Digestive tract hemorrhage risk assessment system based on deep learning

Abstract

The invention relates to the technical field of medical data mining, in particular to a digestive tract hemorrhage risk assessment system based on deep learning, which comprises the following steps: firstly, preprocessing non-structural text data, and determining keywords and non-key segmented words in the text data; then, in combination with risk characteristics represented by text data changes and the influence of aspirin on the bleeding risk, form risk characteristic values are determined on the dimension of a text form; then combining semantic information of the text data with text semantics of all risk links corresponding to risks occurring in the digestive tract to carry out correlation analysis, and determining semantic risk characteristic values on the dimension of the semantic information; and finally, in combination with the formal risk feature values and the semantic risk feature values, determining the attention weight of each text evaluation dimension, and adjusting the neural network model according to the attention weights, so that the accuracy of the output digestive tract hemorrhage risk evaluation coefficient is higher.

Description

Technical Field

[0001] This invention relates to the field of medical data mining technology, specifically to a deep learning-based gastrointestinal bleeding risk assessment system. Background Technology

[0002] Current technologies for assessing the risk of gastrointestinal bleeding (GTB) by combining semantic information from patients with deep learning typically input a vector composed of the frequency of extracted GTB-related keywords directly into a neural network model to output a GTB risk assessment coefficient. However, these existing technologies do not consider the semantic information of the keywords themselves or the variations in textual data that indicate the risk of GTB. This results in low accuracy of the output GTB risk assessment coefficient after directly inputting the vector of extracted GTB-related keyword frequency into the neural network model. Summary of the Invention

[0003] To address the issue of low accuracy in existing technologies that combine semantic information of patients with gastrointestinal bleeding with deep learning methods to output gastrointestinal bleeding risk assessment coefficients, this application aims to provide a deep learning-based gastrointestinal bleeding risk assessment system. The specific technical solution adopted is as follows: The first aspect of this application provides a deep learning-based gastrointestinal bleeding risk assessment system, comprising: The data acquisition and preprocessing module is used to acquire each text data of patients with gastrointestinal bleeding in each text evaluation dimension from the medical database; and extract keywords and all non-keywords from the text data. The risk feature value determination module is used to determine the formal risk feature value of each text evaluation dimension based on the recording time interval between the current text data and the previous text data, the aspirin dosing interval of the current text data, the change in the number of non-key words in the sentence containing the keyword, and the stability of the changes of various keywords; and to determine the semantic risk feature value of each text evaluation dimension based on the correlation between the overall vector features of the keyword word vectors and the preset benchmark word vectors of each risk link. The gastrointestinal bleeding risk assessment module is used to determine the attention weight of each text assessment dimension based on the formal risk feature value and the semantic risk feature value; adjust the neural network model according to the attention weight, and output the gastrointestinal bleeding risk assessment coefficient of the patient at the current moment.

[0004] Furthermore, the process of extracting keywords and all non-keyword segments from the text data includes: Each piece of text data is input into the BioBERT model for word segmentation and keyword extraction, and the output includes keywords and non-keyword segments corresponding to gastrointestinal bleeding; the non-keyword segments are other words besides the keywords.

[0005] Furthermore, the process of obtaining the formal risk characteristic value includes: Under each text evaluation dimension, each keyword in the current text data is taken as the target word in turn; The number of descriptive words for the target word in each text data is determined by the average number of non-keywords in all sentences containing the target word in each text data; the set of all kinds of keywords in each text data is taken as the corresponding keyword set; the symptom description coefficient of the target word in the current text data is determined by the overlap of the keyword set between the current text data and the previous text data and the relative reduction of the number of descriptive words. The time interval coefficient of the current text data is determined by negatively correlated with the time interval between the recording time of the current text data and the recording time of the previous text data. The risk development coefficient of the current text data is determined based on the time interval between the recording time of the current text data and the date of the patient's most recent aspirin administration. The weighted structural risk feature value of the target word in the current text data is determined based on the product of the risk development coefficient, the time interval coefficient, and the symptom description coefficient. For each text evaluation dimension, the corresponding formal risk feature value is determined based on the mean of the weighted structural risk feature values ​​of all keywords in the current text data.

[0006] Furthermore, the process of obtaining the symptom description coefficient includes: Based on the difference between the number of descriptive words for the target word in the previous text data and the number of descriptive words for the target word in the current text data, determine the amount of reduction in descriptive words for the target word in the current text data; A reference set is determined based on the intersection between the keyword set of the current text data and the keyword set of the corresponding previous text data; the keyword combination stability of the current text data is determined based on the ratio between the number of keyword types in the reference set and the number of keyword types in the keyword set of the current text data. By performing a positive correlation mapping between the product of the reduction in descriptive words and the stability of the keyword combination, the symptom description coefficient of the target word in the current text data is determined.

[0007] Furthermore, the process of obtaining the risk development coefficient includes: The time when a patient with gastrointestinal bleeding takes aspirin is taken as the time of administration; the time interval between the current text data recording time and each time of administration is taken as the corresponding judgment time length. When the determination time length is less than or equal to the preset first time length threshold, the normalized value of the determination time length is used as the reference development coefficient for the corresponding administration time. When the determination time length is greater than a preset first time threshold and less than or equal to a preset second time threshold, the preset highest risk characteristic value is used as the reference development coefficient for the corresponding administration time. When the determination time length is greater than the preset second time length threshold, the negative correlation mapping value of the determination time length is used as the reference development coefficient of the corresponding administration time. The maximum value of the reference development coefficients at all times of administration is taken as the danger development coefficient of the current text data.

[0008] Furthermore, the process of obtaining the semantic risk feature value includes: In each text evaluation dimension, word2vec is used to convert all keywords in the current text data into word vectors, and the word vector of each keyword is determined. The semantic similarity is determined based on the cosine similarity between the word vector of each keyword and the word vector of each other keyword; in the current text data, hierarchical clustering is performed using the negative correlation mapping value of the semantic similarity between keywords as the clustering distance to obtain at least two keyword clusters. The center vector is determined based on the mean vector of word vectors of all keywords in each keyword cluster; the risk score is determined based on the mean of all semantic similarities between all keywords in each keyword cluster. Based on the vector similarity distribution between the central vector and the preset benchmark word vectors of each risk link, the risk association strength of each keyword cluster is determined; The local risk feature value of each keyword cluster is determined by multiplying the risk association strength with the risk score; the semantic risk feature value of each text evaluation dimension is determined by the mean of the local risk feature values ​​of all keyword clusters.

[0009] Furthermore, the process of obtaining the risk association strength includes: Calculate the cosine similarity between the center vector and the preset benchmark word vector of each risk link to determine the link matching degree between each keyword cluster and each risk link; determine the corresponding risk association strength based on the maximum link matching degree between each keyword cluster and all risk links.

[0010] Furthermore, the process of obtaining the attention weights includes: The attention weight for each text evaluation dimension is determined based on the product of the formal risk feature value and the semantic risk feature value.

[0011] Furthermore, the process of obtaining the gastrointestinal bleeding risk assessment coefficient includes: In each text evaluation dimension, the frequency of each keyword in the current text data is used as the corresponding frequency feature value; the frequency feature values ​​of all keywords in the current text data are arranged in the order of their occurrence in all text data to determine the feature vector of the current text data in each text evaluation dimension. The product of the feature vector of the current text data in each text evaluation dimension and the corresponding attention weight is used as the weighted vector; the joint feature vector is determined by concatenating the weighted vectors of all text evaluation dimensions. The joint feature vector is input into a trained multilayer fully connected neural network, which outputs a risk assessment index for the digestive tract.

[0012] Furthermore, the output of the gastrointestinal bleeding risk assessment coefficient for the patient at the current moment also includes: When the gastrointestinal bleeding risk assessment index is greater than the preset risk threshold, a bleeding warning is issued; No bleeding warning will be issued when the risk assessment of gastrointestinal bleeding is less than or equal to a preset risk threshold.

[0013] Secondly, this application provides a computer device including a memory and a processor. The memory is used to store computer program code, and the processor is used to call and run the computer program code from the memory to execute a system as described in the first aspect of this application or any embodiment of the first aspect.

[0014] Thirdly, this application provides a computer program product, which includes computer program code that, when executed, performs a system as described in the first aspect of this application or any embodiment thereof.

[0015] Fourthly, this application provides a computer-readable storage medium that stores computer program code, which, when executed, performs a system as described in the first aspect of this application or any embodiment thereof.

[0016] This application has the following beneficial effects: This application first preprocesses unstructured text data to identify keywords and non-keywords. Then, combining the risk characteristics represented by changes in the text data and the impact of aspirin on bleeding risk, formal risk feature values ​​are determined at the text form dimension. Next, a correlation analysis is performed between the semantic information of the text data and the text semantics of each risk stage corresponding to the gastrointestinal risk, determining semantic risk feature values ​​at the semantic information dimension. Finally, combining formal and semantic risk feature values, the attention weight for each text evaluation dimension is determined, and the neural network model is adjusted based on the attention weight to improve the accuracy of the output gastrointestinal bleeding risk assessment coefficient. This invention further enhances the influence of key modal data on the final prediction result, reduces noise interference from weak modal data, and further improves the accuracy of the prediction result. Attached Figure Description

[0017] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a structural diagram of a deep learning-based gastrointestinal bleeding risk assessment system provided in one embodiment of the present invention; Figure 2 This is a schematic diagram of a computer device structure provided in one embodiment of the present invention. Detailed Implementation

[0019] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a deep learning-based gastrointestinal bleeding risk assessment system proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment, and specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined with "first" or "second" may explicitly or implicitly include one or more of that feature.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0021] The following description, in conjunction with the accompanying drawings, details a specific scheme for a deep learning-based gastrointestinal bleeding risk assessment system provided by this invention.

[0022] This application provides a deep learning-based gastrointestinal bleeding risk assessment system. Please refer to [link to relevant documentation]. Figure 1 The diagram illustrates a structural diagram of a deep learning-based gastrointestinal bleeding risk assessment system according to an embodiment of the present invention. The system includes: a data acquisition and preprocessing module 101, a risk feature value determination module 102, and a gastrointestinal bleeding risk assessment module 103.

[0023] The data acquisition and preprocessing module 101 is used to acquire each piece of text data of patients with gastrointestinal bleeding in each text evaluation dimension from the medical database; and to extract keywords and all non-keywords from the text data.

[0024] In one specific implementation of this invention, the text assessment dimension includes clinical documents such as chief complaint, physical examination report, doctor's assessment record, discharge record, and medical record. The implementer can adjust the text assessment dimension according to the specific implementation environment. In the medical database, each text data for each text assessment dimension of the patient with gastrointestinal bleeding is collected at each recording time and the corresponding recording time is marked. In this embodiment, the current text data is the text data whose recording time is closest to the current time. Further, this embodiment determines the attention weight based on the risk characteristics reflected in the text data. This allows the neural network model, after adjusting the attention weight, to enhance the influence of key modality data on the final prediction result, reduce noise interference from weak modality data, further improve the accuracy of the prediction result, and make the final gastrointestinal bleeding risk assessment coefficient more accurate.

[0025] Analyzing text data typically requires extracting keywords and word segments based on natural language processing methods. Preferably, in some possible implementations of this invention, the process of extracting keywords and all non-keyword word segments from text data includes: Each text data is input into the BioBERT model for word segmentation and keyword extraction, and the output includes keywords and non-keyword segments corresponding to gastrointestinal bleeding; non-keyword segments are other words besides keywords.

[0026] Before inputting text data into the BioBERT model, it needs to be fine-tuned. Specifically: use historical text data for training as the fine-tuning dataset, and divide it into training and validation sets in a 7:3 ratio; use a manually labeled dataset labeling system, including entities or treatment conclusions such as drugs, pathological entities, organ parts, and pathological diagnosis opinions, and label entities in the BIOES format; use dmis-lab / biobert-base-cased-v1.1 provided in HuggingFace Transformers as the pre-trained model; perform minor fine-tuning on the BioBERT base model, with the following settings: set the task type to sequence labeling (Named Entity Recognition); set the input format to [CLS]+token1+token2+…+[SEP]; the output label is the BIO label corresponding to each token; define the loss function as cross-entropy loss, set the optimizer AdamW, fine-tune the epochs to 3-5, and set the batch size to 16-32. Fine-tuning of the BioBERT model was achieved by using a labeled training set and monitoring the F1-score on the validation set for early stopping. Each text data point from the current patient was input into the fine-tuned BioBERT model for word segmentation and keyword extraction, outputting a formatted JSON object of keywords related to gastrointestinal bleeding, thus identifying the keywords and all non-keyword segments in each text data point.

[0027] In other possible implementations of this invention, the jieba word segmentation method can be used to segment each piece of text data, and then the Latent Dirichlet Allocation (LDA) model can be used to extract keywords, or the TF-IDF algorithm can be used to extract keywords. The LDA model, TF-IDF algorithm and jieba word segmentation method are technical means well known to those skilled in the art, and will not be further limited or described here.

[0028] The risk feature value determination module 102 is used to determine the formal risk feature value of each text evaluation dimension based on the recording time interval between the current text data and the previous text data, the aspirin dosing interval of the current text data, the change in the number of non-key words in the sentence containing the keyword, and the stability of the changes of various keywords; and to determine the semantic risk feature value of each text evaluation dimension based on the correlation between the overall vector features of the keyword word vectors in each text evaluation dimension and the preset benchmark word vectors of each risk link.

[0029] After extracting keywords from text data, before analyzing the semantic features of each keyword in the text data, it is necessary to first consider that the writing style or writing pattern of clinical documents in medical data itself contains rich information. When doctors record text information of patients with gastrointestinal bleeding, even if they do not directly mention terms such as "black stool" or "vomiting blood," their writing behavior will produce systematic differences due to the severity of the condition. This difference is reflected in the surface features of the text: when faced with a more serious case of gastrointestinal bleeding, doctors' cognitive resources are mainly focused on the patient's rescue, resulting in a specific pattern in the writing of text data: the length of descriptive sentences becomes shorter (a certain physiological risk represented by a keyword appears continuously, indicating that this risk has not been cured during the treatment process, and in subsequent records, the changes compared to the previous one are generally recorded, rather than re-describing the risk represented by this keyword), and the interval between text data records is short (reflecting rapid changes in the condition).

[0030] In medical texts, keywords do not exist in isolation. When multiple identical keywords appear in two consecutive text datasets, indicating the same keyword combination pattern, it often signifies multiple concurrent factors interacting, leading to a significant increase in risk. Therefore, when evaluating the descriptive features of keywords, the co-occurrence characteristics of keyword combinations should be considered to further improve the accuracy of feature evaluation. Furthermore, considering that the risk of gastrointestinal bleeding varies depending on the time elapsed after aspirin administration, the aspirin dosing interval in the current text data can be used to assess non-semantic bleeding risk features. Therefore, this embodiment of the invention further determines the formal risk feature value for each text evaluation dimension based on the recording time interval between the current and previous text data, the aspirin dosing interval in the current text data, the change in the number of non-keyword segments in the sentence containing the keyword, and the stability of various keyword changes. This ensures that a larger formal risk feature value corresponds to a more significant bleeding risk feature in non-semantic features, thus increasing the reference value of the corresponding text evaluation dimension data.

[0031] Preferably, in some possible implementations of the embodiments of the present invention, the process of obtaining the formal risk feature value includes: Under each text evaluation dimension, each keyword in the current text data is taken as the target word in turn; The number of descriptive words for the target word in each text dataset is determined by the average number of non-keyword segments in all sentences containing the target word. The set of all types of keywords in each text dataset is taken as the corresponding keyword set. The symptom description coefficient of the target word in the current text dataset is determined based on the overlap of the keyword sets between the current and previous text datasets and the relative decrease in the number of descriptive words. In a specific implementation of this invention, the process of obtaining the symptom description coefficient includes: Based on the difference between the number of descriptive words for the target word in the previous text data and the number of descriptive words for the target word in the current text data, determine the reduction in descriptive words for the target word in the current text data; based on the intersection between the keyword set of the current text data and the keyword set of its corresponding previous text data, determine the reference set; based on the ratio between the number of keyword types in the reference set and the number of keyword types in the keyword set of the current text data, determine the stability of the keyword combination in the current text data.

[0032] Firstly, when gastrointestinal bleeding is severe, doctors will focus their cognitive resources on patient resuscitation, resulting in shorter descriptive sentences. Therefore, the greater the reduction in descriptive words, that is, the greater the reduction in descriptive words for the target word in the current text data, the higher the risk of gastrointestinal bleeding. In addition, the more similar the keyword combinations between the current text data and the previous text data, that is, the closer the keyword set of the reference set is to the current text data, the more it corresponds to the situation where multiple concurrent factors lead to a greater increase in risk, and the higher the corresponding risk of gastrointestinal bleeding.

[0033] Therefore, a positive correlation mapping is further performed between the product of the reduction in descriptive words and the stability of keyword combinations to determine the symptom description coefficient of the target words in the current text data. This ensures that the larger the symptom description coefficient, the higher the risk of gastrointestinal bleeding reflected by the corresponding text evaluation dimension, and the more attention should be paid to the corresponding text evaluation dimension.

[0034] In one specific implementation of this invention, the process of obtaining the symptom description coefficient is expressed by the following formula: ;in, For the first Target words in the current text data of each text evaluation dimension Symptom description coefficient; For the first The target words in the previous text data of the current text data for each text evaluation dimension The number of descriptive words and target words in the current text data The difference between the number of descriptive words, i.e., the target word The reduction in descriptive words in the current text data; For the first The number of keyword types in the intersection between the keyword set of the current text data and the keyword set of the corresponding previous text data under each text evaluation dimension; For the first The number of keyword types in the keyword set of the current text data under each text evaluation dimension; For the first Stability of keyword combinations in the current text data under each text evaluation dimension; for The function, whose purpose is to normalize, can be replaced by other normalization functions, such as linear normalization functions, which will not be elaborated further here.

[0035] It should be noted that, in order to ensure that the calculation results are meaningful, when performing fractional operations, if the denominator is 0, a parameter adjustment factor greater than 0 needs to be added to the denominator to prevent the denominator from being 0. The value of the parameter adjustment factor is set by the implementer according to the actual situation, and is set to 0.1 in this application.

[0036] By negatively correlating the time interval between the current text data's recording time and the previous text data's recording time, a time interval coefficient for the current text data is determined. The smaller the time interval between the recording times of adjacent text data, the shorter the text data recording interval, reflecting faster changes in the patient's condition. Therefore, the smaller the time interval coefficient of the current text data under the corresponding text assessment dimension, the higher the risk of gastrointestinal bleeding at the current moment, and the more attention should be paid to the corresponding text assessment dimension.

[0037] In one specific implementation of this invention, the method for negatively correlated mapping of the time interval between the recording time of the current text data and the recording time of the previous text data is as follows: normalize the time interval between the recording time of the current text data and the recording time of the previous text data to determine the reference interval length; and take the difference between the real number 1 and the reference interval length as the negative correlation mapping result, that is, the time interval coefficient of the current text data.

[0038] Further considering the impact of aspirin use, the risk development coefficient of the current text data is determined based on the time interval between the recording time of the current text data and the date of the patient's most recent aspirin use. In a specific implementation of this invention, the process of obtaining the risk development coefficient includes: The time of each aspirin administration for a patient with gastrointestinal bleeding is taken as the administration time; the time interval between the current text data recording time and each administration time is taken as the corresponding judgment time length; when the judgment time length is less than or equal to a preset first time length threshold, the normalized value of the judgment time length is taken as the reference development coefficient for the corresponding administration time; when the judgment time length is greater than the preset first time length threshold and less than or equal to a preset second time length threshold, the preset highest risk characteristic value is taken as the reference development coefficient for the corresponding administration time; when the judgment time length is greater than the preset second time length threshold, the negative correlation mapping value of the judgment time length is taken as the reference development coefficient for the corresponding administration time; the maximum value of the reference development coefficients for all administration times is taken as the risk development coefficient for the current text data.

[0039] In one specific implementation of this invention, the method for determining the negative correlation mapping of time length involves: normalizing the difference between the determined time length and a preset second time length threshold to determine a reference difference; and using the difference between the real number 1 and the reference difference as the negative correlation mapping result of the determined time length. It should be noted that, unless otherwise specified, the normalization method in this embodiment of the invention employs linear normalization, which will not be further elaborated here.

[0040] In one specific implementation of this invention, the first duration threshold is set to 31 days, the second duration threshold is set to 90 days, and the preset highest risk characteristic value is set to 1. According to the corresponding paper on the prediction of non-variceal upper gastrointestinal bleeding in patients taking enteric-coated aspirin based on aspirin risk scores, the first year of aspirin use is a high-risk period for gastrointestinal bleeding, with the relative risk of upper gastrointestinal bleeding being highest between days 31 and 90. The traumatic effect is cumulative over time; the bleeding risk is in an upward phase but has not reached its peak between days 0 and 30, while some patients show some adaptation within 90 days to one year, and the risk decreases compared to the peak phase. Therefore, the maximum value is selected as the risk development coefficient based on the time period in which each judgment duration falls. A higher risk development coefficient indicates a higher risk of gastrointestinal bleeding reflected in the current text data corresponding to the text assessment dimension, and thus requires more attention to the corresponding text assessment dimension.

[0041] Finally, based on the correlation, the weighted structural risk feature value of the target word in the current text data is determined by multiplying the risk development coefficient, time interval coefficient, and symptom description coefficient. It should be noted that, in addition to multiplication, implementers can also use other methods to determine the weighted structural risk feature value depending on the specific implementation environment. For example, the sum of the risk development coefficient, time interval coefficient, and symptom description coefficient can be normalized to determine the weighted structural risk feature value; thus, the larger the weighted structural risk feature value, the more attention should be paid to the text evaluation dimension of the target word.

[0042] Since the current text data under each text evaluation dimension typically corresponds to multiple keywords, further, under each text evaluation dimension, the corresponding formal risk feature value is determined based on the mean of the weighted structural risk feature values ​​of all keywords in the current text data. This ensures that the larger the formal risk feature value, the higher the attention should be paid to the corresponding text evaluation dimension at the non-semantic level, and a greater attention weight should be given to the features of the text evaluation dimension to improve the accuracy of the gastrointestinal bleeding risk assessment coefficient output by the neural network model.

[0043] In one specific implementation of this invention, the process of obtaining the formal risk feature value is expressed by the following formula: ;in, For the first Formal risk feature values ​​for each text assessment dimension; For the first The number of keyword categories in the current text data for each text evaluation dimension; For the first In the current text data of the text evaluation dimension, the first... Symptom description coefficients for various keywords; For the first The time interval coefficient of the current text data under each text evaluation dimension; For the first Risk development coefficient of current text data under each text evaluation dimension; For the first In the current text data of the text evaluation dimension, the first... Weighted structural risk eigenvalues ​​for various keywords.

[0044] Further analysis requires considering the semantic features of each text assessment dimension. Stronger semantic associations among keywords within a single dimension indicate more consistent risk signals and a higher likelihood of corresponding gastrointestinal bleeding risk. For example, the strong association between deep ulcers and vascular exposure in PACS clearly points to severe mucosal damage. Furthermore, for the current text data of each assessment dimension, the more similar the semantic features of the corresponding keywords are to the semantic features of each risk factor representing bleeding risk, the more significant the bleeding risk characteristic of the corresponding text assessment data at the current moment, and the more attention should be paid to that text assessment dimension. Therefore, based on the correlation between the overall word vector features of keywords in each text assessment dimension and the preset benchmark word vectors of each risk factor, the semantic risk feature value of each text assessment dimension is determined. A higher semantic risk feature value indicates a higher gastrointestinal bleeding risk reflected by the corresponding text assessment dimension, and thus, a greater focus on that text assessment dimension at the semantic analysis level.

[0045] Preferably, in some possible implementations of the embodiments of the present invention, the process of obtaining semantic risk feature values ​​includes: In each text evaluation dimension, all keywords in the current text data are converted into word vectors using word2vec to determine the word vector of each keyword. It should be noted that word2vec is a word vector conversion method well known to those skilled in the art, and will not be further limited or elaborated here.

[0046] The semantic similarity is determined based on the cosine similarity between the word vector of each keyword and the word vector of every other keyword. In the current text data, hierarchical clustering is performed using the negative correlation mapping value of the semantic similarity between keywords as the clustering distance, resulting in at least two keyword clusters. In a specific implementation of this invention, the reciprocal of the semantic similarity is used as the clustering distance for hierarchical clustering. It should be noted that hierarchical clustering is a well-known technique among those skilled in the art and will not be elaborated further here. Through hierarchical clustering, keywords with similar word vectors, i.e., semantically similar, are clustered into the same keyword cluster. Under each text evaluation dimension, the higher the semantic similarity between keywords in each keyword cluster, the stronger the semantic association of keywords within a single dimension, and the higher the risk of gastrointestinal bleeding reflected by the corresponding text evaluation dimension. Therefore, the risk score is further determined based on the mean of all semantic similarities between all keywords in each keyword cluster, so that the higher the risk score, the more attention needs to be paid to the corresponding text evaluation dimension.

[0047] Further, based on the mean vector of word vectors of all keywords in each keyword cluster, the corresponding center vector is determined; based on the vector similarity distribution between the center vector and the preset benchmark word vectors of each risk link, the risk association strength of each keyword cluster is determined; the process of obtaining the risk association strength includes: calculating the cosine similarity between the center vector and the preset benchmark word vectors of each risk link, and determining the link matching degree between each keyword cluster and each risk link.

[0048] The central vector represents the semantic features of each keyword cluster as a whole, and each risk link can directly or indirectly reflect the state of gastrointestinal bleeding. In a specific implementation of this invention, the types of risk links include four links: mucosal injury, drug-driven, symptom warning, and protective effect. Among them, the keywords of the mucosal injury link include "ulcer", "erosion", and "vascular exposure"; the keywords of the drug-driven link include "NSAIDs combination" and "high-dose aspirin"; the keywords of the symptom warning link include "melena", "hematemesis", and "anemia"; and the keywords of the protective effect include "regularity". "PPI" and "Hp eradication"; and in this embodiment of the invention, the preset benchmark word vector for each risk link is the average vector of all word vectors corresponding to all keywords in each risk link. The preset benchmark word vector obtained by the average vector represents the overall semantic features of each risk link. Therefore, for each central vector, the smaller the cosine similarity between it and a certain preset benchmark word vector, the more the semantic features of the overall keywords of the keyword cluster of the central vector are consistent with the semantic features of the keywords of the risk link corresponding to the preset benchmark word vector. That is, the more likely the corresponding keyword cluster is to correspond to the risk features of gastrointestinal bleeding.

[0049] Therefore, the risk association strength is further determined based on the maximum value of the link matching degree between each keyword cluster and all risk links. The greater the risk association strength, the more likely the corresponding keyword cluster is to correspond to the state of gastrointestinal bleeding, and the higher the risk of gastrointestinal bleeding reflected. Therefore, the corresponding text evaluation dimension needs to be paid more attention to.

[0050] Finally, based on the correlation, the local risk feature value of each keyword cluster is determined by multiplying the risk association strength with the risk score; the semantic risk feature value of each text evaluation dimension is determined by the mean of the local risk feature values ​​of all keyword clusters; the larger the obtained semantic risk feature value, the higher the risk of gastrointestinal bleeding reflected by the corresponding text evaluation dimension, and the more attention needs to be paid to the corresponding text evaluation dimension at the semantic analysis level.

[0051] In one specific implementation of this invention, the process of obtaining semantic risk feature values ​​includes: ;in, For the first Semantic risk feature values ​​for each text evaluation dimension; For the first The number of keyword clusters in the current text data for each text evaluation dimension; For the first In the current text data of the text evaluation dimension, the first... The average of all semantic similarities between all keywords in a keyword cluster, which is also the corresponding risk score; For the first In the current text data of the text evaluation dimension, the first... The maximum value of the matching degree between a keyword cluster and all risk links, which is also the corresponding risk association strength; For the first In the current text data of the text evaluation dimension, the first... Local risk feature values ​​of keyword clusters.

[0052] The gastrointestinal bleeding risk assessment module 103 is used to determine the attention weight of each text assessment dimension based on the formal risk feature value and the semantic risk feature value; adjust the neural network model according to the attention weight, and output the gastrointestinal bleeding risk assessment coefficient of the patient at the current moment.

[0053] Finally, based on the dimensional attention requirements represented by the formal risk feature value and semantic risk feature value of each text evaluation dimension, the attention weight of each text evaluation dimension is determined according to the formal risk feature value and semantic risk feature value. Thus, when adjusting the neural network model according to the attention weight of each text evaluation dimension, the text evaluation dimension with a larger attention weight has a greater impact on the output result.

[0054] Preferably, in a specific implementation of this invention, the attention weight for each text evaluation dimension is determined based on the product of the formal risk feature value and the semantic risk feature value; such that the larger the attention weight, the greater the impact of the corresponding text evaluation dimension on the output of the neural network model; the process of obtaining the attention weight is expressed by the formula: ;in, For the first Attention weights for each text evaluation dimension; For the first Semantic risk feature values ​​for each text evaluation dimension; For the first The text assessment dimension is a form of risk feature value.

[0055] Finally, the neural network model is adjusted according to the attention weights of each text evaluation dimension to output the gastrointestinal bleeding risk assessment coefficient for the patient at the current moment. Preferably, in some possible implementations of this invention, the process of obtaining the gastrointestinal bleeding risk assessment coefficient includes: In each text evaluation dimension, the frequency of each keyword in the current text data is used as the corresponding frequency feature value. The frequency feature values ​​of all keywords in the current text data are arranged in the order of their occurrence across all text data to determine the feature vector for each text evaluation dimension. Keywords that do not appear in the current text data have a frequency feature value of 0, but are still present in their corresponding feature vectors. The product of the feature vector of the current text data in each text evaluation dimension and its corresponding attention weight is used as the weighted vector; that is, each element in the feature vector is weighted by the attention weight to determine the corresponding weighted vector, thereby improving the impact of the output results for each text evaluation dimension.

[0056] Then, the weighted vectors of all text evaluation dimensions are concatenated to determine the joint feature vector; the joint feature vector is then input into the trained multi-layer fully connected neural network to output the risk assessment index of the digestive tract. In one specific implementation of this invention, the multi-layer fully connected neural network employs a three-layer fully connected neural network; the first layer is fully connected + BatchNorm + ReLU + Dropout; the second layer is fully connected + BatchNorm + ReLU + Dropout; and the third layer is fully connected + Sigmoid output; the input dimension is the fused feature dimension, and the output is a single predicted probability value; the gastrointestinal bleeding results of the data are manually labeled based on multimodal data of historical gastrointestinal bleeding patients; the joint feature vector of these multimodal data is extracted and divided into training and validation sets in a 7:3 ratio; binary cross-entropy is used as the loss function, and the AdamW optimizer is used for end-to-end joint training to train the prediction neural network; then, the joint feature vector of the gastrointestinal bleeding patient at the current moment is input into the trained three-layer fully connected neural network to output the gastrointestinal bleeding risk probability, and this gastrointestinal bleeding risk probability is used as the gastrointestinal bleeding risk assessment coefficient of the gastrointestinal bleeding patient at the current moment.

[0057] In one specific implementation of this invention, after outputting the gastrointestinal bleeding risk assessment coefficient for the patient at the current moment, the method further includes: when the gastrointestinal bleeding risk assessment index is greater than a preset risk threshold, it indicates that the risk of gastrointestinal bleeding is high at the current moment, and a bleeding warning is issued; when the gastrointestinal bleeding risk assessment is less than or equal to the preset risk threshold, no bleeding warning is issued; wherein the preset risk threshold is set to 0.75, which can be adjusted according to the specific implementation environment, so that it can react according to the magnitude of the gastrointestinal bleeding risk assessment index, making the assessment result more specific.

[0058] In summary, a deep learning-based gastrointestinal bleeding risk assessment system first preprocesses unstructured text data to identify keywords and non-keywords. Then, combining the risk characteristics represented by changes in the text data and the impact of aspirin on bleeding risk, formal risk feature values ​​are determined at the text form dimension. Next, a correlation analysis is performed between the semantic information of the text data and the semantics of the various risk stages corresponding to the gastrointestinal risk, determining semantic risk feature values ​​at the semantic information dimension. Finally, combining the formal and semantic risk feature values, the attention weight for each text assessment dimension is determined, and the neural network model is adjusted based on the attention weight to improve the accuracy of the output gastrointestinal bleeding risk assessment coefficient. This invention further enhances the influence of key modal data on the final prediction results, reduces noise interference from weak modal data, and further improves the accuracy of the prediction results.

[0059] This application also provides a computer device; please refer to [link / reference]. Figure 2 The illustration shows a schematic diagram of a computer device structure provided by an embodiment of the present invention. The computer device includes a memory 201, a processor 202, and a computer program 203 stored in the memory 201 and running on the processor 202. When the processor 202 executes the computer program 203, the computer device can execute any of the deep learning-based gastrointestinal bleeding risk assessment systems described above.

[0060] This application also provides a computer program product that, when run on a computer device, enables the computer device to execute any of the deep learning-based gastrointestinal bleeding risk assessment systems described above.

[0061] This application also provides a computer-readable storage medium storing computer program code. When the computer program code is run on a computer device, the computer device can execute any of the deep learning-based gastrointestinal bleeding risk assessment systems described above.

[0062] In the embodiments provided in this application, it should be understood that the computer device, computer program product and computer-readable storage medium provided are all used to execute the corresponding system provided above, and therefore the beneficial effects they can achieve can be referred to the beneficial effects of the system provided above, which will not be repeated here.

[0063] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0064] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. A deep learning-based gastrointestinal bleeding risk assessment system, characterized in that, The system includes: The data acquisition and preprocessing module is used to acquire each text data of patients with gastrointestinal bleeding in each text evaluation dimension from the medical database; and extract keywords and all non-keywords from the text data. The risk feature value determination module is used to determine the formal risk feature value of each text evaluation dimension based on the recording time interval between the current text data and the previous text data, the aspirin dosing interval of the current text data, the change in the number of non-key words in the sentence containing the keyword, and the stability of the changes of various keywords; and to determine the semantic risk feature value of each text evaluation dimension based on the correlation between the overall vector features of the keyword word vectors and the preset benchmark word vectors of each risk link. The gastrointestinal bleeding risk assessment module is used to determine the attention weight of each text assessment dimension based on the formal risk feature value and the semantic risk feature value; adjust the neural network model according to the attention weight, and output the gastrointestinal bleeding risk assessment coefficient of the patient at the current moment.

2. The gastrointestinal bleeding risk assessment system based on deep learning according to claim 1, characterized in that, The process of extracting keywords and all non-keyword segments from the text data includes: Each piece of text data is input into the BioBERT model for word segmentation and keyword extraction, and the output includes keywords and non-keyword segments corresponding to gastrointestinal bleeding; the non-keyword segments are other words besides the keywords.

3. The gastrointestinal bleeding risk assessment system based on deep learning according to claim 1, characterized in that, The process of obtaining the formal risk characteristic value includes: Under each text evaluation dimension, each keyword in the current text data is taken as the target word in turn; The number of descriptive words for the target word in each text data is determined by the average number of non-keywords in all sentences containing the target word in each text data; the set of all kinds of keywords in each text data is taken as the corresponding keyword set; the symptom description coefficient of the target word in the current text data is determined by the overlap of the keyword set between the current text data and the previous text data and the relative reduction of the number of descriptive words. The time interval coefficient of the current text data is determined by negatively correlated with the time interval between the recording time of the current text data and the recording time of the previous text data. The risk development coefficient of the current text data is determined based on the time interval between the recording time of the current text data and the date of the patient's most recent aspirin administration. The weighted structural risk feature value of the target word in the current text data is determined based on the product of the risk development coefficient, the time interval coefficient, and the symptom description coefficient. For each text evaluation dimension, the corresponding formal risk feature value is determined based on the mean of the weighted structural risk feature values ​​of all keywords in the current text data.

4. The gastrointestinal bleeding risk assessment system based on deep learning according to claim 3, characterized in that, The process of obtaining the symptom description coefficient includes: Based on the difference between the number of descriptive words for the target word in the previous text data and the number of descriptive words for the target word in the current text data, determine the amount of reduction in descriptive words for the target word in the current text data; A reference set is determined based on the intersection between the keyword set of the current text data and the keyword set of the corresponding previous text data; the keyword combination stability of the current text data is determined based on the ratio between the number of keyword types in the reference set and the number of keyword types in the keyword set of the current text data. By performing a positive correlation mapping between the product of the reduction in descriptive words and the stability of the keyword combination, the symptom description coefficient of the target word in the current text data is determined.

5. A deep learning-based gastrointestinal bleeding risk assessment system according to claim 3, characterized in that, The process of obtaining the risk development coefficient includes: The time when a patient with gastrointestinal bleeding takes aspirin is taken as the time of administration; the time interval between the current text data recording time and each time of administration is taken as the corresponding judgment time length. When the determination time length is less than or equal to the preset first time length threshold, the normalized value of the determination time length is used as the reference development coefficient for the corresponding administration time. When the determination time length is greater than a preset first time threshold and less than or equal to a preset second time threshold, the preset highest risk characteristic value is used as the reference development coefficient for the corresponding administration time. When the determination time length is greater than the preset second time length threshold, the negative correlation mapping value of the determination time length is used as the reference development coefficient of the corresponding administration time. The maximum value of the reference development coefficients at all times of administration is taken as the danger development coefficient of the current text data.

6. The gastrointestinal bleeding risk assessment system based on deep learning according to claim 1, characterized in that, The process of obtaining the semantic risk feature value includes: In each text evaluation dimension, word2vec is used to convert all keywords in the current text data into word vectors, and the word vector of each keyword is determined. The semantic similarity is determined based on the cosine similarity between the word vector of each keyword and the word vector of each other keyword; in the current text data, hierarchical clustering is performed using the negative correlation mapping value of the semantic similarity between keywords as the clustering distance to obtain at least two keyword clusters. The center vector is determined based on the mean vector of word vectors of all keywords in each keyword cluster; the risk score is determined based on the mean of all semantic similarities between all keywords in each keyword cluster. Based on the vector similarity distribution between the central vector and the preset benchmark word vectors of each risk link, the risk association strength of each keyword cluster is determined; The local risk feature value of each keyword cluster is determined by multiplying the risk association strength with the risk score; the semantic risk feature value of each text evaluation dimension is determined by the mean of the local risk feature values ​​of all keyword clusters.

7. A deep learning-based gastrointestinal bleeding risk assessment system according to claim 6, characterized in that, The process of obtaining the risk association strength includes: Calculate the cosine similarity between the center vector and the preset benchmark word vector of each risk link to determine the link matching degree between each keyword cluster and each risk link; determine the corresponding risk association strength based on the maximum link matching degree between each keyword cluster and all risk links.

8. The gastrointestinal bleeding risk assessment system based on deep learning according to claim 1, characterized in that, The process of obtaining the attention weights includes: The attention weight for each text evaluation dimension is determined based on the product of the formal risk feature value and the semantic risk feature value.

9. A deep learning-based gastrointestinal bleeding risk assessment system according to claim 1, characterized in that, The process of obtaining the gastrointestinal bleeding risk assessment coefficient includes: In each text evaluation dimension, the frequency of each keyword in the current text data is used as the corresponding frequency feature value; the frequency feature values ​​of all keywords in the current text data are arranged in the order of their occurrence in all text data to determine the feature vector of the current text data in each text evaluation dimension. The product of the feature vector of the current text data in each text evaluation dimension and the corresponding attention weight is used as the weighted vector; the joint feature vector is determined by concatenating the weighted vectors of all text evaluation dimensions. The joint feature vector is input into a trained multilayer fully connected neural network, which outputs a risk assessment index for the digestive tract.

10. A deep learning-based gastrointestinal bleeding risk assessment system according to claim 1, characterized in that, The output of the gastrointestinal bleeding risk assessment coefficient for the patient at the current moment also includes: When the gastrointestinal bleeding risk assessment index is greater than the preset risk threshold, a bleeding warning is issued; No bleeding warning will be issued when the risk assessment of gastrointestinal bleeding is less than or equal to a preset risk threshold.

Images