Case-based keyword extraction and matching system and method

Abstract

The invention belongs to the technical field of medical data processing, and discloses a keyword extraction and matching system and method based on a case, and the method comprises the following steps: S1, manually marking keywords in a case text, and obtaining a digital data text; s2, examining the digital data text obtained in the step S1; s3, inputting the digital data text into the hybrid model to extract keywords from the digital data text; s4, standardizing the medical nouns to obtain a standard library; s5, matching the extracted keywords with a standard library, and outputting matching items corresponding to the keywords; the key information in the unstructured text is understood and accurately recognized by using the advanced NLP technology and combining medical professional knowledge and an artificial intelligence algorithm, the processing accuracy and efficiency are improved, and the problem that the key information is difficult to automatically and accurately extract due to complex medical terms and different expression habits of doctors is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical data processing, and in particular to a case-based keyword extraction and matching system and method. Background Art

[0002] With the modernization of medical practice, electronic health record (EHR) systems have become the core of medical information management. These systems have accumulated a huge amount of medical data, including a large amount of unstructured text data, such as case complaints, physician notes, and diagnostic reports, which have become valuable resources for medical research and clinical decision-making. These texts not only record the patient's symptoms and treatment process, but also contain the doctor's professional judgment and the logic of treatment decisions.

[0003] However, there are significant challenges in making effective use of this unstructured data. Although natural language processing (NLP) and machine learning technologies have made significant progress in other fields, there are still many challenges in applying these technologies directly to the medical field. First, medical terminology itself is extremely complex and changeable. The same disease and symptoms may be described in completely different ways in different doctors' notes; and the processing of medical texts requires not only a high degree of precision and a deep understanding of professional terminology, but also must take into account legal requirements for data privacy and security. In addition, the widespread use of abbreviations and colloquialisms further increases the difficulty of text understanding. Incorrect spellings and grammatical errors are also common problems. These factors combined make it particularly difficult to automatically and accurately extract key information from these texts. Summary of the invention

[0004] The present invention aims to provide a case-based keyword extraction and matching system and method, which uses advanced NLP technology and combines medical expertise and artificial intelligence algorithms to understand and accurately identify key information in unstructured texts, improve the accuracy and efficiency of extraction, and solve the problem of difficulty in automatically and accurately extracting key information due to complex medical terminology and different expression habits of doctors.

[0005] In order to achieve the above object, the present invention provides the following technical solutions:

[0006] A case-based keyword extraction and matching method comprises the following steps:

[0007] S1, manually marking keywords in the case text, and digitally storing the case text and the marked content to obtain digital data text;

[0008] S2, reviewing the digitized data text obtained in step S1;

[0009] S3, inputting the digitized data text into the hybrid model to train the model and extract keywords from the digitized data text;

[0010] S4, standardize medical terms to obtain a standard library;

[0011] S5. Match the keywords extracted in step S3 with the standard library in step S4, and output matching items corresponding to the keywords.

[0012] Furthermore, the digital data text obtained in the review step S1 includes:

[0013] S101, digitize the data of the digital data text; S102, use cluster analysis to identify abnormal text; S103, manually review the abnormal text to ensure that the text is indeed abnormal to avoid false alarms.

[0014] Furthermore, the hybrid model of step S3 is a pre-trained model and a bidirectional long short-term memory network-conditional random field model.

[0015] Furthermore, the step of matching the keywords extracted in step S3 with the standard library in step S4 includes:

[0016] S501, calculating the similarity between the keywords extracted in step S3 and the medical terms in the standard library in step S4;

[0017] S502, determining whether the similarity result is greater than a similarity threshold, if so, taking the medical term with the largest similarity result as a matching item and outputting it;

[0018] S503: Manually check the accuracy of the matching items. If the accuracy is lower than a preset value, increase the similarity threshold.

[0019] A case-based keyword extraction and matching system, the system comprising:

[0020] A storage module is used to digitally store case text and marked content to obtain digital data text;

[0021] A review module, used for reviewing the digitized data text obtained in step S1;

[0022] An extraction module, used to extract keywords from digitized data text;

[0023] A processing module is used to standardize medical terms and obtain a standard library;

[0024] The matching module is used to match the keywords extracted in step S3 with the medical term standard library in step S4, and output matching items corresponding to the keywords.

[0025] Furthermore, the review module includes:

[0026] A word conversion module is used to convert the data of the digital data text into digitized data;

[0027] Outlier detection module, used to identify abnormal documents.

[0028] Furthermore, the matching module includes:

[0029] A calculation module, used to calculate the similarity between the keywords extracted in step S3 and the medical terms in the standard library in step S4;

[0030] A judgment module, used to judge whether the similarity result is greater than a first preset threshold, and if so, take the medical term with the largest similarity result as a matching item;

[0031] Output module, used to output matching items.

[0032] The beneficial effects of the technical solution are:

[0033] 1. The present invention discloses a case-based keyword extraction and matching system and method, which uses advanced natural language processing (NLP) technology and combines medical expertise and artificial intelligence algorithms to understand and extract key information from unstructured texts to improve processing accuracy and efficiency. The method can accurately understand and process complex medical terms, adapt to different expression habits of doctors, and ensure that the data processing process meets privacy protection standards. It can understand medical expertise and adapt to data processing for special needs in the medical field. By combining manual labeling and advanced machine learning models, the present invention can accurately identify key information in a wide range of medical texts, thereby supporting more accurate diagnostic decisions, disease monitoring, and treatment effect evaluation.

[0034] 2. The case-based keyword extraction and matching system and method of the present invention can effectively utilize the information in unstructured medical texts to greatly improve the quality of medical services, promote the development of personalized medicine, and provide strong data support for public health research. In addition, with the advancement of medical AI technology, the application potential of the method proposed in the present invention is huge, not only limited to single medical record processing, but also can be extended to the automated extraction and integration of medical knowledge, as well as support for complex medical decision-making processes. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a process framework diagram of a keyword extraction and matching method based on cases and chief complaints of the present invention;

[0036] Figure 2 This is a deep learning model process framework diagram of the hybrid model in step S3 in Example 1 of the present invention. DETAILED DESCRIPTION

[0037] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments:

[0038] Example 1

[0039] like Figure 1 As shown, a case-based keyword extraction and matching method includes the following steps:

[0040] S1, manually marking keywords in the case text, and digitally storing the case text and the marked content to obtain digital data;

[0041] By cooperating with the hospital information department, batches of desensitized case data are exported from the hospital information system (HIS) and stored in the local server; professional medical personnel extract the keywords required for disease diagnosis according to the chief complaint in the case text; and organize them into a two-column table, the first column is the case text, and the second column is the keyword content. Finally, the data file is stored in the local server in CSV format to obtain digital data. An example of the table content is shown in Table 1 below;

[0042] Table 1 Keyword tag table of case text

[0043]

[0044]

[0045]

[0046]

[0047] S2, reviewing the digitized data text obtained in step S1;

[0048] Review of digitized data texts, including,

[0049] S101. First, the text data is digitized. For example, each word is converted into a fixed-length dense vector using a pre-trained word embedding model (Word2Vec). The vector representation of a document can be obtained by averaging or weighted averaging all word vectors in a document. Once the text is converted into numerical form, traditional outlier detection techniques can be used to identify abnormal documents.

[0050] S102, using cluster analysis to identify abnormal texts; specifically, using distance similarity measurement: calculating the average distance similarity between each document vector and all other document vectors, and setting a threshold, and determining documents that are significantly different from other documents and reach the threshold as abnormal values;

[0051] S103. Manual review of abnormal texts. For texts identified as abnormal by the algorithm, manual review is performed to ensure that these texts are indeed abnormal and not false positives.

[0052] S3, inputting the digitized data into a hybrid model of a pre-trained model (MC-BERT) and a bidirectional long short-term memory network (BiLSTM)-conditional random field (CRF) model to train the model and extract keywords from the digitized data;

[0053] The processed text is preprocessed by python. Since the text type is 'str', each word is converted into a corresponding fixed-length vector representation by creating a word table Word2Vec and input into the model in the form of a tensor. The main components of the model are MC-BERT, BiLSTM and CRF.

[0054] MC-BERT (Medical Concept BERT) is a natural language processing model customized for the medical field. It is optimized and pre-trained based on the Bidirectional Encoder Representations from Transformers (BERT) architecture. BERT itself is a pre-trained deep learning model developed by Google that aims to understand the context of natural language text. By pre-training on large-scale text data, the BERT model is able to capture the complex relationships between words and the deep structure of language, thereby achieving excellent performance in various NLP tasks. In the medical field, due to the complexity of professional terminology and its wide application, the model needs to have a deeper understanding of medical text. Although the conventional BERT model performs well in general language understanding tasks, its effect may be limited when faced with medical texts full of professional terminology and details. Therefore, MC-BERT is pre-trained on a large number of medical texts, including case reports, medical research papers, clinical guidelines, etc., so as to learn the vocabulary, terminology and their usage context of medical professions. Through pre-training, MC-BERT can understand the relationship between complex medical terms, such as the connection between symptoms and diseases, and the dependence between treatment methods and treatment effects. MC-BERT retains the architecture of the original BERT model, including a multi-layer Transformer encoder that uses a self-attention mechanism to process text. Like BERT, MC-BERT also supports multiple input formats and can process single text fragments or pairs of text fragments to adapt to different NLP tasks. The key difference is that MC-BERT is pre-trained on a large professional corpus in the medical field, allowing it to perform better on specific medical NLP tasks.

[0055] BiLSTM (Bidirectional Long Short-Term Memory) is a special recurrent neural network (RNN) for processing sequence data. It improves on the traditional Long Short-Term Memory (LSTM) network to simultaneously consider the forward and backward context information of each element in the sequence. This structure is particularly suitable for fields such as natural language processing (NLP), where the meaning of a word may be strongly affected by the words before and after it. The traditional LSTM effectively solves the gradient vanishing or gradient exploding problems encountered by the standard recurrent neural network (RNN) when processing long sequences by introducing gating mechanisms (including forget gates, input gates, and output gates), enabling the network to learn long-distance dependencies in the sequence. BiLSTM is an extension of LSTM. By combining a forward LSTM and a reverse LSTM in parallel, it can simultaneously capture the forward and backward context information of each element in the sequence. This bidirectional architecture makes BiLSTM more efficient in capturing the temporal dynamic characteristics of the sequence. For each element in the sequence, the network can not only consider its historical information (forward LSTM part), but also utilize its future information (reverse LSTM part).

[0056] Conditional Random Field (CRF) is a statistical modeling method that is widely used in tasks such as named entity recognition (NER) and part-of-speech tagging (POStagging) in natural language processing (NLP). As a discriminative model, CRF aims to build a conditional probability model from input sequence to output sequence. Different from the generative model, it does not directly model the distribution of the observation sequence, but focuses on the conditional probability of labeling the sequence given the observation sequence. CRF can take into account the dependency between adjacent labels in the sequence and use global information for optimization, so as to achieve better performance in sequence labeling tasks. Among CRF models, the most common one is the linear chain conditional random field (Linear-chain CRF), whose model structure is specially designed for sequence data, assuming that the label at the current position only depends on the labels of the previous and next positions. Given an input sequence X = (x1, x2, ..., x n ) and the corresponding label sequence Y = (y1, y2, ..., y n ), the linear chain conditional random field model aims to learn the conditional probability P(X|Y). This probability can be expressed as follows:

[0057]

[0058] Among them, f kis the transfer feature function, which depends on the adjacent labels. l is the state feature function, which depends on the label and observation of the current position. k and u l is the weight of the feature function. Z(X) is the normalization factor, which ensures that the sum of the probabilities over all possible label sequences is 1, and is calculated as follows:

[0059]

[0060] Where Y' represents all possible tag sequences.

[0061] The most likely tag sequence Y * , that is: Y * = argmax Y P(Y|X).

[0062] Among them, Figure 2 As shown, the steps of deep learning of the hybrid model are:

[0063] Step 1, Input and processed text: First, the input text data is provided to the system.

[0064] The second step is MC-BERT preprocessing: word segmentation. The input text is first segmented by MC-BERT's word segmenter to divide the text into smaller units (such as words or subwords). The segmented text is converted into word vectors through MC-BERT's pre-trained embedding layer to provide input for subsequent deep learning tasks.

[0065] Step 3, MC-BERT’s Transformer encoder layer: The text vector is processed through MC-BERT’s multi-layer Transformer structure, where each layer performs self-attention and feed-forward network operations to capture the complex semantics and contextual relationships in the text;

[0066] Step 4, feature extraction: The high-level feature representation obtained from the last layer of MC-BERT is extracted and used as the basis for subsequent sequence labeling tasks;

[0067] Step 5, BiLSTM layer: The features extracted by MC-BERT are input into the BiLSTM layer, which further processes these features through its bidirectional structure to capture long-distance dependencies in the sequence data;

[0068] Step 6, CRF layer: The output of BiLSTM is then fed into the conditional random field (CRF) layer, which uses the output of the previous step and the transition probability between labels to calculate the optimal label sequence for the entire sequence;

[0069] Step 7, Output: The system outputs the predicted label for each tag, completing the entire process from raw text to precise annotation.

[0070] S4, standardize medical terms to obtain a standard library;

[0071] Based on the guidelines, industry standards, expert consensus and authoritative textbooks issued by official academic institutions, we collect and organize standardized medical terms related to electrocardiogram examinations, extract entities and form a standard library of disease characteristic terms, mainly including the chief complaint, symptoms, signs, clinical diagnosis, physiological normal / abnormal, pathological medical terms, and similar names that have appeared. Each medical term is encoded for unique identification to form an entity standard library. Disease: Disease is the core entity of our knowledge graph, and all other entities are related to disease. The International Classification of Diseases, 10th Edition (ICD-11) is used as the disease terminology standard, and a controlled dictionary is used to map the original diagnostic terms to ICD-11 terms. Corresponding disease entities have been created for all standardized diagnostic diseases and their parent terms in ICD-11.

[0072] S5, matching the keywords extracted in step S3 with the medical term standard library in step S4, and outputting matching items corresponding to the keywords;

[0073] S501, using a function to calculate the edit distance (Levenshtein distance) for calculating the similarity between the keywords extracted in step S3 and the standard library medical terms in step S4;

[0074] S502, for each keyword, traverse each noun and its variants in the medical terminology library, set a similarity threshold, and only when the calculated similarity exceeds the similarity threshold, it is considered a successful match. For each keyword, select the medical noun with the highest similarity as its matching item and output it;

[0075] S503, manually checking the accuracy of the matching items, and if the accuracy is lower than a preset value, raising the first preset threshold;

[0076] Analyze the preliminary matching results, manually check the accuracy of some matching items, and adjust the similarity threshold or optimize the similarity calculation method based on the results of the manual check to improve the matching accuracy.

[0077] Example 2

[0078] A case-based keyword extraction and matching system, the system comprising:

[0079] A storage module is used to digitally store case text and marked content to obtain digital data text;

[0080] A review module, used for reviewing the digitized data text obtained in step S1;

[0081] An extraction module, used to extract keywords from digitized data text;

[0082] A processing module is used to standardize medical terms and obtain a standard library;

[0083] The matching module is used to match the keywords extracted in step S3 with the medical term standard library in step S4, and output matching items corresponding to the keywords.

[0084] The review modules include:

[0085] A word conversion module is used to convert the data of the digital data text into digitized data;

[0086] Outlier detection module, used to identify abnormal documents.

[0087] The matching modules include:

[0088] A calculation module, used to calculate the similarity between the keywords extracted in step S3 and the medical terms in the standard library in step S4;

[0089] A judgment module, used to judge whether the similarity result is greater than a first preset threshold, and if so, take the medical term with the largest similarity result as a matching item;

[0090] Output module, used to output matching items.

[0091] In summary, the present invention is a case-based keyword extraction and matching system and method, which uses advanced natural language processing (NLP) technology and combines medical expertise and artificial intelligence algorithms to understand and extract key information in unstructured texts to improve the accuracy and efficiency of processing; and the method can accurately understand and process complex medical terms, adapt to the different expression habits of doctors, and ensure that the data processing process meets the standards of privacy protection, and can understand medical expertise and adapt to data processing for special needs in the medical field; by combining manual labeling and advanced machine learning models, the present invention can accurately identify key information in a wide range of medical texts, thereby supporting more accurate diagnostic decisions, disease monitoring and treatment effect evaluation; by effectively utilizing information in unstructured medical texts, the quality of medical services is greatly improved, the development of personalized medicine is promoted, and strong data support is provided for public health research. In addition, with the advancement of medical AI technology, the application potential of the method proposed in the present invention is huge, not only limited to single medical record processing, but also extended to the automated extraction and integration of medical knowledge, and support for complex medical decision-making processes.

[0092] The above is only an embodiment of the present invention, and the common knowledge such as the known specific technical solutions or characteristics in the solution is not described in detail here. It should be pointed out that for those skilled in the art, several modifications and improvements can be made without departing from the technical solution of the present invention, which should also be regarded as the protection scope of the present invention, and these will not affect the effect of the implementation of the present invention and the practicality of the patent. The scope of protection required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.

Claims

1. A case-based keyword extraction and matching method, characterized in that: The following steps are involved: S1, manually marking keywords in the case text, and digitally storing the case text and the marked content to obtain digital data text; S2, reviewing the digitized data text obtained in step S1; S3, inputting the digitized data text into the hybrid model to train the model and extract keywords from the digitized data text; S4, standardize medical terms to obtain a standard library; S5. Match the keywords extracted in step S3 with the standard library in step S4, and output matching items corresponding to the keywords.

2. The case-based keyword extraction and matching method according to claim 1, characterized in that: The digital data text obtained in the review step S1 includes: S101, digitize the data of the digital data text; S102, use cluster analysis to identify abnormal text; S103, manually review the abnormal text to ensure that the text is indeed abnormal to avoid false alarms.

3. The case-based keyword extraction and matching method according to claim 1, characterized in that: The hybrid model of step S3 is a pre-trained model and a bidirectional long short-term memory network-conditional random field model.

4. The case-based keyword extraction and matching method according to claim 1, characterized in that: The keywords extracted in step S3 are matched with the standard library in step S4. include, S501, calculating the similarity between the keywords extracted in step S3 and the medical terms in the standard library in step S4; S502, determining whether the similarity result is greater than a similarity threshold, if so, taking the medical term with the largest similarity result as a matching item and outputting it; S503: Manually check the accuracy of the matching items. If the accuracy is lower than a preset value, increase the similarity threshold.

5. A case-based keyword extraction and matching system, characterized in that: The system comprises: A storage module is used to digitally store case text and marked content to obtain digital data text; A review module, used for reviewing the digitized data text obtained in step S1; An extraction module, used to extract keywords from digitized data text; A processing module is used to standardize medical terms and obtain a standard library; The matching module is used to match the keywords extracted in step S3 with the medical term standard library in step S4, and output matching items corresponding to the keywords.

6. A case-based keyword extraction and matching system according to claim 5, characterized in that: The review module includes: A word conversion module is used to convert the data of the digital data text into digitized data; Outlier detection module, used to identify abnormal documents.

7. The case-based keyword extraction and matching system according to claim 5, characterized in that: The matching module comprises: A calculation module, used to calculate the similarity between the keywords extracted in step S3 and the medical terms in the standard library in step S4; A judgment module, used to judge whether the similarity result is greater than a first preset threshold, and if so, take the medical term with the largest similarity result as a matching item; Output module, used to output matching items.

Images