Device and method for automatically generating fishbone diagram of quality control circle in hospital adverse event management

Abstract

The present invention provides a device and method for automatically generating a fishbone diagram of a quality control circle in hospital adverse event management, and belongs to the field of medical information management technology. The data acquisition module collects text descriptive data of adverse events in the hospital database, the data preprocessing module performs data cleaning and long text segmentation on the text descriptive data to obtain valid data, the deep learning module trains and tests the valid data, performs entity recognition on the main causes of the adverse events, and realizes the adverse event entity extraction function. The fishbone diagram generation module automatically generates the fish head and fishbone content of the fishbone diagram from the entity extraction results, and realizes the visual presentation of the adverse event analysis results. The fishbone diagram sending module is used to push the fishbone diagram, so that the medical and nursing management personnel of each department can view the latest fishbone diagram in time. The present invention can realize the intelligent and logical analysis of the causes of the entire adverse event.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical information management, and in particular relates to a device and method for automatically generating a quality control circle fishbone diagram in hospital adverse event management. Background Art

[0002] Adverse events are injuries caused by medical treatment, which are contrary to the natural course of the disease, such as falls, unplanned extubation, etc. The occurrence of adverse events is a secondary injury to patients, prolonging the patient's hospitalization time, causing a waste of medical resources, worsening the doctor-patient relationship, and bringing extremely negative impacts on the reputation of medical staff and hospitals. In order to reduce the occurrence of adverse events, hospitals need to collect adverse event information and promptly analyze and feedback the causes of adverse events.

[0003] Currently, some artificial intelligence technical methods have been applied in the field of hospital adverse event management, but they mainly focus on the automatic identification of adverse events. For example, Chinese patent (CN110704526A) discloses a method and device for acquiring information of adverse events. The invention collects structured source data, semi-structured source data, structured text mixed source data and text source data, identifies preset information items according to their corresponding identification methods, maps the information items into preset standard fields, and obtains reporting information of adverse events. The entire reporting process supports the collection and identification of multiple types of source data; Chinese patent (CN112382388A) discloses an early warning method for pressure sore adverse events. The invention obtains early warning indicators for pressure sore adverse events, and at the same time extracts features of the early warning indicator system for pressure sore adverse events constructed based on random up- and down-sampling text convolutional neural network, and finally uses the algorithm of support vector machine to establish a prediction model and output early warning results; Chinese patent (CN112766903A) discloses a method, device, equipment and medium for identifying adverse events. The method for identifying adverse events provided by the invention can select different recognition models according to different text types, and then perform corresponding adverse event recognition on the text to be recognized according to the selected recognition model.

[0004] Existing research mainly focuses on methods for automatically identifying / predicting adverse events, without the need for manual screening and identification, to avoid missed / delayed reporting of adverse events. These methods can identify adverse events well, but in terms of quality control circle management and analysis of adverse events, manual analysis and classification methods are still used. Artificial intelligence methods are not used to conduct intelligent and logical analysis of the causes of adverse events, and the entity content of the causes of adverse events cannot be automatically extracted. Summary of the invention

[0005] In view of the deficiencies in the prior art, the present invention provides a device and method for automatically generating a fishbone diagram of a quality control circle in hospital adverse event management, which collects textual descriptive data of adverse events reported by hospitals, and automatically generates a fishbone diagram of a quality control circle in adverse event management after data processing and recognition, thereby realizing intelligent and logical analysis of the causes of the entire adverse event.

[0006] The present invention achieves the above technical objectives through the following technical means.

[0007] The method for automatically generating a fishbone diagram of a quality control circle in hospital adverse event management comprises the following steps:

[0008] Step (1), read the entity extraction results of the saved detection data set D″′, use the clustering algorithm, set the word segmentation length threshold, and generate a text matrix;

[0009] Step (2), randomly select K points as cluster centers, use the fishbone diagram to automatically generate the distance calculation method to calculate the distance d between each point and the K cluster centers, find the cluster center closest to a point, and assign the point to the nearest cluster center. At this time, the text matrix is ​​divided into K clusters, where K is the number of clusters;

[0010] Step (3): Use the fishbone diagram to automatically generate the mean point calculation method to calculate the mean point of each cluster. The mean point of each cluster is taken as the cluster center of each cluster;

[0011] Step (4), calculate the distance d′ from each point to all cluster centers, and assign the point to the nearest cluster;

[0012] Step (5), repeatedly iterate steps (3) and (4) until the mean point no longer changes, otherwise go to step (3);

[0013] Step (6), finally generate clusters C = {C1, C2, C3, ..., C k ,...,C K}, where Ck is the category number;

[0014] Step (7), perform word frequency statistics on the subclusters contained in cluster class C. Class Ck contains F phrases, and the word frequency statistics are arranged in descending order as {C k1 ,C k2 ,C k3 ,...,C kf ...,C kF}, where C kf C k Phrases with frequency ranking f in the class, 1≤f≤F, 1≤k≤K;

[0015] Step (8), clustering is completed;

[0016] Step (9), the fishbone diagram generation module reads the reported adverse event category and automatically generates the content of the fish head of the fishbone diagram;

[0017] Step (10), sort the results in descending order and extract the phrases of each cluster {C k1 ,C k2 ,C k3 ,...,C kg ...,C kG}, where G is the number of phrases to be extracted, 1≤g≤G, 1≤G≤F, initialization k=1, g=1;

[0018] Step (11), read C k1 For the main title of cluster Ck, the fishbone diagram generation module automatically generates the main bone content of the fishbone diagram;

[0019] Step (12), g = g + 1, read C kg For the subtitle of cluster Ck, the fishbone diagram generation module automatically generates the middle bone content of the fishbone of the fishbone diagram;

[0020] Step (13), if 1≤g≤G, go to step (12), otherwise go to step (14);

[0021] Step (14), k=k+1;

[0022] Step (15), if 1≤k≤K, go to step (11), otherwise go to step (16);

[0023] Step (16), generating fishbone contents of the fishbone diagram according to the contents of the large bones and the contents of the middle bones;

[0024] Step (17), generating a fishbone diagram according to the content of the fish head and the content of the fish bones.

[0025] Furthermore, the entity extraction result of the detection data set D″′ is achieved by the following method:

[0026] 1) Clean the current hospital adverse event detection dataset D″ to obtain a valid dataset D″′, which is then input into the trained deep learning model;

[0027] 2) In the word embedding layer, the dataset D″′ is segmented, and the long text of the adverse event text data ds′m in the dataset D″′ is segmented into ds′ m ={w″1,w″2,w″3,...,w″ i ,...,w″ I}, where w″ i represents the i-th sentence divided in ds′m, I represents the total number of sentences divided in a certain adverse event text, 1≤i≤I;

[0028] 3) Then w″ i The word segmentation process gets w″ i ={word′1, word′2, word′3, ..., word′j, ..., word′J}, where word′j represents w″ i The jth word divided in , J represents the total number of words divided in the sentence, 1≤j≤J;

[0029] 4) Convert the word word'j after word segmentation to a word vector x' j ={word′j1, word′j2, word′j3,..., word′j e , ..., word′j E}, where x′ j is the word vector expression of word word′j, word′j e is a value of 0 or 1, 1≤e≤E, E represents the dimension of the word vector; i After the word vectors of all words are divided, they are converted into w″′ i Indicates that w″′ i ={x′1,x′2,x′3,...,x′ j ,...,x′ J},x′ j ∈R o , R represents the word vector space, o is the dimension of character embedding;

[0030] 5) w″′ i ={x′1,x′2,x′3,...,x′ j ,...,x′ J} As the input feature of the bidirectional long short-term memory network layer, read and save the state sequence at each moment from left to right (h * 1,h * 2,h * 3,...,h * J ), read and save the state sequence at each moment from right to left (h *′ J ,...,h *′ 3,h *′ 2,h *′ 1) At time t, the output of each position is concatenated into h′ tj =[h * j :h *′ j ], get the complete sequence state of sentence t where d ω′ is the dimension of the word vector, and ω′ is a parameter matrix learned through training;

[0031] 6) In the attention mechanism layer, the tanh function and soft max function are used to obtain the weight matrix λ′=tanh(H′) and α′=softmax(ω′ T λ′), the sentence vector is represented by σ′=H′α′ T , the final vector representation of the sentence used for classification is H *′ =tanh(σ′), where ω′ T is the transpose of a detected parameter matrix;

[0032] 7) H *′ As the input feature of the classification output layer, the softmax classifier is used to obtain the optimal predicted sequence label Y′={y′1,y′2,y′3,...,y′ j ,...,y′ J};

[0033] 8) Compare the predicted sequence label Y′ with the sample label and calculate the precision Q, recall R1 and F1 value;

[0034] 9) Evaluate the accuracy Q, recall R1 and F1 value. If the Q, R1 and F1 values ​​are all greater than the threshold a, go to 10), otherwise go to 2) and reset the parameter tuning;

[0035] 10) The current detection process ends;

[0036] 11) Save the entity extraction results of the detection data set D″′.

[0037] Furthermore, the Said Said Among them, TP refers to the number of samples with positive predicted labels whose results are positive, FP refers to the number of samples with negative predicted labels whose results are positive, TN refers to the number of samples with negative predicted labels whose results are negative, and FN refers to the number of samples with positive predicted labels whose results are negative.

[0038] Furthermore, the training process of the deep learning model is:

[0039] 1) From the hospital database, collect various adverse event data sets D = {D1, D2, D3, ..., D s ,...,D S}, where Ds is the adverse event category number, 1≤s≤S, S is the total number of categories in the adverse event text dataset, and D s={ds1,ds2,ds3,...,ds n ,...,ds N}, dsn is the text data number of the adverse event contained in the Ds class, 1≤n≤N, N is D s The total number of adverse event text data included in the class;

[0040] 2) Clean the dataset D to obtain a valid dataset D′. The text data D′ contained in each category of the dataset s ={ds1,ds2,ds3,...,ds m ,...,ds M}, where ds m is the text data number of the adverse event contained in the D′s class, 1≤m≤M, M is the number of D′ after cleaning s The total number of text data in the adverse events included in the class;

[0041] 3) In the word embedding layer, the valid data set D′ is segmented. First, ds m Split long text into ds m ={w1, w2, w3, ..., w i , ..., w I}, where w i Representative ds m The i-th sentence divided in , I represents the total number of sentences divided in a certain adverse event text, 1≤i≤I;

[0042] 4) Then w i Word segmentation processing, get w i ={word1,word2,word3,...,wordj,...,wordJ}, where wordj represents w i The jth word divided in , J represents the total number of words divided in a sentence, 1≤j≤J;

[0043] 5) Convert the word wordj after word segmentation into word vectors to obtain word vector xj = {wordj1, wordj2, wordj3, ..., wordj e ,...,wordj E}, where x j is the word vector expression of word wordj, wordj e is a value of 0 or 1, 1≤e≤E, E represents the dimension of the word vector; i After all the words are divided, the word vectors are transformed and used as w′ i It means that w′ i ={x1,x2,x3,...,x j,...,x J}, x j ∈R o , R represents the word vector space, o is the dimension of character embedding;

[0044] 6) w′ i ={x1,x2,x3,...,x j ,...,x J} As the input feature of the bidirectional long short-term memory network layer, read and save the state sequence (h1,h2,h3,...,h J ), read and save the state sequence at each moment from right to left in reverse (h′ J ,...,h′3,h′2,h′1), at time t, the output of each position is concatenated into h tj =[h j :h′ j ], and obtain the complete sequence state H = {h t1 ,h t2 ,h t3 ,...,h tj ,...,h tJ}, where d ω is the dimension of the word vector, and ω is a parameter matrix learned through training;

[0045] 7) In the attention mechanism layer, the tanh function and soft max function are used to obtain the weight matrix λ = tanh (H) and α = softmax (ω T λ), the sentence vector is represented by σ=Hα T , the final vector representation of the sentence used for classification is H * = tanh(σ), where ω T is the transpose of the parameter matrix learned through training;

[0046] 8) H * As the input feature of the classification output layer, the softmax classifier is used to obtain the optimal prediction sequence label Y = {y1, y2, y3, ..., y j , ..., y j}; The training process ends.

[0047] Furthermore, the fishbone diagram is automatically generated and the distance is calculated by:

[0048] Suppose we calculate the distance between text 1 and text 2, and the text 1 vector is represented by {v1,v2,v3,...,v i ,...,v p}, the text2 vector is represented as {u1,u2,u3,...,u i ,...,u p}, where p represents the number of text objects, 1≤i≤p, then the distance calculation formula is

[0049] Furthermore, the method for automatically generating and calculating the mean point of the fishbone diagram is:

[0050] Assume that the computation cluster C k The mean point C k The document in the class represents {T1,T2,T3,...,T b ,...,T B}, the document vector is represented as {T 1k ,T 2k ,T 3k ,...,T bk ,...,T Bk}, then cluster C k The mean point of is: Among them C k The document number in the class is T b , B is returned to C k The number of documents in the class, 1≤b≤B.

[0051] A device for automatically generating a fishbone diagram of a quality control circle in hospital adverse event management, comprising a data acquisition module, a data preprocessing module, a deep learning module, a fishbone diagram generation module and a fishbone diagram sending module connected in sequence; the deep learning module comprises a deep learning training submodule and a deep learning detection submodule, both of which use a deep learning model, and the deep learning model comprises a word embedding layer, a bidirectional long short-term memory network layer, an attention mechanism layer and a classification output layer;

[0052] The data collection module is used to collect textual descriptive data of adverse events in the hospital database;

[0053] The data preprocessing module is used to perform data cleaning and long text segmentation on the text descriptive data to obtain valid data;

[0054] The deep learning module is used to train and test valid data, perform entity recognition on the main causes of adverse events, and realize the adverse event entity extraction function;

[0055] The fishbone diagram generation module is used to automatically generate the fish head and fishbone content of the fishbone diagram from the entity extraction results, so as to realize the visual presentation of the adverse event analysis results;

[0056] The fishbone diagram sending module is used to push the fishbone diagram, so that the medical and nursing management personnel of each department can view the latest fishbone diagram in time.

[0057] The beneficial effects of the present invention are as follows: the present invention extracts adverse event text data from a hospital database, adopts deep learning technology, first trains a deep learning model, and then generates an optimal prediction sequence label through detection, which can automatically extract the entity content of the cause of the adverse event, replacing the work of manual analysis and classification; the present invention adopts a clustering algorithm, sorts the entity extraction results by keyword frequency, automatically generates the fish head and fish bones (including large bones and middle bones) of the fishbone diagram, and intelligently generates the fishbone diagram, avoiding errors that may be caused by manual analysis and improving the recognition accuracy; finally, the fishbone diagram sending module automatically generates a picture of the fishbone diagram and pushes it, helping medical personnel to quickly and directly understand the main causes of the adverse events, forming a good feedback loop and reducing the incidence of adverse events. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is a schematic diagram of the structure of the device for automatically generating the fishbone diagram of the quality control circle in hospital adverse event management according to the present invention;

[0059] Figure 2 A flow chart showing effective data training for the deep learning training submodule of the present invention;

[0060] Figure 3 A flow chart for detecting valid data for the deep learning training submodule of the present invention;

[0061] Figure 4 A flow chart of a method for automatically generating a quality control circle fishbone diagram in hospital adverse event management according to the present invention;

[0062] Figure 5 The fishbone diagram is generated by the automatic generation method of the quality control circle fishbone diagram in the management of adverse events such as falls of hospital patients according to the present invention.

[0063] In the figure: 1-data acquisition module, 2-data preprocessing module, 3-deep learning module, 4-fishbone diagram generation module, 5-fishbone diagram sending module. DETAILED DESCRIPTION

[0064] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0065] like Figure 1As shown, the automatic generation device of the fishbone diagram of the quality control circle in the hospital adverse event management of the present invention includes a data acquisition module 1, a data preprocessing module 2, a deep learning module 3, a fishbone diagram generation module 4 and a fishbone diagram sending module 5. The data acquisition module 1 is connected to the data preprocessing module 2, the data preprocessing module 2 is connected to the deep learning module 3, the deep learning module 3 is connected to the fishbone diagram generation module 4, and the fishbone diagram generation module 4 is connected to the fishbone diagram sending module 5.

[0066] The data collection module 1 is used to collect textual descriptive data of adverse events in the hospital database;

[0067] The data preprocessing module 2 is used to clean the text descriptive data and segment the long text to obtain valid data;

[0068] The deep learning module 3 includes a deep learning training submodule and a deep learning detection submodule. Both the deep learning training submodule and the deep learning detection submodule use a deep learning model (including a word embedding layer, a bidirectional long short-term memory network layer, an attention mechanism layer, and a classification output layer) to train and test effective data, respectively, to perform entity recognition on the main causes of adverse events, and to realize the adverse event entity extraction function;

[0069] The fishbone diagram generation module 4 runs a fishbone diagram automatic generation method, which is used to automatically generate the fish head and fishbone contents of the fishbone diagram through a clustering algorithm based on the entity extraction results, so as to realize the visual presentation of the adverse event analysis results;

[0070] The fishbone diagram sending module 5 is used to automatically push the fishbone diagram through the hospital's official account, so that medical and nursing management personnel of various departments can view the latest fishbone diagram in time.

[0071] like Figure 2 As shown in the figure, the process of deep learning training submodule training valid data is as follows:

[0072] Step S1: Collect various adverse event data sets D = {D1, D2, D3, ..., D s ,...,D S}, where D s is the adverse event category number, 1≤s≤S, S is the total number of categories in the adverse event text dataset, D s ={ds1,ds2,ds3,...,ds n ,...,ds N}, dsn is D s The text data number of the adverse event contained in the class, 1≤n≤N, N is D sThe total number of adverse event text data included in the category; adverse event categories include unplanned extubation, patient falls, other skin injuries, pressure sores, pressure injuries, accidents, medication errors, and other events;

[0073] Step S2: Clean the dataset D and delete the duplicate text data in the dataset D to obtain a valid dataset D′. The categories of the dataset remain unchanged, but the amount of text data contained in each category is reduced. s ={ds1,ds2,ds3,...,ds m ,...,ds M}, where dsm is D′ s The text data number of the adverse event contained in the class, 1≤m≤M, M is D′ after cleaning s The total number of text data in the adverse events included in the class;

[0074] Step S3: In the word embedding layer, use the existing word segmentation tools and the pre-prepared main feature dictionary and regular expressions to segment the valid data set D′. First, split the long text dsm into ds m ={w1,w2,w3,...,w i ,...,w I}, where w i represents the i-th sentence divided in dsm, I represents the total number of sentences divided in a certain adverse event text, 1≤i≤I;

[0075] Step S4: Then w i Word segmentation processing, get w i ={word1,word2,word3,...,wordj,...,wordJ}, where wordj represents w i The jth word divided in , J represents the total number of words divided in a sentence, 1≤j≤J;

[0076] Step S5: Convert the word wordj after word segmentation to a word vector x j ={wordj1, wordj2, wordj3,..., wordj e , ..., wordj E}, where x j is the word vector expression of word wordj, wordj e is a value of 0 or 1, E represents the word vector dimension, 1≤e≤E; i After all the words are divided, the word vectors are transformed and used as w′ i It means that w′ i ={x1,x2,x3,...,xj ,...,x J}, x j ∈R o , R represents the word vector space, o is the dimension of character embedding;

[0077] Step S6: In the bidirectional long short-term memory network layer, w′ i ={x1,x2,x3,...,x j ,...,x J} as the input feature of this layer, read and save the state sequence (h1,h2,h3,...,h J ) and read and save the state sequence (h′) at each moment from right to left. J ,...,h′3,h′2,h′1), at time t, the output of each position is concatenated into h tj =[h j :h′ j ], and obtain the complete sequence state H = {h t1 ,h t2 ,h t3 ,...,h tj ,...,h tJ}, where d ω is the dimension of the word vector, and ω is a parameter matrix learned through training;

[0078] Step S7: In the attention mechanism layer, use the tanh function and the softmax function to obtain the weight matrix λ = tanh (H) and α = softmax (ω T λ), where ω T is the transpose of the parameter matrix obtained through training and learning, and the vector representation of the sentence is σ=Hα T , the final vector representation of the sentence used for classification is H * =tanh(σ);

[0079] Step S8: In the classification output layer, H * As the input feature of this layer, a softmax classifier is used to obtain the optimal predicted sequence label Y = {y1, y2, y3, ..., y j ,...,y J};

[0080] Step S9: The current training process ends.

[0081] like Figure 3 As shown in the figure, the process of deep learning detection submodule detecting valid data is as follows:

[0082] Step R1: The current hospital adverse event detection data set is D″. The data set D″ is cleaned, and the repeated text data in the data set D″ is deleted to obtain a valid data set D″′, which is input into the trained deep learning model;

[0083] Step R2: In the word embedding layer, the existing word segmentation tools and the pre-prepared main feature dictionary and regular expressions are used to segment the data set D″′. The adverse event text data in the data set D″′ is numbered as ds′m. First, the long text of ds′m is segmented into ds′ m ={w″1,w″2,w″3,...,w″ i ,...,w″ I}, where w″ i represents the i-th sentence divided in ds′m, I represents the total number of sentences divided in a certain adverse event text, 1≤i≤I;

[0084] Step R3: Then adjust w″ i The word segmentation process gets w″ i ={word′1, word′2, word′3, ..., word′j, ..., word′J}, where word′j represents w″ i The jth word divided in , J represents the total number of words divided in the sentence, 1≤j≤J;

[0085] Step R4: Convert the word word′j after word segmentation to a word vector x′ j ={word′j1, word′j2, word′j3,..., word′j e , ..., word′j E}, where x′ j is the word vector expression of word word′j, word′j e is a value of 0 or 1, E represents the dimension of the word vector, 1≤e≤E; the sentence w″ i After the word vectors of all words are divided, they are converted into w″′ i Indicates that w″′ i ={x′1,x′2,x′3,...,x′ j ,...,x′ J},x′ j ∈R o , R represents the word vector space, o is the dimension of character embedding;

[0086] Step R5: In the bidirectional long short-term memory network layer, w″′ i ={x′1,x′2,x′3,...,x′ j ,...,x′J} as the input feature of this layer, read forward from left to right and save the state sequence at each moment (h * 1,h * 2,h * 3,...,h * J ) and read and save the state sequence at each moment from right to left in reverse (h *′ J ,...,h *′ 3,h *′ 2,h *′ 1) At time t, the output of each position is concatenated into h′ tj =[h * j :h *′ j ], and get the complete sequence state H′={h′ t1 ,h′ t2 ,h′ t3 ,...,h′ tj ,...,h′ tJ}, where d ω′ is the dimension of the word vector, and ω′ is a parameter matrix learned through training;

[0087] Step R6: In the attention mechanism layer, use the tanh function and the softmax function to obtain the weight matrix λ′=tanh(H′) and α′=softmax(ω′ T λ′), where ω′ T is the transpose of the parameter matrix obtained by detection, and the vector representation of the sentence is σ′=H′α′ T , the final vector representation of the sentence used for classification is H *′ =tanh(σ′);

[0088] Step R7: In the classification output layer, H *′ As the input feature of this layer, a softmax classifier is used to obtain the optimal predicted sequence label Y′={y′1,y′2,y′3,...,y′ j ,...,y′ J};

[0089] Step R8: Compare the predicted sequence label Y′ with the sample label, and calculate the accuracy Q, recall R1 and F1 value. Where TP refers to the number of samples with positive predicted labels, FP refers to the number of samples with negative predicted labels, TN refers to the number of samples with negative predicted labels, and FN refers to the number of samples with positive predicted labels. TP+FP+TN+FN=J;

[0090] Step R9: Evaluate the accuracy Q, recall R1 and F1 value. If the values ​​of Q, R1 and F1 are all greater than a (a is the threshold value, which is an empirical value), go to step R10; otherwise, go to step R2 and reset the parameter tuning.

[0091] Step R10: The current detection process ends;

[0092] Step R11: Save the entity extraction result of the detection data set D″′;

[0093] Step R12: Generate a fishbone diagram using an automatic fishbone diagram generation method;

[0094] Step R13: Send the fishbone diagram via WeChat Enterprise Account;

[0095] Step R14: The operation ends.

[0096] Fishbone diagram automatic generation method:

[0097] Step T1: Read the saved file, use the clustering algorithm, set the word segmentation length threshold, and generate a text matrix;

[0098] Step T2: Randomly select K points as cluster centers, use the fishbone diagram to automatically generate distance calculation method to calculate the distance d between each point and the K cluster centers, find the cluster center closest to the point, and assign it to this cluster center. At this time, the text matrix is ​​divided into K clusters;

[0099] Step T3: Use the fishbone diagram to automatically generate the mean point calculation method to calculate the mean point X of each cluster, and use the mean point of each cluster as the cluster center of each cluster;

[0100] Step T4: Use the fishbone diagram to automatically generate the distance calculation method to calculate the distance d′ from each point to all cluster centers in turn, and assign it to the nearest cluster;

[0101] Step T5: Repeat steps T3 and T4 until the mean point no longer changes, otherwise go to step T3;

[0102] Step T6: Finally generate cluster class C = {C1, C2, C3, ..., C k ,...,C K}, where C k is the category number, K is the number of clusters;

[0103] Step T7: Perform word frequency statistics on the subclusters contained in cluster class C. k There are F phrases in the text, and the frequency statistics are arranged in descending order as {C k1 ,C k2 ,C k3 ,...,C kf ...,C kF}, where C kf C k Phrases with frequency ranking f in the class, 1≤f≤F, 1≤k≤K;

[0104] Step T8: clustering ends;

[0105] Step T9: The fishbone diagram generation module reads the reported adverse event category and automatically generates the content of the fish head of the fishbone diagram;

[0106] Step T10: Arrange the results in descending order and extract the phrases of each cluster class in turn. k1 ,C k2 ,C k3 ,...,C kg ...,C kG}, where G is the number of phrases to be extracted, 1≤g≤G, 1≤G≤F, initialization k=1, g=1;

[0107] Step T11: Read C k1 Cluster C k The main title of the fishbone diagram, the fishbone diagram generation module automatically generates the main bone content of the fishbone diagram;

[0108] Step T12: g=g+1, read C kg Cluster C k The fishbone diagram generation module automatically generates the middle bone content of the fishbone diagram;

[0109] Step T13: If 1≤g≤G, go to step T12, otherwise go to step T14;

[0110] Step T14: k=k+1;

[0111] Step T15: If 1≤k≤K, go to step T11, otherwise go to step T16;

[0112] Step T16: generating fishbone contents of the fishbone diagram according to the contents of the large bones and the contents of the middle bones;

[0113] Step T17: generating a fishbone diagram according to the content of the fish head and the content of the fishbone;

[0114] Step T18: The operation ends.

[0115] Fishbone diagram automatic generation distance calculation method:

[0116] Suppose we calculate the distance between text 1 and text 2, and the text 1 vector is represented by {v1,v2,v3,...,v i ,...,v p}, the text2 vector is represented as {u1,u2,u3,...,u i ,...,u p}, where p represents the number of objects in the text, 1≤i≤p, then the distance calculation formula is:

[0117]

[0118] The fishbone diagram automatically generates the method for calculating the mean point:

[0119] Assume that the computation cluster C k The mean point C k The document in the class represents {T1,T2,T3,...,T b ,...,T B}, where C k The document number in the class is T b , B is returned to C k The number of documents in the class, 1≤b≤B; the document vector is represented by {T 1k , T 2k , T 3k , ..., T bk , ..., T Bk}, then cluster C k The mean point of is:

[0120]

[0121] Example: Automatic generation process of fishbone diagram of quality control circle in management of adverse events such as patient falls

[0122] (I) Take the training process of a piece of data from a patient fall adverse event as an example:

[0123] In the data collection module 1, the patient fall adverse event data set is collected from the hospital database, s = 2, D2 = {d21, d22, d23, ..., d2 n ,...,d2 N}, 1≤n≤N, N=234 is the total number of adverse event text data included in category D2;

[0124] In the data preprocessing module 2, the data set is cleaned and the duplicate text data in the data set is deleted to obtain a valid data set. The category of the data set remains unchanged, but the number of text data contained in the category is reduced. D2 = {d21, d22, d23, ..., d2 m,...,d2 M} , 1≤m≤M, M=220 is the total number of text data in the adverse events contained in the D2 category after cleaning;

[0125] Enter deep learning module 3. In the word embedding layer, use the existing word segmentation tools and the pre-prepared main feature dictionary and regular expressions to segment the valid data set. m = 1, select the first adverse event data {patient wakes up in the morning, on 2017-08-09 07:00 He got out of bed and went to the toilet on his own, wearing slippers. In a hurry to go to the toilet, he accidentally fell, his buttocks landed on the ground, and the back of his head landed on the armrest of the accompanying chair. The patient was immediately placed on the bed and reported to the doctor; physical examination: the patient was conscious; blood pressure: 136 / 78mmHg; pulse: 82 times / min; respiration: 18 times / min; there was a 3*4cm subcutaneous hematoma on the lower left side of the back of the head, which was painful to touch, and there was no other obvious trauma. The patient complained of no discomfort such as dizziness and vertigo. The doctor advised to apply local ice packs for cold compresses, closely observe the changes in the condition, and give psychological comfort; the original fall score was 75 points, and the current fall score is 100 points. The education of anti-fall measures was strengthened again, and the patient was advised not to get out of bed alone. The patient and his family expressed understanding}, that is, d21, and the long text was segmented into d21={w1,w2,w3,...,w i ,...,w 27}, i=1, select the first sentence after segmentation {patient wakes up in the morning}, that is, w1;

[0126] Then w1 is segmented to obtain w1 = {patient, morning, after};

[0127] The words after word segmentation are transformed into word vectors, and the word vectors x1 = {0, 0, 1} are obtained, which is the word vector expression of the word {patient}, x2 = {0, 1, 0} is the word vector expression of the word {morning}, and x3 = {1, 0, 0} is the word vector expression of the word {after}. E = 3 represents the word vector dimension. The word vectors of all words divided by sentence w1 are transformed and represented by w′1 as w′1 = {x1, x2, x3}, that is, o represents the dimension of character embedding is 3×3;

[0128] In the bidirectional long short-term memory network layer, As the input feature of this layer, the state sequence (h1, h2, h3) at each moment is read forward from left to right and saved, and the state sequence (h′3, h′2, h′1) at each moment is read reversely from right to left and saved. At time t, the output of each position is spliced ​​into h t1 =[h1:h′1], h t2 =[h2:h′2], h t3 = [h3:h′3], and get the complete sequence state of sentence at time t H = {h t1 ,ht2 ,h t3};

[0129] In the attention mechanism layer, the tan h function and soft max function are used to obtain the weight matrix λ = tanh (H) and α = softmax (ω T λ), where ω T is the transpose of the parameter matrix obtained through training and learning, and the vector representation of the sentence is σ=Hα T , the final vector representation of the sentence used for classification is H * =tanh(σ);

[0130] In the classification output layer, H * As the input feature of this layer, a softmax classifier is used to obtain the optimal predicted sequence label Y = {O, CL-1, O}, where O means that the word vector does not belong to any label, and CL-1 means that the word vector belongs to the first CL label;

[0131] The current training process ends.

[0132] (II) The current detection process of the hospital patient fall adverse event dataset is as follows:

[0133] In the data collection module 1, the adverse event detection data set of the fall type of the current hospital patients is collected as D″, and the data preprocessing module 2 is transferred to clean the data set D″, delete the repeated text data in the data set D″, and obtain the valid data set D″′, which is input into the trained deep learning model;

[0134] Enter deep learning module 3. In the word embedding layer, use the existing word segmentation tools and the pre-prepared main feature dictionary and regular expressions to perform word segmentation on the data set D″′; select an adverse event text data in the data set D″′ as {the patient was admitted to the hospital due to "paroxysmal chest tightness for 5 years, recurring for 5 hours" and suspected of "coronary heart disease", with a history of "ventricular premature beats, short-term ventricular tachycardia, atrial premature beats, short-term atrial tachycardia, hypertension grade 3 (very high risk group), left kidney lesion"; hospitalization accident score 8 points, fall score 60 points, self-care ability score 60 points, heart function grade 2, bed rail protection, family members accompany the bedside, September 14 02:05 patient wants to get up The patient urinated, did not call family members, did not press the call bell, and then accidentally slipped from the end of the bed to the ground. The nurse on duty heard the family members pressing the bell and immediately rushed to the ward and found that the patient was sitting on the ground, with family members beside him, and the bed rails were in place and pulled up. The patient did not complain of pain or other discomfort, was conscious, and both pupils were equal in size and round, with a diameter of 2.5mm, and there was a light reflex. There was no nausea or vomiting, blood pressure was 128 / 67mmHg, and there was no obvious trauma on the body. The doctor on duty was reported to the doctor on duty, who assisted the family members to help the patient to the bed after checking, continued to observe the changes in the condition, strengthened the anti-fall education, and filled in the adverse event report}, and segmented the long text into 35 sentences, and selected a sentence after segmentation {the patient accidentally slipped from the end of the bed to the ground};

[0135] The word segmentation process yields the words {after, accidentally, from, end of bed, slipped to the ground};

[0136] The words after word segmentation are transformed into word vectors, and the word vectors x′1={0, 0, 0, 0, 1} are obtained, which is the word vector expression of the word {after}, x′2={0, 0, 0, 1, 0} is the word vector expression of the word {careless}, x′3={0, 0, 1, 0, 0} is the word vector expression of the word {from}, x′4={0, 1, 0, 0, 0} is the word vector expression of the word {end of the bed}, x′5={1, 0, 0, 0, 0} is the word vector expression of the word {slipped to the ground}, and E=5 represents the word vector dimension; the word vectors of all words divided into sentences are transformed and expressed as

[0137] In the bidirectional long short-term memory network layer, As the input feature of this layer, the state sequence (h * 1,h * 2,h * 3,h * 4,h * 5) and read and save the state sequence at each moment from right to left in reverse (h *′ 5,h *′ 4,h *′ 3,h *′ 2, h *′)1, at time t, the output of each position is spliced ​​into h′ t1 =[h * 1:h *′ 1], h′ t2 =[h * 2:h *′ 2], h′ t3 =[h * 3:h *′ 3], h′ t4 =[h * 4:h *′ 4], h′ t5 =[h * 5:h *′ 5], we get the complete sequence state H′={h′ t1 ,h′ t2 ,h′ t3 ,h′ t4 ,h′ t5};

[0138] In the attention mechanism layer, the tanh function and soft max function are used to obtain the weight matrix λ′=tanh(H′) and α′=softmax(ω′ T λ′), where ω′ T is the transpose of the parameter matrix obtained through training and learning, and the vector representation of the sentence is σ′=H′α′ T , the final vector representation of the sentence used for classification is H *′ =tanh(σ′);

[0139] In the classification output layer, H *′ As the input feature of this layer, a softmax classifier is used to obtain the optimal predicted sequence label Y = {O, CL-1, O, O, CL-2}, where O means that the word vector does not belong to any label, CL-1 means that the word vector belongs to the first CL label, and CL-2 means that the word vector belongs to the second CL label;

[0140] Compare the predicted sequence label result Y′ of the dataset D″′ with the sample label, and calculate Q as 87.46%, R1 as 79..03%, and F1 as 83.03%;

[0141] Select a=75%, Q, R1, F1 values ​​are all greater than a, go to the next step;

[0142] The current detection process ends;

[0143] The entity extraction results of the detection data set D″′ are saved.

[0144] Enter the fishbone diagram generation module 4 and use the fishbone diagram automatic generation method to generate the fishbone diagram Figure 5 ;

[0145] Go to fishbone diagram sending module 5 and send fishbone diagram via WeChat Enterprise Account Figure 5 ;

[0146] The run is finished.

[0147] (III) The current method for automatically generating a fishbone diagram of adverse events related to falls in hospitals is as follows:

[0148] In the fishbone diagram generation module 4, the saved file is read, a clustering algorithm is used, a word segmentation length threshold is set, and a text matrix is ​​generated;

[0149] Randomly select K points as cluster centers. In the example of the present invention, K=4 is selected. The distance d between each point and the K cluster centers is calculated by using the fishbone diagram automatic generation distance calculation method. The cluster center closest to the point is found and assigned to this cluster center. At this time, the text matrix is ​​divided into K clusters.

[0150] The fishbone diagram is used to automatically generate the mean point calculation method to calculate the mean point X of each cluster, and the mean point of each cluster is used as the cluster center of each cluster;

[0151] The fishbone diagram is automatically generated and the distance calculation method is used to calculate the distance d′ from each point to all cluster centers in turn, and it is assigned to the nearest cluster;

[0152] Repeat the above steps until the mean point no longer changes;

[0153] Finally, the generated cluster class C = {C1, C2, C3, ..., C k , ..., C K}, where C k is the category number, K is the number of clusters, that is, C = {C1, C2, C3, C4};

[0154] The word frequency statistics of the subclusters contained in cluster class C are performed separately. k There are F phrases in the text, and the frequency statistics are arranged in descending order as {C k1 ,C k2 ,C k3 ,...,C kf ...,C kF}, that is, {C 11 ,C 12 ,C 13 ,...,C 1f ...,C 1F},{C 21 ,C 22 ,C 23 ,...,C 2f ...,C 2F},{C 31,C 32 ,C 33 ,...,C 3f ...,C 3F},{C 41 ,C 42 ,C 43 ,...,C 4f ...,C 4F};

[0155] Clustering is finished.

[0156] The fishbone diagram generation module reads {the cause of the patient's fall} and automatically generates the content of the fish head of the fishbone diagram;

[0157] Arrange the results in descending order and extract the phrases of each cluster class in turn {C k1 ,C k2 ,C k3 ,...,C kg ...,C kG}, select G=5 to set the number of phrases to be extracted, 1≤g≤G, 1≤G≤F, initialize k=1, g=1;

[0158] Read C k1 Cluster C k The main title of the fishbone diagram, the fishbone diagram generation module automatically generates the main bone content of the fishbone diagram;

[0159] Read C kg Cluster C k The fishbone diagram generation module automatically generates the middle bone content of the fishbone diagram;

[0160] After the inner loop operation, the contents of the fish bones of the fishbone diagram are generated according to the contents of the big bone {not found by the nurse} and the contents of the middle bone {not reminded, fall risk not assessed, inspection lacking, not taken seriously};

[0161] After the outer loop operation, the fishbone content of the fishbone diagram is generated according to the content of the big bone {weak legs} and the content of the middle bone {unsteady walking, weak lower limbs, dizziness, poor mental state}, the fishbone content of the fishbone diagram is generated according to the content of the big bone {slippery floor} and the content of the middle bone {water on the ground, floor drain, wet pants, messy bathroom}, and the fishbone content of the fishbone diagram is generated according to the content of the big bone {falling out of bed} and the content of the middle bone {narrow treatment bed, no anti-slip slippers, no guardrails, high bed};

[0162] A fishbone diagram is generated according to the content of the fish head and the content of the fishbone, such as Figure 5 A fishbone diagram generated by the automatic generation method of fishbone diagram for quality control circles in the management of adverse events such as patient falls.

[0163] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention without departing from the principles and intent of the present invention.

Claims

1. A method for automatically generating a fishbone diagram of a quality control circle in hospital adverse event management, characterized in that: Includes steps: Step (1), reading the entity extraction results of the stored detection data set D″, using a clustering algorithm, setting a word segmentation length threshold, and generating a text matrix, wherein the text matrix represents a set of multiple vector points; The entity extraction result of the detection data set D″ is achieved by the following method: 1) Clean the current hospital adverse event detection data set D″ to obtain the first valid data set D″′, which is input into the trained deep learning model; 2) In the word embedding layer, the first valid data set D″′ is segmented, and the long text of the adverse event text data ds′m in the first valid data set D″′ is segmented into ds′ m ={w″1,w″2,w″3,...,w″ i ,...,w″ I }, where w″ i represents the i-th sentence divided in ds′m, I represents the total number of sentences divided in a certain adverse event text, 1≤i≤I; 3) Then w″ i The word segmentation process gets w″ i ={word′1, word′2, word′3, ..., word′j, ..., word′J}, where word′j represents w″ i The jth word divided in , J represents the total number of words divided in the sentence, 1≤j≤J; 4) Convert the word word'j after word segmentation to a word vector x' j ={word′j1, word′j2, word′j3,..., word′j e , ..., word′j E }, where x′ j is the word vector expression of word word′j, word′j e is a value of 0 or 1, 1≤e≤E, E represents the word vector dimension; i After the word vectors of all words are divided, they are converted into w″′ i Indicates that w″′ i ={x′1,x′2,x′3,...,x′ j ,...,x′ J },x′ j ∈R o , R represents the word vector space, o is the dimension of character embedding; 5) w″′ i ={x′1,x′2,x′3,...,x′ j ,...,x′ J } As the input feature of the bidirectional long short-term memory network layer, read and save the state sequence at each moment from left to right (h * 1,h * 2,h * 3,...,h * J ), read and save the state sequence at each moment from right to left (h *′ J ,...,h *′ 3,h *′ 2,h *′ 1) At time t, the output of each position is spliced ​​into h′ tj =[h * j :h *′ j ], and get the complete sequence state H′={h′ t1 ,h′ t2 ,h′ t3 ,...,h′ tj ,...,h′ tJ }, H′∈R dω′×J , where d ω′ is the dimension of the word vector, and ω′ is a parameter matrix learned through training; 6) In the attention mechanism layer, the tanh function and softmax function are used to obtain the weight matrix λ′=tanh(H′) and α′=softmax(ω′ T λ′), the sentence vector is represented by σ′=H′α′ T , the final vector representation of the sentence used for classification is H *′ =tanh(σ′), where ω′ T is the transpose of a detected parameter matrix; 7) H *′ As the input feature of the classification output layer, the softmax classifier is used to obtain the optimal predicted sequence label Y′={y′1,y′2,y′3,...,y′ j ,...,y′ J }; 8) Compare the predicted sequence label Y′ with the sample label and calculate the precision Q, recall R1 and F1 value; 9) Evaluate the accuracy Q, recall R1 and F1 value. If the Q, R1 and F1 values ​​are all greater than the threshold a, go to 10), otherwise go to 2) and reset the parameter tuning; 10) The current detection process ends; 11) Save the entity extraction results of the detection data set D″; Step (2), randomly select K' vector points as cluster centers, use the fishbone diagram to automatically generate the distance calculation method to calculate the distance d between each vector point and the K' cluster centers, find the cluster center closest to each vector point, and assign each vector point to the nearest cluster center. At this time, the text matrix is ​​divided into K' clusters, where K' is the number of clusters; Step (3): Use the fishbone diagram to automatically generate the mean point calculation method to calculate the mean point of each cluster. The mean point of each cluster is taken as the cluster center of each cluster; Step (4), calculate the distance d′ from each point to all cluster centers, and assign the point to the nearest cluster; Step (5), repeatedly iterate steps (3) and (4) until the mean point no longer changes, otherwise go to step (3); Step (6), finally generate clusters C = {C1, C2, C3, ... C k ,…C K' }, where C k is the category number; Step (7), count the word frequencies of the subclusters contained in cluster class C. k There are F phrases in the text, and the frequency statistics are arranged in descending order as {C k1 ,C k2 ,C k3 ,...,C kf ...,C kF }, where C kf C k Phrases with frequency ranking f in the class, 1≤f≤F, 1≤k≤K'; Step (8), clustering is completed; Step (9), the fishbone diagram generation module reads the reported adverse event category and automatically generates the content of the fish head of the fishbone diagram; Step (10), sort the results in descending order and extract the phrases of each cluster {C k1 ,C k2 ,C k3 ,...,C kg ...,C kG }, where G is the number of phrases to be extracted, 1≤g≤G, 1≤G≤F, initialization k=1, g=1; Step (11), read C k1 For the main title of cluster Ck, the fishbone diagram generation module automatically generates the main bone content of the fishbone diagram; Step (12), g = g + 1, read C kg For the subtitle of cluster Ck, the fishbone diagram generation module automatically generates the middle bone content of the fishbone of the fishbone diagram; Step (13), if 1≤g≤G, go to step (12), otherwise go to step (14); Step (14), k=k+1; Step (15), if 1≤k≤K', go to step (11), otherwise go to step (16); Step (16), generating fishbone contents of the fishbone diagram according to the contents of the large bones and the contents of the middle bones; Step (17), generating a fishbone diagram according to the content of the fish head and the content of the fish bones.

2. The method for automatically generating a quality control circle fishbone diagram in hospital adverse event management according to claim 1, characterized in that: Said Said Said Among them, TP refers to the number of samples with positive predicted labels whose results are positive, FP refers to the number of samples with negative predicted labels whose results are positive, and FN refers to the number of samples with positive predicted labels whose results are negative.

3. The method for automatically generating a quality control circle fishbone diagram in hospital adverse event management according to claim 1, characterized in that: The training process of the deep learning model is: 1) From the hospital database, collect various adverse event data sets D = {D1, D2, D3, ..., D s ,...,D S }, where Ds is the adverse event category number, 1≤s≤S', S' is the total number of categories in the adverse event text dataset, D s ={ds1,ds2,ds3,...,ds n ,...,ds N }, ds n is the number of the adverse event text data contained in the Ds category, 1≤n≤N, and N is the total number of adverse event text data contained in the Ds category; 2) Clean the data set D to obtain a second valid data set D′. The text data D′ contained in each category of the second valid data set s ={d″ s1 ,d″ s2 ,d″ s3 ,…,d″ sm ,…,d″ sM }, where d″ sm is the text data number of the adverse events contained in the D′s category, 1≤m≤M, M is the total number of text data of the adverse events contained in the D′s category after cleaning; 3) At the word embedding layer, the second valid data set D′ is segmented. First, d″ sm Split long text into d″ sm ={w1,w2,w3,…,w i ,…,w I }, where w i Represents d″ sm The i-th sentence divided in , I represents the total number of sentences divided in a certain adverse event text, 1≤i≤I; 4) Then w i Word segmentation processing, get w i ={word1,word2,word3,...,wordj,...,wordJ}, where wordj represents w i The jth word divided in , J represents the total number of words divided in a sentence, 1≤j≤J; 5) Convert the word j after word segmentation to a word vector to obtain the word vector x j ={wordj1,wordj2,wordj3,...,wordj e ,...,wordj E }, where x j is the word vector expression of word wordj, wordj e is a value of 0 or 1, 1≤e≤E, E represents the dimension of the word vector; i After all the words are divided, the word vectors are transformed and used as w′ i It means that w′ i ={x1,x2,x3,...,x j ,...,x J }, x j ∈R o , R represents the word vector space, o is the dimension of character embedding; 6) w′ i ={x1,x2,x3,...,x j ,...,x J } As the input feature of the bidirectional long short-term memory network layer, read and save the state sequence (h1,h2,h3,...,h J ), read and save the state sequence at each moment from right to left in reverse (h′ J ,...,h′3,h′2,h′1), at time t, the output of each position is concatenated into h tj =[h j :h′ j ], and obtain the complete sequence state H = {h t1 ,h t2 ,h t3 ,...,h tj ,...,h tJ }, where d ω is the dimension of the word vector, and ω is a parameter matrix learned through training; 7) In the attention mechanism layer, the tanh function and softmax function are used to obtain the weight matrix λ = tanh(H) and α = softmax(ω T λ), the sentence vector is represented by σ=Hα T , the final vector representation of the sentence used for classification is H * = tanh(σ), where ω T is the transpose of the parameter matrix learned through training; 8) H * As the input feature of the classification output layer, the softmax classifier is used to obtain the optimal prediction sequence label Y = {y1, y2, y3, ..., y j ,…y j }; The training process ends.

4. The method for automatically generating a quality control circle fishbone diagram in hospital adverse event management according to claim 1, characterized in that: The method for automatically generating and calculating the distance of the fishbone diagram is as follows: Suppose we calculate the distance between text 1 and text 2, and the text 1 vector is represented by {v1,v2,v3,…,v i′ ,…,v p }, the text2 vector is represented as {u1,u2,u3,…,u i′ ,…,u p }, where p represents the number of objects in the text, 1≤i′≤p, then the distance calculation formula is 5. The method for automatically generating a quality control circle fishbone diagram in hospital adverse event management according to claim 1, characterized in that: The method for automatically generating and calculating the mean point of the fishbone diagram is as follows: Assume that the computation cluster C k The mean point C k The document in the class represents {T1,T2,T3,...,T b ,...,T B }, the document vector is represented as {T 1k ,T 2k ,T 3k ,...,T bk ,...,T Bk }, then cluster C k The mean point of is: Among them, C k The document number in the class is T b , B is returned to C k The number of documents in the class, 1≤b≤B.

6. A device for realizing the method for automatically generating a quality control circle fishbone diagram in hospital adverse event management as described in any one of claims 1 to 5, characterized in that: It comprises a data acquisition module (1), a data preprocessing module (2), a deep learning module (3), a fishbone diagram generating module (4) and a fishbone diagram sending module (5) which are connected in sequence; The data collection module (1) is used to collect textual descriptive data of adverse events in a hospital database; The data preprocessing module (2) is used to clean the text descriptive data and segment the long text to obtain valid data; The deep learning module (3) is used to train and test effective data, perform entity recognition on the main causes of adverse events, and realize the adverse event entity extraction function; The fishbone diagram generation module (4) is used to automatically generate the fish head and fishbone content of the fishbone diagram from the entity extraction results, so as to realize the visual presentation of the adverse event analysis results; The fishbone diagram sending module (5) is used to push the fishbone diagram, so that the medical and nursing management personnel of each department can check the latest fishbone diagram in time.

7. The device according to claim 6, characterized in that The deep learning module (3) comprises a deep learning training submodule and a deep learning detection submodule. Both the deep learning training submodule and the deep learning detection submodule use a deep learning model. The deep learning model comprises a word embedding layer, a bidirectional long short-term memory network layer, an attention mechanism layer and a classification output layer.

Images