CN113808693A - Medicine recommendation method based on graph neural network and attention mechanism - Google Patents

Medicine recommendation method based on graph neural network and attention mechanism Download PDF

Info

Publication number
CN113808693A
CN113808693A CN202111061579.9A CN202111061579A CN113808693A CN 113808693 A CN113808693 A CN 113808693A CN 202111061579 A CN202111061579 A CN 202111061579A CN 113808693 A CN113808693 A CN 113808693A
Authority
CN
China
Prior art keywords
information
data
drug
neural network
gru
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202111061579.9A
Other languages
Chinese (zh)
Inventor
万健
岳魏琦
张蕾
洪高枫
郑慧琳
史斌彬
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Zhejiang Lover Health Science and Technology Development Co Ltd
Zhejiang University of Science and Technology ZUST
Original Assignee
Zhejiang Lover Health Science and Technology Development Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Zhejiang Lover Health Science and Technology Development Co Ltd filed Critical Zhejiang Lover Health Science and Technology Development Co Ltd
Priority to CN202111061579.9A priority Critical patent/CN113808693A/en
Publication of CN113808693A publication Critical patent/CN113808693A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H10/00—ICT specially adapted for the handling or processing of patient-related medical or healthcare data
    • G16H10/60—ICT specially adapted for the handling or processing of patient-related medical or healthcare data for patient-specific data, e.g. for electronic patient records
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H70/00—ICT specially adapted for the handling or processing of medical references
    • G16H70/40—ICT specially adapted for the handling or processing of medical references relating to drugs, e.g. their side effects or intended usage

Landscapes

  • Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Software Systems (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Mathematical Physics (AREA)
  • Biophysics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Public Health (AREA)
  • Primary Health Care (AREA)
  • Medical Informatics (AREA)
  • Epidemiology (AREA)
  • Toxicology (AREA)
  • Pharmacology & Pharmacy (AREA)
  • Medicinal Chemistry (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Chemical & Material Sciences (AREA)
  • Measuring And Recording Apparatus For Diagnosis (AREA)
  • Medical Treatment And Welfare Office Work (AREA)

Abstract

The invention discloses a medicine recommendation method based on a graph neural network and an attention mechanism. The invention takes the structural characteristics of the doctor seeing situation or the medication information of each patient as a node, adopts the graph neural network to capture the relationship among the structural characteristics and learns the high-order characteristics containing the medical system knowledge. Meanwhile, the attention mechanism is used for better modeling the historical medical records of the user, the medicine interaction knowledge is introduced, and the accuracy and the safety of medicine recommendation are effectively improved.

Description

Medicine recommendation method based on graph neural network and attention mechanism
Technical Field
The invention belongs to the technical field of computer application, and relates to a medicine recommendation method based on a graph neural network and an attention mechanism.
Background
The development of modern medical technology has led to the widespread use of electronic medical records, which accumulate large amounts of clinical data, such as vital signs, clinical summary, disease diagnosis, prescription drugs, etc. Meanwhile, the deep learning technology provides a new technical means for mining and utilizing medical data, and is a research hotspot at present. The combined medicine recommendation algorithm based on the electronic medical record can assist a doctor to make a safe and effective prescription according to the change characteristics of the illness state of a patient, the medicine attributes and the action relation among a large number of medicines, and has important research value.
Early drug recommendation techniques were mostly rule based. The related experts extract the medication rules based on the medical information such as diagnosis, disease classification, symptoms, detection results and the like of the patients, and the defects are that the maintenance is complex and the updating and the expansion are difficult. The medicine recommendation technology under deep learning embeds information of physical signs, diagnosis, past medicine and the like of a patient into a low-dimensional space, and uses the embedded representation for recommendation, so that the recommendation accuracy is improved. However, they also have many problems, including data sparsity, inability to effectively utilize historical medical record information of patients, ignoring medical ontology information that medical codes imply, and the like.
A graph neural network is a neural network that acts directly on the graph structure. It has the following characteristics: ignoring the input order of the nodes; in the calculation process, the representation of the node is influenced by the neighbor nodes around the node, and the connection of the graph is unchanged; the representation of graph structure allows reasoning based on graphs. Therefore, the graph neural network becomes a great research hotspot and is widely applied to the fields of social networks, recommendation systems, financial wind control, physical systems, molecular chemistry, life science, knowledge maps, traffic prediction and the like.
The introduction of the graph attention mechanism to graph neural networks has found widespread use in many areas. The graph neural network better realizes the weighted aggregation of the neighbors by learning the weights of the neighbors, further filters noise neighbors, improves the model performance and can realize certain explanation on the result.
Disclosure of Invention
The invention aims to provide a medicine recommendation method which can effectively relieve the sparsity of medical data, effectively utilize the historical case information of a patient and give consideration to the safety of medicines in view of the defects of the prior art.
In order to achieve the above object, the present invention provides a drug recommendation method based on a graph neural network and an attention mechanism, comprising the following steps:
step 1, acquiring historical electronic medical record data, and performing structured processing:
acquiring historical clinic situations of a patient and medication information corresponding to the clinic situations to construct an electronic medical record, wherein the clinic situations comprise diagnosis data and operation condition data; the patient's electronic medical record is denoted as p ═ x1,x2,...,xt-1]T is the current number of visits by the patient, where the ith visit by the patient is denoted as xi＝[di,pi,mi],i＝1,2,...,t-1，diDiagnostic data, p, representing the patient's i-th visitiSurgical condition data, m, representing the ith visit in a patient's medical recordiData representing the medication at the i-th visit.
Step 2, constructing three graph neural networks for learning the structural characteristics of the patient treatment condition and the medication information; the inputs of the three graph neural networks are respectively diagnosis data, operation condition data and medication data of a patient, and the corresponding outputs are respectively de、pe、me；
The three graph neural networks adopt the same structure and specifically comprise nodes and edges; the nodes comprise leaf nodes and non-leaf nodes, the leaf nodes are input data, namely one of diagnosis data, operation condition data and medication data of a patient, and the non-leaf nodes are medical attribution classifications of the leaf nodes; the edge is the medical classification relation of two nodes;
each non-leaf node is represented as the sum of its own vector representation and all its sub-nodes, calculated using the GAT graph attention machine:
Figure BDA0003256807290000021
Figure BDA0003256807290000022
wherein g isnDenotes the nth non-leaf node, K denotes the total number of attentions, ReLU and LeakyReLU denote non-linear functions, ch (n) denotes the vector representation of the nth non-leaf node itself and all its children,
Figure BDA0003256807290000023
weight calculation coefficient, W, representing the current non-leaf node itself and all its children nodes under the kth attentionkRepresents the learning parameters of the non-leaf nodes at the kth attention, e*Vector representation representing nodes, a representing a learnable matrix, aTIs a transpose thereof.
Each leaf node is represented as the sum of the vector representations of itself and all its ancestor nodes, again calculated using the GAT graph attention machine mechanism:
Figure BDA0003256807290000024
Figure BDA0003256807290000025
wherein c'nRepresents the nth leaf node, an (n) represents the vector representation of the nth leaf node itself and all its ancestor nodes,
Figure BDA0003256807290000026
weight calculation coefficient, W ', representing the current leaf node itself and all ancestor nodes thereof under the k-th attention'kThe learning parameters of the k-th attention leaf node are represented.
Step 3, constructing two attention mechanismsThe input of the GRU network model is the output result d of the step 2e、peThe corresponding outputs are respectively k with history informationd、kp；
The two GRU network models with attention mechanisms adopt the same structure and respectively comprise two parallel GRU networks and attention mechanism modules connected with the outputs of the two parallel GRU networks;
the two parallel GRU models are hidden layer output information for acquiring historical visiting situations (i.e., diagnosis or surgical situation information) by adopting different activation functions, and are specifically as follows:
H＝GRU1(r) (5)
Wh＝softmax(Fh(H)) (6)
H＝GRU2(r) (7)
W′htanh(Fh′(H′)) (8)
h and H' respectively represent hidden layer information output by a first GRU network model and a second GRU network model, and Wh，W′hRespectively representing attention mechanism weights obtained by the first GRU network and the second GRU network through softmax and tanh activation functions, Fh，F′hRespectively representing the learnable linear transformation matrix functions of the first GRU network and the second GRU network, wherein r represents deOr pe；
The attention mechanism module calculates k with historical information according to formula (9)d、kpI.e. by
Figure BDA0003256807290000031
For diagnostic information with historical information for different time scales,
Figure BDA0003256807290000032
the surgical condition information with historical information is obtained in different time scales.
Figure BDA0003256807290000033
Wherein t represents the total number of patient visits, Wh(i)，W′h(i) Respectively represent attention mechanism weights obtained by the softmax and tanh activation functions corresponding to the ith visit,
Figure BDA0003256807290000034
representing element-by-element multiplication; k represents kdOr kp；
Step 4, constructing two memory neural networks MANN with the same structure; wherein the key-value pair stored in the first memory neural network is' the ith visit diagnosis data fusion information
Figure BDA0003256807290000035
"-" graph neural network medication information
Figure BDA0003256807290000036
"; the key-value pair stored in the second memory neural network is' the fusion information of the condition of the operation of the ith visit
Figure BDA0003256807290000037
"-" graph neural network medication information
Figure BDA0003256807290000038
”；
The ith treatment situation and the historical treatment situation are multiplied by each other in a contraposition way, and the weight of the ith treatment situation is calculated
Figure BDA0003256807290000039
Figure BDA00032568072900000310
Wherein
Figure BDA0003256807290000041
Or
Figure BDA0003256807290000042
The information of the treatment condition with historical information is shown in the ith treatment;
by weight
Figure BDA0003256807290000043
Obtaining historical medication vector
Figure BDA0003256807290000044
The ith dose was as follows:
Figure BDA0003256807290000045
further obtaining keys of memory neural network
Figure BDA0003256807290000046
Figure BDA0003256807290000047
Wherein
Figure BDA0003256807290000048
Representing learning weight, key
Figure BDA0003256807290000049
Corresponding value is
Figure BDA00032568072900000410
According to
Figure BDA00032568072900000411
Can obtain
Figure BDA00032568072900000412
And
Figure BDA00032568072900000413
step 5, constructing a drug interaction knowledge base
Introduction of knowledge of drug interactions, use of adjacency matrix ACThe adjacency matrix A represents the coexistence relationship of medicines in the electronic medical recordDIndicating drug interaction relationship. And (4) learning the medicine co-occurrence relation and the medicine interaction relation by adopting a graph convolution neural network, and combining the medicine interaction and co-occurrence relation with the key value pair in the step (4) to generate a recommended medicine list.
5-1 step 4
Figure BDA00032568072900000414
And
Figure BDA00032568072900000415
combining to obtain a query vector containing historical medical record information, ith diagnosis information and ith operation information
Figure BDA00032568072900000416
Figure BDA00032568072900000417
Wherein WsRepresenting contrast weights for diagnostic and surgical information.
5-2, constructing a medicine coexistence relation matrix and a medicine interaction relation matrix in the electronic medical record
A*＝D-1(A*+I)D-1 (14)
Wherein D represents A*The diagonal matrix of the transform, D-1 being the inverse thereof, I being an identity matrix, A*Representing the drug coexistence relationship matrix A in the electronic medical recordCOr drug interaction relationship matrix AD。
5-3 learning relationships between drugs using graph convolution neural networks, incorporating drug interactions and co-occurrence relationships into an embedded representation, resulting in a representation matrix of drug co-occurrencesZCAnd a representation matrix Z of a drug interaction mapD：
ZC＝ACtanh(ACme)WC (15)
ZD＝ADtanh(ADme)WD (16)
Wherein, WC，WDParameter matrices, m, for the drug contribution and drug interaction maps, respectivelyeThe set of key-value pairs found in step 4.
Based on matrix ZC、ZDAnd query vector
Figure BDA00032568072900000418
Calculating attention λi：
Figure BDA00032568072900000419
Wherein, WCDRepresenting the drug co-existence relationship and the contrast weight of the interaction.
Finally obtaining a recommended medicine list yi：
Figure BDA0003256807290000051
Wherein, WyRepresenting the weight coefficients at the time of calculation of the recommended medication list.
The recommendation probability y of the drug in the drug listiAnd if the probability is greater than the recommendation probability threshold value rho, recommending the medicine.
It is another object of the present invention to provide a graph neural network and attention mechanism based drug recommendation device, comprising
The data preprocessing module is used for carrying out structuralized processing on the historical treatment condition of the patient and the medication information corresponding to the treatment condition to construct corresponding electronic medical record data;
graph neural network modelA block to learn structural characteristics of patient encounter and medication information; the inputs of the three graph neural networks are respectively diagnosis data, operation condition data and medication data of a patient, and the corresponding outputs are respectively de、pe、me；
The GRU network model module with attention mechanism is used for extracting characteristics of the diagnosis data and the operation condition data output by the GRU network module, and then combining the characteristics to the current diagnosis situation to obtain the diagnosis data and the operation condition data with historical information.
The memory neural network MANN module is used for constructing key value pairs of the ith diagnosis and diagnosis data fusion information output by the GRU network model module with the attention mechanism and the medication information of the graph neural network, and key value pairs of the ith diagnosis and operation condition fusion information output by the GRU network model module with the attention mechanism and the medication information of the graph neural network;
and the drug interaction knowledge base module is used for learning the drug co-occurrence relationship and the drug interaction relationship by adopting a graph convolution neural network, combining the drug interaction and the co-occurrence relationship to the drug embedded expression of the memory neural network MANN module and generating a recommended drug list.
A further object of the present invention is a computer-readable storage medium having stored thereon a computer program which, when executed in a computer, causes the computer to carry out the above-mentioned method.
Yet another object of the present invention is a computing device comprising a memory having stored therein executable code and a processor that, when executing the executable code, implements the method described above.
The invention has the following advantages: the invention takes the structural characteristics of the doctor seeing situation or the medication information of each patient as a node, adopts the graph neural network to capture the relationship among the structural characteristics and learns the high-order characteristics containing the medical system knowledge. Meanwhile, the attention mechanism is used for better modeling the historical medical records of the user, the medicine interaction knowledge is introduced, and the accuracy and the safety of medicine recommendation are effectively improved.
Drawings
FIG. 1 is a drug recommendation process based on a graph neural network and attention mechanism.
FIG. 2 is a tree diagram of a medical code encoding architecture.
Fig. 3 is a comparison of F1 values for different visits with different methods.
Detailed Description
The present invention is further analyzed with reference to the following specific examples.
The invention provides a medicine recommendation method based on a graph neural network and an attention mechanism, which comprises the following steps of:
step 1, acquiring historical electronic medical record data, and performing structured processing:
acquiring historical clinic situations of a patient and medication information corresponding to the clinic situations to construct an electronic medical record, wherein the clinic situations comprise diagnosis data and operation condition data; the patient's electronic medical record is denoted as p ═ x1,x2,...,xt-1]T is the current number of visits by the patient, where the ith visit by the patient is denoted as xi＝[di,pi,mi],i＝1,2,...,t-1，diDiagnostic data, p, representing the patient's i-th visitiSurgical condition data, m, representing the ith visit in a patient's medical recordiData representing the medication at the i-th visit.
Step 2, constructing three graph neural networks for learning the structural characteristics of the patient treatment condition and the medication information; the inputs of the three graph neural networks are respectively diagnosis data, operation condition data and medication data of a patient, and the corresponding outputs are respectively de、pe、me；
The three graph neural networks adopt the same structure and specifically comprise nodes and edges; the nodes comprise leaf nodes and non-leaf nodes, the leaf nodes are input data, namely one of diagnosis data, operation condition data and medication data of a patient, and the non-leaf nodes are medical attribution classifications of the leaf nodes; the edge is the medical classification relation of two nodes;
each non-leaf node is represented as the sum of its own vector representation and all its sub-nodes, calculated using the GAT graph attention machine:
Figure BDA0003256807290000061
Figure BDA0003256807290000062
wherein g isnDenotes the nth non-leaf node, K denotes the total number of attentions, ReLU and LeakyReLU denote non-linear functions, ch (n) denotes the vector representation of the nth non-leaf node itself and all its children,
Figure BDA0003256807290000063
weight calculation coefficient, W, representing the current non-leaf node itself and all its children nodes under the kth attentionkRepresents the learning parameters of the non-leaf nodes at the kth attention, e*Vector representation representing nodes, a representing a learnable matrix, aTIs a transpose thereof.
Each leaf node is represented as the sum of the vector representations of itself and all its ancestor nodes, again calculated using the GAT graph attention machine mechanism:
Figure BDA0003256807290000071
Figure BDA0003256807290000072
wherein c'nRepresents the nth leaf node, an (n) represents the vector representation of the nth leaf node itself and all its ancestor nodes,
Figure BDA0003256807290000073
indicating the current leaf at the kth attentionWeight calculation coefficients, W ', of the child nodes themselves and all ancestor nodes thereof'kThe learning parameters of the k-th attention leaf node are represented.
Step 3, constructing two GRU network models with attention mechanisms, and inputting the output results de and p in the step 2 respectivelyeThe corresponding outputs are respectively k with history informationd、kp；
The two GRU network models with attention mechanisms adopt the same structure and respectively comprise a GRU network and an attention mechanism module connected with the output of the GRU network;
the GRU model uses an attention mechanism to incorporate hidden layer output information of historical encounter situations (i.e. diagnostic or surgical condition information) into the current information representation, and the specific calculation method is as follows:
H＝GRU1(r) (5)
Wh＝softmax(Fh(H)) (6)
H′＝GRU2(r) (7)
W′h＝tanh(Fh′(H′)) (8)
wherein, H, H' respectively represent the hidden layer information output by the first and second GRU network models, Wh，W′hRespectively, the attention mechanism weights obtained by the softmax and tanh activation functions, Fh，F′hRepresenting the learnable linear transformation matrix function of the first and second GRU network models, r represents deOr pe；
The attention mechanism module calculates k with historical information according to formula (9)d、kpI.e. by
Figure BDA0003256807290000074
For diagnostic information with historical information for different time scales,
Figure BDA0003256807290000075
the surgical condition information with historical information is obtained in different time scales.
Figure BDA0003256807290000076
Wherein t represents the total number of patient visits, Wh(i)，Wh(i) Indicating the attention mechanism weight obtained by the softmax and tanh activation functions corresponding to a particular clinic,
Figure BDA0003256807290000077
representing element-by-element multiplication; k is a radical of*Represents kdOr kp；
Step 4, constructing two memory neural networks MANN with similar structures; wherein the key-value pair stored in the first memory neural network is' the ith visit diagnosis data fusion information
Figure BDA0003256807290000078
"-" graph neural network medication information
Figure BDA0003256807290000079
"; the key-value pair stored in the second memory neural network is' the fusion information of the condition of the operation of the ith visit
Figure BDA00032568072900000710
"-" graph neural network medication information
Figure BDA00032568072900000711
”；
The ith treatment situation and the historical treatment situation are multiplied by each other in a contraposition way, and the weight of the ith treatment situation is calculated
Figure BDA0003256807290000081
Figure BDA0003256807290000082
Wherein
Figure BDA0003256807290000083
Or
Figure BDA0003256807290000084
The operation condition information with historical information is shown in the ith visit;
by weight
Figure BDA0003256807290000085
Obtaining historical medication vector
Figure BDA0003256807290000086
The ith dose was as follows:
Figure BDA0003256807290000087
further obtaining keys of memory neural network
Figure BDA0003256807290000088
Figure BDA0003256807290000089
Wherein
Figure BDA00032568072900000810
Representing learning weight, key
Figure BDA00032568072900000811
Corresponding value is
Figure BDA00032568072900000812
Step 5, constructing a drug interaction knowledge base
Introduction of knowledge of drug interactions, use of adjacency matrix ACRepresenting drugs in an electronic medical recordCoexistence relationship of objects, adjacency matrix ADIndicating drug interaction relationship. And (4) learning the medicine co-occurrence relation and the medicine interaction relation by adopting a graph convolution neural network, and combining the medicine interaction and co-occurrence relation with the medicine embedded expression obtained in the step (4) to generate a recommended medicine list.
5-1 will step four
Figure BDA00032568072900000813
And
Figure BDA00032568072900000814
combining to obtain a query vector containing historical medical record information, ith diagnosis information and ith operation information
Figure BDA00032568072900000815
Figure BDA00032568072900000816
Wherein, WsRepresenting contrast weights for diagnostic and surgical information.
5-2, constructing a medicine coexistence relation matrix and a medicine interaction relation matrix in the electronic medical record
A*＝D-1(A*+I)D-1 (14)
Wherein D represents A*The diagonal matrix of the transform, D-1 being the inverse thereof, I being an identity matrix, A*Representing the drug coexistence relationship matrix A in the electronic medical recordCOr drug interaction relationship matrix AD。
5-3 learning relationships between drugs using graph convolution neural networks, incorporating drug interactions and co-occurrence relationships into the embedded representation, resulting in a representation matrix Z of drug co-occurrencesCAnd a representation matrix Z of a drug interaction mapD：
zC＝ACtanh(ACme)WC (15)
ZD＝ADtanh(ADme)WD (16)
Wherein, WC，WDAnd me is the structural characteristic of the drug output in the step 2.
Based on matrix ZC、ZDAnd query vector
Figure BDA0003256807290000091
Calculating attention λi:
Figure BDA0003256807290000092
Wherein, WCDRepresenting the drug co-existence relationship and the contrast weight of the interaction.
Finally obtaining a recommended medicine list yi：
Figure BDA0003256807290000093
Wherein, WyRepresenting the weight coefficients at the time of calculation of the recommended medication list.
The obtained medicine list is a group of one-dimensional matrixes with absolute values smaller than 1, the horizontal and vertical coordinates respectively represent the medicine type and the recommendation probability, and when the recommendation probability of the medicine in the medicine list is larger than a preset threshold value of 0.5, the medicine is recommended.
The experimental process comprises the following steps:
the experiment used electronic medical record data from the MIMIC-III (medical Information Mark for Intelligent Care) database, which was a free public Intensive care data set published by the institute of technology, Massachusetts, institute of technology, and computing physiology laboratory. The present invention uses the diagnostic, surgical and prescription data in the database to screen patients for medications received within 24 hours of entering the ICU.
To measure the accuracy of the recommendation, the invention uses a Jaccard similarity coefficient (Jaccard), namely the size of the intersection of the real drug and the recommended drug, divided by the size of the union, an average F1 value (F1), namely the harmonic mean of the accuracy and the recall rate, and a precision calling curve (PRAUC) as the measurement index of the accuracy.
To measure the safety of the recommended drugs, the drug interaction rate DDI, i.e., the ratio of DDI drugs contained in the recommended combination drug, is used.
Compared with the current six effective methods, the Nearst method recommends the combined medicine which is the same as the previous diagnosis according to the similarity of the current diagnosis and the previous diagnosis; the LR method is L2 regularized logistic regression, using multiple heat vectors to represent input data, and binary classification to process multi-label outputs. The Leap method uses a recurrent neural network to model tag dependencies, and uses a content-based attention mechanism to capture mappings between tag instances. The RETAIN method is based on a sequence data drug combination of a two-layer attention network model that selects important clinical variables in past visits. The GAMENET method is a method for integrating historical medications and drug interaction DDIs using a graph-volume network via a storage module. The PREMIER method learns patient history representations using the attention mechanism, in combination with the graph attention mechanism, for drug interactions.
Table 1 model comparison experiment
Figure BDA0003256807290000094
Figure BDA0003256807290000101
Table 1 is the performance of various methods on the data set for the drug recommendation task. Experimental results show that the model of the method can achieve the best effect in all methods. In particular, the methods proposed by the present invention are 0.97%, 0.89% and 0.93% higher than the latest method (PREMIER) in Jaccard, PRAUC and F1 scores, respectively. Meanwhile, the method gives consideration to drug interaction, and the lowest DDI is 0.0705 under the condition of the interaction rate of the first 40 drugs in all similar deep learning methods. In addition, the average medicine quantity recommended by the invention is 14.98, and the average medicine quantity closest to the real medical record in comparison with each deep learning method is 14.68.
GRAD-mkg represents the experimental results of the present invention with the knowledge base of drug interactions removed. Under the condition of no drug interaction knowledge base, the accuracy of the recommended drugs does not change greatly, but the DDI in the model recommended drugs becomes high and reaches 0.767, which shows that the knowledge of the interaction relationship among the drugs is combined with the query vector with the historical visit information in the method, so that the interaction rate in the recommended drugs is reduced, and the medication safety is improved. The Grad-tree represents a model generated by learning structural features of the patient's visit and medication information using only the neural network of the graph. After the medical code body structure is not embedded, the accuracy rate of drug recommendation is obviously reduced, which shows that the graph neural network used in the method has the coding capability of high-order structural features, can enrich the embedded representation of the medical body, makes up the problem of sparse training data to a certain extent, and improves the accuracy rate of drug recommendation.
Because the number of visits of each patient is different, the influence of the number of visits in the past should be considered. The present invention is superior to all other methods for different timing lengths. As shown in FIG. 3, the present invention has the highest F1 value in all categories by visit number. Particularly, for data with more times of visits, the accuracy can still be kept higher than that of other methods, which shows that the method has better modeling capability for long-time dependence in patient medical records.
TABLE 2 comparative experiments on DDI of different degrees
Figure BDA0003256807290000102
Figure BDA0003256807290000111
Further experiments were also performed in the present invention with respect to the effect on drug interactions. The first 40, 60, 80, 100 DDI types were used, respectively, to investigate the impact of the present invention and those compared methods when considering the use of different numbers of DDIs. As the results are shown in Table 2, although the Δ DDI ratio rises from-18.48 to-0.26%, when the number of DDI types considered is changed from 40 to 100, GRAD is the only algorithm capable of achieving DDI reduction, and the Δ DDI ratio is always larger than zero regardless of the DDI types. This indicates that the present invention can reduce the interaction rate of recommended drugs after introducing knowledge of drug interactions, and is safer.
It can thus be seen that the present invention has the following advantages: the proposed medicine recommendation algorithm based on the graph neural network and the attention mechanism takes the structural characteristics of the treatment condition or the medication information of each patient as a node, captures the relationship among the nodes by adopting the graph neural network, and learns the high-order characteristics containing the medical classification relationship. Meanwhile, the attention mechanism is used for better modeling the historical medical records of the user, the medicine interaction knowledge is introduced, and the accuracy and the safety of medicine recommendation are effectively improved.

Claims (7)

1. A medicine recommendation method based on a graph neural network and an attention mechanism is characterized by comprising the following steps:
step 1, acquiring historical electronic medical record data, and performing structured processing:
acquiring historical clinic situations of a patient and medication information corresponding to the clinic situations to construct an electronic medical record, wherein the clinic situations comprise diagnosis data and operation condition data; the patient's electronic medical record is denoted as p ═ x1，x2，...，xt-1]T is the current number of visits by the patient, where the ith visit by the patient is denoted as xi＝[di，pi，mi]，i＝1，2，...，t-1，diDiagnostic data, p, representing the patient's i-th visitiSurgical condition data, m, representing the ith visit in a patient's medical recordiMedication data representing the ith visit;
step 2, constructing three graph neural networks for learning about patientsStructural characteristics of the diagnosis and medication information; the inputs of the three graph neural networks are respectively diagnosis data, operation condition data and medication data of a patient, and the corresponding outputs are respectively de、pe、me；
Step 3, constructing two GRU network models with attention mechanism, and inputting the output results d of the step 2e、peThe corresponding outputs are respectively k with history informationd、kp；
Step 4, constructing two memory neural networks MANN with the same structure; wherein the key-value pair stored in the first memory neural network is' the ith visit diagnosis data fusion information
Figure FDA0003256807280000011
"-" graph neural network medication information
Figure FDA0003256807280000012
"; the key-value pair stored in the second memory neural network is' the fusion information of the condition of the operation of the ith visit
Figure FDA0003256807280000013
"-" graph neural network medication information
Figure FDA0003256807280000014
”；
Step 5, constructing a drug interaction knowledge base
5-1 step 4
Figure FDA0003256807280000015
And
Figure FDA0003256807280000016
combining to obtain a query vector containing historical medical record information, ith diagnosis information and ith operation information
Figure FDA0003256807280000017
Figure FDA0003256807280000018
Wherein WsContrast weights representing diagnostic and surgical information;
5-2, constructing a medicine coexistence relation matrix and a medicine interaction relation matrix in the electronic medical record
A*＝D-1(A*+I)D-1 (14)
Wherein D represents A*The diagonal matrix of the transform, D-1 being the inverse thereof, I being an identity matrix, A*Representing the drug coexistence relationship matrix A in the electronic medical recordCOr drug interaction relationship matrix AD；
5-3 learning relationships between drugs using graph convolution neural networks, incorporating drug interactions and co-occurrence relationships into the embedded representation, resulting in a representation matrix Z of drug co-occurrencesCAnd a representation matrix Z of a drug interaction mapD：
ZC＝ACtanh(ACme)WC (15)
ZD＝ADtanh(ADme)WD (16)
Wherein, WC，WDParameter matrices, m, for the drug contribution and drug interaction maps, respectivelyeA set of key-value pairs obtained in step 4;
based on matrix ZC、ZDAnd query vector
Figure FDA0003256807280000021
Calculating attention λi：
Figure FDA0003256807280000022
Wherein, WCDA contrast weight representing drug co-existence relationship and interaction;
finally obtaining a recommended medicine list yi：
Figure FDA0003256807280000023
Wherein, WyRepresenting a weight coefficient at the time of calculation of the recommended medicine list;
the recommendation probability y of the drug in the drug listiAnd if the probability is greater than the recommendation probability threshold rho, recommending the corresponding medicine.
2. The drug recommendation method based on graph neural networks and attention mechanism as claimed in claim 1, wherein in step 2, the three graph neural networks adopt the same structure and comprise nodes and edges; the nodes comprise leaf nodes and non-leaf nodes, the leaf nodes are input data, namely one of diagnosis data, operation condition data and medication data of a patient, and the non-leaf nodes are medical attribution classifications of the leaf nodes; the edge is the medical classification relation of two nodes;
each non-leaf node is represented as the sum of its own vector representation and all its sub-nodes, calculated using the GAT graph attention machine:
Figure FDA0003256807280000024
Figure FDA0003256807280000025
wherein g isnDenotes the nth non-leaf node, K denotes the total number of attentions, ReLU and LeakyReLU denote non-linear functions, ch (n) denotes the vector representation of the nth non-leaf node itself and all its children,
Figure FDA0003256807280000026
weight calculation coefficient, W, representing the current non-leaf node itself and all its children nodes under the kth attentionkRepresents the learning parameters of the non-leaf nodes at the kth attention, e*Vector representation representing nodes, a representing a learnable matrix, aTTransposing the same;
each leaf node is represented as the sum of the vector representations of itself and all its ancestor nodes, again calculated using the GAT graph attention machine mechanism:
Figure FDA0003256807280000027
Figure FDA0003256807280000028
wherein c'nRepresents the nth leaf node, an (n) represents the vector representation of the nth leaf node itself and all its ancestor nodes,
Figure FDA0003256807280000031
weight calculation coefficient, W ', representing the current leaf node itself and all ancestor nodes thereof under the k-th attention'kThe learning parameters of the k-th attention leaf node are represented.
3. The method for recommending drugs based on graph neural network and attention mechanism according to claim 1 or 2, wherein in step 3, the two GRU network models with attention mechanism are of the same structure, and each of the two GRU network models comprises two parallel GRU networks and an attention mechanism module connected with the outputs of the two parallel GRU networks;
the two parallel GRU models are specifically a first GRU network and a second GRU network, the first GRU network and the second GRU network respectively adopt different activation functions to acquire hidden information of historical visiting situations, and the method specifically comprises the following steps:
H＝GRU1(r) (5)
Wh＝softmax(Fh(H)) (6)
H′＝GRU2(r) (7)
W′h＝tanh(Fh′(H′)) (8)
h and H' respectively represent hidden layer information output by a first GRU network model and a second GRU network model, and Wh，W′hRespectively representing attention mechanism weights obtained by the first GRU network and the second GRU network through softmax and tanh activation functions, Fh，F′hRespectively representing the learnable linear transformation matrix functions of the first GRU network and the second GRU network, wherein r represents deOr pe；
The attention mechanism module calculates k with historical information according to formula (9)d、kpI.e. by
Figure FDA0003256807280000032
For diagnostic information with historical information for different time scales,
Figure FDA0003256807280000033
the operation condition information with history information of different time scales is obtained;
Figure FDA0003256807280000034
wherein t represents the total number of patient visits, Wh(i)，W′h(i) Respectively represent attention mechanism weights obtained by the softmax and tanh activation functions corresponding to the ith visit,
Figure FDA0003256807280000035
representing element-by-element multiplication; k is a radical of*Represents kdOr kp。
4. A graph-based neural network and attention machine as claimed in claim 1 or 2The prepared medicine recommending method is characterized in that the memory neural network carries out counterpoint multiplication on the ith diagnosis condition and the historical diagnosis condition in the step 4, and the weight of the ith diagnosis condition is calculated
Figure FDA0003256807280000036
Figure FDA0003256807280000037
Wherein
Figure FDA0003256807280000038
Or
Figure FDA0003256807280000039
The information of the treatment condition with historical information is shown in the ith treatment;
by weight
Figure FDA00032568072800000310
Affecting historical dose
Figure FDA00032568072800000311
The ith dose was as follows:
Figure FDA00032568072800000312
further obtaining keys of memory neural network
Figure FDA0003256807280000041
Figure FDA0003256807280000042
Wherein
Figure FDA0003256807280000043
Representing learning weight, key
Figure FDA0003256807280000044
Corresponding value is
Figure FDA0003256807280000045
According to
Figure FDA0003256807280000046
Can obtain
Figure FDA0003256807280000047
And
Figure FDA0003256807280000048
5. the medicine recommending device based on the graph neural network and the attention mechanism is characterized by comprising the following components:
the data preprocessing module is used for carrying out structuralized processing on the historical treatment condition of the patient and the medication information corresponding to the treatment condition to construct corresponding electronic medical record data;
the figure neural network module is used for learning structural characteristics of the patient treatment condition and the medication information; the inputs of the three graph neural networks are respectively diagnosis data, operation condition data and medication data of a patient, and the corresponding outputs are respectively de、pe、me；
The GRU network model module with attention mechanism is used for extracting characteristics of the diagnosis data and the operation condition data output by the GRU network module, and then combining the characteristics to the current diagnosis situation to obtain the diagnosis data and the operation condition data with historical information;
the memory neural network MANN module is used for constructing key value pairs of the ith diagnosis and diagnosis data fusion information output by the GRU network model module with the attention mechanism and the medication information of the graph neural network, and key value pairs of the ith diagnosis and operation condition fusion information output by the GRU network model module with the attention mechanism and the medication information of the graph neural network;
and the drug interaction knowledge base module is used for learning the drug co-occurrence relationship and the drug interaction relationship by adopting a graph convolution neural network, combining the drug interaction and the co-occurrence relationship to the drug embedded expression of the memory neural network MANN module and generating a recommended drug list.
6. A computer-readable storage medium, having stored thereon a computer program which, when executed on a computer, causes the computer to perform the method of any one of claims 1 to 5.
7. A computing device comprising a memory having stored therein executable code and a processor that, when executing the executable code, implements the method of any of claims 1-5.
CN202111061579.9A 2021-09-10 2021-09-10 Medicine recommendation method based on graph neural network and attention mechanism Pending CN113808693A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111061579.9A CN113808693A (en) 2021-09-10 2021-09-10 Medicine recommendation method based on graph neural network and attention mechanism

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111061579.9A CN113808693A (en) 2021-09-10 2021-09-10 Medicine recommendation method based on graph neural network and attention mechanism

Publications (1)

Publication Number Publication Date
CN113808693A true CN113808693A (en) 2021-12-17

Family

ID=78940750

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111061579.9A Pending CN113808693A (en) 2021-09-10 2021-09-10 Medicine recommendation method based on graph neural network and attention mechanism

Country Status (1)

Country Link
CN (1) CN113808693A (en)

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114386528A (en) * 2022-01-18 2022-04-22 平安科技（深圳）有限公司 Model training method and device, computer equipment and storage medium
CN115458045A (en) * 2022-09-15 2022-12-09 哈尔滨工业大学 Drug pair interaction prediction method based on heterogeneous information network and recommendation system
CN115631852A (en) * 2022-11-02 2023-01-20 北京大学重庆大数据研究院 Certificate type recommendation method and device, electronic equipment and nonvolatile storage medium
CN116110588A (en) * 2022-05-10 2023-05-12 北京理工大学 Medical time sequence prediction method based on dynamic adjacency matrix and space-time attention
CN116153528A (en) * 2023-04-20 2023-05-23 武汉纺织大学 Drug recommendation method based on attention mechanism and global retrieval
CN116631612A (en) * 2023-06-09 2023-08-22 广东工业大学 Graph convolution herbal medicine recommendation method and computer based on multi-graph fusion
WO2023202508A1 (en) * 2022-04-18 2023-10-26 浙江大学 Cognitive graph-based general practice patient personalized diagnosis and treatment scheme recommendation system

Cited By (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114386528B (en) * 2022-01-18 2024-05-14 平安科技（深圳）有限公司 Model training method and device, computer equipment and storage medium
CN114386528A (en) * 2022-01-18 2022-04-22 平安科技（深圳）有限公司 Model training method and device, computer equipment and storage medium
WO2023202508A1 (en) * 2022-04-18 2023-10-26 浙江大学 Cognitive graph-based general practice patient personalized diagnosis and treatment scheme recommendation system
CN116110588A (en) * 2022-05-10 2023-05-12 北京理工大学 Medical time sequence prediction method based on dynamic adjacency matrix and space-time attention
CN116110588B (en) * 2022-05-10 2024-04-26 北京理工大学 Medical time sequence prediction method based on dynamic adjacency matrix and space-time attention
CN115458045A (en) * 2022-09-15 2022-12-09 哈尔滨工业大学 Drug pair interaction prediction method based on heterogeneous information network and recommendation system
CN115458045B (en) * 2022-09-15 2023-05-23 哈尔滨工业大学 Drug pair interaction prediction method based on heterogeneous information network and recommendation system
CN115631852B (en) * 2022-11-02 2024-04-09 北京大学重庆大数据研究院 Certificate type recommendation method and device, electronic equipment and nonvolatile storage medium
CN115631852A (en) * 2022-11-02 2023-01-20 北京大学重庆大数据研究院 Certificate type recommendation method and device, electronic equipment and nonvolatile storage medium
CN116153528B (en) * 2023-04-20 2023-07-14 武汉纺织大学 Drug recommendation method based on attention mechanism and global retrieval
CN116153528A (en) * 2023-04-20 2023-05-23 武汉纺织大学 Drug recommendation method based on attention mechanism and global retrieval
CN116631612B (en) * 2023-06-09 2024-03-19 广东工业大学 Graph convolution herbal medicine recommendation method and computer based on multi-graph fusion
CN116631612A (en) * 2023-06-09 2023-08-22 广东工业大学 Graph convolution herbal medicine recommendation method and computer based on multi-graph fusion

Similar Documents

Publication Publication Date Title
CN113808693A (en) Medicine recommendation method based on graph neural network and attention mechanism
Liang et al. Text feature extraction based on deep learning: a review
Sahakyan et al. Explainable artificial intelligence for tabular data: A survey
Xia et al. Complete random forest based class noise filtering learning for improving the generalizability of classifiers
Srinivasu et al. From blackbox to explainable AI in healthcare: existing tools and case studies
Che et al. Deep computational phenotyping
Kim et al. Multi-modal stacked denoising autoencoder for handling missing data in healthcare big data
Kollmannsberger et al. Deep learning in computational mechanics
Jiang et al. A hybrid intelligent model for acute hypotensive episode prediction with large-scale data
Huang et al. A scalable approach for statistical learning in semantic graphs
CN113223656A (en) Medicine combination prediction method based on deep learning
Comito et al. AI-driven clinical decision support: enhancing disease diagnosis exploiting patients similarity
CN113793696A (en) Similarity-based method, system, terminal and readable storage medium for predicting occurrence frequency of side effects of new drug
Wang et al. Fuzzy-clustering and fuzzy network based interpretable fuzzy model for prediction
Baghaei et al. Deep representation learning: Fundamentals, perspectives, applications, and open challenges
CN113808734A (en) Causal medical diagnosis method based on deep learning
Rampogu A review on the use of machine learning techniques in monkeypox disease prediction
CN116072298B (en) Disease prediction system based on hierarchical marker distribution learning
Xu et al. Masked graph neural networks for unsupervised anomaly detection in multivariate time series
Fekihal et al. Self-organizing map approach for identifying mental disorders
Kuo et al. Mining surprising patterns and their explanations in clinical data
CN116543853A (en) Drug interaction prediction method, computer device, and medium
Che et al. Time series feature learning with applications to health care
Kaur et al. A review on Naive Baye’s (NB), J48 and K-means based mining algorithms for medical data mining
CN117352189B (en) Abnormal behavior evaluation method, system and equipment based on high-order topological structure

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination