CN-121983226-A - Personalized medicine recommendation method and system for fusing adverse drug reactions

CN121983226ACN 121983226 ACN121983226 ACN 121983226ACN-121983226-A

Abstract

The invention belongs to the technical field of intelligent medical treatment, and particularly relates to a personalized medicine recommendation method and system for fusing adverse drug reactions, wherein the method comprises the following steps: inputting longitudinal electronic health record multi-source data of a patient, extracting and fusing time sequence characteristics and medicine similarity characteristics, and generating comprehensive information representation of the patient. The local drug information representation and the drug information representation are generated by a drug molecule bipartite graph and a drug information box, respectively, based on the patient's comprehensive information representation. And after the external adverse reaction information is introduced and processed by the encoder, the comprehensive information of the patient is used for representing and weighting, so that the adverse reaction characteristics of the medicine are obtained. And splicing the local drug information representation, the drug information representation and the adverse drug reaction characteristics, and outputting final drug combination recommendation through predictive scoring. The invention fully utilizes various useful information, captures complementarity and relativity among different medicine information sources, improves the quality of multi-source data fusion, and optimizes the accuracy and safety of medicine recommendation.

Inventors

ZHANG YUANYUAN
WANG ZHENNUO
DU SHUANG
WANG SHAOQIANG
JIANG MINGJIAN

Assignees

青岛理工大学

Dates

Publication Date: 20260505
Application Date: 20260105

Claims (8)

1. The personalized medicine recommendation method for fusing adverse drug reactions is characterized by comprising the following steps of: S1, patient representation: acquiring longitudinal electronic health record EHR multisource data of a patient, extracting and fusing time sequence characteristics and medicine similarity characteristics, and generating comprehensive information representation of the patient ; S2, drug information integration: Comprises two steps of medicine molecular bipartite graph and medicine information box, the step of medicine molecular bipartite graph includes decomposing medicine molecule to construct medicine molecular bipartite graph, and representing comprehensive information of patient Performing dimension transformation to obtain local function vector, and converting the local function vector into local drug information representation ; The drug information box comprises the steps of extracting an EHR graph and a drug interaction DDI graph through EHR multi-source data, acquiring drug co-occurrence and interaction information from the EHR graph and the DDI graph, and modeling to obtain unified representation of drug knowledge Simultaneously, all historical patient information of multiple follow-up is stored in dynamic memory, and comprehensive information of the patient is represented And unified representation Z of drug knowledge queries the history of dynamic memory through an attention mechanism to generate personalized drug information features Fusing the personalized medicine information features to generate medicine information representation ; S3, encoding adverse drug reaction: The adverse drug reaction information introduced into the external database is input into the encoder for characteristic coding processing, and is represented by the comprehensive information of the patient The characteristics of the coding treatment are weighted and summed to obtain the final adverse reaction characteristics of the medicine ; S4, recommending the medicine: Representing drug information With local drug information representation Adverse drug reaction characteristics Splicing is carried out, and prediction scoring is carried out on the spliced characteristic expression, so that final drug combination recommendation is obtained.
2. The method of claim 1, wherein the longitudinal electronic health record EHR data includes a plurality of time series of visits including a corresponding disease signature, a surgical signature, and a drug signature, wherein the time series modeling of the disease signature and the surgical signature is performed to obtain a disease signature representation and a surgical signature representation of the current visit of the patient, wherein the disease signature representation and the surgical signature representation of the current visit of the patient are time series signatures, wherein the drug similarity between the longitudinal visit of the patient is simultaneously extracted and enhanced to obtain a full drug embedding signature, i.e., a drug similarity signature, and wherein the obtained disease signature representation and the surgical signature representation are concatenated with the full drug embedding signature to obtain a comprehensive information representation of the final patient 。
3. The method of claim 2, wherein constructing the composite information representation of the patient The method comprises the following steps: Step1, respectively carrying out feature coding and embedding representation on disease features and operation features to obtain disease embedded vectors and operation embedded vectors; Step2, respectively integrating time sequence features of the disease embedded vector and the operation embedded vector by using RNN and a transducer to obtain a disease feature representation and an operation feature representation of the current patient visit, wherein the disease feature representation and the operation feature representation of the current patient visit are time sequence features; Step3, constructing a graph structure reflecting the medicine similarity among visits based on all medicine feature sets used by patients in each visit, carrying out feature enhancement through a graph convolution network GCN to obtain enhanced visit representation, and carrying out averaging and aggregation on the node dimensions of all visits to obtain comprehensive medicine embedding features, namely medicine similarity features; step 4. Cascading the features obtained from Step2 and Step3 to obtain a comprehensive feature representation of the patient 。
4. The method of claim 1, wherein the step of bipartite mapping the drug molecules comprises decomposing the drug molecules and constructing a bipartite mapping Comprehensive information representation of patients via fully connected neural networks Performing dimension transformation to obtain local function vector The calculation formula is as follows: ; Wherein, the Is a linear transformation matrix, a function Comprehensive information representation for patients The transformation is performed such that the first and second parameters, Representing the local functional vector of the current patient in treating the disease, Representing a Sigmoid function; local function vector generation through a masking neural network Conversion to a local drug information representation The calculation formula is as follows: ; Wherein, the ". As indicated above, represents the element bitwise product, function For local function vector-based Generating a local drug information representation ; As a parameter matrix, the elements of the parameter matrix are bipartite And (5) masking.
5. The method according to claim 1, wherein the step of a pharmaceutical information box comprises in particular: s2.1, extracting an EHR image and a drug interaction DDI image through EHR multi-source data, and respectively establishing an adjacent matrix Adjacency matrix Adding self-connection to the adjacent matrix and carrying out symmetrical normalization, wherein the formula is as follows: ; ; Wherein, the Is a diagonal matrix of the type, Is an identity matrix; The characteristic matrix of the drug co-occurrence is represented, Representing a drug co-occurrence feature matrix; S2.2, modeling an EHR graph and a DDI graph respectively by using a two-layer graph rolling network GCN, learning the embedded representation of the medicine under the co-occurrence and interaction relationship, and carrying out weighted fusion on medicine embedded matrixes of the EHR graph and the DDI graph to obtain unified representation of medicine knowledge The form is as follows: ; Wherein, the Drug-embedding matrices from EHR and DDI maps, respectively; Is a matrix of trainable hidden weight parameters, Is a learnable adjustment parameter; s2.3, arranging all historical patient information of multiple follow-up times in time sequence, taking each patient as a key, taking a corresponding medication vector as a value, and forming a key value pair set Stored in dynamic memory, expressed as: ; Wherein, the Represent the first The patient's position of the patient is, Represent the first The current general characteristics of the individual patient are, Represent the first The medication that is currently being recommended by the individual patient, The number of current visits to the patient is indicated, ; S2.4 will be Comprehensive information representation of final patient of secondary visit Unified representation of drug knowledge Querying relevant history records in dynamic memory through an attention mechanism to generate personalized medicine information characteristics: ; ; Wherein, the Is based on comprehensive information representation of patient Unified representation of drug knowledge Is a function of the content of the (c) and the (c), Then it is a double search result based on the time dimension, Representation of Corresponding to the medication vector of all historical patient information for multiple follow-ups, Representation of Key of (c) corresponding to current integrated characteristics of the patient for all historical patient information for multiple follow-up; s2.5, representing the personalized medicine information characteristics and the comprehensive information of the patient Performing dimension transformation and information aggregation to obtain drug information representation, wherein the expression is as follows: ; Wherein, the In order to represent the information of the medicament, Comprehensive information representation expressed as patient And (5) performing dimension transformation and information aggregation by using the conversion function.
6. The method of claim 1, wherein the step of encoding for adverse drug reactions in S3 comprises: s3.1, taking a pre-constructed adverse drug reaction characteristic matrix o as input, and converting the high-dimensional and sparse adverse drug reaction characteristic matrix o into usable low-dimensional representation through characteristic coding by a multi-layer perceptron, wherein the specific coding structure comprises a full-connection network and an intermediate nonlinear activation function ReLU, and the specific coding process is as follows: ; ; Wherein the method comprises the steps of And Is the weight of the two-layer perceptron, Representing the process vector obtained in the encoding process, b1 and b2 are corresponding offset vectors, and the process vector is obtained after encoding Representing a corresponding low-dimensional ADR feature embedding for each drug; s3.2 Using the comprehensive information representation of the final patient Feature embedding Weighted summation is carried out to obtain the adverse reaction characteristics of the medicine The expression is as follows: ; Wherein, the Representing weight matrix, representing the comprehensive information of the final patient Mapped to ADR space.
7. The method of claim 1, wherein the final pharmaceutical combination recommendation comprises characterizing an adverse drug reaction With medication information presentation And Adding and reducing the dimension of the products, and then obtaining the output of the integrated medicament recommendation through sigma scaling of a Sigmoid function The specific formula is as follows: ; Wherein, the sign ∈ indicates the concatenation operation of the vector, +., (. Quadrature.) is a Sigmoid function for limiting the output value between 0 and 1; I function utilization threshold Output to a model Performing binarization processing, predicting that medicine with probability greater than or equal to 0.5 is recommended as effective medicine, converting result into 0 or 1, and further obtaining medicine combination recommendation ; 。
8. The personalized medicine recommendation system for fusing the adverse drug reactions is characterized by comprising a patient representation module, a medicine information integration module, a medicine adverse reaction module and a medicine recommendation module; the patient representation module acquires longitudinal electronic health record EHR multisource data of the patient, extracts and fuses time sequence characteristics and medicine similarity characteristics, and generates comprehensive information representation of the patient ; The medicine information integrating module comprises a medicine molecule bipartite graph module and a medicine information box module, wherein the medicine molecule bipartite graph module is used for decomposing medicine molecules to construct a medicine molecule bipartite graph and representing comprehensive information of a patient Performing dimension transformation to obtain local function vector, and converting the local function vector into local drug information representation The drug information box module extracts an EHR image and a drug interaction DDI image through EHR multi-source data, acquires drug co-occurrence and interaction information from the EHR image and the DDI image, and models the drug co-occurrence and interaction information to obtain unified representation of drug knowledge Simultaneously, all historical patient information of multiple follow-up is stored in dynamic memory, and comprehensive information of the patient is represented Unified representation of drug knowledge Querying the history record of dynamic memory by means of attention mechanism to generate personalized medicine information features Fusing the personalized medicine information features to generate medicine information representation ; The adverse drug reaction coding module introduces external adverse drug reaction information, inputs the information into the coder for characteristic coding processing, and utilizes the comprehensive information of the patient to represent The characteristics of the coding treatment are weighted and summed to obtain the final adverse reaction characteristics of the medicine ; The medicine recommending module represents medicine information With local drug information representation Adverse drug reaction characteristics Splicing the characteristic expressions into uniform characteristic expressions, and carrying out predictive scoring on the spliced characteristic expressions to obtain final drug combination recommendation.

Description

Personalized medicine recommendation method and system for fusing adverse drug reactions Technical Field The invention relates to the field of intelligent medical treatment, in particular to a personalized medicine recommendation and medication safety control technology based on electronic health records, and specifically relates to a personalized medicine recommendation method and a personalized medicine recommendation system for fusing medicine adverse reactions by combining medicine similarity and adverse reaction information with a time sequence module. Background With the rapid development of medical data acquisition technology and Electronic Health Record (EHR) systems, the diversity and complexity of medical data is increasing. Different sources and different types of medical data (e.g., diagnosis, surgery, medicine, etc.) together form a longitudinal health profile of a patient, and these multi-modal, multi-view data often contain rich and complementary information. How to integrate and utilize the heterogeneous medical data to realize safe and personalized medicine recommendation is an important research topic in the current intelligent medical field. The traditional medicine recommendation method mainly comprises two major types, namely an example method for recommending based on current treatment information, wherein the method only depends on current static clinical diagnosis of a patient to recommend medicines and neglects dynamic change of historical health state, and the other type is a time sequence method based on longitudinal patient record, and a richer patient representation is constructed by mining historical health data of the patient to realize personalized medicine recommendation. In recent years, deep learning techniques such as recurrent neural networks, graph neural networks, and the like have been widely applied to modeling of longitudinal EHR data, capable of capturing complex timing dependencies and global context information in multiple patient visits. The longitudinal method still has a plurality of key problems that firstly, most of the existing longitudinal methods adopt static calculation flows, are difficult to adapt to histories with different lengths, particularly, the new patients or the histories have poor performance under the condition of less history data (namely cold start problems), secondly, although part of the work already considers Drug-Drug interaction (Drug-Drug Interactions, DDI), the system integration of information on adverse Drug reaction (Adverse Drug Reaction, ADR) is still insufficient, so that the recommended result has hidden danger in terms of safety, and furthermore, the existing method has limited diversity and personalized treatment aspects of Drug recommendation, and is difficult to combine accuracy and safety. To solve the above problems, some studies have attempted to introduce multi-source heterogeneous information into a drug recommendation system. For example, safeDrug method uses the molecular diagram and DDI diagram to control the safety of drug combinations, COGNet models the correlation between drug recommendations through multiple diagram information, and Trans-GAHNet fuses the longitudinal clinical and drug information of patients to improve the recommendation performance. However, these methods remain limited in terms of cold start, model complexity and practical clinical application, and depth fusion and dynamic weight assignment to drug adverse reaction information are not yet sufficient. In summary, the electronic health record-based drug recommendation system is an important technology for improving the clinical intelligent medication level. The existing method relies on simple feature fusion of patient history information, lacks dynamic selection and weighting mechanisms of multi-source information such as longitudinal health data, medicine similarity and adverse reaction, and the like, and partially utilizes effective information to influence recommendation accuracy and safety. An innovative method capable of dynamically integrating multi-mode characteristics such as multi-source medical information, drug similarity and adverse reaction and considering recommendation accuracy and safety is needed in the current drug recommendation field. Disclosure of Invention Aiming at the defects of the prior art, the invention provides a personalized medicine recommendation method and a personalized medicine recommendation system for fusing adverse reactions of medicines, and aims to solve the technical problems of realizing dynamic selection and weighting of longitudinal data, multi-source information and medicine safety of patients in medicine recommendation, effectively utilizing unique characteristics of each patient and unique molecular representation of different medicines, and improving the accuracy of personalized medicine recommendation. In order to solve the technical problems, the invention provides a personalized medicine rec