Methods and systems for normalizing and mapping indicator names for heterogeneous medical texts
By extracting target entity triples from heterogeneous medical texts, combining literal morphology and dimensional logic, and using orthogonal projection technology to decouple core word meanings and contextual modifications, rule and semantic fusion is performed, solving the problem of poor correction ability for the normalization of indicators in heterogeneous medical texts and achieving highly accurate indicator mapping.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- WEST CHINA HOSPITAL SICHUAN UNIV
- Filing Date
- 2026-03-30
- Publication Date
- 2026-07-17
AI Technical Summary
Existing technologies have poor correction capabilities when normalizing heterogeneous medical text indicators. They are unable to cope with non-standard complex expressions, inverted word order, or newly emerging terms, and cannot use auxiliary information such as units of measurement for logical verification, leading to mismatches and incorrect associations.
A medical entity boundary recognition model is used to extract indicator names, units of measurement, and numerical attribute fields. Combined with a standard medical terminology database and a deep learning network, the probability vector of rule and semantic matching is calculated through literal morphological similarity, unit conversion consistency, and orthogonal projection algorithms. The Hadamard product and weighted correction are then performed to select the best standard terminology for mapping.
It achieves high-accuracy normalization of heterogeneous medical texts, reduces the impact of noise from modifying words, improves the accuracy and discriminative power of indicator normalization, and can handle difficult terms.
Smart Images
Figure CN121936473B_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of text processing, and in particular relates to a method and system for normalizing and mapping indicator names for heterogeneous medical texts. Background Technology
[0002] Due to varying information system construction standards and data entry habits across different medical institutions, the same clinical laboratory indicator often exists in multiple forms across different data sources. These forms may include different full names, abbreviations, English abbreviations, aliases, and even typos, frequently accompanied by differences or missing units of measurement. Traditional indicator normalization methods often rely on manually maintained large dictionaries or fuzzy string matching techniques based on edit distance. While useful for processing standard and normative texts, these methods have poor generalization capabilities and struggle to handle non-standard, complex expressions, inverted word order, or newly emerging terminology. Furthermore, they cannot utilize auxiliary information such as units of measurement for logical verification, making it difficult to meet the needs of large-scale medical data governance.
[0003] Most models vectorize the entire text, failing to distinguish between the core semantic meaning and contextual modifiers in indicator names. This leads to core features being easily obscured by modifier information in high-dimensional space, resulting in mismatches. Furthermore, different units of measurement can cause incorrect associations of indicators with entirely different numerical meanings. Existing fusion models often employ linear weighting strategies combined with rules and semantic features, lacking mechanisms for handling ambiguous cases and failing to utilize external knowledge such as standard terminology definitions for deep, fine-grained verification and error correction. Therefore, there is an urgent need for a heterogeneous medical text indicator normalization mapping method that can deeply integrate literal form, dimensional logic, and decoupled semantic features, and possess corrective capabilities. Summary of the Invention
[0004] This addresses the issue of poor correction capabilities in existing technologies when processing heterogeneous medical text indicator normalization mapping.
[0005] In the first aspect, this invention proposes a method for normalizing and mapping indicator names for heterogeneous medical texts, comprising the following steps:
[0006] The medical entity boundary recognition model is used to perform semantic parsing on the acquired heterogeneous medical text and extract the target entity triples consisting of the indicator name field, the unit of measurement field, and the numerical attribute field.
[0007] Based on the standard medical terminology database, a set of candidate standard terms associated with the indicator name field is retrieved, the literal similarity between the target entity and each candidate standard term is calculated, and the unit conversion consistency coefficient of the measurement unit field is combined to generate a rule matching probability vector of the first dimension.
[0008] The index name field is mapped to a high-dimensional vector space. The initial feature vector is decomposed into a main feature component representing the core word meaning and an auxiliary feature component representing the contextual modification using an orthogonal projection algorithm. The cosine similarity between the main feature component and the vector of each term in the candidate standard term set is calculated to obtain the semantic matching probability vector of the second dimension.
[0009] The fusion confidence score is obtained by calculating the Hadamard product of the rule matching probability vector and the semantic matching probability vector. If the fusion confidence score is within a preset fuzzy range, the definition text of the corresponding candidate standard term is mapped to a definition feature vector. The semantic correlation between the auxiliary feature component and the definition feature vector is calculated as a supplementary weight to perform weighted correction on the fusion confidence score. Based on the ranking result of the corrected confidence score, the standard term with the highest score is selected as the index normalization mapping output of the heterogeneous medical text.
[0010] Optionally, the step of using a medical entity boundary recognition model to perform semantic parsing on the acquired heterogeneous medical text and extracting target entity triples consisting of an indicator name field, a unit of measurement field, and a numerical attribute field includes:
[0011] Input heterogeneous medical text into a deep learning network model based on BERT-BiLSTM-CRF;
[0012] BERT pre-trained layers are used to convert characters in heterogeneous medical texts into word embedding vector sequences.
[0013] The word embedding vector sequence is input into a bidirectional long short-term memory network layer to extract the contextual feature vector of the text;
[0014] The context feature vector is decoded by a conditional random field layer, and the BIO tag sequence corresponding to each character is output.
[0015] Based on the BIO tag sequence, consecutive tag fields are extracted into indicator name field, unit of measurement field, and numerical attribute field.
[0016] Optionally, the calculation of the literal morphological similarity between the target entity and each candidate standard term includes:
[0017] Calculate the Levinstein edit distance between the target entity's indicator name field string and the candidate standard term name string;
[0018] The maximum value between the length of the indicator name field string and the length of the candidate standard term name string is used as the normalization factor.
[0019] The Levenstein edit distance is normalized using the normalization factor, and the difference between 1 and the normalization result is calculated to obtain a distance score representing literal similarity.
[0020] Optionally, generating the rule matching probability vector of the first dimension by combining the unit conversion consistency coefficient of the unit of measurement field includes:
[0021] Preconstruct a standard unit conversion matrix containing different unit categories, where units in the same row are mutually convertible units;
[0022] Determine whether the extracted unit of measurement field and the default unit of measurement corresponding to the candidate standard term are located in the same row of the standard unit conversion matrix;
[0023] If yes, then set the dimension conversion consistency coefficient to 1; if no, then set the dimension conversion consistency coefficient to 0.
[0024] Multiply the distance score of the literal similarity by the unit conversion consistency coefficient to obtain the element value in the rule matching probability vector.
[0025] Optionally, the step of using an orthogonal projection algorithm to decompose the initial feature vector into a main feature component representing the core word meaning and an auxiliary feature component representing contextual modification includes:
[0026] A basis vector set composed of general medical core terms is constructed, and the basis vector set is orthogonalized using the Gram-Schmidt process to generate a core subspace;
[0027] Calculate the projection vector of the initial feature vector of the indicator name field onto the core subspace, and determine the projection vector as the principal feature component;
[0028] Calculate the vector difference between the initial feature vector and the principal feature component, and determine the resulting difference vector as the auxiliary feature component.
[0029] Optionally, calculating the cosine similarity between the main feature component and the vectors of each term in the candidate standard term set to obtain the semantic matching probability vector of the second dimension includes:
[0030] Calculate the dot product between the principal feature components and the candidate standard term vectors;
[0031] Calculate the magnitude of the principal feature component and the magnitude of the candidate standard term vector respectively;
[0032] Divide the dot product by the product of the two moduli to obtain the cosine similarity value;
[0033] The cosine similarity values of all candidate standard terms are normalized using the Softmax function with a temperature parameter, and the semantic matching probability vector is output.
[0034] Optionally, the step of calculating the Hadamard product of the rule matching probability vector and the semantic matching probability vector to obtain the fusion confidence score, and if the fusion confidence score is within a preset fuzzy range, then the definition text of the corresponding candidate standard term is mapped to a definition feature vector, including:
[0035] Multiply the corresponding elements of the two probability vectors to obtain the fused confidence vector, and determine whether the maximum value in the vector is within the closed interval of 0.2 to 0.7;
[0036] If so, it is determined to be within the preset fuzzy range, and all explanatory definition texts of the candidate standard terms with the corresponding maximum value are extracted;
[0037] The explanatory definition text is segmented and stop words are removed. The word vector of each segment is found using the Word2Vec model, and the average value of all segmented word vectors is calculated as the definition feature vector.
[0038] Optionally, the step of calculating the semantic correlation between the auxiliary feature components and the defined feature vector as supplementary weights to weight and correct the fusion confidence includes:
[0039] The dot product of the auxiliary feature components and the defined feature vector is used as a scalar of semantic relevance.
[0040] The semantic relevance scalar is multiplied by a preset adjustment factor coefficient to obtain the supplementary weight value;
[0041] The supplementary weight values are added to the original fusion confidence to obtain the corrected fusion confidence.
[0042] In another aspect, the present invention also proposes an indicator name normalization and mapping system for heterogeneous medical texts, comprising the following modules:
[0043] The extraction module is used to perform semantic parsing on the acquired heterogeneous medical text using a medical entity boundary recognition model, and extract the target entity triplet consisting of the indicator name field, the unit of measurement field, and the numerical attribute field.
[0044] The generation module is used to retrieve a set of candidate standard terms associated with the indicator name field based on a standard medical terminology library, calculate the literal similarity between the target entity and each candidate standard term, and generate a rule matching probability vector of the first dimension by combining the unit of measurement field with the dimensional conversion consistency coefficient.
[0045] The first calculation module is used to map the index name field to a high-dimensional vector space, decompose the initial feature vector into a main feature component representing the core word meaning and an auxiliary feature component representing the contextual modification using an orthogonal projection algorithm, calculate the cosine similarity between the main feature component and the vector of each term in the candidate standard term set, and obtain the semantic matching probability vector of the second dimension.
[0046] The second calculation module is used to calculate the Hadamard product of the rule matching probability vector and the semantic matching probability vector to obtain the fusion confidence. If the fusion confidence is within a preset fuzzy range, the definition text of the corresponding candidate standard term is mapped to the definition feature vector, and the semantic correlation between the auxiliary feature component and the definition feature vector is calculated as a supplementary weight to perform weighted correction on the fusion confidence. Based on the ranking result of the corrected confidence, the standard term with the highest score is selected as the index normalization mapping output of the heterogeneous medical text.
[0047] Preferably, the step of using a medical entity boundary recognition model to perform semantic parsing on the acquired heterogeneous medical text and extracting target entity triples consisting of an indicator name field, a unit of measurement field, and a numerical attribute field includes:
[0048] Input heterogeneous medical text into a deep learning network model based on BERT-BiLSTM-CRF;
[0049] BERT pre-trained layers are used to convert characters in heterogeneous medical texts into word embedding vector sequences.
[0050] The word embedding vector sequence is input into a bidirectional long short-term memory network layer to extract the contextual feature vector of the text;
[0051] The context feature vector is decoded by a conditional random field layer, and the BIO tag sequence corresponding to each character is output.
[0052] Based on the BIO tag sequence, consecutive tag fields are extracted into indicator name field, unit of measurement field, and numerical attribute field.
[0053] Preferably, the calculation of the literal morphological similarity between the target entity and each candidate standard term includes:
[0054] Calculate the Levinstein edit distance between the target entity's indicator name field string and the candidate standard term name string;
[0055] The maximum value between the length of the indicator name field string and the length of the candidate standard term name string is used as the normalization factor.
[0056] The Levenstein edit distance is normalized using the normalization factor, and the difference between 1 and the normalization result is calculated to obtain a distance score representing literal similarity.
[0057] Preferably, the step of generating a rule-matching probability vector of the first dimension by combining the unit conversion consistency coefficient of the unit of measurement field includes:
[0058] Preconstruct a standard unit conversion matrix containing different unit categories, where units in the same row are mutually convertible units;
[0059] Determine whether the extracted unit of measurement field and the default unit of measurement corresponding to the candidate standard term are located in the same row of the standard unit conversion matrix;
[0060] If yes, then set the dimension conversion consistency coefficient to 1; if no, then set the dimension conversion consistency coefficient to 0.
[0061] Multiply the distance score of the literal similarity by the unit conversion consistency coefficient to obtain the element value in the rule matching probability vector.
[0062] Preferably, the step of decomposing the initial feature vector into a main feature component representing the core word meaning and an auxiliary feature component representing contextual modification using an orthogonal projection algorithm includes:
[0063] A basis vector set composed of general medical core terms is constructed, and the basis vector set is orthogonalized using the Gram-Schmidt process to generate a core subspace;
[0064] Calculate the projection vector of the initial feature vector of the indicator name field onto the core subspace, and determine the projection vector as the principal feature component;
[0065] Calculate the vector difference between the initial feature vector and the principal feature component, and determine the resulting difference vector as the auxiliary feature component.
[0066] Preferably, calculating the cosine similarity between the main feature component and the vectors of each term in the candidate standard term set to obtain the semantic matching probability vector of the second dimension includes:
[0067] Calculate the dot product between the principal feature components and the candidate standard term vectors;
[0068] Calculate the magnitude of the principal feature component and the magnitude of the candidate standard term vector respectively;
[0069] Divide the dot product by the product of the two moduli to obtain the cosine similarity value;
[0070] The cosine similarity values of all candidate standard terms are normalized using the Softmax function with a temperature parameter, and the semantic matching probability vector is output.
[0071] Preferably, the step of calculating the Hadamard product of the rule matching probability vector and the semantic matching probability vector to obtain the fusion confidence score, and if the fusion confidence score is within a preset fuzzy range, then the definition text of the corresponding candidate standard term is mapped to a definition feature vector, including:
[0072] Multiply the corresponding elements of the two probability vectors to obtain the fused confidence vector, and determine whether the maximum value in the vector is within the closed interval of 0.2 to 0.7;
[0073] If so, it is determined to be within the preset fuzzy range, and all explanatory definition texts of the candidate standard terms with the corresponding maximum value are extracted;
[0074] The explanatory definition text is segmented and stop words are removed. The word vector of each segment is found using the Word2Vec model, and the average value of all segmented word vectors is calculated as the definition feature vector.
[0075] Preferably, the step of calculating the semantic correlation between the auxiliary feature components and the defined feature vector as supplementary weights to weight and correct the fusion confidence includes:
[0076] The dot product of the auxiliary feature components and the defined feature vector is used as a scalar of semantic relevance.
[0077] The semantic relevance scalar is multiplied by a preset adjustment factor coefficient to obtain the supplementary weight value;
[0078] The supplementary weight values are added to the original fusion confidence to obtain the corrected fusion confidence.
[0079] This invention achieves standardized representation of unstructured data by extracting target entity triples. At the rule level, by utilizing the consistency check of unit conversion, it considers not only literal similarity but also eliminates interfering terms that are similar in name but different in substance at the attribute level. Using orthogonal projection technology, the text vector is decoupled into a principal component representing the core meaning and an auxiliary component representing the context. This allows the core semantics to be focused during the matching process, reducing the noise impact of modifying words on similarity calculation. Feature complementarity is achieved by calculating the Hadamard product of the rules and semantic probability vectors. For fuzzy matching intervals, the weight is adjusted using the correlation between the auxiliary component and the standard terminology definition text. This allows for the discovery of hidden contextual clues when dealing with difficult terms with low discriminative power, improving the accuracy of index normalization. Attached Figure Description
[0080] Figure 1 A flowchart of the first embodiment;
[0081] Figure 2 This is a schematic diagram of the hierarchical structure of the BERT-BiLSTM-CRF model;
[0082] Figure 3 This is a schematic diagram of the rule and semantic fusion correction mechanism. Detailed Implementation
[0083] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0084] In the first embodiment, the present invention proposes a method for normalizing and mapping indicator names for heterogeneous medical texts, such as... Figure 1 This includes the following steps:
[0085] S1. Use the medical entity boundary recognition model to perform semantic parsing on the acquired heterogeneous medical text and extract the target entity triplet consisting of the indicator name field, the unit of measurement field, and the numerical attribute field.
[0086] Specifically, a training corpus containing heterogeneous medical texts from multiple sources is constructed, and the indicator names, units of measurement, and numerical attributes are manually labeled using the BIO sequence labeling system to form a labeled dataset. The BioBERT pre-trained model is selected as the embedding layer to convert heterogeneous medical texts into word vector sequences and input them into a bidirectional long short-term memory network. The bidirectional contextual features of the text are extracted, and the feature sequences are decoded using a conditional random field layer to output the optimal label sequence. Based on the start and end markers in the label sequence, the corresponding text segments are segmented and assigned to the indicator name field, the unit of measurement field, and the numerical attribute field, respectively, thereby combining to generate structured target entity triples.
[0087] In an optional embodiment, the step of using a medical entity boundary recognition model to perform semantic parsing on the acquired heterogeneous medical text and extracting target entity triples consisting of an indicator name field, a unit of measurement field, and a numerical attribute field includes:
[0088] Input heterogeneous medical text into a deep learning network model based on BERT-BiLSTM-CRF;
[0089] BERT pre-trained layers are used to convert characters in heterogeneous medical texts into word embedding vector sequences.
[0090] The word embedding vector sequence is input into a bidirectional long short-term memory network layer to extract the contextual feature vector of the text;
[0091] The context feature vector is decoded by a conditional random field layer, and the BIO tag sequence corresponding to each character is output.
[0092] Based on the BIO tag sequence, consecutive tag fields are extracted into indicator name field, unit of measurement field, and numerical attribute field.
[0093] The BERT-BiLSTM-CRF network model includes BERT pre-trained layers, bidirectional long short-term memory network layers, and conditional random field layers, such as... Figure 2 The Conditional Random Field (CRF) layer is a global constraint and decoding model for sequence labeling tasks. It uses the emission scores of candidate labels corresponding to each character output by the BiLSTM layer as input parameters and initializes and trains a global state transition matrix within the model. This matrix learns and records the legality constraint weights for transitions between adjacent labels, such as assigning high weights to the transition from 'B-indicator' to 'I-indicator' and using negative penalty scores to prevent the 'I' label from directly appearing after the 'O' label. In the decoding phase, this layer uses a Viterbi dynamic programming algorithm to jointly solve for the local emission score of each character and the state transition scores between adjacent characters. It searches from all possible combinations of paths to find the legal sequence with the highest global probability score that strictly conforms to the BIO naming boundary rules, thus effectively eliminating label logic conflicts and gaps caused by local predictions. The training set for the medical entity boundary recognition model is a heterogeneous collection of medical texts containing BIO format annotation information. During the training phase, the CRF loss function is used to represent the difference between the predicted label sequence and the true label sequence, and the model parameters are iteratively updated using a backpropagation algorithm.
[0094] A pre-trained BERT-Base-Chinese model was constructed and loaded. Heterogeneous medical texts were tokenized, with a maximum sequence length (MaxSequenceLength) of 128. Padding was applied to any insufficient lengths, resulting in a word embedding vector sequence with dimensions [Batch_Size, 128, 768]. This sequence was then fed into a bidirectional LSTM layer with 256 hidden units and a dropout ratio of 0.5 to prevent overfitting. The bidirectional network detected forward and backward semantic dependencies of the text, outputting context feature vectors with dimensions [Batch_Size, 128, 512]. A state transition matrix was learned through a CRF layer to constrain the validity of label predictions, and the Viterbi algorithm was used to decode the globally optimal BIO label sequence. Although the extracted numerical attribute fields do not directly participate in subsequent indicator name normalization calculations, these fields, as clinical observations, will be bound to the normalized standard indicator names and units to form the structured medical data output.
[0095] S2, Based on the standard medical terminology database, retrieve the set of candidate standard terms associated with the indicator name field, calculate the literal similarity between the target entity and each candidate standard term, and combine the unit conversion consistency coefficient of the measurement unit field to generate the rule matching probability vector of the first dimension.
[0096] Specifically, the SNOMED CT and LOINC internationally recognized medical standard terminology sets are pre-imported to construct a standard medical terminology database. A keyword-based inverted index is established, and the extracted indicator name field is used as the query term. Fuzzy retrieval is performed in the inverted index to obtain the top K most relevant standard terms, forming a candidate standard terminology set. For each candidate standard term in this set, the Levinstein edit distance between the term and the indicator name field is calculated. The edit distance is then converted into a literal similarity value between 0 and 1 using a normalization formula. A dimensional dictionary containing common medical units and conversion ratios is constructed. The dictionary checks whether there is a direct or indirect conversion path between the target entity's unit of measurement field and the recommended unit of the candidate standard term. If such a path exists, the dimensional conversion consistency coefficient is set to 1; otherwise, it is set to 0. The literal similarity and the dimensional conversion consistency coefficient are weighted and summed or multiplied according to a preset ratio to obtain the rule score for each candidate term, which is then combined to generate the first-dimensional rule matching probability vector.
[0097] In an optional embodiment, calculating the literal morphological similarity between the target entity and each candidate standard term includes:
[0098] Calculate the Levinstein edit distance between the target entity's indicator name field string and the candidate standard term name string;
[0099] The maximum value between the length of the indicator name field string and the length of the candidate standard term name string is used as the normalization factor.
[0100] The Levenstein edit distance is normalized using the normalization factor, and the difference between 1 and the normalization result is calculated to obtain a distance score representing literal similarity.
[0101] In the technical implementation, a planning algorithm is used to construct a two-dimensional matrix D[i][j] to calculate the Levenstein edit distance, where D[i][j] represents the minimum number of single-character edit operations required to convert the first i characters of the source string into the first j characters of the target string. Assuming the target entity name is "fasting blood glucose" with length L1=4, and the candidate standard term is "fasting blood glucose value" with length L2=5, the calculated edit distance Dist is 1. The normalization factor is Max(L1,L2)=5. The literal morphological similarity score is calculated using the formula Score=1-(Dist / Max(L1,L2)), which is 0.8. If the two are completely identical, Dist is 0, and the score is 1; this process transforms discrete character differences into continuous numerical indicators within the interval [0,1].
[0102] In an optional embodiment, generating a rule-matching probability vector of the first dimension by combining the unit of measurement field with the dimensional conversion consistency coefficient includes:
[0103] Preconstruct a standard unit conversion matrix containing different unit categories, where units in the same row are mutually convertible units;
[0104] Determine whether the extracted unit of measurement field and the default unit of measurement corresponding to the candidate standard term are located in the same row of the standard unit conversion matrix;
[0105] If yes, then set the dimension conversion consistency coefficient to 1; if no, then set the dimension conversion consistency coefficient to 0.
[0106] Multiply the distance score of the literal similarity by the unit conversion consistency coefficient to obtain the element value in the rule matching probability vector.
[0107] A standard unit conversion matrix, stored in the form of a two-dimensional array or hash table, is pre-constructed in the system. Each row of the matrix represents an independent dimensional category. For example, the first row of the matrix is defined as mass concentration ['mg / L', 'g / L', 'ng / mL', 'ug / L'], and the second row is defined as molar concentration ['mmol / L', 'umol / L']. During processing, the unit field of the target entity is extracted and compared with the default unit of the candidate terms. Boolean matching is performed by traversing the matrix. If the two are not in the same row, it indicates a conflict in dimensional meaning, and the dimensional conversion consistency coefficient is forced to be set to 0; if they are in the same row, the coefficient is set to 1. The aforementioned literal similarity score is multiplied by this coefficient to obtain the rule matching probability. If the coefficient is 0, the matching probability of that item is zero, and erroneous mappings that are literally similar but have contradictory meanings are eliminated.
[0108] S3, map the index name field to a high-dimensional vector space, and use the orthogonal projection algorithm to decompose the initial feature vector into a main feature component representing the core word meaning and an auxiliary feature component representing contextual modification. Calculate the cosine similarity between the main feature component and the vector of each term in the candidate standard term set to obtain the semantic matching probability vector of the second dimension.
[0109] Specifically, the indicator name field is input into a BERT deep neural network model fine-tuned with medical corpus, and the mean vector of the last hidden state is extracted as the high-dimensional initial feature vector. A large number of core medical term vectors without modifiers are collected, and principal component analysis is used to construct a low-dimensional orthogonal subspace representing the core semantics and the corresponding orthogonal projection matrix. The initial feature vector is multiplied by the orthogonal projection matrix to obtain the projection vector on the subspace, which is the principal feature component. The principal feature component is subtracted from the initial feature vector to obtain the difference vector, which is the auxiliary feature component representing contextual modification. Similarly, using the BERT model, each term in the candidate standard term set is transformed into a standard term vector, and the cosine similarity value between the principal feature component and each standard term vector is calculated. All similarity values are normalized through the Softmax function to obtain the semantic matching probability vector of the second dimension.
[0110] The BERT deep neural network model fine-tuned based on the medical corpus is based on the BERT-Base-Chinese architecture, including a 12-layer Transformer encoder, 768-dimensional hidden layers, and a 12-head self-attention mechanism. A fine-tuning dataset based on the LOINC / HL7 standard indicator terminology library is selected, supplemented by electronic medical records and SNOMED CT standard terminology texts. The masked language model is used as the core training objective. The learning rate is set to 5e-5, the batch size to 16, and the number of fine-tuning iterations to 10 rounds to complete the adaptation to the medical field, so that the semantic encoding of medical indicator names such as fasting blood glucose and serum creatinine is more in line with clinical scenarios.
[0111] In an optional embodiment, the step of using an orthogonal projection algorithm to decompose the initial feature vector into a main feature component representing the core word meaning and an auxiliary feature component representing contextual modification includes:
[0112] A basis vector set composed of general medical core terms is constructed, and the basis vector set is orthogonalized using the Gram-Schmidt process to generate a core subspace;
[0113] Calculate the projection vector of the initial feature vector of the indicator name field onto the core subspace, and determine the projection vector as the principal feature component;
[0114] Calculate the vector difference between the initial feature vector and the principal feature component, and determine the resulting difference vector as the auxiliary feature component.
[0115] The Word2Vec network model employs a skip-word model or continuous bag-of-words model architecture, consisting of an input layer, a projection layer, and an output layer. The training set utilizes a large-scale general medical text corpus. The training process maximizes conditional probabilities through a sliding window context prediction mechanism and reduces computational complexity using negative sampling or hierarchical softmax techniques. The model's input is the one-hot encoded vector of medical vocabulary, and the output is a densely distributed representation of that vocabulary in the vector space.
[0116] A basis vector set is constructed by filtering high-frequency medical entity names such as protein, glucose, cell, and enzyme. , , , The V is represented using fixed-dimensional vectors generated by Word2Vec or BERT. The Gram-Schmidt orthogonalization process is applied to transform V into an orthogonal basis matrix U, and the space spanned by U is the core semantic subspace S. To ensure the consistency of the closed-loop computation space, the extracted indicator name fields are independently encoded, independent of the NER model's context vector, to generate initial feature vectors x of the same dimension. Principal feature components... Through projection matrix Calculated, i.e. This component retains the essence of the indicator; the auxiliary feature component is calculated as follows: This component contains modifiers such as fasting, post-meal, and intravenous information. This orthogonal decomposition decouples composite concepts in the vector space.
[0117] In an optional embodiment, calculating the cosine similarity between the main feature component and the vectors of each term in the candidate standard term set to obtain the semantic matching probability vector of the second dimension includes:
[0118] Calculate the dot product between the principal feature components and the candidate standard term vectors;
[0119] Calculate the magnitude of the principal feature component and the magnitude of the candidate standard term vector respectively;
[0120] Divide the dot product by the product of the two moduli to obtain the cosine similarity value;
[0121] The cosine similarity values of all candidate standard terms are normalized using the Softmax function with a temperature parameter, and the semantic matching probability vector is output.
[0122] Let the principal eigencomponents obtained from the decomposition be vectors. The candidate standard term vector is B, which is obtained by inputting the text of each standard term in the candidate standard term set into a model that is from the same source as the basis vector set V and performing the same processing. The cosine similarity calculation formula is... The calculation results are located in the interval [-1, 1]. To enhance high similarity terms and suppress noise in low similarity terms, a Softmax function with a temperature parameter T is used for normalization. The temperature parameter T is preferably set between 0.05 and 0.1. A lower temperature coefficient makes the probability distribution sharper, amplifying the probability value of the best match, giving the model higher discriminative power in semantic matching decisions, and ensuring that the sum of the output semantic matching probability vectors is strictly 1.
[0123] S4, calculate the Hadamard product of the rule matching probability vector and the semantic matching probability vector to obtain the fusion confidence. If the fusion confidence is within a preset fuzzy range, map the definition text of the corresponding candidate standard term to a definition feature vector, calculate the semantic correlation between the auxiliary feature component and the definition feature vector as a supplementary weight, and perform weighted correction on the fusion confidence. Based on the ranking result of the corrected confidence, select the standard term with the highest score as the index normalization mapping output of the heterogeneous medical text.
[0124] Specifically, the elements at corresponding positions in the rule matching probability vector and the semantic matching probability vector are multiplied one by one to obtain a fusion confidence vector containing the preliminary scores of each candidate word. A lower and upper threshold for the fuzzy interval are set, and the fusion confidence vector is iterated. When the fusion confidence value of a candidate standard term falls between the lower and upper thresholds, a correction mechanism is triggered. The detailed explanatory definition text of the candidate standard term is retrieved from the standard medical terminology database, and the definition text is encoded into a fixed-length definition feature vector using the Sentence-BERT sentence vector model. The dot product of the auxiliary feature components obtained above and the definition feature vector is calculated as the semantic relevance. The semantic relevance is multiplied by a preset adjustment coefficient and added to the original fusion confidence to complete the weighted correction of the confidence of the candidate standard term. Figure 3.
[0125] Sort the confidence values corresponding to all terms in the candidate standard term set in descending order, and select the candidate standard term with the highest confidence value; extract the standard code and standard name of the standard term, and output the standard code and standard name as the only normalized mapping result of the original metric name in the heterogeneous medical text, and store it in the structured database to complete the mapping process.
[0126] In an optional embodiment, calculating the Hadamard product of the rule matching probability vector and the semantic matching probability vector to obtain the fused confidence. If the fused confidence is within a preset fuzzy interval, map the definition text of the corresponding candidate standard term to a definition feature vector, including:
[0127] Multiply the corresponding elements of the two probability vectors to obtain a fused confidence vector, and determine whether the maximum value in the vector is within the closed interval of 0.2 to 0.7;
[0128] If so, it is determined that it is within the preset fuzzy interval, and all explanatory definition texts of the candidate standard term corresponding to the maximum value are extracted;
[0129] Segment the explanatory definition text and remove stop words, find the word vectors of each segment through the Word2Vec model, and calculate the average value of all segment word vectors as the definition feature vector.
[0130] Perform the Hadamard product operation, that is , where R is the rule probability vector and S is the semantic probability vector. The preset fuzzy interval threshold is [0.2, 0.7]. If the maximum value of the calculated fused confidence falls within this interval, it indicates that there is ambiguity in the current match and enhanced verification needs to be triggered. Call the knowledge base interface to obtain the LOINC or ICD standard definition text of the candidate term, such as "Measure the concentration of glucose in the blood...". Use the Chinese word segmentation tool to segment the definition text, and remove meaningless words such as "of", "is", "in" according to the stop word list. Use the pre-trained model homologous to the above to find the vectors of the remaining content words, and calculate the arithmetic mean MeanPooling of the vectors. The generated 300-dimensional vector is the definition feature vector representing the deep semantics of the term.
[0131] In an optional embodiment, calculating the semantic association degree between the auxiliary feature component and the definition feature vector as a supplementary weight to perform weighted correction on the fused confidence, including:
[0132] Calculate the dot product of the auxiliary feature component and the definition feature vector as the semantic association degree scalar;
[0133] The semantic relevance scalar is multiplied by a preset adjustment factor coefficient to obtain the supplementary weight value;
[0134] The supplementary weight values are added to the original fusion confidence to obtain the corrected fusion confidence.
[0135] Using the auxiliary feature components obtained above With defining eigenvectors Perform the dot product operation, that is The dot product result indicates whether modifiers in the original text are implied in the definition of the standard term. Input adjustment factor. The preferred range for the factor is 0.1 to 0.3, and the supplementary weights are calculated. The corrected formula is as follows: ,in, This is the corrected fusion confidence level. This is the fusion confidence level before correction. If the modifier is highly relevant to the definition, the dot product is positive, and the confidence level increases; if the modifier conflicts with the definition, the confidence level decreases. If the fusion confidence level above is not within the preset fuzzy range, this step is skipped, and the original fusion confidence level is used as the corrected confidence level, completing the causal logic loop.
[0136] To verify the effectiveness of this invention, the following hardware was used: a server configured with an NVIDIA Tesla V100 graphics processor and 64GB of memory. The software environment was based on the Linux operating system, Python programming language, and PyTorch deep learning framework. The dataset consisted of 50,000 annotated corpora containing heterogeneous medical texts, divided into training, validation, and test sets in an 8:1:1 ratio. The model parameters were set as follows: maximum sequence length of 128, batch size of 32, learning rate of 2e-5, and 20 iterations.
[0137] Experimental group: Using the complete scheme described in this invention, including the BERT-BiLSTM-CRF model, orthogonal projection decomposition and fuzzy interval correction, the precision rate was 94.5%, the recall rate was 93.8%, and the F1 score was 94.1%.
[0138] Ablation group 1: The orthogonal projection algorithm was removed, and the similarity was calculated using only the original vectors without feature vector decomposition. The precision was 91.2%, the recall was 89.5%, and the F1 score was 90.3%.
[0139] Ablation group 2: The steps of removing ambiguous regions and defining feature vectors were corrected, with a precision of 92.5%, a recall of 91.0%, and an F1 score of 91.7%.
[0140] Control group: The BERT pre-trained layer was replaced with the traditional Word2Vec static word vector layer, without subsequent feature decomposition and fuzzy correction. The precision was 86.4%, the recall was 84.2%, and the F1 score was 85.3%.
[0141] The experimental group showed the best overall performance, with an F1 score of 94.1%, higher than all other groups. Ablation group 2 performed second best, with an F1 score of 91.7%.
[0142] In a second embodiment, the present invention also provides an indicator name normalization and mapping system for heterogeneous medical texts, comprising the following modules:
[0143] The extraction module is used to perform semantic parsing on the acquired heterogeneous medical text using a medical entity boundary recognition model, and extract the target entity triplet consisting of the indicator name field, the unit of measurement field, and the numerical attribute field.
[0144] The generation module is used to retrieve a set of candidate standard terms associated with the indicator name field based on a standard medical terminology library, calculate the literal similarity between the target entity and each candidate standard term, and generate a rule matching probability vector of the first dimension by combining the unit of measurement field with the dimensional conversion consistency coefficient.
[0145] The first calculation module is used to map the index name field to a high-dimensional vector space, decompose the initial feature vector into a main feature component representing the core word meaning and an auxiliary feature component representing the contextual modification using an orthogonal projection algorithm, calculate the cosine similarity between the main feature component and the vector of each term in the candidate standard term set, and obtain the semantic matching probability vector of the second dimension.
[0146] The second calculation module is used to calculate the Hadamard product of the rule matching probability vector and the semantic matching probability vector to obtain the fusion confidence. If the fusion confidence is within a preset fuzzy range, the definition text of the corresponding candidate standard term is mapped to the definition feature vector, and the semantic correlation between the auxiliary feature component and the definition feature vector is calculated as a supplementary weight to perform weighted correction on the fusion confidence. Based on the ranking result of the corrected confidence, the standard term with the highest score is selected as the index normalization mapping output of the heterogeneous medical text.
[0147] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can be referred to each other.
[0148] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A method for normalizing and mapping indicator names for heterogeneous medical texts, characterized in that, Includes the following steps: The heterogeneous medical text is semantically parsed using a medical entity boundary recognition model to extract target entity triples consisting of an indicator name field, a unit of measurement field, and a numerical attribute field. The medical entity boundary recognition model is a deep learning network model based on BERT-BiLSTM-CRF. Based on a standard medical terminology database, a set of candidate standard terms associated with the indicator name field is retrieved. The literal similarity between the target entity and each candidate standard term is calculated, and the consistency coefficient of the unit of measurement field is combined to generate a rule matching probability vector for the first dimension. The calculation of the literal similarity between the target entity and each candidate standard term includes: calculating the Lewinstein edit distance between the indicator name field string of the target entity and the candidate standard term name string; obtaining the maximum value between the length of the indicator name field string and the length of the candidate standard term name string as a normalization factor; normalizing the Lewinstein edit distance using the normalization factor; and calculating the difference between 1 and the normalization result to obtain a distance score representing the literal similarity. The index name field is mapped to a high-dimensional vector space. The initial feature vector is decomposed into a main feature component representing the core word meaning and an auxiliary feature component representing the contextual modification using an orthogonal projection algorithm. The cosine similarity between the main feature component and the vector of each term in the candidate standard term set is calculated to obtain the semantic matching probability vector of the second dimension. The fusion confidence score is obtained by calculating the Hadamard product of the rule matching probability vector and the semantic matching probability vector. If the fusion confidence score is within a preset fuzzy range, the definition text of the corresponding candidate standard term is mapped to a definition feature vector. The semantic correlation between the auxiliary feature component and the definition feature vector is calculated as a supplementary weight to perform weighted correction on the fusion confidence score. Based on the ranking result of the corrected confidence score, the standard term with the highest score is selected as the index normalization mapping output of the heterogeneous medical text.
2. The method according to claim 1, characterized in that, The method utilizes a medical entity boundary recognition model to perform semantic parsing on the acquired heterogeneous medical text, extracting target entity triples composed of an indicator name field, a unit of measurement field, and a numerical attribute field, including: Input heterogeneous medical text into a deep learning network model based on BERT-BiLSTM-CRF; BERT pre-trained layers are used to convert characters in heterogeneous medical texts into word embedding vector sequences. The word embedding vector sequence is input into a bidirectional long short-term memory network layer to extract the contextual feature vector of the text; The context feature vector is decoded by a conditional random field layer, and the BIO tag sequence corresponding to each character is output. Based on the BIO tag sequence, consecutive tag fields are extracted into indicator name field, unit of measurement field, and numerical attribute field.
3. The method according to claim 2, characterized in that, The step of generating the first-dimensional rule-matching probability vector by combining the unit of measurement field with the dimensional conversion consistency coefficient includes: Preconstruct a standard unit conversion matrix containing different unit categories, where units in the same row are mutually convertible units; Determine whether the extracted unit of measurement field and the default unit of measurement corresponding to the candidate standard term are located in the same row of the standard unit conversion matrix; If yes, then set the dimension conversion consistency coefficient to 1; if no, then set the dimension conversion consistency coefficient to 0. Multiply the distance score of the literal similarity by the unit conversion consistency coefficient to obtain the element value in the rule matching probability vector.
4. The method according to claim 1, characterized in that, The method of decomposing the initial feature vector into main feature components representing the core word meaning and auxiliary feature components representing contextual modification using the orthogonal projection algorithm includes: A basis vector set composed of general medical core terms is constructed, and the basis vector set is orthogonalized using the Gram-Schmidt process to generate a core subspace; Calculate the projection vector of the initial feature vector of the indicator name field onto the core subspace, and determine the projection vector as the principal feature component; Calculate the vector difference between the initial feature vector and the principal feature component, and determine the resulting difference vector as the auxiliary feature component.
5. The method according to claim 1, characterized in that, The calculation of the cosine similarity between the main feature component and the vectors of each term in the candidate standard term set, to obtain the semantic matching probability vector of the second dimension, includes: Calculate the dot product between the principal feature components and the candidate standard term vectors; Calculate the magnitude of the principal feature component and the magnitude of the candidate standard term vector respectively; Divide the dot product by the product of the two moduli to obtain the cosine similarity value; The cosine similarity values of all candidate standard terms are normalized using the Softmax function with a temperature parameter, and the semantic matching probability vector is output.
6. The method according to claim 1, characterized in that, The calculation of the Hadamard product of the rule matching probability vector and the semantic matching probability vector yields the fusion confidence score. If the fusion confidence score falls within a preset fuzzy range, the definition text of the corresponding candidate standard term is mapped to a definition feature vector, including: Multiply the corresponding elements of the two probability vectors to obtain the fused confidence vector, and determine whether the maximum value in the vector is within the closed interval of 0.2 to 0.7; If so, it is determined to be within the preset fuzzy range, and all explanatory definition texts of the candidate standard terms with the corresponding maximum value are extracted; The explanatory definition text is segmented and stop words are removed. The word vector of each segment is found using the Word2Vec model, and the average value of all segmented word vectors is calculated as the definition feature vector.
7. The method according to claim 4, characterized in that, The step of calculating the semantic correlation between the auxiliary feature components and the defined feature vector as supplementary weights to weight and correct the fusion confidence includes: The dot product of the auxiliary feature components and the defined feature vector is used as a scalar of semantic relevance. The semantic relevance scalar is multiplied by a preset adjustment factor coefficient to obtain the supplementary weight value; The supplementary weight values are added to the original fusion confidence to obtain the corrected fusion confidence.
8. A system for normalizing and mapping indicator names for heterogeneous medical texts, characterized in that, Includes the following modules: The extraction module is used to perform semantic parsing on the acquired heterogeneous medical text using a medical entity boundary recognition model, and extract target entity triples consisting of an indicator name field, a unit of measurement field, and a numerical attribute field. The medical entity boundary recognition model is a deep learning network model based on BERT-BiLSTM-CRF. The generation module is used to retrieve a set of candidate standard terms associated with the indicator name field based on a standard medical terminology database, calculate the literal similarity between the target entity and each candidate standard term, and generate a rule matching probability vector of the first dimension by combining the unit conversion consistency coefficient of the measurement unit field. The calculation of the literal similarity between the target entity and each candidate standard term includes: calculating the Lewinstein edit distance between the indicator name field string of the target entity and the candidate standard term name string; obtaining the maximum value between the length of the indicator name field string and the length of the candidate standard term name string as a normalization factor; normalizing the Lewinstein edit distance using the normalization factor, and calculating the difference between 1 and the normalization result to obtain a distance score representing the literal similarity. The first calculation module is used to map the index name field to a high-dimensional vector space, decompose the initial feature vector into a main feature component representing the core word meaning and an auxiliary feature component representing the contextual modification using an orthogonal projection algorithm, calculate the cosine similarity between the main feature component and the vector of each term in the candidate standard term set, and obtain the semantic matching probability vector of the second dimension. The second calculation module is used to calculate the Hadamard product of the rule matching probability vector and the semantic matching probability vector to obtain the fusion confidence. If the fusion confidence is within a preset fuzzy range, the definition text of the corresponding candidate standard term is mapped to the definition feature vector, and the semantic correlation between the auxiliary feature component and the definition feature vector is calculated as a supplementary weight to perform weighted correction on the fusion confidence. Based on the ranking result of the corrected confidence, the standard term with the highest score is selected as the index normalization mapping output of the heterogeneous medical text.