Multi-dimensional data classification and grading dynamic identification system in medical health field

Abstract

The invention discloses a multi-dimensional data classification and grading dynamic identification system in the field of medical health, and the system comprises a data preparation module which is used for collecting medical data, metadata of the medical data, and health information data elements; the classification and grading rule sorting module is used for defining a set of classification and grading rules conforming to medical business characteristics and regularly examining and updating the classification and grading rules; the field name standardization module is used for performing field name identification and field content correction on the metadata information of the data acquired by the data preparation module and the health information data element; and the classification and grading model module is used for respectively carrying out data classification and data grading on the data information after the field names are standardized. And the comprehensive judgment module is used for giving a final classification and grading result to the data set. By adopting the method and the device, the field identification can be more accurately carried out, so that the data classification is accurate, and the privacy of the patient is protected.

Description

Technical Field

[0001] This invention relates to the field of medical information processing technology, and in particular to a multi-dimensional data classification and grading dynamic identification system in the medical and health field. Background Technology

[0002] The healthcare industry generates diverse and complex data across numerous application scenarios. However, the storage and access of this data are subject to various legal and regulatory constraints, requiring compliance in both data collection and open sharing. While healthcare data provides crucial support for the construction of smart hospitals and holds immense value, it also contains a significant amount of personal privacy information, posing substantial data security risks. Therefore, effectively leveraging the value of healthcare data while simultaneously conducting data security governance to prevent data breaches and protect patient privacy is paramount in the era of big data in healthcare. Summary of the Invention

[0003] The purpose of this invention is to provide a multi-dimensional data classification and grading dynamic identification system in the medical and health field to solve the problems mentioned in the background art. When carrying out data classification and grading, firstly, data classification and grading rules should be formulated based on data management and usage needs, combined with national, industry, and local policies, standards, and requirements, to conform to the characteristics of medical business. Secondly, in the past, the classification and grading of internal enterprise data resources mostly relied on manual annotation or the use of text similarity algorithms for similarity matching, which either wastes a lot of manpower or has a high error rate. With the development of natural language processing algorithms, semantic recognition-based methods have gradually become a hot topic in text similarity research. Therefore, this study combines text similarity recognition algorithms with semantic recognition algorithms, using multiple natural language algorithms in a progressive manner. Based on the data content stored in the fields, the study corrects the best-matched health information data elements through data element value range rules and keyword features, thereby achieving more accurate field recognition. Simultaneously, in the data classification model, various privacy information recognition rules and sensitive record recognition rules are customized based on the stored content. The security level of data elements is refined according to the specific information stored, resulting in more accurate data classification. Furthermore, considering the potential data security risks caused by data fusion of personal information, key data in the diagnosis and treatment process, and other data, classification rules and security levels based on combined data elements are formulated. By judging the combined data elements, the dataset classification is dynamically adjusted, thereby more securely protecting patient privacy.

[0004] To achieve the above-mentioned objectives, this invention provides a multi-dimensional data classification and grading dynamic identification system in the medical and health field, comprising a data preparation module, a classification and grading rule sorting module, a field name standardization module, a classification and grading model module, and a comprehensive judgment module, wherein:

[0005] The data preparation module is used to collect medical data, medical data metadata, and health information data elements. The medical data metadata includes data attribute information, as well as the data organization, data domains and their relationships.

[0006] The classification and grading rules module is used to clarify a set of classification and grading rules that conform to the characteristics of medical business based on professional standards or industry regulations in the medical field, and to regularly review and update the classification and grading rules.

[0007] The field name standardization module is used to identify field names and correct field content for the metadata information and health information data elements collected by the data preparation module.

[0008] The classification and grading model module is used to classify and grade data information after the field names have been standardized.

[0009] The comprehensive judgment module is used to comprehensively consider the results of the classification and grading model module from the field-level grading, data record grading, data volume grading model and table business attribute classification results according to the data classification and grading guidelines in the medical and health field, and give the dataset a final classification and grading result.

[0010] Furthermore, the field name standardization module includes a field recognition model unit and a field content correction model unit, wherein:

[0011] The field name recognition model unit is used to recognize field names in the metadata information of medical data.

[0012] The field content correction model unit, based on the data content stored in the field, corrects the best matched health information data elements through the value range rules and keyword features of the data elements.

[0013] Furthermore, the identification process includes the following steps:

[0014] Step S1: Using Ratcliff / Obershelp similarity measure, the similarity between the table fields of medical data and the health information data elements is calculated, and the best health information data elements that meet the threshold conditions are matched.

[0015] Step S2: For fields that did not find a match in the previous step, first use the jieba word segmentation tool to segment the health information data elements, then use TF-IDF (term frequency-inverse file frequency) to calculate the word vector of the field in the corpus, and finally compare the cosine similarity between the table field of the medical data and the word vector of the health information data elements to find the matching results that meet the threshold conditions.

[0016] Step S3: For fields that did not match results in the previous step, the RoBERTA word vector deep semantic similarity analysis model is used for further identification. The similarity score between the field and the health information data element is calculated, and a threshold is set to filter out the health information data elements with the most similar meaning.

[0017] Furthermore, the classification and grading model module includes a data classification model unit and a data grading model unit, wherein:

[0018] The data classification model unit classifies the business attributes of a table based on the feature information contained in the table name and the feature information of the key fields contained in the table.

[0019] The data classification model unit is used to classify data information by field level, data record level, and data volume level.

[0020] Furthermore, the data classification model units include field-level hierarchical identification models, data record hierarchical models, and data volume hierarchical models, among which:

[0021] The field-level hierarchical recognition model is used to classify individual fields and combined fields based on the hierarchical information of labeled health information data elements.

[0022] The data record classification model dynamically identifies the sensitivity of each record in the medical data table based on the data information in the medical data table, combined with the sensitive information identification model, and classifies the record level for security.

[0023] The data volume classification model classifies data volume based on the number of people included in the medical data table.

[0024] Furthermore, the field-level hierarchical recognition model includes data element annotation sub-units, field hierarchical sub-units, and combined field hierarchical model sub-units, wherein:

[0025] The data element annotation sub-unit uses manual annotation to annotate health information data elements according to the hierarchical rules, forming a hierarchical rule knowledge base, which is continuously improved and supplemented.

[0026] The field classification sub-unit associates the standardized field names with the health information data elements in the data element annotation sub-unit to obtain the classification category;

[0027] The combined field hierarchical model subunit formulates hierarchical rules and security levels based on the data security risks caused by data fusion, and dynamically adjusts its hierarchical level by judging the combined data elements corresponding to the combined fields.

[0028] Furthermore, the data classification model unit first makes a strict judgment based on the field-level classification and data record classification results, applying the highest and most stringent criteria. Second, referring to the data classification and grading guidelines in the medical and health field, the dataset is divided into three levels—core, important, and general—based on the amount of data involved. The classification and grading of the data information table is then comprehensively judged to obtain the classification and grading results of the medical information table.

[0029] Compared with existing technologies, this system and method have the following advantages:

[0030] 1) A method combining multiple text similarity recognition algorithms and semantic recognition algorithms is adopted, using multiple natural language algorithms in a progressive manner to achieve more accurate field recognition;

[0031] 2) Based on the data content stored in the fields, the best matching health information data elements are corrected through the value range rules and keyword features of the data elements;

[0032] 3) Customize various privacy information identification rules and sensitive information identification rules based on the stored content, refine the security level of data elements according to the specific information stored, and thus classify data more accurately;

[0033] 4) Based on the potential data security risks caused by the fusion of personal information, key data in the diagnosis and treatment process, etc., we have formulated hierarchical rules and security levels based on combined data elements. By judging the combined data elements corresponding to the combined fields, we can dynamically adjust their hierarchical levels to more securely protect patients' personal privacy.

[0034] 5) By analyzing and identifying the dataset from multiple dimensions, including field name, field content, field combination, data record, data volume, and table business attributes, the dataset can be classified and graded more comprehensively. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the structure of a multi-dimensional data classification and grading dynamic recognition system in the field of medical and health care. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] like Figure 1The diagram shown is a structural schematic of the present invention. Taking a medical institution as an example, the clinical diagnosis and treatment database of the hospital is used as the original data source to implement the medical data classification and grading system provided by the present invention. Specifically, it includes a data preparation module, a classification and grading rule sorting module, a field name standardization module, a classification and grading model module, and a comprehensive judgment module, wherein:

[0038] The data preparation module collects the datasets that need to be classified and graded from the platform database. Taking the medical record table as an example, the number of samples collected is 137,095, and the collected fields include 42 fields such as age at visit, date of visit, department code, department name, resting heart rate upon admission, height upon admission, systolic blood pressure upon admission, diastolic blood pressure upon admission, body temperature upon admission, ward code upon admission, ward name upon admission, weight upon admission, and attending physician.

[0039] Classification and Grading Rules Review Module: Based on professional standards or industry regulations in the medical field, a set of classification and grading rules that conform to the characteristics of medical business is defined. The data classification and grading rules are shown in Table 1. Among them, the general data grading rules are shown in Table 2 based on the importance and risk level of the data, as well as the potential damage and impact on the subjects of personal health and medical data. The classification and grading rules are reviewed and updated regularly.

[0040] Table 1. Data Classification and Grading Dimensions:

[0041]

[0042] Table 2 General Data Classification Table:

[0043]

[0044] The field name standardization module mainly includes a field recognition model unit and a field content correction model unit.

[0045] The field name recognition model unit, based on the metadata information of the medical data, first uses the Ratcliff / Obershelp similarity measure to calculate the similarity between the table fields of the medical data and the health information data elements. The calculation steps are as follows:

[0046] Step S101: Determine the current longest matching block. Given two strings... and ,

[0047] ,

[0048] ,

[0049] Define the common subsequence length matrix ,express and The length of the longest common subsequence.

[0050] Using a dynamic programming recurrence relation, it can be expressed as follows:

[0051] ,

[0052] Among them, if the character and If a match is found, the length of the longest common subsequence is increased by 1. If no match is found, it is removed from the list. or Choose the maximum value from the two options.

[0053] Initialize the boundary conditions as follows:

[0054] ,

[0055] The final result is that the maximum matching block length is .

[0056] Step S102: Calculate the similarity using the following formula:

[0057] ,

[0058] in, The total length of the matched character blocks. For string Total length, For string The total length.

[0059] The calculation results are shown in Table 3. The table shows the top three health information data elements that match the similarity. Then, the data element with the highest score of 0.75 or higher is selected as the best matching data element. Table 4 shows the best matching data element in the final algorithm.

[0060] Table 3. Examples of top 3 matching results from the Ratcliff / Obershelp algorithm:

[0061]

[0062] Table 4. Examples of best matching results from the Ratcliff / Obershelp algorithm:

[0063]

[0064] For fields that did not meet the threshold condition (i.e., similarity score less than 0.75) in the previous step, the TF-IDF statistical model is used for further identification. First, the health information data elements are segmented using the jieba word segmentation tool. Then, TF-IDF is used to calculate the segmented vectors of the fields in the corpus. Finally, the cosine similarity between the table fields of medical data and the segmented vectors of health information data elements is compared to find the best matching result that meets the threshold condition (i.e., similarity score greater than or equal to 0.75). The formulas for calculating TF-IDF and cosine similarity are as follows:

[0065] Step S201, calculate TF (term frequency), as follows:

[0066] ,

[0067] in, For the current word segmentation, The current field name, For word segmentation In field name The number of times it appears in For field name The total number of middle participles.

[0068] Step S202, calculate IDF (Inverse Document Frequency), as follows:

[0069] ,

[0070] Where N is the total number of health data elements.

[0071] This represents the number of health data elements containing the word 't'.

[0072] Step S203, calculate the TF-IDF value, as shown below:

[0073] ,

[0074] Step S204, each field name One can be used Vector representation, as shown in the formula below:

[0075] ,

[0076] in, The total size of the vocabulary, for each word In the document There is a corresponding one in each. value.

[0077] Step S205: Calculate the similarity using the cosine similarity theorem. The formula is as follows:

[0078] ,

[0079] in, and For field name of vector,

[0080] Let be the Euclidean norm of a vector, calculated using the following formula:

[0081] ,

[0082] The dot product of two vectors is calculated using the following formula:

[0083] ,

[0084] The cosine similarity result ranges from [0,1], and the closer it is to 1, the more similar the two vectors are.

[0085] The algorithm results are shown in Table 5:

[0086] Table 5. Examples of best matching results from the TF-IDF algorithm:

[0087]

[0088] For fields that did not find a match in the previous step, the RoBERTA word embedding deep semantic similarity analysis model is used for further identification. The similarity score between the field and the health information data elements is calculated, and a threshold (similarity score greater than or equal to 0.75) is set to filter out the health information data elements with the most similar meaning. Then, a classification result is obtained based on the data element annotation units. Fields that do not meet the threshold condition are manually confirmed. The specific steps are as follows:

[0089] Step S301, Input representation, the specific process is as follows:

[0090] First, the string is tokenized, for the input sentence... Using BERT tokenizer, sentences are broken down into subwords:

[0091] ,

[0092] in Because a word can be broken down into multiple subwords.

[0093] Then, the input sequence is constructed, which includes the following parts:

[0094] ,

[0095] in, It is a sequence start marker, used for learning global information of the sentence; It is a sequence end marker used to distinguish sentence boundaries; assuming If the number of tokens is [number], then the input length is [length]. .

[0096] Step S302, embedding layer computation, the core task is to find the longest matching block between two strings and calculate the similarity based on the total length of the matching block. Its calculation formula is as follows:

[0097] Each token Mapped to a fixed dimension The embedding vector is represented as:

[0098] ,

[0099] in It refers to the size of the vocabulary list. It is the embedding vector obtained by looking up a table.

[0100] Position embedding Location of each token Location information is provided, represented as:

[0101] ,

[0102] in It is usually a fixed-position code calculated using sine and cosine functions, but it can also be trainable parameters.

[0103] chinese_roberta_wwm_ext is mostly used for single-segment tasks, where only a single segment is used, represented as:

[0104] ,

[0105] in This refers to the first sentence, all The values ​​are the same.

[0106] Final input sequence Transformed into an embedding matrix :

[0107] .

[0108] Step S303, construct the Transformer encoder, and... The embedding matrix is ​​input into the multi-layer encoder of the Transformer.

[0109] The core of Transformer is multi-head self-attention, where each token pays attention to information from other tokens in the sequence. First, a linear transformation is performed, and the input embedding... Projected as , represented as

[0110] ,

[0111] in It is a trainable matrix. It is the dimension of attention for each head.

[0112] Then, the similarity between tokens is calculated using dot products to generate attention weights, which are represented as follows:

[0113] ,

[0114] in, It is an attention score matrix. It is a scaling factor to prevent the score from becoming too large.

[0115] The result of concatenating multiple attention heads is represented as follows:

[0116] ,

[0117] in, It's the number of heads that attract attention. . It is the output weight.

[0118] In the feedforward network, the attention result at each location passes through two fully connected layers, as shown below:

[0119] ,

[0120] in, .

[0121] Residual connections and layer normalization are applied to each submodule, as follows:

[0122] ,

[0123] go through After layer Transformer encoding, the output sequence , represented as:

[0124]

[0125] Step S304: Embedded word vector extraction.

[0126] For each token Use the corresponding hidden state directly. As its contextual semantic representation:

[0127]

[0128] Reuse Hidden state As a representation of a complete sentence, it is shown as follows:

[0129]

[0130] In step S305, the similarity value between the embedded word vectors is calculated using cosine similarity, and a threshold is set to match the best result.

[0131] The results of the algorithm are shown in Table 6:

[0132] Table 6. Examples of best matching results from the RoBERTA algorithm:

[0133]

[0134] The field content correction model unit, based on the data content stored in the field, identifies the field content through regular expression rules such as the value range rules and keyword features of data elements, such as the data element identification rules for ID card, name, address, and disease diagnosis name. This prevents mismatches between field names and stored field content information, and corrects the best matched health information data elements. The identification rules are shown in Table 7.

[0135] Table 7 shows some examples of recognition rules based on field content:

[0136]

[0137] The classification and grading model module includes a data classification model unit and a data grading model unit;

[0138] The data classification model unit identifies the feature information contained in the table name and the feature information of the key fields contained in the table, and classifies the business attributes of the table. It is mainly divided into personal attribute data, health status data, medical application data, medical payment data, health resource data and public health data. The specific business attribute classification rules are shown in Table 8.

[0139] Table 8. Example of business attribute classification rules:

[0140]

[0141] The data classification model unit is divided into field-level classification identification model, data record classification model, and data volume classification model. The field-level classification identification model includes data element annotation sub-units, field classification sub-units, and combined field classification model sub-units. Specifically:

[0142] The data element annotation subunit formulates data element annotation rules based on classification and grading rules. Health and wellness information data elements are annotated manually. Annotators manually grade the health and wellness information data elements according to the annotation rules, forming a grading rule knowledge base, which is continuously improved and supplemented. Annotation examples are shown in Table 9.

[0143] Table 9. Examples of annotations for health information data elements:

[0144]

[0145] The field classification sub-unit associates the standardized field names with the health information data elements in the data element labeling unit to obtain the classification category. An example of field classification is shown in Table 10.

[0146] Table 10: Field Hierarchy Example:

[0147]

[0148] The combined field hierarchical model subunit, based on the potential data security risks arising from the fusion of personal information, key data during the diagnosis and treatment process, etc., has formulated hierarchical rules and security levels based on combined data elements, such as general level 5 data element combinations, general level 4 data element combinations, and general level 3 data element combinations. The hierarchical level is dynamically adjusted by judging the combined data elements corresponding to the combined fields. The hierarchical rules are shown in Table 11:

[0149] Table 11. Rules for Hierarchical Classification of Combined Fields:

[0150]

[0151] The data record grading model dynamically identifies the sensitivity of each record in the medical data table based on the data information, combined with sensitive information identification models, such as disease diagnosis and surgical record identification models. It then categorizes the records according to their security level. The sensitive information identification rules are shown in Table 12.

[0152] Table 12 Examples of rules for identifying sensitive information:

[0153]

[0154] In the data volume classification model, the data volume is classified according to the number of people included in the medical data table; the specific classification rules are shown in Table 13: Table 13 Data Volume Classification Rules:

[0155]

[0156] In the comprehensive judgment module, based on the table business attribute classification model in the data classification model unit, and the field-level hierarchical identification model, data record hierarchical model, and data volume hierarchical model in the data hierarchical model unit, the table business attribute category, field-level hierarchical, data record hierarchical, and data volume hierarchical are identified respectively. First, the field-level hierarchical and data record hierarchical results are judged with the highest and strictest judgment. Second, referring to the data classification and hierarchical guidelines in the medical and health field, as shown in Table 14, the dataset is divided into three levels—core, important, and general—based on the amount of important information such as sensitive personal information, physiological condition data, and biometric data. The classification and hierarchical of the data information table are then comprehensively judged to obtain the classification and hierarchical results of the medical information table, as shown in Table 15.

[0157] Table 14 Comprehensive Judgment Rules for Medical Data Classification and Grading:

[0158]

[0159] Table 15 Classification and grading results of the medical visit information form:

[0160]

[0161] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-dimensional data classification and grading dynamic identification system in the field of medical and health care, characterized in that, It includes a data preparation module, a classification and grading rule sorting module, a field name standardization module, a classification and grading model module, and a comprehensive judgment module, among which: The data preparation module is used to collect medical data, medical data metadata, and health information data elements. The medical data metadata includes data attribute information, as well as the data organization, data domains and their relationships. The classification and grading rules module is used to clarify a set of classification and grading rules that conform to the characteristics of medical business based on professional standards or industry regulations in the medical field, and to regularly review and update the classification and grading rules. The field name standardization module is used to identify field names and correct field content for the metadata information and health information data elements collected by the data preparation module. The classification and grading model module is used to classify and grade data information after the field names have been standardized. The comprehensive judgment module is used to comprehensively consider the results of the classification and grading model module from the field-level grading, data record grading, data volume grading model and table business attribute classification results according to the data classification and grading guidelines in the medical and health field, and give the dataset a final classification and grading result.

2. The multi-dimensional data classification and grading dynamic identification system in the medical and health field according to claim 1, characterized in that, The field name standardization module includes a field recognition model unit and a field content correction model unit, wherein: The field name recognition model unit is used to recognize field names in the metadata information of medical data; The field content correction model unit, based on the data content stored in the field, corrects the best matched health information data elements through the value range rules and keyword features of the data elements.

3. The multi-dimensional data classification and grading dynamic identification system in the medical and health field according to claim 2, characterized in that, The identification process includes the following steps: Step S1: Using Ratcliff / Obershelp similarity measure, the similarity between the table fields of medical data and the health information data elements is calculated, and the best health information data elements that meet the threshold conditions are matched. Step S2: For fields that did not find a match in the previous step, first use the jieba word segmentation tool to segment the health information data element, then use TF-IDF to calculate the word segmentation vector of the field in the corpus, and finally compare the cosine similarity between the table field of the medical data and the word segmentation vector of the health information data element to find the matching results that meet the threshold conditions. Step S3: For fields that did not match results in the previous step, the RoBERTA word vector deep semantic similarity analysis model is used for further identification. The similarity score between the field and the health information data element is calculated, and a threshold is set to filter out the health information data elements with the most similar meaning.

4. The multi-dimensional data classification and grading dynamic identification system in the medical and health field according to claim 1, characterized in that, The classification and grading model module includes a data classification model unit and a data grading model unit, wherein: The data classification model unit classifies the business attributes of a table based on the feature information contained in the table name and the feature information of the key fields contained in the table. The data classification model unit is used to classify data information by field level, data record level, and data volume level.

5. A multi-dimensional data classification and grading dynamic identification system in the medical and health field according to claim 4, characterized in that, The data classification model unit includes a field-level hierarchical identification model, a data record hierarchical model, and a data volume hierarchical model, among which: The field-level hierarchical recognition model is used to classify individual fields and combined fields based on the hierarchical information of labeled health information data elements. The data record classification model dynamically identifies the sensitivity of each record in the medical data table based on the data information in the medical data table, combined with the sensitive information identification model, and classifies the record level for security. The data volume classification model classifies data volume based on the number of people included in the medical data table.

6. A multi-dimensional data classification and grading dynamic identification system in the medical and health field according to claim 5, characterized in that, The field-level hierarchical recognition model includes data element annotation sub-units, field hierarchical sub-units, and combined field hierarchical model sub-units, wherein: The data element annotation sub-unit uses manual annotation to annotate health information data elements according to the hierarchical rules, forming a hierarchical rule knowledge base, which is continuously improved and supplemented. The field classification sub-unit associates the standardized field names with the health information data elements in the data element annotation sub-unit to obtain the classification category; The combined field hierarchical model subunit formulates hierarchical rules and security levels based on the data security risks caused by data fusion, and dynamically adjusts its hierarchical level by judging the combined data elements corresponding to the combined fields.

7. A multi-dimensional data classification and grading dynamic identification system in the medical and health field according to claim 4, characterized in that, The data classification model unit first makes a strict judgment based on the field-level classification and data record classification results, taking the highest level. Second, referring to the data classification and grading guidelines in the medical and health field, the dataset is divided into three levels: core, important, and general, based on the amount of data involved. The classification and grading of the data information table is then comprehensively judged to obtain the classification and grading results of the medical information table.

Images