to maintain the recall ability, the number of sparse tables needs to be increased to improve the space hit rate.

[0054] The entire formula multiplied by the number of objects P reflects the parallel mapping requirement under the multi-object anchor point space aggregation modeling. Since each object vector needs to be mapped in all sparse tables, the actual number of hash tables required by the system is not static, but a function of the number of objects. If each sparse table is regarded as a slice of the semantic space in a certain anchor point dimension, the minimum required number of M represents the ability of the system to completely cover the semantic space of the patient retrieval entity in the limited anchor point structure. When M is insufficient, some keywords input by the patient may be missing corresponding anchor points in some or all tables, resulting in the loss of dimensional reference points when the system aggregates the central anchor point, thereby causing semantic ambiguity or recommendation deviation.

[0055] The underlying logic of this formula comes from the recall analysis principle of the local sensitive hashing algorithm in the approximate search of high-dimensional vector space. In the traditional LSH theory, in order to effectively identify two approximate object vectors in a high-dimensional space, a sufficient number of hash tables must be constructed to ensure that the two vectors are mapped to the same or similar positions in at least one table. Corresponding to the system, if each keyword vector in the patient demand cannot be effectively mapped in the sparse anchor point space, the corresponding medical service features will be weakened in the subsequent hospital comparison, ultimately affecting the recommendation accuracy. Therefore, in order to ensure the overall matching performance of the system, the number of sparse tables must be dynamically calculated and set according to the keyword density, the hash table collision probability, and the allowed matching error control range of the system.

[0056] In the medical application scenario, the practical significance of this formula is particularly important. Due to the strong professional and fine features of medical services, different patients may use different terms to describe the same disease, and this language difference will be presented as a multi-dimensional distribution in the vector space. If the system uses a uniform table number setting without considering the input size, matching confidence, and keyword distribution difference, it is easy to cause mismatching or false matching problems in actual operation, thereby seriously affecting the credibility of system recommendation and user experience. By introducing the above sparse table number constraint formula, the system can realize dynamic optimization of the sparse space structure, so that the system can maintain stable semantic coverage ability under different user input conditions.

[0057] In addition, this formula can also guide the optimization of resource allocation in the system. When facing large-scale high-concurrency queries, the system can use this formula to calculate the minimum sparse table demand of the current request, load the corresponding number of sparse tables through an asynchronous table caching mechanism, avoid wasting computing resources caused by loading all hash tables, and achieve a balance between resource use and retrieval performance. Combined with the cache heat analysis mechanism, this formula can also be used for dynamic table number pruning, that is, for frequently accessed common keyword combinations, the corresponding table set can be set to have higher redundancy to improve query speed, and for low-frequency requests, the table number is compressed to save system resources.

[0058] Reference Figure 3 and Figure 4 Further, the confidence is:

[0059]

[0060] where Γ c (w i ) is the confidence score of the ith keyword w i belongs to the big data keyword category c; is the jth standard reference keyword in the big data keyword category c; is a keyword wi The embedding space semantic similarity between ; The standard reference keywords for the class c The document frequency of the keyword w in the medical database; n is the number of standard reference keywords in the keyword class.

[0061] The formula is Γ c (w i ) represents the confidence score of the i-th keyword w i belonging to the class c judged by the system. This score is used to evaluate the semantic consistency of the keyword with the target class, and the higher the value, the closer the semantic features of the keyword to the semantic center of the class c, and the higher the confidence of the system in its classification. This index is not only used for subsequent keyword screening and vector modeling, but also used to limit irrelevant keywords from entering the semantic index structure to avoid semantic pollution. In the formula, is the semantic similarity value of the keyword w i and the j-th standard reference keyword in the class c. This similarity is calculated in the semantic embedding space, usually by projecting the vocabulary into a high-dimensional semantic space in advance through a medical special language model such as BioBERT, Med-BERT or medical semantic Word2Vec model, and then obtaining the value through cosine similarity calculation. Since each class c usually contains multiple standard reference words, these words together constitute the semantic boundary of the class, so the system summarizes the similarity of the keyword w i with these words.

[0062] In this confidence formula, the numerator is a weighted sum term, the core of which is to multiply each semantic similarity term by the logarithm of the frequency of the reference keyword in the medical database, i.e. This weighting mechanism reflects the system's preference for high-frequency medical keywords. In actual medical databases, some standard keywords such as "diabetes", "coronary heart disease", "CT examination", "hospitalization reimbursement" and others have a much higher frequency than rare terms. By introducing the logarithmic weighting mechanism of document frequency, the system can enhance the confidence score of words close to the common semantic center, while suppressing the risk of misclassification of words with accidental semantic similarity but ambiguous semantic attribution. The denominator is the square root of the sum of the similarity values, which serves as a normalization function, so that the overall score is not affected by the number of reference words n. The reason for using the square root of the sum is to build a consistent normalization standard with the vector norm, so that the attribution score can be calculated between multiple classes. The addend 10 -6 is a numerical stability constant to avoid zero values in the denominator, which does not participate in the substantive adjustment of the confidence score.

[0063] In the specific operation of the anchor-matching-based medical data analysis and processing system, the confidence score calculation mainly occurs after the user completes the text input and before the keyword extraction module outputs. The system calculates the confidence score Γ for each category of all possible medical-related words in the user's text. c (w i The system calculates and retains the highest-scoring keyword within each category as the representative keyword for that category. For example, if a patient's input text contains keywords such as "cirrhosis," "interferon treatment," "medical insurance reimbursement," and "Shenzhen Nanshan," the system might categorize "cirrhosis" as a disease because it has high semantic similarity and frequency of occurrence with standard reference terms under disease categories such as "chronic liver disease," "abnormal liver function," and "viral hepatitis." "Interferon treatment" might be categorized as a treatment method; "medical insurance reimbursement" as a cost model; and "Shenzhen Nanshan" as a geographical location. The final retained keywords for each category represent the optimal result based on the confidence level calculation.

[0064] The introduction of this mechanism resolves the classification conflicts inherent in traditional keyword extraction when faced with semantic redundancy and ambiguity. Since users often use free-form expressions when describing medical needs, and synonyms, generalized terms, and specific terms appear alternately, relying solely on surface keyword matching is insufficient to accurately determine category attribution. By introducing a fusion model of semantic similarity and document frequency, the system can improve the stability and accuracy of structured input while maintaining semantic understanding. Furthermore, the confidence score is not only used for single classification but also serves as a moderating factor for subsequent vector weight modeling. During system vectorization, each retained keyword is transformed into a semantic vector, participating in the projection into the sparse anchor space. The confidence score Γ... c (w i The score can be used as a vector weighting coefficient to determine the semantic weight of the keyword in the vector combination. The higher the score, the more important the keyword, and the greater its influence on subsequent anchor point projection and center anchor point aggregation. Conversely, even if a keyword with a low score is included in the model, its weight is appropriately reduced to minimize its disturbance to the final result.

[0065] Furthermore, the family of sparse inner product approximation functions is:

[0066]

[0067] in, For the i-th object vector f i The anchor point in the m-th sparse inner product approximation table; T is the hash value of the sparse inner product approximation table; f i,k Let be the value of the i-th object vector in the k-th dimension; is the random projection weight of the k-th dimension for the t-th hash function in the m-th sparse inner product approximation table; ρ i,k is the i-th object vector; i is the normalized confidence bias in the k-th dimension; is the bit-wise concatenation to generate the hash anchor index value; d is the dimension of the object vector.

[0068] This formula describes how to generate the i-th object vector f i After T times of hash mapping, a discrete hash value is generated as the sparse anchor index of the vector in the m-th sparse inner product approximation table This sparse anchor value is composed of multiple binary hash functions, representing the position of the object in the hash table, which is used for subsequent similarity retrieval and central anchor aggregation operations. In the formula, the dimension d of each object vector f i is defined, i.e., its representation dimension in the high-dimensional semantic space. This vector is derived from the structured semantic expression obtained by the system through keyword extraction, semantic classification and vector embedding after the user inputs the medical demand. Each dimension f i,k of the object vector corresponds to the projection value of a keyword semantic feature in the k-th dimensional space.

[0069] Each sparse table is composed of T independent hash functions, and each hash function compresses the high-dimensional vector into a single binary bit. These hash functions are controlled by a set of previously randomly generated projection vector parameters . For the t-th hash function of the m-th sparse table, its role is to project the weighted sum of the components of the input vector in all dimensions to form a scalar judgment result. If the result is positive, it outputs 1, otherwise -1, which is mapped to binary representation by the sign function sign(·) and used as the hash bit at that position. Unlike traditional hash projection, the system introduces a normalized confidence bias factor ρ i,k in the hash function, which is used to adjust the contribution weight of different dimension components in the projection. This bias value represents the confidence uncertainty of the i-th object in the k-th dimension during the construction of the semantic vector, which may be due to factors such as semantic ambiguity, context ambiguity or fuzzy class membership boundary during word vector modeling. The system introduces this mapping function to control the confidence bias within a stable range and simultaneously implement a smooth penalty on uncertain components. The function form ensures that when the confidence bias is small, i.e., the dimension component is reliable, its contribution to the overall projection is large; otherwise, when there is a large bias in the dimension, its influence on the hash projection is suppressed, reducing the influence of semantic disturbance on the consistency of anchor generation.

[0070] After the hash function sums the product of each dimension component, it is converted to binary symbol by the sign function, and then multiplied by 2 t-1, which forms an integer hash bit value. All T-bit hash values are combined into a final anchor index value by bitwise concatenation operation This construction method can make the system maintain high-dimensional semantic distribution information while encoding the results as low-dimensional sparse representation, thereby facilitating fast query, aggregation and matching in multiple sparse hash tables. During system operation, each keyword vector will be mapped into multiple sparse hash tables, each table having different random projection vectors to achieve multi-angle segmentation of the semantic space. In this way, the same object vector can obtain different anchor indexes in different tables, thereby improving overall coverage and query robustness. When multiple object vectors are mapped to similar positions, the system can aggregate their anchor positions to form a semantic center anchor for subsequent hospital matching. This design allows even if the user input keyword expression is not completely consistent, as long as its semantics are similar, its vector mapping result will be close in the anchor space, thereby achieving semantic fault-tolerant matching. In addition, since each hash function only performs weighted summation and sign judgment, its computational complexity is low and easy to deploy in parallel, suitable for fast indexing in massive high-dimensional semantic vectors of medical databases. The introduced normalized confidence function further enhances the system's adaptability to input uncertainty and semantic ambiguity, enabling the system to maintain stable matching performance in open semantic scenarios.

[0071] Further, the center anchor position is:

[0072]

[0073] wherein z = {z1, z2, …, z M} is a center anchor candidate vector, z1 is a first center anchor candidate, z2 is a second center anchor candidate, and z M is an Mth center anchor candidate; is a hash density frequency representing an anchor in the mth sparse inner product approximation table; is a discrete integer vector space composed of all anchor spaces.

[0074] The formula is essentially a least squares type optimization model with a weighted regularization term, used to find a center anchor vector z = {z1, z2, …, z M in the sparse space, so that the overall distance between it and the anchor set obtained by mapping all object vectors is minimized. Each dimension z q of the center needle point vector C m ​denotes the candidate position of the center anchor in the m-th sparse approximation table, and the whole vector constitutes a discrete point in sparse space, which represents the structured semantic barycenter of the whole medical demand entity.

[0075] The outer minimization operation of the formula denotes the selection of a set of optimal center anchor positions from all possible combinations of anchor positions. Since the anchor point space in the sparse table is a discrete integer set, the optimization process is carried out in the discrete space Z M , rather than in the continuous Euclidean space. Each element of this discrete space is an M-dimensional vector, representing the candidate anchor position on all M sparse tables.

[0076] The inner layer of the formula is a double summation structure. The first summation index i traverses each object vector in the set of object vectors, i.e., the vector representation of each keyword extracted from the patient demand; the second summation index m traverses the number of sparse inner product approximation tables, and each object vector has an anchor value The system attempts to find a center anchor z m such that the distance between it and the object anchors is minimized in all tables.

[0077] The calculation of each distance term uses a weighted squared distance form, where the distance calculation term is , i.e., the absolute difference between the candidate center anchor and the object anchor, which characterizes the relative positional difference between them in the hash space of the sparse table. This difference is weighted by a normalization factor with a logarithmic penalty term, which is where denotes the hash density frequency of the anchor in the m-th sparse table, i.e., the frequency of the anchor appearing in all hospital data.

[0078] The core purpose of introducing the logarithmic weighting term of the hash density frequency is to achieve distance penalty adjustment for "popular" anchor positions. If an anchor frequently appears in the table, it means that the information it represents has generalization or general matching ability, and the system will reduce the penalty degree of the distance error related to it to prevent such anchors from dominating the matching results; on the contrary, if an anchor is rare, its information is often more discriminative, and the system will accordingly increase the penalty weight of deviation from it to preserve the semantic expression ability of key semantic anchors.

[0079] The entire objective function penalizes each anchor deviation through a squared weighting structure, forming an optimization objective surface. The system finally selects z = {z1,..., z M} as the center anchor position C of the medical demand as a wholeq , which represents the semantic center of gravity of the demand in the sparse space. The center of gravity is the reference vector for the system to match the hospital anchor points in the medical database in the subsequent stage, reflecting the joint structural expression of all user input keywords in the multi-anchor point space.

[0080] At the system application level, the significance of the center needle point calculation lies not only in aggregating input information, but also in compressing high-dimensional semantic vectors into low-dimensional index structures, and enhancing the robustness of the system when facing partial input missing, ambiguous expression or semantic conflict. This aggregation method does not rely on vector linear superposition or mean operation, but uses the statistical characteristics of needle point distribution in sparse space to construct a minimization objective function, thereby having stronger discrete matching ability and semantic reconciliation ability. Especially in the face of medical semantics with synonymous expressions, pragmatic ambiguity or word frequency bias, the system can automatically adjust the anchor point aggregation logic to avoid the dominance of a keyword in the overall semantic judgment, and improve the fairness of matching.

[0081] Since the center needle point vector is composed of a set of discrete index values, its result can be directly used to compare with the center needle points of each hospital in the hospital database. The system can determine the closest hospital anchor point by simple sparse space distance measures such as Hamming distance, hash collision rate or discrete point sequence difference, thereby completing the final recommendation. This mechanism makes the system free from the computational complexity and instability brought by floating point vector comparison in high-dimensional semantic space, while maintaining high accuracy and greatly improving matching efficiency.

[0082] Further, the best hospital is:

[0083]

[0084] where H * is the number of the best hospital; is the coordinate of the hospital center anchor point in the mth sparse inner product approximation table; is the anchor point position in the mth sparse inner product approximation table of hospital h; δ(·) is the anchor point matching score function, which is the inverse of Hamming distance; is the query frequency of hospital h in the mth sparse inner product approximation table near the hospital center anchor point; GeoD(h, u) is the geographical distance between hospital h and the location u of the patient; is the hospital set.

[0085] In this formula, the hospital set H contains all the hospital records with structured anchor point information in the medical database, each record has an anchor point position vector in multiple sparse inner product approximation tables These anchor positions represent the structural positions covered by the hospital in the semantic space, which are the multi-dimensional encoding results of the hospital's service capability. The core of the formula is to represent the anchor points of each candidate hospital h, and the center anchor vector generated by the user's demand Structural matching in sparse space is performed to measure the consistency of the hospital with the user's demand in multiple semantic dimensions. The basic calculation unit of the matching degree is , which is the similarity evaluation function between the position of the center needle point in the mth sparse table and the corresponding anchor position of the hospital. This function uses the reciprocal form of Hamming distance, that is:

[0086]

[0087] Here, the Hamming distance Ham(·) represents the number of bit differences between two sparse space anchors in the discrete index space, reflecting the degree of deviation between them at the structural encoding level. When the two anchors are exactly the same, the distance is 0, and the matching degree reaches the maximum; if there are multiple bit differences, the matching degree decreases accordingly. The purpose of using the reciprocal form is to build a mechanism that gets a higher score for closer anchor combinations, ensuring that structural consistency plays a central role in the recommendation process. Each matching degree item is also multiplied by a hotness factor , where represents the number of times hospital h is searched in the history search near the position of the center needle point in the mth anchor table. This value is automatically recorded by the system during runtime through logs, indicating whether the hospital frequently appears in similar semantic anchor matching requests in history, which is an implicit adaptability indicator of the hospital to this type of demand. The introduction of this term gives the system the ability to perceive user behavior, making the recommendation mechanism not only rely on the static representation of structural matching, but also include experience feedback in dynamic interaction, improving the system's adaptability to real user demand scenarios.

[0088] The matching score of each semantic dimension in the formula is also divided by a geographic cost item 1+GeoD(h,u), where GeoD(h,u) represents the geographic distance between hospital h and the current patient's location u. This distance calculation is based on the spatial Euclidean distance or route cost estimation between the patient's location information (such as IP resolution, GPS positioning, or user manually entered address) and the registered geographic coordinates of the hospital. By introducing the geographic distance term as a penalty factor for the score, the system can automatically balance the semantic matching degree of medical services and the convenience of patient visits. Among hospitals with similar service capabilities, the system will tend to recommend hospitals closer to the patient to improve actual service accessibility and satisfaction.

[0089] The entire scoring function is calculated for all M anchor tables, and the sum of the scores of all dimensions is taken as the comprehensive score of the hospital. The final selected hospital number with the highest score is taken as the recommendation result, that is:

[0090] H * = the hospital that best matches semantics, has high search frequency, and has the optimal geographical distance;

[0091] The recommendation strategy realizes a multi-factor fusion decision mechanism of semantic structure priority, behavior frequency weighting, and distance penalty regulation. The system ensures the structural consistency of medical needs through sparse structure hash matching, learns the user group behavior preferences through the frequency adjustment mechanism, and ensures the actual availability of the results through the geographical factor constraint. Especially when facing multiple candidate hospitals with similar structural expressions but significant regional location differences, the system can dynamically balance service capacity and distance convenience to make a more individual demand-oriented recommendation decision. In addition, the function also has high interpretability. The system can output matching degree score, historical frequency, and geographical distance dimension indicators when displaying the recommendation results, which helps users understand the recommendation reasons and provides behavior feedback data support for the operators of the medical platform, which is conducive to further optimizing service layout and supply and demand scheduling.

[0092] Further, each medical demand big data keyword corresponds to a big data keyword category; the number of extracted medical demand big data keywords exceeds a set big data keyword quantity threshold; and the number of medical demand big data keywords under each big data keyword category is at most one.

[0093] The system stipulates that each medical demand big data keyword must be uniquely attributed to a big data keyword category. This design is derived from the "single category constraint" principle in semantic modeling, aiming to avoid the confusion caused by the ambiguous attribution of keywords to structural expressions. In actual medical texts, users may input words such as "director", "top three hospitals", "weekend", "medical insurance", "liver function", and "CT". These words have different semantic attribution possibilities in different contexts. For example, "director" may be a doctor's title or an administrative position; "weekend" can refer to time or imply scheduling strategies. To avoid confusion caused by such polysemous words, the system uses a semantic confidence evaluation mechanism of keyword category context to automatically determine the most likely category to which each keyword belongs in the current semantic environment, and excludes other possible attributions, so that each keyword is uniquely labeled as a sub-item under a category in the extraction result. This mechanism not only improves the orthogonality of the category in vector representation, but also enhances the dimension independence in anchor aggregation, which is conducive to subsequent sparse space modeling.

[0094] Secondly, the system stipulates that the number of extracted medical demand big data keywords must exceed the set big data keyword quantity threshold. This threshold mechanism is mainly used to ensure that the input semantic content has sufficient information density, so that the system has sufficient dimensionality when constructing the structured semantic space representation. If too few keywords are extracted, it will lead to insufficient vector dimensionality, too high anchor point sparsity, and semantic center aggregation distortion, thereby affecting the stability and interpretability of subsequent hospital matching. Therefore, the system sets a keyword quantity lower limit. Only when the number of extracted legal keywords is greater than the keyword quantity lower limit, the system will start the anchor point mapping and center needle point calculation process; otherwise, it will prompt the user that the input information is incomplete or unclear, and guide the user to supplement the description. This mechanism also avoids false triggering of invalid matching or low semantic resolution error recommendations to some extent. Further, the system also stipulates that the number of medical demand big data keywords under each big data keyword category is at most one. This constraint is a sparsity control of the structured input dimension of the system, and is the basis for constructing a stable and non-redundant anchor point space representation. In this system, the semantic vector space is not a dense continuous space, but a sparse structure space. Each category dimension corresponds to an independently mappable semantic channel in the sparse table. If there are multiple keywords under a certain category, it will cause overlap, interference, or even self-cancellation in the space representation of this category dimension, thereby destroying the structural comparability in the anchor point space. Therefore, after keyword extraction, the system sorts the candidate keywords in each category by confidence score, and only keeps the highest scorer as the representative keyword of that category to participate in the subsequent mapping. This mechanism not only enhances the representativeness of each category anchor point, but also reduces the subsequent calculation complexity, and improves the interpretability and efficiency of the overall matching.

[0095] The above three constraint mechanisms jointly construct the semantic control logic of the system input end. From "each keyword is uniquely classified" to "the number of extracted keywords exceeds the threshold" to "only one keyword is retained for each category", it constitutes a three-in-one input preprocessing framework for semantic ambiguity elimination, information density guarantee, and dimension independence maintenance. This framework not only ensures that the system can still extract structured semantic units from complex, variable, and non-standard natural language inputs, but also provides a highly standardized, sparsity-controllable, and dimensionally separable structured input for subsequent sparse inner product mapping, anchor aggregation, distance evaluation, and hospital recommendation, etc. core modules, laying a solid foundation for vectorization modeling and index structure generation of medical demand.

[0096] The following is a specific implementation scheme example of an anchor point matching-based medical data analysis and processing system based on the present application. This example covers the entire process of user input medical demand text, keyword extraction and classification, semantic vector construction, anchor point generation, center needle point calculation, matching hospital scoring and recommendation, and is accompanied by clear numerical values and intermediate calculation results, facilitating understanding of the running mechanism of the system in engineering practice.

[0097] A patient is located in a certain district of a certain city, and the following medical demand description text is input: "Recently often abdominal distension, want to find a third-grade class-A hospital to do a gastroscopy, preferably can make an appointment on weekends, medical insurance can be reimbursed, and the doctor's qualification should be higher."

[0098] The system extracts the following keywords from the natural language in this paragraph according to the semantic model and entity recognition tool:

[0099] Symptom description class: abdominal distension;

[0100] Examination item class: gastroscopy;

[0101] Hospital level class: third-grade class-A;

[0102] Consultation period class: weekends;

[0103] Medical insurance category: medical insurance;

[0104] Expert qualification class: high doctor qualification (mapped to chief physician)

[0105] The number of keywords is 6, which exceeds the system-set keyword quantity threshold P1=4, meeting the conditions.

[0106] Set each keyword to a 5-dimensional semantic vector (simplified to 128 dimensions for ease of display) through a medical word vector model (such as BioWord2Vec). Construct the following object vector set:

[0107] f1=abdominal distension=[0.4,0.1,0.0,0.2,0.3];

[0108] f2=gastroscopy=[0.3,0.5,0.1,0.0,0.2];

[0109] f3=third-grade class-A=[0.6,0.0,0.2,0.1,0.1];

[0110] f4=weekends=[0.2,0.3,0.4,0.1,0.0];

[0111] f5=medical insurance=[0.1,0.2,0.0,0.5,0.3];

[0112] f6=chief physician=[0.5,0.2,0.1,0.0,0.4].

[0113] Suppose the system uses M=4 sparse inner product approximation tables, each with T=3 hash bits, i.e. each hash table has 3 independent hash functions to generate anchor points for each object vector. Take the first object f1 as an example, using the following simulated random weights: In the first hash table, the random weight corresponding to the first hash function is:

[0114] Assume the setting confidence bias is: p 1,k = 0.1 (all dimensions).

[0115] Calculate the normalization correction term: the factor for each dimension is

[0116] Projection results:

[0117]

[0118] s1 = 0.12 - 0.02 + 0 + 0.08 + 0.06 = 0.24 · 0.645 ≈ 0.1548;

[0119]

[0120] Repeat 3 times, and get the anchor point position in a table as three binary 101 (i.e. decimal 5). In this way, the anchor point mapping is completed for all M = 4 tables and all P = 6 objects, and the anchor point set is obtained

[0121] For each hash table m, the system performs weighted distance optimization centering on the anchor point position of all objects in the table. Taking the first hash table as an example, the anchor point values of the six objects are: 4, 5, 5, 6, 5, and 4.

[0122] Let the anchor point hash density frequency be:

[0123] HDF1(4) = 500;

[0124] HDF1(5) = 1200;

[0125] HDF1(6) = 800;

[0126] Find the center position z1 ∈ {4, 5, 6}, and calculate the cost:

[0127] Take z1 = 5 as an example:

[0128]

[0129] Use the approximate value:

[0130]

[0131] The cost is approximately:

[0132]

[0133] Repeat the process, and select the z1 that minimizes the cost as the center needle point coordinate of the first dimension The same reason for the remaining dimensions, eventually get C q =[5,...].

[0134] Suppose there is a hospital A in the hospital database, whose four-dimensional anchor point coordinates are a A =[5,4,6,5], and another hospital B is [6,5,5,6]. The center anchor point is C q =[5,4,5,5], and the patient's location is in a certain area, and the geographical distance of the two hospitals is respectively:

[0135] GeoD(A, u) = 2km;

[0136] GeoD(B, u) = 8km;

[0137] The query frequency is respectively:

[0138] Calculate the score of hospital A:

[0139]

[0140] For example, the first item: log(1+300)≈5.71; 1+GeoD(A, u) = 3; the score is about

[0141] The total score after adding the four items is about 6.85. Although the heat of hospital B is high, the score is about 5.20 because of the long distance. The final output recommendation result is hospital A, which is the best recommended hospital for the user.

[0142] The above shows and describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above examples, and the above examples and descriptions in the specification are only the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection claimed by the present application is defined by the appended claims and their equivalents.

Claims

1. A medical data analysis and processing system based on anchor point matching, characterized in that, The system comprises a medical database, a patient data input unit, a big data keyword extraction unit, a retrieval analysis unit and a screening unit; the medical database is established through big data collection, wherein each piece of data corresponds to a hospital; each piece of data comprises a plurality of medical demand big data keywords with different keyword categories and is stored in a sparse inner product approximate table form; the hospital central anchor point of the hospital corresponding to each piece of data is determined according to the sparse inner product approximate table corresponding to all medical demand big data keywords of each piece of data; the patient data input unit is used for providing the patient with a medical demand description text; the big data keyword extraction unit is used for extracting medical demand big data keywords from the medical demand description text; the retrieval analysis unit is used for grouping all medical demand big data keywords into a retrieval entity, performing feature extraction on the retrieval entity, representing the features in a vector form according to a vector space model to obtain an object vector set, determining the number of sparse inner product approximate tables and constructing a sparse inner product approximate table according to a sparse inner product approximate function family; each object vector in the object vector set is mapped through the sparse inner product approximate table to obtain the anchor point position of the object vector; the screening unit is used for calculating the central anchor point position common to all object vectors in the retrieval entity, calculating the hospital central anchor point closest to the central anchor point position, and taking the hospital corresponding to the hospital central anchor point as the best hospital of the patient.

2. The anchor point matching based medical data analysis and processing system of claim 1, wherein, The big data keyword extraction unit extracts medical demand big data keywords from the medical demand description text, calculates the confidence of the medical demand big data keywords, and determines whether the medical demand big data keywords can be classified into a specific big data keyword category.

3. The anchor point matching based medical data analysis and processing system of claim 2, wherein, The number M of sparse inner product approximate tables satisfies the following constraint relationship: wherein δ is a tolerance for the missing matching rate, P is the number of object vectors in the object vector set, and P1 is a big data keyword quantity threshold.

4. The anchor point matching based medical data analysis and processing system of claim 3, wherein, The confidence is: wherein, Γ c (w i ) is the confidence score of the ith keyword w i belonging to the big data keyword category c; is the jth standard reference keyword in the big data keyword category c; is the embedding space semantic similarity between the keyword w i and ; is the standard reference keyword ; n is the number of standard reference keywords in the big data keyword category.

5. The anchor point matching based medical data analysis and processing system of claim 4, wherein, The sparse inner product approximate function family is: wherein, is the mth anchor point in the sparse inner product approximation table for the ith object vector f i ; T is the number of hash bits of the sparse inner product approximation table; f i,k is the value of the ith object vector in the kth dimension; is the random projection weight of the kth dimension corresponding to the tth hash function in the mth sparse inner product approximation table; p i,k is the normalized confidence bias of the ith object vector f i in the kth dimension; is the bitwise concatenation to generate the hash anchor point index value; d is the dimension of the object vector.

6. The anchor point matching based medical data analysis and processing system of claim 5, wherein, Central anchor position Is: where z = {z1, z2, …, z M} is the center anchor candidate vector, z1 is the first center anchor candidate, z2 is the second center anchor candidate, z M is the Mth center anchor candidate; is the hash density frequency of the anchor in the mth sparse inner product approximation table; is the discrete integer vector space composed of all anchor points.

7. The anchor point matching based medical data analysis and processing system of claim 6, wherein, The best hospital is: where H * is the number of best hospitals; is the coordinate of the hospital center anchor in the m-th sparse inner product approximation table; is the anchor position in the m-th sparse inner product approximation table of hospital h; δ(·) is the anchor matching score function, which is the inverse of Hamming distance; is the query frequency of hospital h in the m-th sparse inner product approximation table near the hospital center anchor; GeoD(h, u) is the geographical distance between hospital h and the location u of the patient; is the set of hospitals.

8. The anchor point matching based medical data analysis and processing system of claim 7, wherein, Each medical demand big data keyword corresponds to a big data keyword category; the number of extracted medical demand big data keywords exceeds the set big data keyword quantity threshold; the number of medical demand big data keywords under each big data keyword category is at most one.

9. The anchor point matching based medical data analysis and processing system of claim 8, wherein, The big data keyword categories at least include position, department name, hospital level, treatment period, examination item, symptom description, treatment method, expert qualification and medical insurance category; the primary key in each piece of data in the medical database is the hospital number of the hospital.

Images