Drug recommendation method and system fusing time sequence heterogeneous graph node information

Abstract

The invention provides a drug recommendation method and system fusing time sequence heterogeneous graph node information, and the method comprises the steps: collecting the drug use information of a current patient at each treatment time node, taking the patient and drugs as nodes, and taking the feedback of the patient to the drugs and the association information between the drugs as the edges between the nodes to construct a time sequence heterogeneous graph data set; performing feature extraction on the time sequence heterogeneous graph data set, and inputting an extraction result into a time perception graph neural network to obtain first drug response prediction probability distribution and second drug response probability distribution output by the time perception graph neural network; and generating a drug recommendation list based on the first drug response prediction probability distribution and the second drug response probability distribution. According to the method, the multi-source heterogeneous data including the electronic medical record of the patient, the medicine components and efficacy and the medicine interaction relationship are fused to realize medicine recommendation, and the accuracy of medicine recommendation is improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and particularly relates to a drug recommendation method and system that integrates temporal heterogeneous graph node information. Background Art

[0002] With the continuous development of personalized medicine, drug recommendation systems play an increasingly important role in clinical decision-making support. Traditional drug recommendation methods mostly rely on content-based recommendations, collaborative filtering, or other static methods. Although these methods can provide preliminary recommendation effects, it is difficult to take into account the dynamic changes of patients, the diversity of drugs, and the influence of time factors.

[0003] Therefore, due to the above reasons, the recommendation accuracy of existing drug recommendation systems is relatively low. Summary of the Invention

[0004] The present invention provides a drug recommendation method and system that integrates temporal heterogeneous graph node information, aiming to solve the defect of relatively low recommendation accuracy of drug recommendation systems in the prior art, and to implement a drug recommendation method and system with higher accuracy.

[0005] The present invention provides a drug recommendation method that integrates temporal heterogeneous graph node information, including:

[0006] Collect the medication information of the current patient at each visit time node, and construct a temporal heterogeneous graph data set with the patient and the drug as nodes, and the feedback of the patient on the drug and the association information between drugs as the edges between nodes;

[0007] Extract features from the temporal heterogeneous graph data set, and input the extraction result into a time-aware graph neural network to obtain the first drug response prediction probability distribution and the second drug response probability distribution output by the time-aware graph neural network, where the first drug response prediction probability distribution represents the prediction probability distribution of the current patient in the disease time series, and the second drug response probability distribution represents the probability distribution of the drug and the current patient in terms of drug efficacy matching;

[0008] Generate a drug recommendation list based on the first drug response prediction probability distribution and the second drug response probability distribution.

[0009] According to the drug recommendation method that integrates temporal heterogeneous graph node information provided by the present invention, the medication information of the current patient includes the medication record of the current patient, the historical health record of the current patient, and the interaction data between the current patient and the drug.

[0010] A drug recommendation method for fusing temporal heterogeneous graph node information. The step of performing feature extraction on the temporal heterogeneous graph data set and inputting the extraction result into a time-aware graph neural network to obtain the first drug response prediction probability distribution and the second drug response probability distribution output by the time-aware graph neural network specifically includes:

[0011] Perform feature extraction on the temporal heterogeneous graph data set to respectively obtain a patient node feature sequence and a drug node feature sequence;

[0012] Input the patient node feature sequence and the drug node feature sequence into the time-aware matching module of the time-aware graph neural network to obtain the first drug response prediction probability distribution output by the time-aware matching module;

[0013] Input the drug node feature sequence into the drug matching module of the time-aware graph neural network to obtain the second drug response probability distribution output by the drug matching module.

[0014] A drug recommendation method for fusing temporal heterogeneous graph node information. The step of inputting the patient node feature sequence and the drug node feature sequence into the time-aware matching module of the time-aware graph neural network to obtain the first drug response prediction probability distribution output by the time-aware matching module specifically includes:

[0015] After inputting the patient node feature sequence into the time-aware matching module, perform feature extraction by the graph convolutional neural network of the time-aware matching module to obtain the spatial features of the patient's response to the drug at each time point;

[0016] Add time position encoding to the spatial features of the patient's response to the drug based on the time information in the patient node feature sequence to obtain a drug response sequence with temporal information;

[0017] Construct the query matrix of the Transformer model in the time-aware matching module based on the drug node feature sequence, construct the key and value of the Transformer model based on the drug response sequence with temporal information, and use them as the input of the Transformer model decoder to obtain the first drug response prediction probability distribution output by the decoder.

[0018] A drug recommendation method for fusing temporal heterogeneous graph node information. The step of inputting the drug node feature sequence into the drug matching module of the time-aware graph neural network to obtain the second drug response probability distribution output by the drug matching module specifically includes:

[0019] After inputting the drug node feature sequence into the drug matching module, the graph attention network of the drug matching module performs feature extraction to obtain the feature space representation of the drug node;

[0020] Construct a similarity matrix between drug nodes based on several similarity calculation methods;

[0021] Determine the second drug response probability distribution based on the similarity matrix and the drug node feature sequence.

[0022] According to a drug recommendation method that fuses temporal heterogeneous graph node information provided by the present invention, the step of generating a drug recommendation list based on the first drug response prediction probability distribution and the second drug response probability distribution specifically includes:

[0023] Perform linear weighting on the first drug response prediction probability distribution and the second drug response probability distribution;

[0024] Generate the drug recommendation list based on the weighted result.

[0025] The present invention also provides a drug recommendation system that fuses temporal heterogeneous graph node information, including:

[0026] An acquisition module, configured to acquire the medication information of the current patient at each visit time node, and construct a temporal heterogeneous graph data set with patients and drugs as nodes and the feedback of patients on drugs and the association information between drugs as the edges between nodes;

[0027] A determination module, configured to perform feature extraction on the temporal heterogeneous graph data set, input the extraction result into a time-aware graph neural network, and obtain the first drug response prediction probability distribution and the second drug response probability distribution output by the time-aware graph neural network, where the first drug response prediction probability distribution represents the prediction probability distribution of the current patient on the disease time series, and the second drug response probability distribution represents the probability distribution of the drug matching the current patient in terms of drug efficacy;

[0028] A generation module, configured to generate a drug recommendation list based on the first drug response prediction probability distribution and the second drug response probability distribution.

[0029] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, and when the processor executes the program, it implements the drug recommendation method that fuses temporal heterogeneous graph node information as described in any one of the above.

[0030] The present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the drug recommendation method for fusing temporal heterogeneous graph node information as described in any one of the above.

[0031] The present invention also provides a computer program product, including a computer program. When the computer program is executed by a processor, it implements the drug recommendation method for fusing temporal heterogeneous graph node information as described in any one of the above.

[0032] The drug recommendation method and system for fusing temporal heterogeneous graph node information provided by the present invention, by introducing temporal heterogeneous node information, better utilizes the patient's historical medical records, drug usage records, patient feedback information, and drug interaction data with the help of a graph neural network, effectively captures the dynamic changes in the patient's health status and the drug effects at different time nodes, improves the time sensitivity of the drug recommendation model, and makes the recommendation results more personalized, thereby generating a more accurate drug recommendation list for each patient and improving the treatment effect and safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0034] Figure 1 is one of the schematic flowcharts of the drug recommendation method for fusing temporal heterogeneous graph node information provided by the present invention;

[0035] Figure 2 is a schematic diagram of the temporal heterogeneous graph dataset in the drug recommendation method for fusing temporal heterogeneous graph node information provided by the present invention;

[0036] Figure 3 is another schematic flowchart of the drug recommendation method for fusing temporal heterogeneous graph node information provided by the present invention;

[0037] Figure 4 is a schematic structural diagram of the drug recommendation system for fusing temporal heterogeneous graph node information provided by the present invention;

[0038] Figure 5 is a schematic structural diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0039] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions in the present invention will be clearly and completely described below with reference to the accompanying drawings in the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0040] The following Figures 1 to 3 introduces the drug recommendation method for fusing temporal heterogeneous graph node information of the present invention, as Figure 1 shown, including:

[0041] Step 101: Collect the medication information of the current patient at each visit time node, and construct a temporal heterogeneous graph data set with patients and drugs as nodes and the feedback of patients on drugs and the association information between drugs as the edges between nodes;

[0042] Optionally, collect the historical medication information of the current patient through an electronic medical record system and organize it according to the visit time nodes.

[0043] It can be understood that, that is, extract the records of the patient at each visit in the electronic pathology system, and organize the medication information of the patient from them, so as to recommend drugs to the patient based on the patient's medication records.

[0044] Optionally, the medication information of the current patient includes the current patient's medication records, the current patient's historical health records, and the interaction data between the current patient and drugs.

[0045] Among them, the current patient's medication records include the medications prescribed by the doctor for the current patient at each visit.

[0046] The current patient's historical health records include the records of the patient's physical health status, denoted as Optionally, it may include the patient's oral records and / or relevant medical examination results.

[0047] The interaction data between the current patient and drugs is the feedback of the patient on the treatment effect and / or adverse reactions of the medications for this visit, denoted as F i ={f i 1 , f i 2 , f i 3 , …, f i n}.

[0048] Further, obtain drug information records from the electronic database of drugs. The drug information records characterize the association relationships between drugs. For example, drugs need to be used in association, have similar properties, or cannot be used simultaneously.

[0049] Integrate the medication information of the patient collected at each medical visit time node in chronological order to form a temporal heterogeneous graph dataset G t ={V, E, t}, as Figure 2 shown.

[0050] It can be understood that each time segment t corresponds to the medical events generated at the time node of each medical treatment of the current patient. In each time segment, the patient, the drugs used by the patient, and the related drugs form nodes, and the feedback of the patient to the drugs and the association relationships between the drugs form the edges between the nodes.

[0051] For example, Figure 2 the node P in represents the patient, and d 1 to d 6 represent drugs. Among them, d 2 , d 3 and d 6 with a gray background represent the drugs used by the patient in this medical event. The connection line between P and the corresponding drug node represents that the feedback to this drug is effective or resistant; the drug nodes d 1 , d 4 and d 5 outside the gray background represent other drugs associated with the drugs used in this medical event. That is, d 1 , d 4 and d 2 have similar efficacy and can be used as alternative drugs. d 5 and d 6 have similar efficacy. In the case where the patient is resistant to d 6 , d 5 should not be used either.

[0052] Through the above method, based on the patient's historical medical records and the electronic database of drugs, the temporal heterogeneous graph dataset can be collected, sorted, and constructed.

[0053] Step 102: Extract features from the temporal heterogeneous graph dataset, and input the extraction results into a time-aware graph neural network to obtain the first drug response prediction probability distribution and the second drug response probability distribution output by the time-aware graph neural network. Among them, the first drug response prediction probability distribution characterizes the prediction probability distribution of the current patient in the disease time series, and the second drug response probability distribution characterizes the probability distribution of the drug matching the current patient in terms of efficacy;

[0054] Furthermore, feature extraction is performed on the data in the temporal heterogeneous graph dataset.

[0055] It can be understood that the data in the temporal heterogeneous graph dataset represents the patient's historical health data, disease status, and temporal information. Therefore, when the features extracted from the data in the temporal heterogeneous graph dataset are used as the input of the neural network model, the neural network model can predict the patient's future response to different drugs based on the above-mentioned multi-dimensional information, so as to achieve more accurate drug recommendations for each patient.

[0056] Corresponding to the constructed temporal heterogeneous graph dataset, the neural network model correspondingly selects the Time-Aware Graph Neural Networks (TAGNN).

[0057] The extracted features are input into the pre-constructed and trained TAGNN model to obtain the first drug response prediction probability distribution representing the current patient's drug response on the disease time series, and the second drug response probability distribution representing the drug efficacy matching with the current patient.

[0058] Among them, the first drug response prediction probability distribution aims to reflect the influence of the patient's medication feedback over time, and the first drug response prediction probability distribution aims to reflect the influence generated by the interaction between the drug and the patient.

[0059] Step 103, generate a drug recommendation list based on the first drug response prediction probability distribution and the second drug response probability distribution.

[0060] Combine the first drug response prediction probability distribution and the second drug response probability distribution to generate a final drug recommendation list.

[0061] Optionally, determine the corresponding first weight and second weight for the first drug response prediction probability distribution and the second drug response probability distribution, use the first weight to weight the probability corresponding to each drug in the first drug response prediction probability distribution to obtain the first weight of each drug; use the second weight to weight the probability corresponding to each drug in the second drug response probability distribution to obtain the second weight of each drug.

[0062] Take the sum of the first weight and the second weight of each drug as the final weight of each drug, sort the final weights in descending order to obtain the corresponding drug recommendation list, which helps doctors select the most appropriate treatment plan.

[0063] By introducing time-series heterogeneous node information, the present invention better utilizes the patient's historical medical records, medication usage records, patient feedback information, and drug interaction data with the help of graph neural networks, effectively captures the dynamic changes in the patient's health status and the medication effects at different time nodes, improves the time sensitivity of the drug recommendation model, and makes the recommendation results more personalized. As a result, a more accurate drug recommendation list is generated for each patient, improving the treatment effect and safety.

[0064] In the drug recommendation method of the present invention that integrates time-series heterogeneous graph node information, the step of performing feature extraction on the time-series heterogeneous graph data set and inputting the extraction result into a time-aware graph neural network to obtain the first drug response prediction probability distribution and the second drug response probability distribution output by the time-aware graph neural network specifically includes:

[0065] Performing feature extraction on the time-series heterogeneous graph data set to respectively obtain a patient node feature sequence and a drug node feature sequence;

[0066] Specifically, after constructing the time-series heterogeneous graph data set, perform grouping and labeling processing on the time-series heterogeneous graph data set, and classify and store the nodes and edges within each time segment in the format of "time - patient - record type". For example, the patient's historical health records, used medications and their interaction relationships, and patient feedback are respectively stored in corresponding folders to form a time-series heterogeneous graph data set with a hierarchical structure, providing data support for subsequent analysis and modeling.

[0067] On this basis, in order to ensure the consistency and integrity of the data, it is necessary to perform data cleaning and standardization processing on the time-series heterogeneous graph data set G t to remove noise data and fill in missing values.

[0068] Specifically, extract and encode the patient information and drug information in the time-series heterogeneous graph data set respectively to construct a preliminary patient node feature sequence and a preliminary drug node feature sequence

[0069] Among them, the patient node feature sequence X p includes features such as the patient's historical health status (such as disease history, allergy history), medication usage records, etc.; the drug node feature sequence X d includes features such as the composition, efficacy, usage frequency, and common interactions of the drug.

[0070] Furthermore, perform Z-score normalization processing on the preliminary patient node feature sequence X p and the preliminary drug node feature sequence X d :

[0071]

[0072] Wherein, μ is the mean value of the feature, and σ is the standard deviation of the feature.

[0073] The normalized X p and X d are respectively used as the patient node feature sequence and the drug node feature sequence obtained by feature extraction, so that each feature value has a similar numerical range, ensuring the stability of the model input, and storing the processed patient and drug feature sequences as structured data files for subsequent modeling and analysis.

[0074] Input the patient node feature sequence and the drug node feature sequence into the time-aware matching module of the time-aware graph neural network to obtain the first drug response prediction probability distribution output by the time-aware matching module;

[0075] In this embodiment, as Figure 3 shown, a time-aware matching module TMmodule is pre-constructed in the TAGNN as the first sub-network model, which is used to determine the first drug response prediction probability distribution representing the drug time-effect matching probability based on the patient node feature sequence and the drug node feature sequence.

[0076] Input the drug node feature sequence into the drug matching module of the time-aware graph neural network to obtain the second drug response probability distribution output by the drug matching module.

[0077] A drug matching module MMmodule is pre-constructed in the TAGNN as the second sub-network model, which is used to determine the second drug response prediction probability distribution representing the drug efficacy matching probability based on the drug node feature sequence.

[0078] In the drug recommendation method of the present invention that fuses the time-series heterogeneous graph node information, the step of inputting the patient node feature sequence and the drug node feature sequence into the time-aware matching module of the time-aware graph neural network to obtain the first drug response prediction probability distribution output by the time-aware matching module specifically includes:

[0079] After inputting the patient node feature sequence into the time-aware matching module, the graph convolutional neural network of the time-aware matching module performs feature extraction to obtain the spatial features of the patient's response to the drug at each time point;

[0080] Specifically, in the time-aware matching module, the patient node feature sequence is fed into the graph convolutional neural network (GCN) for feature extraction, and the patient features and drug features at each moment are mapped to obtain the spatial features F p (t) of the patient's response to the drug at each time point:

[0081] F p (t) = σ(W g · GCN(X p (t), A p ) + b g );

[0082] Wherein, A p is the adjacency matrix between patient nodes, used to represent the correlation relationship between time steps; W g and b g are the learnable weight matrix and bias vector respectively, and σ is the ReLU activation function.

[0083] Add temporal position encoding to the spatial features of the drug response for the patient based on the temporal information in the patient node feature sequence, to obtain a drug response sequence with temporal information;

[0084] Further, add temporal position encoding Pos(t) based on the temporal information in the patient node feature sequence to capture the dynamic changes related to time:

[0085]

[0086] Wherein, t represents the time node, i represents the dimension index in the temporal position encoding vector, and df represents the dimension of the hidden layer in the Transformer model.

[0087] On this basis, concatenate the spatial features F p (t) of the patient's drug response and the temporal position encoding Pos(t) together to obtain a drug response sequence F' p (t), F' p (t) = F p (t) + Pos(t), as the input of the subsequent module.

[0088] Construct the query matrix of the Transformer model in the time-aware matching module based on the drug node feature sequence, and construct the key and value of the Transformer model based on the drug response sequence with temporal information, as the input of the Transformer model decoder, to obtain the first drug response prediction probability distribution output by the decoder.

[0089] After obtaining F' p (t), input it into the encoder of the Transformer in the time-aware matching module to learn the influence of the response at past moments on the current drug response:

[0090] Q = K = V = F' p (t);

[0091]

[0092] Wherein, Q represents Query, K represents Key, and V represents Value; H t represents the historical hidden state.

[0093] And extract multiple feature relationships based on the multi-head attention mechanism:

[0094] H t = Concat(head 1 , …, head h W O );

[0095] Wherein, h is the number of attention heads, and W O is the output projection matrix.

[0096] Secondly, add residual connections and normalization to obtain:

[0097] H t = LayerNorm(H t + F' p (t));

[0098] And further process it through a feed-forward neural network layer to obtain:

[0099] H t = MLP(LayerNorm(H t ).

[0100] Meanwhile, use the drug feature sequence as the input of the decoder of the Transformer, construct the Query of the Transformer model, and input it into the decoder of the Transformer to obtain the output:

[0101] Q = X d ,

[0102]

[0103] Output = Concat(head 1 , …, head h W O );

[0104] Output = MLP(LayerNorm(Output));

[0105] On this basis, combine the finally output response prediction sequence with an activation function for non-linear mapping, and finally send it into a fully connected layer to obtain the final first drug response prediction probability distribution Pt :

[0106] P t = σ(W t ·Output + b t );

[0107] Wherein, W t is a learnable weight coefficient, b t is a bias matrix, and σ is an activation function.

[0108] In the method for drug recommendation that fuses the information of nodes in the temporal heterogeneous graph of the present invention, the step of inputting the drug node feature sequence into the drug matching module of the time-aware graph neural network to obtain the second drug response probability distribution output by the drug matching module specifically includes:

[0109] After inputting the drug node feature sequence into the drug matching module, the graph attention network of the drug matching module performs feature extraction to obtain a feature space representation of the drug node;

[0110] Specifically, in the drug matching module, the drug node feature sequence is sent to the graph attention network (GAT) for feature extraction:

[0111]

[0112] Wherein, a ij is the attention weight between node i and node j, a and W are learnable parameters, is the feature space representation of the drug node, and the correlation relationship in the constructed drug graph network is modeled to perform feature fusion for each node.

[0113] Construct a similarity matrix between drug nodes based on several similarity calculation methods;

[0114] According to the feature space representation of the drug nodes learned by the model, a similarity matrix between drug nodes is determined in combination with several similarity calculation methods.

[0115] Optionally, in this embodiment, two different similarity calculation methods, cosine similarity and Euclidean distance, are used to calculate the similarity between different drugs:

[0116]

[0117] And a similarity matrix S between drug nodes is obtained according to the similarity between drugs d :

[0118] S d= α·CosSim + (1 - α)·(1 - normalize(EuclideanDist)).

[0119] Determine the second drug response probability distribution based on the similarity matrix and the drug node feature sequence.

[0120] Based on the similarity matrix S between different drug nodes d and the drug node sequence of the patient Perform feature aggregation, combine with a fully connected layer, and finally output the second drug response distribution probability of the current patient's disease:

[0121] F d = F d * S d ;

[0122] P d = σ(W d ·F d + b d );

[0123] In the formula, W d is a learnable weight coefficient, b d is a bias matrix, and σ is an activation function.

[0124] In the drug recommendation method of the present invention that fuses temporal heterogeneous graph node information, the step of generating a drug recommendation list based on the first drug response prediction probability distribution and the second drug response probability distribution specifically includes:

[0125] Perform linear weighting on the first drug response prediction probability distribution and the second drug response probability distribution;

[0126] Generate the drug recommendation list based on the weighted result.

[0127] In this embodiment, based on the obtained first drug response prediction probability distribution P t and the second drug response probability distribution P d perform linear weighting to implement comprehensive matching score calculation and obtain a comprehensive score S ij :

[0128] S ij = βP t + (1 - β)P d ;

[0129] In the formula, β is a regulation parameter that balances the influence of time and drug characteristics on the score.

[0130] Based on the comprehensive score, use the Top - K method to obtain the top K drugs with the highest scores in descending order and generate a drug recommendation list List:

[0131] List = Top_K(S ij );

[0132] Based on the obtained drug recommendation list, doctors can select the final treatment plan according to the specific conditions of the patient (such as the patient's medical history, drug tolerance, past drug reactions, etc.), improving the treatment effect and reducing unnecessary side effects.

[0133] The present invention uses the TAGNN model to update the patient's medical status and drug response in real time, so it can dynamically adjust the drug recommendation list. Different from traditional static recommendation systems, this dynamic adjustment mechanism allows timely adjustment of the treatment plan when the patient's condition changes, ensuring the timeliness and effectiveness of drug use. This is particularly important in clinical practice, especially for patients with chronic diseases or rapidly changing conditions, which can significantly improve the treatment effect and reduce adverse reactions.

[0134] The drug recommendation system that fuses the information of nodes in the temporal heterogeneous graph provided by the present invention will be described below. The drug recommendation system that fuses the information of nodes in the temporal heterogeneous graph described below can be correspondingly referred to the drug recommendation method that fuses the information of nodes in the temporal heterogeneous graph described above.

[0135] As Figure 4 shown, the drug recommendation system that fuses the information of nodes in the temporal heterogeneous graph includes a collection module 401, a determination module 402, and a generation module 403;

[0136] The collection module 401 is used to collect the medication information of the current patient at each visit time node, and construct a temporal heterogeneous graph data set with patients and drugs as nodes and the patient's feedback on drugs and the association information between drugs as the edges between nodes;

[0137] Optionally, collect the historical medication information of the current patient through the electronic medical record system and organize it according to the visit time node.

[0138] It can be understood that, that is, extract the records of the patient's each visit in the electronic pathology system, and organize the medication information of the patient from them, so as to recommend drugs to the patient based on the patient's medication records.

[0139] Optionally, the medication information of the current patient includes the current patient's medication records, the current patient's historical health records, and the interaction data between the current patient and drugs.

[0140] Among them, the current patient's medication records include the drugs prescribed by the doctor for the current patient at each visit.

[0141] The current patient's historical health records include the records of the patient's physical health status, denoted as Optionally, it may include the patient's oral records and / or relevant medical examination results.

[0142] The current patient-drug interaction data is the feedback of the patient on the therapeutic effect and / or adverse reactions of the drugs used in this visit, denoted as F i ={f i 1 , f i 2 , f i 3 , …, f i n}.

[0143] Furthermore, drug information records are obtained from the electronic database of drugs. The drug information records characterize the association relationships between drugs, such as the need for associated use between drugs, similar drug properties, or the inability to be used simultaneously.

[0144] The medication information of the patient collected at each visit time node is integrated in chronological order to form a temporal heterogeneous graph dataset G t ={V, E, t}, as Figure 2 shown.

[0145] It can be understood that each time segment t corresponds to the medical events generated at the time node of each diagnosis and treatment of the current patient. In each time segment, the patient, the drugs used by the patient, and the related drugs form nodes, and the feedback of the patient on the drugs and the association relationships between the drugs form the edges between the nodes.

[0146] For example, Figure 2 the node P in represents the patient, and d 1 to d 6 represent drugs. Among them, the d 2 , d 3 and d 6 with a gray background represent the drugs used by the patient in this medical event. The connection line between P and the corresponding drug node indicates that the feedback on the drug is effective or resistant; the drug nodes d 1 , d 4 and d 5 outside the gray background represent other drugs associated with the drugs used in this medical event, that is, d 1 , d 4 and d 2 have similar drug effects and can be used as alternative drugs. d 5 and d 6 have similar drug effects, and in the case where the patient is resistant to d 6 , d 5 should not be used either.

[0147] In the above manner, based on the patient's historical medical records and the electronic database of drugs, a temporal heterogeneous graph dataset can be collected, organized, and constructed.

[0148] A determination module 402 is configured to extract features from the temporal heterogeneous graph dataset, and input the extraction result into a time-aware graph neural network to obtain a first drug response prediction probability distribution and a second drug response probability distribution output by the time-aware graph neural network, where the first drug response prediction probability distribution represents the prediction probability distribution of the current patient over the disease time series, and the second drug response probability distribution represents the probability distribution of the drug matching the current patient in terms of drug efficacy.

[0149] Furthermore, features are extracted from the data in the temporal heterogeneous graph dataset.

[0150] It can be understood that the data in the temporal heterogeneous graph dataset represents the patient's historical health data, disease status, and temporal information. Thus, when the features extracted from the data in the temporal heterogeneous graph dataset are used as the input of the neural network model, the neural network model can predict the patient's future responses to different drugs based on the above multi-dimensional information, so as to achieve more accurate drug recommendations for each patient in a targeted manner.

[0151] Corresponding to the constructed temporal heterogeneous graph dataset, the neural network model correspondingly selects a time-aware graph neural network (TAGNN).

[0152] The extracted features are input into a pre-constructed and trained TAGNN model to obtain a first drug response prediction probability distribution representing the current patient over the disease time series and a second drug response probability distribution representing the drug matching the current patient in terms of drug efficacy output by the model.

[0153] Among them, the first drug response prediction probability distribution aims to reflect the influence of the patient's medication feedback over time, and the first drug response prediction probability distribution aims to reflect the influence generated by the interaction between the drug and the patient.

[0154] A generation module 403 is configured to generate a drug recommendation list based on the first drug response prediction probability distribution and the second drug response probability distribution.

[0155] Combine the first drug response prediction probability distribution and the second drug response prediction probability distribution to generate a final drug recommendation list.

[0156] Optionally, determine corresponding first and second weights for the first drug response prediction probability distribution and the second drug response prediction probability distribution. Use the first weight to weight the probability corresponding to each drug in the first drug response prediction probability distribution to obtain the first weight of each drug; use the second weight to weight the probability corresponding to each drug in the second drug response prediction probability distribution to obtain the second weight of each drug.

[0157] Take the sum of the first weight and the second weight of each drug as the final weight of each drug, sort the final weights in descending order to obtain the corresponding drug recommendation list, and help the doctor select the most appropriate treatment plan.

[0158] By introducing temporal heterogeneous node information, the present invention better utilizes the patient's historical medical records, drug usage records, patient feedback information, and drug interaction data with the help of a graph neural network, effectively captures the dynamic changes in the patient's health status and the drug effects at different time nodes, improves the time sensitivity of the drug recommendation model, and makes the recommendation results more personalized, thereby generating a more accurate drug recommendation list for each patient and improving the treatment effect and safety.

[0159] Figure 5 An example of the physical structure diagram of an electronic device is shown as Figure 5 shown. The electronic device may include: a processor 510, a communication interface 520, a memory 530, and a communication bus 540. Among them, the processor 510, the communication interface 520, and the memory 530 complete communication with each other through the communication bus 540. The processor 510 can call the logical instructions in the memory 530 to execute the drug recommendation method that fuses temporal heterogeneous graph node information. The method includes: collecting the drug usage information of the current patient at each visit time node, constructing a temporal heterogeneous graph data set with patients and drugs as nodes and the patient's feedback on drugs and the association information between drugs as the edges between nodes; performing feature extraction on the temporal heterogeneous graph data set, inputting the extraction result into a time-aware graph neural network to obtain the first drug response prediction probability distribution and the second drug response probability distribution output by the time-aware graph neural network, where the first drug response prediction probability distribution represents the prediction probability distribution of the current patient in the disease time series, and the second drug response probability distribution represents the probability distribution of the drug and the current patient in terms of drug efficacy matching; generating a drug recommendation list based on the first drug response prediction probability distribution and the second drug response probability distribution.

[0160] In addition, when the logical instructions in the above-mentioned memory 530 are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs.

[0161] On the other hand, the present invention also provides a computer program product. The computer program product includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the drug recommendation method for fusing temporal heterogeneous graph node information provided by the above-mentioned various methods. The method includes: collecting the medication information of the current patient at each medical visit time node, constructing a temporal heterogeneous graph data set with patients and drugs as nodes and the patient's feedback on drugs and the association information between drugs as the edges between nodes; performing feature extraction on the temporal heterogeneous graph data set, and inputting the extraction result into a time-aware graph neural network to obtain the first drug response prediction probability distribution and the second drug response probability distribution output by the time-aware graph neural network. Among them, the first drug response prediction probability distribution represents the prediction probability distribution of the current patient in the disease time series, and the second drug response probability distribution represents the probability distribution of the drug and the current patient in terms of drug efficacy matching; generating a drug recommendation list based on the first drug response prediction probability distribution and the second drug response probability distribution.

[0162] In another aspect, the present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements a drug recommendation method for fusing temporal heterogeneous graph node information provided by the above-mentioned various methods. The method includes: collecting the medication information of the current patient at each visit time node, and constructing a temporal heterogeneous graph data set with patients and drugs as nodes and the feedback of patients on drugs and the association information between drugs as the edges between nodes; performing feature extraction on the temporal heterogeneous graph data set, and inputting the extraction result into a time-aware graph neural network to obtain the first drug response prediction probability distribution and the second drug response probability distribution output by the time-aware graph neural network. Among them, the first drug response prediction probability distribution represents the prediction probability distribution of the current patient in the disease time series, and the second drug response probability distribution represents the probability distribution of the drug matching the current patient in terms of drug efficacy; generating a drug recommendation list based on the first drug response prediction probability distribution and the second drug response probability distribution.

[0163] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art can understand and implement it without creative efforts.

[0164] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0165] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments or equivalently replace some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A drug recommendation method integrating temporal heterogeneous graph node information, characterized in that: include: Collect the medication information of the current patient at each visit time node, and construct a time series heterogeneous graph dataset with patients and drugs as nodes, and patients' feedback on drugs and the association information between drugs as edges between nodes; Performing feature extraction on the time series heterogeneous graph data set, inputting the extraction result into a time-aware graph neural network, and obtaining a first drug response prediction probability distribution and a second drug response probability distribution output by the time-aware graph neural network, wherein the first drug response prediction probability distribution represents the prediction probability distribution of the current patient on the disease time series, and the second drug response probability distribution represents the probability distribution of drug efficacy matching with the current patient; A drug recommendation list is generated based on the first drug response prediction probability distribution and the second drug response probability distribution.

2. The drug recommendation method integrating temporal heterogeneous graph node information according to claim 1 is characterized in that: The medication information of the current patient includes the medication record of the current patient, the historical health record of the current patient, and the interaction data between the current patient and the drug.

3. The drug recommendation method integrating temporal heterogeneous graph node information according to claim 1 is characterized in that: The step of extracting features from the temporal heterogeneous graph data set, inputting the extraction results into the time-aware graph neural network, and obtaining the first drug response prediction probability distribution and the second drug response probability distribution output by the time-aware graph neural network specifically includes: Performing feature extraction on the temporal heterogeneous graph data set to obtain a patient node feature sequence and a drug node feature sequence respectively; Inputting the patient node feature sequence and the drug node feature sequence into a time-aware matching module of a time-aware graph neural network to obtain the first drug response prediction probability distribution output by the time-aware matching module; The drug node feature sequence is input into a drug matching module of a time-aware graph neural network to obtain the second drug response probability distribution output by the drug matching module.

4. The drug recommendation method integrating temporal heterogeneous graph node information according to claim 3 is characterized in that: The step of inputting the patient node feature sequence and the drug node feature sequence into a time-aware matching module of a time-aware graph neural network to obtain the first drug response prediction probability distribution output by the time-aware matching module specifically includes: After the patient node feature sequence is input into the time-aware matching module, the graph convolutional neural network of the time-aware matching module performs feature extraction to obtain the spatial features of the patient's response to the drug at each time point; Adding time position coding to the spatial characteristics of the patient's response to the drug based on the time information in the patient node feature sequence to obtain a drug response sequence with time sequence information; A query matrix of the Transformer model in the time-aware matching module is constructed based on the drug node feature sequence, and keys and values ​​of the Transformer model are constructed based on the drug response sequence with timing information, which are used as input to the Transformer model decoder to obtain the first drug response prediction probability distribution output by the decoder.

5. The drug recommendation method integrating temporal heterogeneous graph node information according to claim 3 is characterized in that: The step of inputting the drug node feature sequence into a drug matching module of a time-aware graph neural network to obtain the second drug response probability distribution output by the drug matching module specifically includes: After the drug node feature sequence is input into the drug matching module, the graph attention network of the drug matching module performs feature extraction to obtain a feature space representation of the drug node; Construct a similarity matrix between drug nodes based on several similarity calculation methods; The second drug response probability distribution is determined based on the similarity matrix and the drug node feature sequence.

6. The drug recommendation method for integrating temporal heterogeneous graph node information according to any one of claims 1 to 5, characterized in that: The step of generating a drug recommendation list based on the first drug response prediction probability distribution and the second drug response probability distribution specifically includes: linearly weighting the first drug response prediction probability distribution and the second drug response probability distribution; The drug recommendation list is generated based on the weighted results.

7. A drug recommendation system integrating temporal heterogeneous graph node information, characterized in that: include: The collection module is used to collect the medication information of the current patient at each consultation time node, and construct a time-series heterogeneous graph dataset with patients and drugs as nodes, and patients' feedback on drugs and the association information between drugs as edges between nodes; A determination module is used to perform feature extraction on the time series heterogeneous graph data set, input the extraction result into the time-aware graph neural network, and obtain a first drug response prediction probability distribution and a second drug response probability distribution output by the time-aware graph neural network, wherein the first drug response prediction probability distribution represents the prediction probability distribution of the current patient on the disease time series, and the second drug response probability distribution represents the probability distribution of the drug and the current patient in terms of drug efficacy matching; A generating module is used to generate a drug recommendation list based on the first drug response prediction probability distribution and the second drug response probability distribution.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the drug recommendation method that integrates the time-series heterogeneous graph node information as described in any one of claims 1 to 6 is implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the drug recommendation method that integrates the time-series heterogeneous graph node information as described in any one of claims 1 to 6 is implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the drug recommendation method that integrates the time-series heterogeneous graph node information as described in any one of claims 1 to 6 is implemented.

Images