Medical data isomerism-oriented patient similarity dynamic fusion method

Abstract

The invention discloses a medical data isomerism-oriented patient similarity dynamic fusion method, which comprises the steps of screening high-density core fields based on a medical database to construct a reference table, configuring a multi-source medical data table as a comparison table, comparing the comparison table with reference table fields, and determining an initial field set for comparison; a decision matrix is generated through a preset field screening mechanism, the initial field set is screened, and effective matching fields are generated; according to a field feature differentiation matching multi-modal similarity algorithm, calculating the similarity of corresponding fields in the reference table and the comparison table; and calculating the matching degree between the patient records by adopting a preset dichotomy model algorithm according to a similarity calculation result so as to realize dynamic fusion of the patient information. Therefore, accurate matching and similarity calculation of heterogeneous medical data and dynamic fusion of patient information are realized.

Description

Technical Field

[0001] The present invention relates to the field of medical data technology, and in particular to a patient similarity dynamic fusion method oriented to medical data heterogeneity. Background Art

[0002] In the healthcare field, patient matching research often encounters inaccurate matching due to the heterogeneity of multi-source medical data. Traditional methods suffer from limitations such as limited data coverage, fragmented matching methods, and limited processing capabilities for complex language structures. These methods result in incomplete medical information and hinder the implementation of life-cycle health management.

[0003] The problem of insufficient patient information matching accuracy caused by the multi-source heterogeneity of medical data is manifested in the following ways: First, patient information is highly sensitive, and data connectivity among hospitals is low. Large-scale medical data is generally difficult to obtain. Without real data as a basis, it is difficult to ensure model accuracy. For example, there is a high rate of discrepancy in field naming across institutions (e.g., the coding rate for "current address" and "residential address" in the address field is inconsistent). Second, data in the medical field has problems such as multi-source data, unstructured information expression, diversified expression of similar information, low data information density, and different data storage standards in different hospitals.

[0004] Among the related technologies, those designed based on small-scale tasks and small-scale data sets, insufficient data volume and data types will lead to poor generalization ability of the model. The application scope remains within the hospital or small scientific research tasks, only matching some simple texts, and lacking actual large-scale patient medical data matching experience; the algorithm based on small-scale text matching is not an industrialized model and is not suitable for application in large-scale text similarity matching tasks; the processing ability for complex language structures and meanings is limited, and text similarity matching technology is usually based on simple models such as word frequency and word vectors, which are difficult to handle complex language structures and semantic meanings, and the matching results may be misjudgments or omissions. Summary of the Invention

[0005] This application aims to solve the problems existing in the prior art in processing medical data heterogeneity and calculating patient similarity by providing a dynamic fusion method for patient similarity oriented to medical data heterogeneity.

[0006] This application provides a dynamic fusion method for patient similarity based on heterogeneous medical data, including: S101, based on the medical database, high-density core fields are screened to build a benchmark table, and a multi-source medical data table is configured as a comparison table to compare the fields with the benchmark table to determine the initial field set for comparison; S102, generating a decision matrix through a preset field screening mechanism, screening the initial field set, and generating valid matching fields; S103, using a multimodal similarity algorithm to perform differential matching based on field features, and calculating the similarity of corresponding fields in the reference table and the comparison table; S104, using a preset binary classification model algorithm, and calculating the matching degree between patient records based on the similarity calculation results, to achieve dynamic fusion of patient information.

[0007] Preferably, the core fields are obtained in the following manner: Perform data density analysis on each field in the medical database and calculate the proportion of non-null values ​​in each field; Select fields whose non-null value ratio is higher than the preset threshold as candidate core fields; Based on the medical business logic and the correlation between fields, fields that can fully reflect the patient's key information are selected from the candidate core fields and used as core fields to build a benchmark table.

[0008] Preferably, determining the initial field set for comparison includes: Collect medical data tables from different data sources as comparison tables; Perform semantic analysis on the field names and meanings of the benchmark table and comparison table, and establish a field semantic association model; Based on the field semantic association model, find the fields with similar semantics in the benchmark table and the comparison table to form an initial field set.

[0009] Preferably, the preset field screening mechanism includes: A1. Calculate the field sparsity and volatility; Among them, the sparsity rate P = U / S × 100%, where U is the number of valid records in the field (non-null values) and S is the total number of records. The sparsity rate is used to quantify the integrity of the field data. The volatility E∈{0,1,2}, where 0 represents an immutable field, 1 represents a low-variability field, and 2 represents a high-variability field, is used to evaluate the stability of the field value. A2. Generate a decision matrix based on field sparsity and volatility to filter out valid matching fields.

[0010] Preferably, the decision matrix ensures that the matching fields have both high coverage and stability, specifically: When P≥90% and E=0, the field is used as the valid matching field; When P ≥ 50% and E ≤ 1, the field is used as the valid matching field; When P<50% and E=0, the field is used as the valid matching field; When P<50% and E=1, choose whether to adopt it based on the field characteristics; When P<10% and E≤2, it is not adopted.

[0011] Preferably, the S104 specifically includes: Collect historical patient record matching data and construct a training dataset containing similarity features of patient record pairs and corresponding matching labels; Select the binary classification model algorithm and use the training dataset to train the binary classification model; The similarity result calculated currently is used as a feature input into the trained binary classification model, and the matching probability between the patient records is output, and the matching probability is used as the matching degree between the patient records; The matching degree is used to determine whether the patient records match. If they match, the corresponding patient records in the comparison table are fused with the patient records in the benchmark table to achieve dynamic fusion of patient information.

[0012] Preferably, the S103 specifically includes: B1. Standardize and pre-process the data content of valid matching fields, including spelling correction, removal of invalid characters, data formatting, data normalization, and null value data standardization; B2. Based on the content features of the effective matching fields, a multimodal similarity algorithm is used to differentially calculate the data similarity of the effective matching fields.

[0013] Preferably, B2 includes: C1. When the content features are in a fixed standardized format, a strict matching algorithm is used. If the content is the same, the similarity is 1, and if the content is different, the similarity is 0. C2. When the content feature is any one of the name field, address field, and other fields, a string similarity algorithm is used.

[0014] One or more technical solutions provided in this application have at least the following technical effects or advantages: By screening high-density core fields to build a benchmark table, the efficiency and accuracy of data processing are effectively improved; through the preset field screening mechanism and multimodal similarity algorithm, accurate matching and similarity calculation of heterogeneous medical data are achieved; through the binary classification model algorithm, dynamic calculation of matching between patient records is achieved, thereby realizing the dynamic fusion of patient information. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 Schematic diagram of the process of a dynamic fusion method of patient similarity oriented to medical data heterogeneity according to an embodiment of the present invention; Figure 2 Schematic diagram of a decision matrix according to an embodiment of the present invention. DETAILED DESCRIPTION

[0016] To facilitate understanding of the present invention, the present application will be described more comprehensively below with reference to the relevant drawings; the drawings show preferred embodiments of the present invention, but the present invention can be implemented in many different forms and is not limited to the embodiments described herein; on the contrary, the purpose of providing these embodiments is to enable a more thorough and comprehensive understanding of the disclosed content of the present invention.

[0017] It should be noted that the terms “vertical”, “horizontal”, “up”, “down”, “left”, “right” and similar expressions used herein are for illustrative purposes only and do not represent the only implementation method.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this invention pertains; the terms used herein in the specification of the present invention are for the purpose of describing specific embodiments only and are not intended to limit the present invention; the term "and / or" used herein includes any and all combinations of one or more of the associated listed items.

[0019] Example 1: Figure 1 It is a flowchart of a patient similarity dynamic fusion method for medical data heterogeneity according to an embodiment of the present invention.

[0020] like Figure 1 As shown in FIG, a patient similarity dynamic fusion method for medical data heterogeneity includes the following steps: S101, based on the medical database, high-density core fields are screened to build a benchmark table, and a multi-source medical data table is configured as a comparison table to compare with the benchmark table fields.

[0021] Specifically, the core fields are set as high-density fields in the entire population information table, including but not limited to ID number, name, gender, date of birth and current address. For example, the core fields can be: household address, contact person, contact number and main diagnosis. As an example, the core fields are determined as follows: Perform data density analysis on each field in the medical database and calculate the proportion of non-null values ​​in each field; Select fields whose non-null value ratio is higher than the preset threshold as candidate core fields; Based on the medical business logic and the correlation between fields, fields that can fully reflect the patient's key information are selected from the candidate core fields to build a benchmark table.

[0022] Based on the need to match missing information in the medical records of multiple visits for the entire population, benchmark table fields for patient information matching were selected from the database. These fields included information such as ID number, name, gender, date of birth, place of birth, household address, current address, contact person, contact number, and main diagnosis, which were used for patient information matching because of their highest data density.

[0023] Specifically, the multi-source medical data table contains four types of structured medical data tables: inpatient medical record homepage, resident death medical certificate, death information, and women's reproductive history, which are used to compare with the benchmark table to determine the fields that can be compared as the initial field set.

[0024] The comparison table is based on data covering the entire lifecycle of a patient, from birth to death, and overlapping fields with the benchmark table. Specifically, the multi-source medical data table covers all medical information about a patient, from birth to hospitalization and even death, and can serve as the basis for establishing a comprehensive population health record. These four aspects overlap with the information covered in the benchmark table, making them suitable for matching tasks.

[0025] In some embodiments, determining an initial set of fields for comparison includes: Collect medical data tables from different data sources as comparison tables; Perform semantic analysis on the field names and meanings of the reference table and comparison table, and establish a field semantic association model (a field mapping relationship model can also be established); Based on the field semantic association model, find the fields with similar semantics in the benchmark table and the comparison table to form an initial field set.

[0026] It should be noted that before comparing the fields in the reference table and the comparison table, it is necessary to establish a field mapping relationship between the reference table and the comparison table, compare the field attributes, and clarify the consistency of the field attributes. Specifically, the field attributes include the field name, code, and data type. The mapping mechanism is: Field name mapping (e.g., "Current Address" in the base table → field "k4" on the medical record front page); Field code mapping (e.g. gender code "xb_dm" → standard code "e2"); Data type alignment (text / numeric / date formats are unified); Exemplarily, the mapping relationship between the fields in the benchmark table and the comparison table can be: ID number-document number, current address-permanent address during lifetime, place of birth-place of birth; thus, a mapping relationship table between the fields in the benchmark table and the comparison table is established, with the aim of establishing a basis for the comparability of data across institutions and eliminating matching errors caused by terminology differences.

[0027] It should be noted that the effective matching fields screened need to be compared with the reference table and the benchmark table fields, and the field attributes of each field in the reference table and the benchmark table need to be determined. The field attributes include name, code, and data type. Select the fields that can be compared, and then execute step S102.

[0028] S102 , extracting an initial set of fields that can be compared from the comparison table and the reference table, generating a decision matrix through a preset field screening mechanism, screening the initial set of fields, and generating valid matching fields.

[0029] Specifically, the preset field filtering mechanisms include: A1. Calculate the field sparsity and volatility.

[0030] The sparsity ratio P = U / S × 100%, where U is the number of valid records in the field (non-null values) and S is the total number of records. The sparsity ratio is used to quantify the completeness of the field data. The higher the P value, the more complete the data. Volatility E∈{0,1,2}, where 0 represents an immutable field (e.g., date of birth), 1 represents a low-variable field (e.g., name), and 2 represents a high-variable field (e.g., contact number). Volatility is used to evaluate the stability of field values.

[0031] A2. Generate a decision matrix based on field sparsity and volatility, such as Figure 2 As shown, filter out valid matching fields.

[0032] The decision matrix ensures high coverage and stability of matching fields, solving the problem of matching failure caused by data sparsity. Specifically, it includes: When P ≥ 90% and E = 0, use (e.g., gender, date of birth); When P ≥ 50% and E ≤ 1, use (such as name, ID number, contact number); When P < 50% and E = 0, use (e.g., blood type, ethnicity); When P < 50% and E = 1, choose whether to adopt it based on the field characteristics (such as address, main diagnosis); When P < 10% and E ≤ 2, it was not used (such as occupation, date of consultation).

[0033] S103: Differentiate and match the multimodal similarity algorithm based on the field features, and use the matched similarity algorithm to calculate the similarities of the corresponding fields in the reference table and the comparison table.

[0034] Specifically, step S103 includes: B1. Standardize and pre-process the data content of valid matching fields, including spelling correction, removal of invalid characters, data formatting, data normalization, and null value data specification, to improve the accuracy and efficiency of matching results, including: Spelling correction: Handling simplified and traditional Chinese characters, name spaces, and mixed pinyin / English writing; Invalid character cleaning: Removing line breaks, tab characters, and delimiters; Format standardization: Unifying the formats of dates, phone numbers, ID numbers (such as YYYY - MM - DD); Semantic normalization: Mapping "male / M" to "male", and "Han" to "Han ethnicity"; Null value specification: Unifying the representation form of null values.

[0035] It should be noted that in actual applications, due to reasons such as input errors, there are often many problems with the collected original data, and these problem data will directly interfere with the correctness of the extracted knowledge. Therefore, in order to improve the accuracy and efficiency of the matching results, it is necessary to perform standardization processing on the data. Some common problems and corresponding solutions in data preprocessing are as follows: 1) Spelling correction: Correcting misspelled words, considering common problems in names, including simplified and traditional Chinese character issues, name spaces, using pinyin, using English, etc.; 2) Removing invalid characters: Removing symbols without actual semantics such as line breaks, tab characters, and delimiters in the text; 3) Data formatting: Converting data such as dates, phone numbers, and ID numbers into a unified standard format; 4) Data normalization: Mapping descriptions with the same meaning to a unified standard expression, such as mapping "male" and "M" for gender to "male", mapping "Han" for ethnicity to "Han ethnicity", and further standardizing allergic drugs and chronic diseases, etc. in combination with medical knowledge base information; 5) Null value data specification: Unifying the structure of null value data to avoid incorrect understanding by the model.

[0036] To further ensure the strict accuracy of data standardization, for texts with common fixed formats, regular expressions are used for batch processing, and for complex medical entities such as drugs, diseases, and symptoms, relevant term dictionaries are introduced for matching.

[0037] B2. Based on the content features of valid matching fields, use a multi - modal similarity algorithm to differentially calculate the data similarity of valid matching fields.

[0038] It can be understood that a suitable similarity calculation method is selected according to the field characteristics: Strict matching: For information with a relatively fixed content format (such as ID numbers, genders, etc.), a strict matching method is used; String similarity algorithms: For information with long content and no semantic sensitivity (such as addresses), string similarity algorithms are used, including the shortest edit distance (Levenshtein distance), Jaro / Jaro-Winkler algorithm, the Chinese improved edit distance algorithm, Jaccard similarity, and cosine similarity.

[0039] Specifically, step B2 includes: C1. When the content features are in a fixed standardized format, a strict matching algorithm is used. If the content is the same, the similarity is 1, and if the content is different, the similarity is 0. The strict matching algorithm is:

[0040] For information with relatively fixed content formats (such as ID card number, gender and other standardized fields), a strict matching algorithm is used, that is, if the content is the same, the similarity is 1, and if the content is different, the similarity is 0.

[0041] C2. When the content feature is long and semantically insensitive information, a string similarity algorithm is used.

[0042] Usually, the string similarity algorithm is set to the shortest edit distance. This similarity calculation is generally used for name matching. When dealing with name matching problems, the name data needs to be preprocessed into pinyin. Here, the xpinyin package in Python is called to pre-process the name data into pinyin, and then the edit distance similarity is used to calculate the name data in the data to be compared.

[0043] The shortest edit distance is the minimum number of editing operations required to transform two strings from one to the other, where editing operations include replacing characters, inserting characters, and deleting characters. The shortest edit distance between two strings can reflect the similarity between the two strings.

[0044] In the embodiment of the present invention, when the content feature is a name field, the string similarity algorithm does not need to pre-process the name into pinyin. Specifically, D1. Set 、 For any two strings, 、 They are the lengths of the two strings, that is, the number of characters (Chinese characters or English characters) contained in the strings. For strings 、 The edit distance of the string 、 Similarity Calculated by the following formula:

[0045] D2. Use character distance separately , Pinyin distance and Wubi distance By performing similarity calculation, we can get the character similarity, pinyin similarity and Wubi similarity of the string, which are recorded as 、 and The three similarities reflect the similarity between two strings from different aspects, but they are not independent of each other. For example, when using the Pinyin input method, the similarity caused by homophones or near-phonetic characters is repeated. Larger and Small, for the use of Wubi input method, similar repetition caused by similar characters, the two strings Smaller and If the difference between the two strings is small (such as completely similar duplicates), then and will be relatively large and close only when the two strings are not similar and repeated. and Based on this, whether the two strings are still The larger the value, the higher the similarity between the two strings.

[0046] D3. Since using either Pinyin distance or Wubi distance alone will result in certain errors, the original character distance is retained and the Pinyin distance and Wubi distance are fused with the character distance to form the fusion similarity. Use the following formula to calculate:

[0047] in, 、 The weight of each part, and , for example, the parameter can be .

[0048] In the embodiment of the present invention, when the content feature is an address field, the string similarity algorithm is set to Jaccard similarity, specifically: E1. Use regular expressions to split non-standard addresses into a three-level administrative structure of province / city / district. Combined with the geocoding database, map the text address to the national standard code and perform tail word segmentation on the remaining address segments.

[0049] Use regular expression splitting method to achieve word segmentation, use regular expression to split address information into province, city, county and township levels, and then split out the specific address at the end. Use regular expression to segment words, because there is no need to load a dictionary, and it is faster than Jieba word segmentation to process regular address text.

[0050] E2. Calculate the set similarity after splitting the address field content by administrative level.

[0051] Jaccard similarity is used to compare the differences and similarities between finite sample sets. In terms of sets, it is the ratio of the intersection to the union. In address text, after word segmentation, the number of identical address phrases is calculated as the ratio of the total number of address phrases in the two sets of address information to be compared. The Jaccard similarity calculation formula is:

[0052] X and Y are phrase sets of two character strings.

[0053] In the embodiment of the present invention, when the content feature is other fields (fields other than fixed format, name, and address), the string similarity algorithm is set to the cosine similarity algorithm, including: As mentioned above, depending on the results of word segmentation, the word segmentation is converted into word vectors through word2vec. The similarity measure of two vectors is reflected by the angle between the vectors. The smaller the angle, the larger the cosine value of the angle, and the stronger the similarity between the strings represented by the two vectors. The formula for calculating cosine similarity can refer to the relevant existing technology, and the present invention will not elaborate on or limit it. For example:

[0054] Therefore, different data fields can use different similarity calculation methods, which is determined by the characteristics of the field and the characteristics of the algorithm.

[0055] S104, using a preset binary classification model algorithm, and calculating the matching degree between patient records based on the similarity calculation results, to achieve dynamic fusion of patient information.

[0056] The specific steps are: Collect historical patient record matching data to construct a training dataset containing similarity features of patient record pairs and corresponding matching labels. Pre-training samples are constructed by manually setting weights based on the sparsity and volatility of field classifications to calculate a group of patients with high similarity. Then, manually label whether the patients are identical or not as training samples for the model. Select a binary classification model algorithm and train it using the training dataset. Use a binary classification model algorithm such as the AdaBoost (Adaptive Boosting) algorithm to pre-train reasonable weights for each information field and calculate whether the patient is a match. The similarity result calculated currently is used as a feature input into the trained binary classification model, and the matching probability between the patient records is output, and the matching probability is used as the matching degree between the patient records; The matching degree is used to determine whether the patient records match. If they match, the corresponding patient records in the comparison table are fused with the patient records in the benchmark table to achieve dynamic fusion of patient information.

[0057] The specific implementation is as follows: 1) Construction of pre-training samples: generating candidate sets (record pairs with similarity greater than 0.8) based on initial weights, and manually annotating a large number of positive and negative samples (same patient / different patient).

[0058] Training patient matching requires weight training, which requires samples. This requires constructing a sample of matching patients. Specifically, given the data, we manually set weights based on the sparsity and volatility of the field classifications to calculate a group of patients with high similarity. We then manually label whether or not these patients are identical, using these as training samples for the model.

[0059] 2) Calculate the similarity between patient records.

[0060] For matching items of two patient information, we select an appropriate binary classification model algorithm based on the field characteristics of the matching items and pre-train reasonable weights for each information field. Here, we use the AdaBoost (Adaptive Boosting) algorithm as an example to design a binary classification model to calculate whether the patients are matched.

[0061] Specifically, the AdaBoost method is an iterative algorithm that adds a new weak classifier in each round until a predetermined, sufficiently low error rate is achieved. Each training example is assigned a weight, indicating the probability of it being selected by a given classifier for inclusion in the training set. If a sample point has been accurately classified, its probability of inclusion in the next training set is reduced; conversely, if a sample point has been inaccurately classified, its weight is increased. In this way, the AdaBoost method can quickly focus on samples with high information content.

[0062] For example, the binary classification model is described as follows: ① Initialize the weight distribution of the training data (each sample): Initially, the weight of each sample is equal.

[0063] For example, initialize the weight distribution of training data: ,in, represents the weight of each sample at the first iteration, represents the weight of the first sample at the first iteration, is the total number of samples.

[0064] ② Iteratively train the weak classifier (decision tree): For the kth iteration, sample points are selected based on the weight of each sample to train the classifier. Based on this classifier, the weights of samples misclassified by it are increased, while the weights of samples correctly classified are decreased. The sample set with the updated weights is then used to train the next classifier. The entire training process continues in this iterative manner.

[0065] For example, Iterations: a) Using weighted distribution The training samples are used to learn and obtain the weak classifier: The performance index of the weak classifier is the value of the following error function To measure:

[0066] b) Calculate weak classifiers Weight , which means Importance in the final classifier:

[0067] along with Reduce, Gradually increases, that is, the classifier with a smaller error rate has a greater role in the final classifier.

[0068] c) Update the weight distribution of the training samples for the next iteration. The weights of samples that are misclassified are increased; the weights of samples that are correctly classified are decreased. The calculation method is as follows:

[0069] in, is the weight of the sample used in the next iteration, In the next iteration, The weight of the samples. Representative The class corresponding to the sample (1 or -1), Represents the weak classifier for the sample The classification of (1 or -1). If the classification is correct, then The value of is 1, otherwise it is -1. is the normalization factor, which is calculated as follows:

[0070] Next, The function acts on the summation result to obtain the final strong classifier :

[0071] in, The function (sign function) is a logical function used to determine the positive or negative sign of a real number.

[0072] According to the characteristics of AdaBoost, the information to be matched is recorded as a sample set The corresponding , given an initial value of 1 / N, to perform initial iterations, and finally obtain a strong classifier .

[0073] ③ Combine the trained weak classifiers into a strong classifier output: After the training process of each weak classifier is completed, the weak classifier with a small classification error rate has a greater voice and plays a greater decisive role in the final classification function, while the weak classifier with a large classification error rate has a smaller voice and plays a smaller decisive role in the final classification function. In other words, weak classifiers with low error rates account for a larger proportion in the final classifier, and vice versa.

[0074] Output the matching probability, and the confidence threshold is set to: >0.95 to be judged as the same patient.

[0075] Therefore, adaptive focusing is performed on difficult-to-match samples, improving robustness in small sample scenarios.

[0076] In summary, data processing efficiency is improved: by filtering high-density core fields to build a benchmark table, the amount of data that needs to be processed is effectively reduced, improving data processing efficiency; at the same time, the preset field screening mechanism and multimodal similarity algorithm further optimize the data processing process, making the calculation of patient similarity faster and more accurate. Improve the accuracy of patient similarity calculation: by multi-dimensionally filtering effective matching fields and differentially matching the multimodal similarity algorithm based on field characteristics, the calculation of patient similarity is made more accurate. This helps medical institutions reduce misjudgments and omissions when identifying patients and matching patient records, thereby improving the accuracy of medical data processing.

[0077] The technical solutions in the above embodiments of the present application have at least the following technical effects or advantages: For the selection of benchmark tables, based on data statistics, we specifically point out the fields that are more reasonable for matching patient information, and establish a new idea of ​​establishing a simple field system for cross-hospital patient information matching; Adopting diversified similarity calculation methods, combined with the characteristics of the field itself and data features, can ensure that patients can perform data similarity calculation according to the actual situation under different data missing conditions; For longer text data such as birthplace, registered address, current address, diagnosis, etc., we introduce more professional standard data for word segmentation and normalization. This makes the word segmentation results more accurate and ensures the accuracy of subsequent similarity calculations.

[0078] By screening high-density core fields to build a benchmark table, the efficiency and accuracy of data processing are effectively improved; through the preset field screening mechanism and multimodal similarity algorithm, accurate matching and similarity calculation of heterogeneous medical data are achieved; through the binary classification model algorithm, dynamic calculation of matching between patient records is achieved, thereby realizing the dynamic fusion of patient information.

[0079] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Various modifications and variations are readily apparent to those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A dynamic fusion method of patient similarity for medical data heterogeneity, characterized by: include: S101, based on the medical database, high-density core fields are screened to build a benchmark table, and a multi-source medical data table is configured as a comparison table to compare the fields with the benchmark table to determine the initial field set for comparison; S102, generating a decision matrix through a preset field screening mechanism, screening the initial field set, and generating valid matching fields; S103, using a multimodal similarity algorithm to perform differential matching based on field features, and calculating the similarity of corresponding fields in the reference table and the comparison table; S104, using a preset binary classification model algorithm, and calculating the matching degree between patient records based on the similarity calculation results, to achieve dynamic fusion of patient information.

2. The patient similarity dynamic fusion method for medical data heterogeneity according to claim 1 is characterized in that: The core fields are obtained as follows: Perform data density analysis on each field in the medical database and calculate the proportion of non-null values ​​in each field; Select fields whose non-null value ratio is higher than the preset threshold as candidate core fields; Based on the medical business logic and the correlation between fields, fields that can fully reflect the patient's key information are selected from the candidate core fields and used as core fields to build a benchmark table.

3. The patient similarity dynamic fusion method for medical data heterogeneity according to claim 2 is characterized in that: The determining of the initial field set for comparison includes: Collect medical data tables from different data sources as comparison tables; Perform semantic analysis on the field names and meanings of the benchmark table and comparison table, and establish a field semantic association model; Based on the field semantic association model, find the fields with similar semantics in the benchmark table and the comparison table to form an initial field set.

4. The patient similarity dynamic fusion method for medical data heterogeneity according to claim 3 is characterized in that: The preset field screening mechanism includes: A1. Calculate the field sparsity and volatility; Among them, the sparsity rate P = U / S × 100%, where U is the number of valid records in the field (non-null values) and S is the total number of records. The sparsity rate is used to quantify the integrity of the field data. The volatility E∈{0,1,2}, where 0 represents an immutable field, 1 represents a low-variability field, and 2 represents a high-variability field, is used to evaluate the stability of the field value. A2. Generate a decision matrix based on field sparsity and volatility to filter out valid matching fields.

5. The patient similarity dynamic fusion method for medical data heterogeneity according to claim 4 is characterized in that: The decision matrix ensures that the matching fields have both high coverage and stability, specifically: When P≥90% and E=0, the field is used as the valid matching field; When P ≥ 50% and E ≤ 1, the field is used as the valid matching field; When P<50% and E=0, the field is used as the valid matching field; When P<50% and E=1, choose whether to adopt it based on the field characteristics; When P<10% and E≤2, it is not adopted.

6. The method for dynamic fusion of patient similarities oriented to medical data heterogeneity according to claim 4, characterized in that: The S104 specifically includes: Collect historical patient record matching data and construct a training dataset containing similarity features of patient record pairs and corresponding matching labels; Select the binary classification model algorithm and use the training dataset to train the binary classification model; The similarity result calculated currently is used as a feature input into the trained binary classification model, and the matching probability between the patient records is output, and the matching probability is used as the matching degree between the patient records; The matching degree is used to determine whether the patient records match. If they match, the corresponding patient records in the comparison table are fused with the patient records in the benchmark table to achieve dynamic fusion of patient information.

7. The method for dynamic fusion of patient similarities oriented to medical data heterogeneity according to claim 5, characterized in that: The S103 specifically includes: B1. Standardize and pre-process the data content of valid matching fields, including spelling correction, removal of invalid characters, data formatting, data normalization, and null value data standardization; B2. Based on the content features of the effective matching fields, a multimodal similarity algorithm is used to differentially calculate the data similarity of the effective matching fields.

8. The method for dynamic fusion of patient similarities oriented to medical data heterogeneity according to claim 7, characterized in that: The B2 includes: C1. When the content features are in a fixed standardized format, a strict matching algorithm is used. If the content is the same, the similarity is 1, and if the content is different, the similarity is 0. C2. When the content feature is any one of the name field, address field, and other fields, a string similarity algorithm is used.

Images