Medical equipment information integration method and system based on natural language processing

Through natural language processing technology, semantic understanding model and query statement generation model are built, which solves the problems of data diversity and integration of medical equipment, realizes efficient and accurate data integration and query, and improves database performance.

CN118916379AInactive Publication Date: 2024-11-08WUXI XINGYU CREATIVE TECHNOLOGY CO LTD

Patent Information

Application Number
CN202411056659.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-02
Publication Date
2024-11-08
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

There are a variety of medical device data types, and existing integration methods are difficult to effectively integrate medical device data from different types and manufacturers.

Method used

Using a natural language processing method, the metadata is parsed and file system is constructed by constructing a semantic understanding model, a relational database index is established, and a query statement generation model is used to realize data query, and the query log update index is analyzed.

Benefits of technology

It improves the efficiency and accuracy of medical equipment data integration, simplifies the data query process, and improves database access performance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118916379A_ABST
    Figure CN118916379A_ABST
Patent Text Reader

Abstract

The invention discloses a medical equipment information integration method and system based on natural language processing, and the method comprises the steps: S1, exporting a data file and corresponding metadata from software of a plurality of pieces of equipment, and carrying out the preprocessing of the data file and the corresponding metadata, and obtaining the preprocessed data file and metadata; s2, constructing a semantic understanding model to analyze the metadata, and filling the analyzed metadata into a relational database; meanwhile, constructing a file system and storing a data file corresponding to the metadata into the file system; s3, establishing an index of the relational database and associating the index with the data table of each device; s4, constructing a query statement generation model, querying medical equipment information, and storing a query log; s5, analyzing the high-frequency content in the query log, and updating the index of the relational database; the invention aims to solve the problem that medical equipment data of different types and different manufacturers are difficult to effectively integrate by an existing integration method due to various types of medical equipment data.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of medical informationization, and in particular to a medical equipment information integration method and system based on natural language processing. Background Art

[0002] Medical device information integration aims to collect, process, and standardize the relevant information of medical devices of different manufacturers and types to improve the quality of medical services and operational efficiency. However, the types of data generated by medical devices are diverse, and traditional methods often rely on independent software or systems of each device for information management, making it difficult to effectively integrate multi-source heterogeneous data of different types and manufacturers. In addition, although most devices contain metadata files when exporting data, which can express the basic information of the data files to a certain extent, the contents of these metadata files are significantly different. Not only are the contents of metadata files different between different devices, but even for the same type of equipment, the contents of their metadata files will be different due to different manufacturers, further increasing the difficulty of information integration. Summary of the invention

[0003] In view of this, the present invention aims to provide a medical device information integration method and system based on natural language processing, aiming to solve the problem that the types of medical device data are diverse and the existing integration methods are difficult to effectively integrate medical device data of different types and manufacturers.

[0004] A method for integrating medical device information based on natural language processing comprises the following steps:

[0005] S1: exporting data files and corresponding metadata from software of multiple devices and preprocessing them to obtain preprocessed data files and metadata;

[0006] S2: Build a semantic understanding model to parse the metadata, and fill the parsed metadata into the relational database; at the same time, build a file system and store the data files corresponding to the metadata into the file system;

[0007] S3: Create an index for the relational database and associate it with the data table of each device;

[0008] S4: Build a query statement generation model, query medical device information, and save the query log;

[0009] S5: Analyze the high-frequency content in the query log and update the index of the relational database;

[0010] Furthermore, the step S1 exports data files and corresponding metadata from software of multiple devices and performs preprocessing to obtain preprocessed data files and metadata, including:

[0011] S11: De-duplicate the exported data files and the corresponding metadata and delete empty files. Meanwhile, fill the missing values ​​with “NaN” to obtain the data files and metadata after preliminary preprocessing.

[0012] S12: For the data file and metadata after preliminary preprocessing, a unique identifier is added as an associated identifier to obtain the preprocessed data file and metadata.

[0013] Furthermore, the step S2 constructs a semantic understanding model to parse the metadata, and fills the parsed metadata into a relational database; at the same time, constructs a file system and stores the data files corresponding to the metadata into the file system, including:

[0014] S21: using regular expressions to convert the preprocessed metadata into table data; the columns of the table data are data categories, and the rows are data quantities;

[0015] S22: Construct a semantic understanding model, which includes a data understanding part and an entity alignment part, wherein the calculation method of the semantic understanding part is:

[0016] f 1 =CNN(x t )

[0017] f 2 =MaxPoling(f 1 )+AveragePooling(f 1 )

[0018] f 3 = Transformer(BiLSTM(x t ))+BiLSTRM(x t )

[0019] f 4 =MLP(Concat(f 2 ,f 3 ))

[0020] Among them, f 1 is the first convolution feature, CNN() is the convolutional neural network operation, x t is the input data, t is the time step index, f 2 is the second convolution feature, MaxPooling() is the maximum pooling operation, AveragePooling() is the global average pooling operation, and f 3 is the semantic temporal feature, Transformer() is the Transformer operation, BiLSTM() is the bidirectional long short-term memory network operation, and f 4To fuse semantic features, Concat() is a concatenation operation and MLP() is a fully connected layer operation;

[0021] The calculation method of the semantic understanding model is uniformly expressed as:

[0022] f 4 =Comprehense(x t )

[0023] Among them, Comprehense() is the semantic understanding model operation;

[0024] Traditional methods usually rely on rule-based text matching methods to parse metadata. These methods are not effective in processing complex and heterogeneous medical device metadata. They are easily unable to parse due to incomplete rule settings. Even if a large number of rules are set, it is still impossible to enumerate all situations in the face of different types of medical devices from different manufacturers. The semantic understanding model of the present invention combines the local feature extraction capability of convolutional neural networks, the feature dimension reduction capability of pooling operations, the global context understanding capability of Transformer, and the temporal modeling capability of bidirectional long short-term memory networks. It can fully and accurately understand the complex semantic information in metadata, improve the accuracy and completeness of parsing, and even when faced with expressions that have never been seen in metadata, it can analyze their meaning and fill them into the relational database.

[0025] S23: The table data is input into the data understanding part of the semantic understanding model in columns, and the fused semantic features are calculated as the column name feature vectors of each device;

[0026] S24: Input the column name feature vector of each device into the entity alignment part of the semantic understanding model to perform entity alignment to obtain the semantic distance between the device column names. The calculation method of the entity alignment part is:

[0027]

[0028]

[0029] in, is the semantic similarity between the column names of the first and second devices, is the column name feature vector of the first device, is the column name feature vector of the second device, ∥∥ is the modulus length operation, is the semantic distance between the first and second device column names, β is the cosine adjustment parameter, exp() is the exponential operation of taking the natural constant, γ is the modulus adjustment parameter, and || is the absolute value operation;

[0030] S25: traverse the semantic distances between all device column names and store them in the distance matrix; then traverse the distance matrix to find all column name feature vector pairs whose semantic distance is less than the category threshold. The category threshold is calculated as follows:

[0031] T=μ-α·σ

[0032] Where T is the category threshold, μ is the average value of all current semantic distances, α is the weight coefficient, and σ is the standard deviation of all current semantic distances;

[0033] For column name feature vectors with semantic distance smaller than the category threshold, they are merged into a new column name feature vector by taking the average of the two vectors and updating the distance matrix after merging.

[0034] For the column name feature vectors whose semantic distance is greater than or equal to the category threshold, they are retained as new independent categories. After all column name feature vectors are processed according to the above steps, the total device attribute column name is obtained;

[0035] For metadata converted into a table format, the column name represents a certain attribute or field described in the metadata; Step S25 calculates the semantic distance between device column names and uses threshold judgment to merge similar column name feature vectors, effectively processing various attributes of diverse medical devices, and providing a data basis for subsequent database construction, data query and log analysis;

[0036] S26: construct a relational database based on the total device attribute column name; at the same time, calculate the similarity between the total device attribute column name and the column name feature vector of each device, match them with the column name with the highest similarity, and then store the corresponding metadata in the relational database;

[0037] S27: construct a file system, and store the data files corresponding to the metadata of each device into the file system; at the same time, copy the file address and store it into a relational database;

[0038] Furthermore, the step S3 establishes an index of the relational database and associates it with the data table of each device, including:

[0039] S31: Select B+ tree as the index type, use the device number as the field, allocate space for each node of the B+ tree, store the device number and the pointer to the data table record;

[0040] S32: Starting from the root node of the B+ tree, sort in ascending order, insert the device number and the pointer to the data table record into the leaf node of the B+ tree, and store them;

[0041] S33: Repeat step S32 until all device numbers are inserted into the B+ tree;

[0042] S34: Store the constructed B+ tree index in a relational database and associate it with the data table of each device.

[0043] Furthermore, the step S4 constructs a query statement generation model, queries medical device information, and saves query logs, including:

[0044] S41: Build a query statement generation model, the calculation method is:

[0045] f SQL =Comprehense(SQL)

[0046] f X =Comprehense(X)

[0047] h trans = Transformer(Concat(f SQL ,f X ))

[0048] z t =sigmoid(W 1 h trans )

[0049]

[0050]

[0051] Among them, f SQL is the SQL fusion semantic feature, f X is the natural language fusion semantic feature, SQL is the SQL query statement, X is the natural language input by the user, and h trans is the splicing feature, z t is the reset update gate under the time step, sigmoid() is the sigmoid function, W 1 To reset the update parameter matrix, is the hidden state at time step t, tanh() is the hyperbolic tangent function, W 2 is the hidden state parameter matrix, h t-1 Generate features for the sentence at time step t-1, h t Generate features for the statements at time step t; SQL stands for "Structured Query Language", which is a standard computer language for accessing and processing databases, allowing users to query, update, and manage data in the database;

[0052] Then generate feature h for the statement at time step t t, use Top-k Sampling to decode and get the SQL query statement;

[0053] Traditional database query methods usually require users to manually write SQL statements, which is time-consuming and labor-intensive and prone to errors, resulting in query failures. The present invention generates a model by constructing a query statement and merging the reset gate and update gate of the gated recurrent unit to obtain a reset update gate, thereby simplifying the calculation structure, reducing the amount of calculation per time step, and occupying less memory during training and inference.

[0054] Then, Transformer is used to encode the fusion semantic features of SQL statements and natural language, so that when the model generates SQL query statements, it can not only fully utilize the understanding ability of natural language, but also capture the temporal dependencies in the query statements, thus realizing the automatic conversion from natural language to SQL statements and greatly improving the query efficiency.

[0055] S42: using the constructed query statement generation model, converting the natural language input by the user into an SQL query statement and sending it to the database, recording and saving the query log; the query log includes the natural language input by the user, the SQL query statement, the execution time of the query statement and the query result.

[0056] Furthermore, the step S5 analyzes the high-frequency content in the query log and updates the index of the relational database, including:

[0057] S51: Count the frequency of each query statement in the log, start from the query statement with the highest frequency and accumulate the frequencies one by one until the cumulative sum of the frequencies exceeds 30% of the total query frequency for the first time, and finally regard the query statements whose cumulative sum of frequencies is greater than 30% as high-frequency query statements;

[0058] S52: Count the execution time of high-frequency query statements, and calculate the index weight according to the frequency and execution time of each query statement. The calculation method is:

[0059]

[0060] Among them, Weight is the index weight, F is the frequency of each high-frequency query statement, Time is the execution time of the high-frequency query statement, max() is the maximum value, and min() is the minimum value;

[0061] S53: Add a new composite index for the fields in the high-frequency query statement according to the index weight, and update it to the entire B+ tree;

[0062] The present invention aims at high-frequency query statements, comprehensively considers two key factors, query frequency and execution time, to calculate index weights, and more accurately identifies indexes that need to be optimized; at the same time, new composite indexes are added to fields in high-frequency query statements, and the entire B+ tree is updated, providing better data access performance for the database.

[0063] The present invention also discloses a medical equipment information integration system based on natural language processing, comprising:

[0064] Data export and preprocessing module: export data files and corresponding metadata from the software of multiple devices and perform preprocessing to obtain preprocessed data files and metadata;

[0065] Semantic understanding and database construction module: Build a semantic understanding model to parse metadata, and fill the parsed metadata into the relational database; at the same time, build a file system and store the data files corresponding to the metadata into the file system;

[0066] Index creation and association module: Create an index for the relational database and associate it with the data tables of each device;

[0067] Query statement generation module: builds a query statement generation model, queries medical device information, and saves query logs;

[0068] Index analysis and update module: Analyze the high-frequency content in the query log and update the index of the relational database.

[0069] Compared with the prior art, the advantages of the present invention are:

[0070] (1) The present invention addresses the problem that existing medical device integration methods are difficult to effectively integrate medical device data of different types and manufacturers. The present invention constructs a semantic understanding model to parse metadata, builds a file system and stores data files corresponding to the metadata in the file system, thereby realizing the associated storage of data files and metadata. At the same time, the present invention analyzes the high-frequency content in the query log and updates the index of the relational database, thereby improving the efficiency of data integration.

[0071] (2) The present invention constructs a semantic understanding model to parse metadata in a semantic understanding manner, combining the local feature extraction capability of convolutional neural networks, the feature dimension reduction capability of pooling operations, the global context understanding capability of Transformer, and the temporal modeling capability of bidirectional long short-term memory networks to accurately understand the semantic information in metadata. Compared with traditional rule-based text matching methods, the present invention can integrate metadata types that have never appeared in historical situations and fill them into relational databases.

[0072] (3) The present invention merges the reset gate and the update gate of the gated cyclic unit to obtain a reset-update gate, thereby simplifying the calculation structure. A query statement generation model is constructed based on this, and query statements of a relational database are automatically generated according to the natural language input by the user, thereby providing a more efficient data query method and improving the efficiency of data query.

[0073] (4) The present invention analyzes the query log, first counting the high-frequency query statements, then based on the statistics, comprehensively considering the frequency and execution time, calculating the index weight of the high-frequency query statements, and finally adding new composite indexes to the fields in the high-frequency query statements based on the index weight, thereby providing the database with better data access performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 A schematic diagram of the process of medical device information integration based on natural language processing provided by the present invention;

[0075] Figure 2 A schematic diagram of the algorithm flow of the semantic understanding model provided by the present invention;

[0076] Figure 3 A schematic diagram of the algorithm flow for splicing feature extraction provided by the present invention. DETAILED DESCRIPTION

[0077] The present invention is further described below in conjunction with the accompanying drawings, but the present invention is not limited in any way. Any changes or substitutions made based on the teachings of the present invention belong to the protection scope of the present invention.

[0078] Embodiment 1: A method for integrating medical device information based on natural language processing, such as Figure 1 As shown, the following steps are included:

[0079] S1: exporting data files and corresponding metadata from software of multiple devices and preprocessing them to obtain preprocessed data files and metadata;

[0080] S11: De-duplicate the exported data files and the corresponding metadata and delete empty files. Meanwhile, fill the missing values ​​with “NaN” to obtain the data files and metadata after preliminary preprocessing.

[0081] S12: For the data file and metadata after preliminary preprocessing, a unique identifier is added as an associated identifier to obtain the preprocessed data file and metadata.

[0082] S2: Build a semantic understanding model to parse the metadata, and fill the parsed metadata into the relational database; at the same time, build a file system and store the data files corresponding to the metadata into the file system;

[0083] S21: using regular expressions to convert the preprocessed metadata into table data; the columns of the table data are data categories, and the rows are data quantities;

[0084] S22: Construct a semantic understanding model, which includes a data understanding part and an entity alignment part, wherein the calculation method of the semantic understanding part is:

[0085] f 1 =CNN(x t )

[0086] f 2 =MaxPooling 1 )+AveragePooling(f 1 )

[0087] f 3 = Transformer(BiLSTM(x t ))+BiLSTM(x t )

[0088] f 4 =MLP(Concat(f 2 ,f 3 ))

[0089] Among them, f 1 is the first convolution feature, CNN() is the convolutional neural network operation, x t is the input data, t is the time step index, f 2 is the second convolution feature, MaxPooling() is the maximum pooling operation, AveragePooling() is the global average pooling operation, and f 3 is the semantic temporal feature, Transformer() is the Transformer operation, BiLSTM() is the bidirectional long short-term memory network operation, and f 4 To fuse semantic features, Concat() is a concatenation operation and MLP() is a fully connected layer operation;

[0090] The calculation method of the semantic understanding model is uniformly expressed as:

[0091] f 4 =Comprehense(x t )

[0092] Among them, Comprehense() is the semantic understanding model operation;

[0093] Traditional methods usually rely on rule-based text matching methods to parse metadata. These methods are not effective in processing complex and heterogeneous medical device metadata. They are easily unable to parse due to incomplete rule settings. Even if a large number of rules are set, it is still impossible to enumerate all situations in the face of different types of medical devices from different manufacturers. The semantic understanding model of the present invention combines the local feature extraction capability of convolutional neural networks, the feature dimension reduction capability of pooling operations, the global context understanding capability of Transformer, and the temporal modeling capability of bidirectional long short-term memory networks. It can fully and accurately understand the complex semantic information in metadata, improve the accuracy and completeness of parsing, and even when faced with expressions that have never been seen in metadata, it can analyze their meaning and fill them into the relational database.

[0094] S23: The table data is input into the data understanding part of the semantic understanding model in columns, and the fused semantic features are calculated as the column name feature vectors of each device;

[0095] S24: Input the column name feature vector of each device into the entity alignment part of the semantic understanding model to perform entity alignment to obtain the semantic distance between the device column names. The calculation method of the entity alignment part is:

[0096]

[0097]

[0098] in, is the semantic similarity between the column names of the first and second devices, is the column name feature vector of the first device, is the column name feature vector of the second device, ∥∥ is the modulus length operation, is the semantic distance between the first and second device column names, β is the cosine adjustment parameter, exp() is the exponential operation of taking the natural constant, γ is the modulus adjustment parameter, and || is the absolute value operation;

[0099] S25: traverse the semantic distances between all device column names and store them in the distance matrix; then traverse the distance matrix to find all column name feature vector pairs whose semantic distance is less than the category threshold. The category threshold is calculated as follows:

[0100] T=μ-α·σ

[0101] Where T is the category threshold, μ is the average value of all current semantic distances, α is the weight coefficient, and σ is the standard deviation of all current semantic distances;

[0102] For column name feature vectors with semantic distance smaller than the category threshold, they are merged into a new column name feature vector by taking the average of the two vectors and updating the distance matrix after merging.

[0103] For the column name feature vectors whose semantic distance is greater than or equal to the category threshold, they are retained as new independent categories. After all column name feature vectors are processed according to the above steps, the total device attribute column name is obtained;

[0104] For metadata converted into a table format, the column name represents a certain attribute or field described in the metadata; Step S25 calculates the semantic distance between device column names and uses threshold judgment to merge similar column name feature vectors, effectively processing various attributes of diverse medical devices, and providing a data basis for subsequent database construction, data query and log analysis;

[0105] For example, there are two types of medical equipment (ultrasound diagnostic equipment and CT machine) data packets, and the corresponding metadata are as follows:

[0106] Ultrasonic diagnostic equipment:

[0107]

[0108]

[0109] It can be seen that there are significant differences between the two in terms of data field names, data structures, and corresponding data files, including "deviceName" and "Name", "examResults" and "scanResults", "measurements" and "dimensions", etc. The traditional data integration method requires listing the rules one by one. When a new field name such as "DeName" appears, the rules need to be added again.

[0110] As for the semantic understanding model, in the data understanding part, the model will calculate the fusion feature vector of "deviceName" and "Name". Then, in the entity alignment stage, it will find that the two vectors are very similar and the semantic distance is very close, so they will be merged into one category. However, no similar field is found for the "radiologistNotes" field, so it will be treated as a new category.

[0111] S26: construct a relational database based on the total device attribute column name; at the same time, calculate the similarity between the total device attribute column name and the column name feature vector of each device, match them with the column name with the highest similarity, and then store the corresponding metadata in the relational database;

[0112] S27: construct a file system, and store the data files corresponding to the metadata of each device into the file system; at the same time, copy the file address and store it into a relational database;

[0113] The results stored in the relational database are as follows:

[0114] deviceName ID patientName examDate examResults files measurements radiologistNotes Ultrasound imaging machine US001 Zhang San 2023-04-01 Hepatic space occupying lesions image1.jpg,image2.jpg length:5.2cm,width:3.8cm - CT machine CT001 Li Si 2023-04-02 Lung shadows slice1.dcm,slice2.dcm length:6.0cm,width:4.5cm,height:2.0cm Further observation is recommended

[0115] S3: Create an index for the relational database and associate it with the data table of each device;

[0116] S31: Select B+ tree as the index type, use the device number as the field, allocate space for each node of the B+ tree, store the device number and the pointer to the data table record;

[0117] S32: Starting from the root node of the B+ tree, sort in ascending order, insert the device number and the pointer to the data table record into the leaf node of the B+ tree, and store them;

[0118] S33: Repeat step S32 until all device numbers are inserted into the B+ tree;

[0119] S34: Store the constructed B+ tree index in a relational database and associate it with the data table of each device.

[0120] S4: Build a query statement generation model, query medical device information, and save the query log;

[0121] S41: Build a query statement generation model, the calculation method is:

[0122] f SQL =Comprehense(SQL)

[0123] f X =Comprehense(X)

[0124] h trans = transformer(Concat(f SQL ,f X ))

[0125] z t =sigmoid(W 1 h trans )

[0126]

[0127]

[0128] Among them, f SQL is the SQL fusion semantic feature, fX is the natural language fusion semantic feature, SQL is the SQL query statement, X is the natural language input by the user, and h trans is the splicing feature, z t is the reset update gate under the time step, sigmoid() is the sigmoid function, W 1 To reset the update parameter matrix, is the hidden state at time step t, tanh() is the hyperbolic tangent function, W 2 is the hidden state parameter matrix, h t-1 Generate features for the sentence at time step t-1, h t Generate features for the statements at time step t; SQL stands for "Structured Query Language", which is a standard computer language for accessing and processing databases, allowing users to query, update, and manage data in the database;

[0129] Then generate feature h for the statement at time step t t , use Top-kSampling to decode and get the SQL query statement;

[0130] Traditional database query methods usually require users to manually write SQL statements, which is time-consuming and labor-intensive and prone to errors, resulting in query failures. The present invention generates a model by constructing a query statement and merging the reset gate and update gate of the gated recurrent unit to obtain a reset update gate, thereby simplifying the calculation structure, reducing the amount of calculation per time step, and occupying less memory during training and inference.

[0131] Then, Transformer is used to encode the fusion semantic features of SQL statements and natural language, so that when the model generates SQL query statements, it can not only fully utilize the understanding ability of natural language, but also capture the temporal dependencies in the query statements, thus realizing the automatic conversion from natural language to SQL statements and greatly improving the query efficiency.

[0132] For example, a hospital database contains a patient information table (Patients), an examination record table (Examinations), and an examination result table (Results); now you need to query the records of patients who have undergone CT examinations in the past month and whose results show abnormalities;

[0133] The traditional method requires writing three SQL statements:

[0134] --Query all patient records within the past month

[0135] SELECT*FROM Patients WHEREDate>'2023-02-01';

[0136] --Query all CT examination records in the past month

[0137] SELECT*FROM Examinations WHERE ExamType='CT'AND Date>'2023-02-01';

[0138] --Query the records of abnormal CT scans in the past month

[0139] SELECT*FROM Results WHERE Result='Abnormal'AND ExamType='CT'ANDDate>'2023-02-01';

[0140] For the present invention, one only needs to input "Please help me search for the records of patients with abnormal CT scans in the past month." The system will automatically parse the natural language, identify key information, and construct the following SQL statement:

[0141] SELECT P.*

[0142] FROM Patients P

[0143] JOIN Examinations E ON P.PatientID=E.PatientID

[0144] JOIN Results R ON E.ExamID=R.ExamID

[0145] WHERE E.ExamType='CT'

[0146] AND R.Result='Abnormal'

[0147] AND E.Date>'2023-02-01';

[0148] The query results are then presented to the user in a table format, including information such as the patient’s name, gender, age, CT examination date, and examination results;

[0149] S42: using the constructed query statement generation model, converting the natural language input by the user into an SQL query statement and sending it to the database, recording and saving the query log; the query log includes the natural language input by the user, the SQL query statement, the execution time of the query statement and the query result.

[0150] S5: Analyze the high-frequency content in the query log and update the index of the relational database;

[0151] S51: Count the frequency of each query statement in the log, start from the query statement with the highest frequency and accumulate the frequencies one by one until the cumulative sum of the frequencies exceeds 30% of the total query frequency for the first time, and finally regard the query statements whose cumulative sum of frequencies is greater than 30% as high-frequency query statements;

[0152] S52: Count the execution time of high-frequency query statements, and calculate the index weight according to the frequency and execution time of each query statement. The calculation method is:

[0153]

[0154] Among them, Weight is the index weight, F is the frequency of each high-frequency query statement, Time is the execution time of the high-frequency query statement, max() is the maximum value, and min() is the minimum value;

[0155] S53: Add a new composite index for the fields in the high-frequency query statement according to the index weight, and update it to the entire B+ tree;

[0156] The present invention aims at high-frequency query statements, comprehensively considers two key factors, query frequency and execution time, to calculate index weights, and more accurately identifies indexes that need to be optimized; at the same time, new composite indexes are added to fields in high-frequency query statements, and the entire B+ tree is updated, providing better data access performance for the database.

[0157] The present invention is particularly suitable for large-scale medical device information integration scenarios with a wide variety of medical devices; the present invention integrates medical device information into a unified relational database through a semantic understanding model; through a query statement generation model, it allows users to conduct comprehensive information queries in a natural language manner; finally, through an index weight mechanism, combined with query content, data access performance is improved to provide strong data protection for medical institutions.

[0158] Embodiment 2: The present invention also discloses a medical device information integration system based on natural language processing, comprising:

[0159] Data export and preprocessing module: export data files and corresponding metadata from the software of multiple devices and perform preprocessing to obtain preprocessed data files and metadata;

[0160] Semantic understanding and database construction module: Build a semantic understanding model to parse metadata, and fill the parsed metadata into the relational database; at the same time, build a file system and store the data files corresponding to the metadata into the file system;

[0161] Index creation and association module: Create an index for the relational database and associate it with the data tables of each device;

[0162] Query statement generation module: builds a query statement generation model, queries medical device information, and saves query logs;

[0163] Index analysis and update module: Analyze the high-frequency content in the query log and update the index of the relational database.

[0164] It should be noted that the serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages and disadvantages of the embodiments. And the terms "including", "comprising" or any other variants thereof in this article are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "including a ..." does not exclude the presence of other identical elements in the process, device, article or method including the element.

[0165] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes a number of instructions for a terminal device (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in multiple embodiments of the present invention.

[0166] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A medical device information integration method based on natural language processing, It is characterized in that The following steps are involved: S1: exporting data files and corresponding metadata from software of multiple devices and preprocessing them to obtain preprocessed data files and metadata; S2: Build a semantic understanding model to parse the metadata, and fill the parsed metadata into the relational database; at the same time, build a file system and store the data files corresponding to the metadata into the file system; S3: Create an index for the relational database and associate it with the data table of each device; S4: Build a query statement generation model, query medical device information, and save the query log; S5: Analyze the high-frequency content in the query log and update the index of the relational database.

2. The medical device information integration method based on natural language processing according to claim 1, It is characterized in that The S1 step includes: S11: De-duplicate the exported data files and the corresponding metadata and delete empty files. Meanwhile, fill the missing values ​​with "NaN" to obtain the data files and metadata after preliminary preprocessing; S12: For the data file and metadata after preliminary preprocessing, a unique identifier is added as an associated identifier to obtain the preprocessed data file and metadata.

3. The medical device information integration method based on natural language processing according to claim 1, It is characterized in that The S2 step includes: Firstly, regular expressions are used to convert the preprocessed metadata into tabular data. Then, a semantic understanding model including the data understanding part and the entity alignment part is constructed, and the tabular data is input to obtain the semantic distance between the device column names. Secondly, the semantic distances between all device column names are traversed, and the column name feature vectors with a semantic distance smaller than the category threshold are merged into a new column name feature vector, and the column name feature vectors with a semantic distance greater than or equal to the category threshold are retained as new independent categories to obtain the total device attribute column name. Finally, a relational database and a file system are constructed based on the total device attribute column name, and the metadata of each device is stored in the relational database, and the data files corresponding to the preprocessed metadata are stored in the file system.

4. The medical device information integration method based on natural language processing according to claim 2, It is characterized in that The S2 step includes: S21: using regular expressions to convert the preprocessed metadata into table data; the columns of the table data are data categories, and the rows are data quantities; S22: Construct a semantic understanding model, which includes a data understanding part and an entity alignment part, wherein the calculation method of the semantic understanding part is: f 1 =CNN(x t ) f 2 =MaxPooling(f 1 )+AveragePooling(f 1 ) f 3 =Transformer(BiLSTM(x t ))+BiLSTM(x t ) f 4 =MLP(Concat(f 2 ,f 3 )) Among them, f 1 is the first convolution feature, CNN() is the convolutional neural network operation, x t is the input data, t is the time step index, f 2 is the second convolution feature, MaxPooling() is the maximum pooling operation, AveragePooling() is the global average pooling operation, and f 3 is the semantic temporal feature, Transformer() is the Transformer operation, BiLSTM() is the bidirectional long short-term memory network operation, and f 4 To fuse semantic features, Concat() is a concatenation operation and MLP() is a fully connected layer operation; The calculation method of the semantic understanding model is uniformly expressed as: f 4 =Understand(x t ) Among them, Comprehense() is the semantic understanding model operation; S23: The table data is input into the data understanding part of the semantic understanding model in columns, and the fused semantic features are calculated as the column name feature vectors of each device; S24: Input the column name feature vector of each device into the entity alignment part of the semantic understanding model to perform entity alignment to obtain the semantic distance between the device column names. The calculation method of the entity alignment part is: in, is the semantic similarity between the column names of the first and second devices, is the column name feature vector of the first device, is the column name feature vector of the second device, ∥∥ is the modulus length operation, is the semantic distance between the first and second device column names, β is the cosine adjustment parameter, exp() is the exponential operation of taking the natural constant, γ is the modulus adjustment parameter, and | | is the absolute value operation; S25: traverse the semantic distances between all device column names and store them in the distance matrix; then traverse the distance matrix to find all column name feature vector pairs whose semantic distance is less than the category threshold. The category threshold is calculated as follows: T=μ-α·σ Where T is the category threshold, μ is the average value of all current semantic distances, α is the weight coefficient, and σ is the standard deviation of all current semantic distances; For column name feature vectors with semantic distance smaller than the category threshold, they are merged into a new column name feature vector by taking the average of the two vectors and updating the distance matrix after merging. For the column name feature vectors whose semantic distance is greater than or equal to the category threshold, they are retained as new independent categories. After all column name feature vectors are processed according to the above steps, the total device attribute column name is obtained; S26: construct a relational database based on the total device attribute column name; at the same time, calculate the similarity between the total device attribute column name and the column name feature vector of each device, match them with the column name with the highest similarity, and then store the corresponding metadata in the relational database; S27: Build a file system, and store the data files corresponding to the metadata of each device into the file system; at the same time, copy the file address and store it into a relational database.

5. The medical device information integration method based on natural language processing according to claim 4, It is characterized in that The S3 step includes: S31: Select B+ tree as the index type, use the device number as the field, allocate space for each node of the B+ tree, store the device number and the pointer to the data table record; S32: Starting from the root node of the B+ tree, sort in ascending order, insert the device number and the pointer to the data table record into the leaf node of the B+ tree, and store them; S33: Repeat step S32 until all device numbers are inserted into the B+ tree; S34: Store the constructed B+ tree index in a relational database and associate it with the data table of each device.

6. The medical device information integration method based on natural language processing according to claim 1, It is characterized in that The S4 step includes: Firstly, the semantic understanding model is used to extract the fused semantic features of SQL statements and natural language input by users. Then, the concatenation operation and Transformer operation are used to fuse the features of the two to obtain the concatenated features. Secondly, the gated recurrent unit is simplified, the reset gate is merged with the update gate, and the neuron link used to operate the previous time step is deleted to obtain the reset update gate.

7. The medical device information integration method based on natural language processing according to claim 5, It is characterized in that The S4 step includes: S41: Build a query statement generation model, the calculation method is: f SQL =Comprehense(SQL) f X =Understand(X) h trans =Transformer(Concat(f SQL ,f X )) z t =sigmoid(W 1 h trans ) Among them, f SQL is the SQL fusion semantic feature, f X is the natural language fusion semantic feature, SQL is the SQL query statement, X is the natural language input by the user, and h trans is the splicing feature, z t is the reset update gate under the time step, sigmoid() is the sigmoid function, W 1 To reset the update parameter matrix, is the hidden state at time step t, tanh() is the hyperbolic tangent function, W 2 is the hidden state parameter matrix, h t-1 Generate features for the sentence at time step t-1, h t Generate features for the sentence at t time steps; Then generate feature h for the statement at time step t t , use Top-k Sampling to decode and get the SQL query statement; S42: using the constructed query statement generation model, converting the natural language input by the user into an SQL query statement and sending it to the database, recording and saving the query log; the query log includes the natural language input by the user, the SQL query statement, the execution time of the query statement and the query result.

8. The medical device information integration method based on natural language processing according to claim 7, It is characterized in that The S5 step includes: S51: Count the frequency of each query statement in the log, start from the query statement with the highest frequency and accumulate the frequencies one by one until the cumulative sum of the frequencies exceeds 30% of the total query frequency for the first time, and finally regard the query statements whose cumulative sum of frequencies is greater than 30% as high-frequency query statements; S52: Count the execution time of high-frequency query statements, and calculate the index weight according to the frequency and execution time of each query statement. The calculation method is: Among them, Weight is the index weight, F is the frequency of each high-frequency query statement, Time is the execution time of the high-frequency query statement, max() is the maximum value, and min() is the minimum value; S53: According to the index weight, a new composite index is added for the fields in the high-frequency query statement, and updated to the entire B+ tree.

9. A medical equipment information integration system based on natural language processing, It is characterized in that include: Data export and preprocessing module: export data files and corresponding metadata from the software of multiple devices and perform preprocessing to obtain preprocessed data files and metadata; Semantic understanding and database construction module: Build a semantic understanding model to parse metadata, and fill the parsed metadata into the relational database; at the same time, build a file system and store the data files corresponding to the metadata into the file system; Index creation and association module: Create an index for the relational database and associate it with the data tables of each device; Query statement generation module: builds a query statement generation model, queries medical device information, and saves query logs; Index analysis and update module: analyzes high-frequency content in query logs and updates the index of relational databases; To realize a medical device information integration method based on natural language processing as described in any one of claims 1-8.

Citation Information

Patent Citations

  • An Internet of things vulnerability information management system based on data virtualization

    CN109542846A

  • Multi-model data distributed storage and hierarchical query method based on semantic classification

    CN112800023A

  • Marine environment data-oriented retrieval method, system and equipment and medium

    CN113535788A

  • Distributed data braiding processing architecture

    CN117435792A

  • Intelligent medical service management system based on big data

    CN117936011A

Cited By

  • Self-adaptive construction method for database index

    CN119357195A

  • A self-adaptive construction method for database index

    CN119357195B