A medical information push method based on post-structural processing of medical records

Abstract

The present invention provides a method for pushing medical information based on post-structural processing of medical records, comprising: receiving an image file of a patient's medical record, and performing OCR text recognition on the image information to obtain text information of the patient's medical record. Performing NLP-based structured conversion on the text information, extracting key information based on preset rules, and generating structured medical record data. Using the structured medical record data to query a predefined database, determining matching information to be pushed, and pushing the information to be pushed to the patient. The present invention improves diagnosis and treatment operations and provides data support, which helps to improve the quality of medical services and patient safety management.

Description

Technical Field

[0001] The present invention relates to the field of medical data integration, and in particular to a medical information push method based on post-structural processing of medical records. Background Art

[0002] With the popularization of medical information construction, especially the rapid development of artificial intelligence technology, the mining of data value and the development of intelligent applications are becoming increasingly important. In the hospital information system, it is usually necessary to obtain clinical data throughout the entire diagnosis and treatment activities based on the patient's historical diagnosis and treatment information in order to follow up on the patient. When the patient is seeing a doctor, the sources of disease-related data include the patient's clinical manifestations, test reports, imaging data, physician diagnosis results recorded in the hospital information system, and information obtained from direct communication with the patient during the visit. Traditional follow-up is based on blank forms manually established by medical staff, and the above-mentioned various types of information are obtained from face-to-face consultations with patients or from the review of medical records, and the blanks in the form are filled in. This method requires a lot of manual time for review and registration, and cannot quickly extract useful information from a large amount of clinical data, and then provide targeted follow-up services to patients, which is not conducive to the promotion of medical development. Summary of the invention

[0003] In order to solve the problems existing in the prior art, the present invention provides a medical information push method based on post-structural processing of medical records, comprising:

[0004] S1. Receive an image file of a patient's medical record, and perform OCR text recognition on the image information to obtain text information of the patient's medical record.

[0005] S2. Perform NLP-based structured conversion on the text information, extract key information based on preset rules, and generate structured medical record data.

[0006] S3. Use the structured medical record data to query a predefined database, determine matching information to be pushed, and push the information to be pushed to the patient.

[0007] Preferably, the step S1 further comprises:

[0008] During the OCR text recognition process, outputting the position of each recognized text sequence in the image file of the patient medical record;

[0009] Based on the positions, highlight marks are displayed on a subset of the output text sequence for which entity words have been detected.

[0010] Preferably, the step S2 further comprises:

[0011] Pre-select a post-structured algorithm model, obtain a large-scale electronic medical record data set, and use the large-scale data set to pre-train the post-structured algorithm model;

[0012] Compare the structured results of the pre-trained model with the annotation results of professional corpus, and update the model based on the comparison results;

[0013] Repeat the above training and updating process until the model converges;

[0014] The unstructured text data from the current medical records is input into the above-mentioned trained post-structured algorithm model to obtain structured text data.

[0015] Preferably, the post-structured algorithm model adopts an algorithm framework based on the attention mechanism; first, the encoding end receives the hidden vector sequence formed by encoding the unstructured source text sequence, and then the decoder extracts the encoded hidden vector sequence information through the attention mechanism to generate a structured text sequence in an autoregressive manner.

[0016] Preferably, the post-structured algorithm model includes a general sub-model and a subject sub-model; the general sub-model is trained based on training text sequences from general medical documents, and the subject sub-model is trained based on training text sequences from various subject areas.

[0017] Preferably, the step S3 further comprises:

[0018] Automatically matching the structured text information with a follow-up template library to obtain matching information to be pushed;

[0019] Distribute the information to be pushed to the patient, and automatically monitor management and data early warning;

[0020] The information to be pushed includes patient education knowledge, dietary guidance, medication guidance, and follow-up visit reminders.

[0021] Preferably, the step S3 further comprises:

[0022] Parsing multiple entity words included in the string of structured medical record text information;

[0023] converting the entity vocabulary into word-grams and marking the word-grams into tokens based on context associated with the entity vocabulary;

[0024] Replacing a token of an entity vocabulary with a predefined token associated with a type of the medical named entity; converting the replaced token into a hash sequence, wherein the hash sequence is a vector formed based on the context;

[0025] Each hash value in the hash sequence is embedded in the vector, and the medical record information is classified based on the similarity between the vector and a vector in classified medical record information in a predefined electronic medical record data set.

[0026] Preferably, the vector similarity calculation is implemented by the following steps:

[0027] Map the high-dimensional feature vector to a low-dimensional vector, and generate an r-dimensional feature value based on the low-dimensional vector;

[0028] Each feature item is weighted in the r-dimensional vector space, the weight value of each dimension in the r-dimensional vector is mapped to a Boolean value according to a preset rule, and multiple Boolean values ​​are concatenated to obtain an r-bit summary value of the medical record text vector;

[0029] A similarity calculation is performed based on the summary value.

[0030] Based on the same inventive concept, the present invention also provides a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above method when executing the computer program.

[0031] The present invention also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the steps of the method described above are implemented.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] The present invention provides a method for pushing medical information based on post-structural processing of medical records, including using technical means such as an OCR text intelligent recognition engine and an NLP natural language intelligent processing engine to output structured medical record data based on authoritative and high-quality medical knowledge, and performing follow-up management of patients according to the medical records. The scheme of the present invention can automatically generate information to be pushed through structured data processing of medical record information, and inform patients of precautions after diagnosis, including diet and medication guidance, follow-up reminders, etc., which significantly saves staff time, provides data support for doctors to improve diagnosis and treatment operations, and helps to improve the quality of medical services and the level of patient safety management. The present invention also indirectly improves the quality of electronic medical record entry to achieve higher electronic medical record rating standards, laying the foundation for future artificial intelligence-assisted diagnosis systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a flowchart of the medical information push method based on post-structural processing of medical records according to the present invention.

[0035] Figure 2A detailed data flow diagram of the medical information push method based on post-structural processing of medical records according to the present invention is shown.

[0036] Figure 3 A schematic diagram of an interface for a patient user to upload medical record image information according to the present invention is shown.

[0037] Figure 4 and Figure 5 A schematic diagram of an interface for identifying medical record image information on the follow-up information platform according to the present invention is shown.

[0038] Figures 6 to 8 A schematic diagram of an interface for automatically generating a follow-up plan based on structured patient medical records on a follow-up information platform according to the present invention is shown. DETAILED DESCRIPTION

[0039] In view of the above-mentioned problems existing in the prior art, the present invention provides a medical information push method based on post-structural processing of medical records, which utilizes a cloud storage platform and a mobile terminal for medical and patient data collection and follow-up. After the patient's medical record is generated, the user's client can scan or photograph the patient's medical record into image information, and then format it through an application. Client users can check and understand the patient's past medical records, test reports and other examination results anytime and anywhere. Then, the follow-up content is standardized according to the specifications, accurately pushed to the patient, and the patient's condition data information is collected. With the help of this system, the patient's precautions after diagnosis, diet and medication guidance, and follow-up reminders can be automatically completed.

[0040] Embodiment 1:

[0041] like Figure 1 As shown, the present invention provides a medical information push method based on post-structural processing of medical records, including:

[0042] S1. Receive an image file of a patient's medical record, and perform OCR text recognition on the image information to obtain text information of the patient's medical record.

[0043] S2. Perform NLP-based structured conversion on the text information, extract key information based on preset rules, and generate structured medical record data.

[0044] S3. Use the structured medical record data to query a predefined database, determine matching information to be pushed, and push the information to be pushed to the patient.

[0045] Figure 2 A detailed data flow diagram of the medical information push method based on post-structural processing of medical records according to the present invention is shown.

[0046] In step S1, the patient user can upload past medical records. Figure 3 As shown, the user client enters the patient file page, clicks the automatic recognition button, selects the medical record picture to be recognized, and clicks OK to upload the file to the background.

[0047] See also Figure 4 and Figure 5 , the OCR recognition module receives the medical record image uploaded by the front-end user, and through the OCR intelligent recognition engine, recognizes the text data in the image into the memory for subsequent processing. In a further preferred embodiment, at least one parameter of the OCR recognition is determined by the historical parameters of the natural language processing model. And through the OCR recognition operation, the position of each text sequence in the medical record image file is output. Based on the position, a highlight mark is displayed on a subset of the output text sequence in which the entity vocabulary has been detected. Then the user selection in the mark is detected, and when the user selection is detected, the detected entity vocabulary is displayed on the selected mark. For example, the entity vocabulary is the terminology of entities such as diseases, symptoms, diagnoses and treatments that can be extracted from the medical record, and the relationship between the entities is obtained, including: the relationship between disease and treatment, the relationship between symptoms and treatment, the relationship between disease and diagnosis, the relationship between symptoms and diagnosis, the relationship between disease and symptoms, the relationship between diseases, and the relationship between symptoms.

[0048] If the original medical record image is a medical record image with handwritten fonts, then in an optional embodiment, such medical record image is redrawn. Specifically, a deep neural network can be used to perform OCR recognition on the scanned image of the handwritten medical record to obtain multiple medical terms in the sentences included in the text in the scanned image. If the OCR recognition is successful, the recognized term and the image coordinates of the term are obtained, and the received recognized term and its image coordinates are sequentially stored in the metadata file. Then, the drawing engine can select the recognized term from the metadata file, assign the term to a predefined medical named entity, and combine the assigned medical named entity with the recognized term and its image coordinates into a sequence. At this time, a sequence related to the predefined medical named entity can be automatically selected from the metadata file, and the corresponding image area of ​​the recognized term can be searched in the stored image coordinates, so that the image area corresponding to the recognized term can be highlighted in the image.

[0049] Wherein, in step S2, after the picture recognition is completed, the scattered unstructured data is input into the natural language processing module, formatted according to the predefined rules, parsed to obtain valid medical record information, and stored in the database. The rule-based extraction method has the characteristics of simple extraction process, and its extraction structure depends on the generation and learning of rules. First, the language rules are constructed by the rule template, and then the dictionary file is constructed, and the regular expression is input to generate the rules. The entity extraction result based on the rule is subjected to regression test. By performing entity recognition and extraction based on the structured processing process based on specific rules, various diagnosis, symptoms, medical advice, inspection, examination, surgery and other information can be quickly and accurately extracted, thereby improving the utilization level of medical records. For example, if the main complaint statement describing the disease of "dizziness for 3 days, with anorexia" is divided into "head" (noun), "dizziness" (verb, behavior phrase), "3" (numeral) and "day" (unit noun), "with anorexia" (verb, behavior phrase) 5 parts according to the category of words. "Head" is the element of the sentence describing "body part", "dizziness" is the element describing "symptoms", "3" is the "numerical value" element, and "day" is the "time unit" element. Under the premise that these elements have been pre-defined, they can be found in the corresponding element classification. In this way, a structured template is formed in order: "[body part][common symptoms][number][time][accompanying symptoms]", and then the structured template is added to the medical record template. When using the medical record template, select and fill in the corresponding items to form a structured medical record.

[0050] In a specific embodiment, the present invention pre-selects a post-structured algorithm model, obtains a large-scale electronic medical record data set, pre-trains the post-structured algorithm model using the large-scale data set, compares the structured results of the pre-trained model with the professional corpus annotation results, updates the model according to the comparison results, and repeats the above training and updating process until the model converges. Then, the unstructured text data from the current medical record is input into the above-trained post-structured algorithm model to obtain structured text data.

[0051] For example, the post-structured algorithm model can adopt an algorithm framework based on the attention mechanism. First, the encoding end encodes the unstructured source text sequence to form a hidden vector sequence, and then the decoder extracts the encoded hidden vector sequence information through the attention mechanism to generate a structured text sequence in an autoregressive manner. Among them, the post-structured algorithm model detects entities from the source text sequence, and each entity includes a label and a value. The value of the entity is extracted from the unstructured source text sequence using the algorithm framework based on the attention mechanism. The entity and the corresponding value are mapped to a mapping table of predefined terms, the value of the entity is converted into the corresponding predefined term, and then structured medical record text information including the entity detected from the input text sequence and the corresponding term is generated.

[0052] The post-structural algorithm model includes a graph structure, which includes nodes and edges; each node corresponds to a text sequence; an edge between two nodes indicates a sequential relationship between two text sequences represented by the two nodes; when detecting the entity, each text sequence of the input text sequence is matched with the text sequence represented in the graph structure. Then the graph structure is updated based on the training text sequence marked with the entity name. The post-structural algorithm model includes a general sub-model and a subject sub-model; the general sub-model is trained based on the training text sequence from general medical documents; and the subject sub-model is trained based on the training text sequence from various subject areas. For the text sequence of the electronic medical record of the patient, the included entities are detected and the document structure of the electronic medical record is analyzed to identify the nested structure therein. Based on the relative position within the nested structure, the detected entity and the nested structure are used to generate a knowledge graph. Specifically, for the level of the nested structure, the main entity in the level is represented and the main concept type of the main entity is found, and the potential relationship with the main concept type is identified based on the part of speech and sentence relationship of the sub-entity. The main entity is connected to the sub-entities using the part-of-speech and concept type metadata.

[0053] When comparing the structured results of the pre-trained model with the professional corpus annotation results and updating the model according to the comparison results, the accuracy of identifying entities from the input text sequence by the post-structured algorithm model is first determined. When the accuracy reaches a predefined threshold, the model is determined to be converged, otherwise, the graph structure is updated, and the training text sequence is adjusted to continue training the post-structured algorithm model.

[0054] In order to further accurately structure the electronic medical record, the embodiment of the present invention can classify the entities in the medical record text, use the existing database to find the optimal set of structured sequences of entity elements, and then retrieve the entity combination. The index table of the entity element database is established by splitting the string value of the phrase, and the phrase is regarded as a segment, that is, n phrases are recorded as sg1, sg2, ..., sg n , indicating that there are n segments. The entity element corresponding to each segment is regarded as a state. Since a phrase may correspond to multiple entity elements, the i-th segment sg i There are i state, denoted as S i，1 , S i，2 , …, S i，ni . Where S i，k represents the kth state of the i-th segment, S i，k The content of the state is the entity element identifier of the state, denoted by x i，k .

[0055] Calculate the index V 1 k,n (s k , P k，n ), where s k represents the segment set consisting of the first k states; P k，n Indicates the state S of the starting segment 1，k The strategy function of the subprocess from the beginning to the end state of the kth segment is expressed as:

[0056] V 1 k,n (s k , P k，n )=V 1 k-1,n (s k , P k，n )+1

[0057] The optimal value function is in S k When given, the index V 1 k,n (s k , P k，n ) k，n The optimal value min(V 1 k,n (s k , P k，n ));

[0058] For medical record structuring, we first need to find the entity element that is the same as the previous and next phrases as the optimal entity element, and then find the optimal segmentation set with the least phrases in the text, that is, the optimal structured phrase sequence. The optimal strategy for calculating the termination segment is P 1，n(S1) = {u1(s1), u2(s2), ..., u n (s n )}. Thus the optimal combination solution is output. k Indicates that the kth segment is in state S k The decision variables at that time.

[0059] The first phrase of the medical record text is used as the initial segmentation, and the segmentation points are found in sequence to divide it into multiple segments. Each part contains multiple phrase segments. For each part, according to the input segments sg1, sg2, ..., sgn, the segmentation boundary is initialized to sg i .loc>sg i-1 .loc+sg i-1 .len; if sg.loc exceeds the segment boundary, add the segment between the current segment and sg to the temporary segment set, and determine V 1 k,n (s k , P k，n ) reaches the minimum value and satisfies sg i-1 .Atr≠sg i-1 .Atr, if all are satisfied, the temporary segmentation set result is added to the optimal segmentation set s. Finally, the optimal segmentation set results obtained by each segmentation are merged to obtain the final segmentation set result, that is, the optimal text combination of medical record structure. Among them, loc, len, and atr represent the position, length, and part of speech of the phrase in the medical record text respectively.

[0060] Among them, in step S3, after the structured medical record text information is obtained through processing by the natural language processing module, Figure 6-Figure 8 As shown, the embodiments of the present invention help patients automatically join the follow-up plan based on structured patient medical records and medical history files, realize automatic matching of the follow-up template library, obtain matching patient education knowledge, and thus accurately distribute the patient education knowledge to the patients. At the same time, it assists doctors in online doctor-patient communication, establishes WeChat, text messages, telephone and other methods for subsequent follow-up of patients, and automatically monitors management status and data warnings.

[0061] In a specific embodiment, multiple entity words included in the string of structured medical record text information can also be parsed. The entity words are converted into word units, and the multiple word units are marked as tokens based on the context associated with the entity words. Then, the tokens of the entity words are replaced with predefined tokens associated with the type of medical named entity. The replaced tokens are converted into a hash sequence, which is a vector formed based on the context. Each hash value in the hash sequence is embedded in the vector. The vector is then processed to classify the medical record information based on the similarity between the vector and the vector in the classified medical record information in a large-scale electronic medical record data set, and the classification of the medical record information is output.

[0062] For the calculation of vector similarity, the high-dimensional feature vector is mapped to a low-dimensional vector, and an r-dimensional feature value is generated based on the low-dimensional vector. Each feature item is weighted in the r-dimensional vector space, and the weight value of each dimension in the r-dimensional vector is mapped to a Boolean value according to the preset rules. The multi-bit Boolean values ​​are connected to obtain the r-bit summary value of the medical record text vector. The similarity is calculated based on the summary value. Specifically:

[0063] Initialize the r-dimensional vector τ and the r-bit Boolean value pes in advance, and then calculate the sentence b in the medical record text set DH. i , use the RSA digest algorithm to get an r-bit digest value.

[0064] Using bol j (s i ) means b i The Boolean value corresponding to the jth bit is defined as the normalization function η(bol j (b i )):

[0065]

[0066] Define τ j Represents the jth dimension of the vector τ, and calculates:

[0067] τ j =τ j +ω(b i )×η(bol j (b i ))

[0068] Among them, j ranges from 1 to r, ω(b i ) represents statement b i The weight value of .

[0069] Determine whether there is an unprocessed statement in the set DH. If so, process the next statement b. i+1Perform iterative calculation; otherwise, if all statements have been processed, calculate the j-th value pes in the Boolean value sequence pes j , if τ j >0, then pes j =1; if τ j ≤0, then pes j =0,j∈[1,r].

[0070] The obtained Boolean value sequence pes is used as the feature value of the current sentence. Then, for the current structured medical record text A and the classified medical record text B, the Boolean value of each sentence is combined to form a statement Boolean value set Φ A and Φ B ,use and Respectively represent the number of elements in each set, Represents the number of related sentences in two sets. The similarity between medical record texts A and B is:

[0071]

[0072] The criterion for the associated statement is that if the Boolean values ​​of the two statements a and b are equal, a With pes b If the XOR operation result of is higher than the predefined threshold η, the two statements are judged to be associated statements.

[0073] In a further aspect, in order to extract key word information from structured medical record text information, the present invention generates a candidate query word vector with word occurrence frequency as an element for each to-be-queried word contained in the structured medical record text information, and stores the generated candidate query word vector in a pre-established vector cache list; based on the candidate query word vector stored in the vector cache list, calculates the probability of each candidate query word being a key query word, that is, the initial probability of the key query word, and stores the calculated initial probability in the vector cache list.

[0074] Then, a test medical record text is obtained, and a test word vector with the word occurrence frequency as an element is generated for each word contained in the test medical record text, and the generated test word vector is stored in the vector cache list. Next, the values ​​of the candidate query word vector and the test word vector stored in the vector cache list are input into a test word generation probability model representing a binomial distribution to calculate the test word generation probability, and the calculated test word generation probability is stored in the vector cache list. Based on the initial probability of the key query word stored in the vector cache list and the test word generation probability, the probability of each candidate query word contained in the test medical record text being selected as a key query word is calculated, and the query word with the highest probability is extracted as the key query word.

[0075] The classified medical record information is classified and annotated based on the following information standards, such as quality, severity, duration, background, signs and symptoms, etc. Finally, the classification results of the medical record information are used to query the matching follow-up information database, and the push action is triggered based on the classification. For example, based on the key medical record information of the patient, a list of drugs and care plans can be determined from the follow-up information database, and the electronic medical records from the patient can be used to consider the contraindications of the patient, the interactions of the selected drugs and care plans, the interactions of drugs with drugs already used by the patient, and exclude drugs that are contraindicated for the patient.

[0076] The above method of the present invention can automatically generate information to be pushed through structured data processing of medical record information, and can inform patients of precautions after diagnosis, including diet and medication guidance, follow-up reminders, etc., which significantly saves staff time, provides data support for doctors to improve diagnosis and treatment operations, and helps to improve the quality of medical services and patient safety management. Through the present invention, the quality of electronic medical record entry can also be indirectly improved to achieve higher electronic medical record rating standards, laying the foundation for future artificial intelligence-assisted diagnosis systems.

[0077] Embodiment 2:

[0078] In order to implement the above method, the present invention also provides a medical information push system based on post-structural processing of medical records, comprising:

[0079] The OCR recognition module 201 is used to receive an image file of a patient's medical record and perform OCR text recognition on the image information to obtain text information of the patient's medical record;

[0080] A natural language processing module 202 is used to perform NLP-based structural conversion on the text information, extract key information based on preset rules, and generate structured medical record data;

[0081] The intelligent push module 203 is used to use the structured medical record data to query a predefined database, determine matching information to be pushed, and push the information to be pushed to the patient.

[0082] The functions implemented by each functional module in this embodiment are the same as those in Embodiment 1 and will not be described in detail here.

[0083] Embodiment 3:

[0084] The present invention also provides a terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above method when executing the computer program.

[0085] Embodiment 4:

[0086] The present invention also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the steps of the method described above are implemented.

[0087] Obviously, the described embodiments are only some embodiments of the present invention, but not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0088] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of complete hardware embodiments, complete software embodiments, or embodiments in combination with software and hardware. Moreover, the present application can adopt the form of computer program products implemented on multiple computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that comprise computer-usable program codes.

[0089] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0090] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0091] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0092] The above are merely embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention are included in the scope of the claims of the present invention to be approved.

Claims

1. A method for pushing medical information based on post-structural processing of medical records, characterized in that: include: S1. Receive an image file of a patient's medical record, and perform OCR text recognition on the image information to obtain text information of the patient's medical record; S2. Performing NLP-based structural transformation on the text information, extracting key information based on preset rules, and generating structured medical record data; S3, using the structured medical record data to query a predefined database, determine matching information to be pushed, and push the information to be pushed to the patient; The step S3 further comprises: Automatically matching the structured text information with a follow-up template library to obtain matching information to be pushed; Distribute the information to be pushed to the patient, and automatically monitor management and data early warning; The information to be pushed includes patient education knowledge, dietary guidance, medication guidance, and follow-up visit reminders; The step S3 further comprises: After obtaining structured medical record text information, the patient is automatically added to the follow-up plan based on the structured patient medical records and medical records, and the follow-up template library is automatically matched to obtain matching patient education knowledge, thereby distributing the patient education knowledge to the patient. At the same time, it assists doctors in online doctor-patient communication, establishes multiple ways to follow up on patients, and automatically monitors management status and data warnings; Parse multiple entity words included in the string of structured medical record text information; convert the entity words into word units, and mark the multiple word units into tokens based on the context associated with the entity words; then, replace the tokens of the entity words with predefined tokens associated with the type of medical named entity; convert the replaced tokens into a hash sequence, which is a vector formed based on the context; embed each hash value in the hash sequence into the vector; then process the vector to classify the medical record information based on the similarity between the vector and the vector in the classified medical record information in the large-scale electronic medical record data set, and output the classification of the medical record information; For the calculation of vector similarity, the high-dimensional feature vector is mapped to a low-dimensional vector, and an r-dimensional feature value is generated based on the low-dimensional vector. Each feature item is weighted in the r-dimensional vector space, and the weight value of each dimension in the r-dimensional vector is mapped to a Boolean value according to the preset rules. The multi-bit Boolean values ​​are connected to obtain the r-bit summary value of the medical record text vector; the similarity is calculated based on the summary value; specifically: Initialize the r-dimensional vector τ and the r-bit Boolean value pes in advance, and then calculate the sentence b in the medical record text set DH. i , use the RSA digest algorithm to get an r-bit digest value; Using bol j (s i ) means b i The Boolean value corresponding to the jth bit is defined as the normalization function η(bol j (b i )): Define τ j Represents the jth dimension of the vector τ, and calculates: t j =t j +ω(b i )×η(bol j (b) i )) Among them, j ranges from 1 to r, ω(b i ) represents statement b i The weight value of Determine whether there is an unprocessed statement in the set DH. If so, process the next statement b. i+1 Perform iterative calculation; otherwise, if all statements have been processed, calculate the j-th value pes in the Boolean value sequence pes j , if τ j >0, then pes j =1; if τ j ≤0, then pes j =0,j∈[1,r]; The obtained Boolean value sequence pes is used as the feature value of the current sentence. Then, for the current structured medical record text A and the classified medical record text B, the Boolean value of each sentence is combined to form a statement Boolean value set Φ A and Φ B ,use and Represents the number of elements in each set, Represents the number of related sentences in two sets. The similarity between medical record texts A and B is: The criterion for the associated statement is that if the Boolean values ​​of the two statements a and b are equal, a With pes b If the XOR operation result of is higher than the predefined threshold η, the two statements are judged to be associated statements.

2. The medical information push method based on post-structural processing of medical records according to claim 1, characterized in that: The step S1 further comprises: During the OCR text recognition process, outputting the position of each recognized text sequence in the image file of the patient medical record; Based on the positions, highlight marks are displayed on a subset of the output text sequence for which entity words have been detected.

3. The medical information push method based on post-structural processing of medical records according to claim 1, characterized in that: The step S2 further comprises: Pre-select a post-structured algorithm model, obtain a large-scale electronic medical record data set, and use the large-scale data set to pre-train the post-structured algorithm model; Compare the structured results of the pre-trained model with the annotation results of professional corpus, and update the model based on the comparison results; Repeat the above training and updating process until the model converges; The unstructured text data from the current medical records is input into the above-mentioned trained post-structured algorithm model to obtain structured text data.

4. The medical information push method based on post-structural processing of medical records according to claim 3 is characterized in that: The post-structured algorithm model adopts an algorithm framework based on the attention mechanism; first, the encoder receives the hidden vector sequence formed by encoding the unstructured source text sequence, and then extracts the encoded hidden vector sequence information through the attention mechanism at the decoder to generate a structured text sequence in an autoregressive manner.

5. The medical information push method based on post-structural processing of medical records according to claim 3 is characterized in that: The post-structured algorithm model includes a general sub-model and a subject sub-model; the general sub-model is trained based on training text sequences from general medical documents, and the subject sub-model is trained based on training text sequences from various subject areas.

6. A terminal device, characterized in that: The method comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method according to any one of claims 1 to 5 when executing the computer program.

7. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

Images