A method for predicting postoperative acute pain

Abstract

The present application relates to a postoperative acute pain prediction method, belonging to the technical field of medical data analysis and processing, solving the problem of insufficient prediction accuracy of postoperative acute pain in the prior art. The method comprises: obtaining multi-omics data and clinical data of a patient to be predicted; constructing a first-level network of each omics based on anchor molecules of each omics; each level network after the second level network is obtained by extending the upper level network based on the molecular interaction relationship; taking each level network as a multi-level molecular interaction network of the patient to be predicted; extracting high-dimensional biological characteristics of the patient to be predicted based on the multi-level molecular interaction network using a trained deep learning model; and predicting the pain type of the patient to be predicted based on the high-dimensional biological characteristics, multi-omics data and clinical data. The accuracy of postoperative acute pain prediction is improved.

Description

Technical Field

[0001] This invention relates to the field of medical data analysis and processing technology, and in particular to a method for predicting postoperative acute pain. Background Technology

[0002] Postoperative acute pain is one of the most common complications in surgical patients, especially those undergoing general anesthesia. Acute pain not only causes intense discomfort but also triggers a series of adverse consequences, including delayed postoperative recovery, increased hospital stay, postoperative cognitive impairment, anxiety and depression, and may even increase the risk of complications such as infection, deep vein thrombosis, and cardiovascular events. Therefore, accurately predicting the risk of postoperative pain before or during the perioperative period has become a key scientific issue in perioperative medicine.

[0003] In current clinical practice, postoperative pain management primarily relies on patient self-reports and assessment scales, and is implemented through empirical adjustments to analgesic dosages. However, due to significant individual differences, patients exhibit vastly different tolerances to pain stimuli and sensitivities to analgesics. The traditional "one-size-fits-all" management approach often leads to insufficient or excessive analgesia, failing to ensure patient comfort and potentially causing drug side effects and dependence risks. In recent years, machine learning methods have been increasingly incorporated into pain prediction and analgesic dosage recommendation research. By integrating multidimensional data such as patient demographics, surgical procedures, anesthetic dosages, and preoperative psychological state, the accuracy of predictions can be improved to some extent. However, these methods still largely depend on clinical and questionnaire data, lacking a direct description of molecular-level mechanisms. Therefore, current technologies have insufficient accuracy in predicting acute postoperative pain and cannot provide a scientific basis for personalized analgesia strategies. Summary of the Invention

[0004] Based on the above analysis, the present invention aims to provide a method for predicting postoperative acute pain, in order to solve the problem of insufficient accuracy in predicting postoperative acute pain that relies solely on clinical information.

[0005] On one hand, embodiments of the present invention provide a method for predicting postoperative acute pain, comprising the following steps:

[0006] Acquire multi-omics and clinical data of patients to be predicted;

[0007] A primary network for each omics is constructed based on anchor molecules for each omics; the secondary network and each subsequent network are obtained by extending the previous network based on molecular interaction relationships; each level of network is used as a multi-level molecular interaction network for the patient to be predicted.

[0008] A trained deep learning model is used to extract high-dimensional biological features of the patient to be predicted based on the multi-level molecular interaction network; and the pain type of the patient to be predicted is predicted based on the high-dimensional biological features, multi-omics data and clinical data.

[0009] Based on further improvements to the above method, the higher-level network is extended based on molecular interaction relationships to obtain the second-level network and each subsequent level network, including:

[0010] Take any one of the second-level networks and each subsequent level network as the current-level network;

[0011] The nodes of the current level network are obtained based on the molecular interaction relationships of the nodes of the previous level network.

[0012] The edges of the current-level network are determined based on the nodes of the current-level network and the corresponding trained graph autoencoder; thus, the current-level network is obtained.

[0013] Based on a further improvement of the above method, nodes of the current level network are obtained from nodes of the previous level network according to molecular interaction relationships, including:

[0014] Use the nodes of the higher-level network as the base nodes of the current-level network;

[0015] The external neighbor molecules of the molecules corresponding to the nodes of the previous level network are selected from the interaction database and used as candidate nodes for the current level network.

[0016] Nodes with high relevance to the base nodes are selected from the candidate nodes and used as extension nodes of the current level network; the base nodes and extension nodes constitute the nodes of the current level network.

[0017] Based on further improvements to the above method, nodes with high relevance to the base nodes are selected from the candidate nodes as extension nodes for the current level network, including:

[0018] The overall relevance score of each candidate node is calculated based on topological affinity, molecular expression consistency, and functional semantic relevance.

[0019] The first number of candidate nodes with the highest overall scores are selected as the expansion nodes for the current level of the network.

[0020] Based on the above method, a further improvement is made to obtain the graph autoencoder corresponding to the trained current-level network in the following way:

[0021] A graph autoencoder is constructed, comprising an encoding module and a decoding module; the encoding module is used to obtain the embedding representation of each node using a multi-layer stacked graph convolutional network; the decoding module is used to predict the probability that there is an edge between node pairs based on the node embedding representation.

[0022] The graph autoencoder is trained based on the constructed training sample set to obtain the graph autoencoder corresponding to the current level network.

[0023] Based on the further improvement of the above method, the training loss of the graph autoencoder corresponding to the current level is calculated in the following way:

[0024] ;

[0025] in, Indicates the reconstruction loss. Indicates loss due to external biological constraints. This represents the weighting parameter.

[0026] Based on further improvements to the above method, the external biological constraint loss is calculated using the following formula:

[0027] ;

[0028] Where N represents the number of samples in the current training batch, This represents the set of pathways or functional categories to which the nodes of the current-level network of the k-th sample belong. Indicates the pathway or functional category in the set. Let represent the set of nodes in the current-level network of the k-th sample that belong to path or functional type p. Represents a set The embedding representation of the i-th node in the array. Represents a set The mean of the embedding representations of all nodes in the array. Represents a set Number of nodes This represents the 2-norm of a matrix.

[0029] Based on the further improvement of the above method, the reconstruction error is calculated using the following formula:

[0030] ;

[0031] Where N represents the number of samples in the current training batch, This represents the number of nodes in the current level network for the k-th sample. Let represent the probability predicted by the graph autoencoder that there is an edge between the i-th node and the j-th node. This indicates whether there is an edge between the i-th node and the j-th node in the true adjacency matrix.

[0032] Based on the further improvement of the above method, the characteristics of each node in each level of the network include the node's attribute characteristics and topological structure characteristics.

[0033] The attribute characteristics of a node include the expression level of the corresponding molecule in the patient and the structural characteristics of the molecule;

[0034] The topological characteristics of a node include its degree centrality, betweenness centrality, and eigenvector centrality.

[0035] Based on further improvements to the above methods, the anchor molecules for each omics are obtained in the following manner:

[0036] A training sample set was constructed by collecting data from multiple postoperative patients; the postoperative patient data included multi-omics data and pain types.

[0037] The samples in the training sample set were divided into two groups according to the type of pain.

[0038] For each molecule in each omics, differential expression analysis was performed between groups to identify anchor molecules for that omics based on the false discovery rate.

[0039] Compared to existing technologies, this invention acquires multi-omics and clinical data from the patient to be predicted, constructs a first-level network for each omics based on anchor molecules, and then extends the network based on molecular interaction relationships to obtain a multi-level molecular interaction network for the patient. A trained deep learning model then extracts high-dimensional biological features based on the patient's multi-level molecular interaction network. Based on these high-dimensional biological features, multi-omics data, and clinical data, the invention predicts the patient's pain type. This not only integrates the full spectrum of multi-omics information but also enhances the biological interpretability of the postoperative acute pain prediction model, thereby improving both prediction accuracy and model interpretability. This helps physicians identify high-risk patients preoperatively, develop analgesia strategies in advance, reduce over-medication or under-analgesia, improve postoperative recovery quality, reduce the risk of complications, and provide an intelligent and scalable solution for clinical pain management.

[0040] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0041] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0042] Figure 1 This is a flowchart of a postoperative acute pain prediction method according to an embodiment of the present invention. Detailed Implementation

[0043] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0044] One specific embodiment of the present invention discloses a method for predicting postoperative acute pain, such as... Figure 1 As shown, it includes the following steps:

[0045] S1. Obtain multi-omics and clinical data of the patients to be predicted;

[0046] S2. Construct a first-level network for each omics based on anchor molecules for each omics; the second-level network and each subsequent level network are obtained by extending the previous level network based on molecular interaction relationships; the network at each level is used as a multi-level molecular interaction network for the patient to be predicted.

[0047] S3. Using a trained deep learning model, extract high-dimensional biological features of the patient to be predicted based on the multi-level molecular interaction network; and predict the pain type of the patient to be predicted based on the high-dimensional biological features, multi-omics data and clinical data.

[0048] During implementation, the pain types include moderate to severe pain and no significant pain.

[0049] Compared with existing technologies, the training method for the postoperative acute pain prediction model provided in this embodiment not only integrates multi-omics full-spectrum information but also enhances the biological interpretability of the postoperative acute pain prediction model. This improves both prediction accuracy and the model's biological interpretability. It helps physicians identify high-risk patients preoperatively, develop analgesia strategies in advance, reduce over-medication or under-analgesia, thereby improving postoperative recovery quality, reducing the risk of complications, and providing an intelligent and scalable solution for clinical pain management.

[0050] During implementation, patient clinical data included demographic information, underlying medical conditions, surgical parameters, and anesthesia management factors. Demographic characteristics included age, sex, body mass index (BMI), smoking and alcohol consumption history, and family medical history to ensure the impact of individual differences on pain perception. Underlying medical conditions covered common chronic diseases such as hypertension, diabetes, and heart disease. Surgical parameters included detailed records of surgical site, surgical procedure, surgical time, intraoperative blood loss, and transfusion status. Anesthesia-related factors included the type of induction drug, maintenance drug dosage, whether multimodal analgesia was used, and the use of adjuvant medications. Z-score standardization was applied to all continuous variables. One-hot coding was used for transformation of categorical variables.

[0051] Current technologies are mostly limited to single-omics differential analysis or correlation studies based on only a few candidate molecules, thus affecting the accuracy of predictions. With the rapid development of multi-omics technologies, proteomics and metabolomics offer novel entry points for elucidating the mechanisms of postoperative pain. Proteins are the direct executors of biological functions, and their expression levels and interaction networks are closely related to pain signal transduction and inflammatory responses; metabolites reflect dynamic biochemical responses under postoperative stress and immune states. Through joint analysis of proteomics and metabolomics, not only can biomarkers closely related to pain be discovered, but complex molecular network regulatory relationships can also be revealed at the systemic level.

[0052] Therefore, the multi-omics approach of this application includes proteomics and metabolomics, and the data includes full-spectrum data from both omics.

[0053] During the procedure, arterial blood was collected from the patient after anesthesia and before surgery, with a blood volume of 5 mL, using EDTA-K2 anticoagulant tubes. Within 30 minutes of blood collection, the sample was centrifuged at 4°C and 3000×g for 10 minutes, and the supernatant plasma was separated. Each sample was aliquoted into 200 μL portions and immediately stored at -80°C. All samples were subjected to a single thaw before entering the testing process; repeated freeze-thaw cycles were not performed. Proteomics was performed using a label-free LC-MS / MS process to complete protein identification, peak extraction, isotope peak normalization, and batch effect correction, forming an expression matrix of the patient's proteins. Metabolomics was performed using UHPLC-MS in dual mode (positive / negative ion). Peak registration, alignment, isotope deconvolution, and internal standard calibration were performed to obtain the patient's metabolite matrix. The proteomics and metabolomics matrices were subjected to logarithmic transformation (log2(x+1)) and Z-score standardization, respectively. Missing values ​​were imputed using the K-nearest neighbor method, with K set to 5, and Euclidean distance was used to measure the feature dimension. After imputation, Z-score standardization was performed again to obtain the full proteomics profile of the patients for statistical testing and subsequent analysis. With full-spectrum metabolomics data.

[0054] In implementation, unlike the traditional approach that relies solely on a fixed database, this invention introduces a dynamic approach in the network construction phase. Based on anchor molecules, it combines an interaction database to construct a biological network centered on these molecules, and expands layer by layer to form a multi-level molecular interaction network. This facilitates the gradual capture of different granular features from local mechanisms to global functions.

[0055] During implementation, for patients to be predicted, a corresponding multi-level molecular interaction network is constructed for each omics.

[0056] Specifically, the anchor molecules for each omics were obtained using the following method:

[0057] S211. Collect data from multiple postoperative patients to construct a training sample set; the postoperative patient data includes the patient's multi-omics data and pain type;

[0058] S212. Divide the samples in the training sample set into two groups according to the type of pain;

[0059] S213. Perform differential expression analysis between groups for each molecule of each omics and determine the anchor molecule of the omics based on the false discovery rate.

[0060] During implementation, the training samples were divided into an experimental group and a control group. The experimental group consisted of samples that experienced moderate to severe pain, while the control group consisted of samples that did not experience significant pain.

[0061] For each protein in proteomics, differential expression analysis is performed between the two groups. For example, a two-tailed independent samples t-test is used to calculate the p-value of the difference between groups, and the Benjamini–Hochberg method is used for multiple test correction to obtain the FDR (false discovery rate). Molecules with an FDR less than a preset threshold are selected as anchor molecules for this proteomics study. For example, the preset threshold is set to 0.05.

[0062] In practice, the ratio of the difference between the two means to the pooled standard deviation can also be used to calculate the effect size Cohen's d. Receiver operating characteristic (ROC) curves are calculated based on the real labels, and the area under the curve (AUROC) is estimated using the DeLong method. The screening thresholds are set as follows: FDR < 0.05, AUROC ≥ 0.75, and absolute effect size ≥ 0.8. Protein molecules satisfying FDR < 0.05 are arranged in descending order of AUROC; if AUROCs are the same, they are arranged in descending order of absolute effect size. A certain number of proteins at the beginning of the sequence are selected as protein anchors, i.e., anchor molecules for proteomics. In a specific embodiment, the anchor molecules for proteomics include DNAJC5, DIABLO, LMAN1, RTN3, PAFAH1B2, ATP5IF1, and THEM6.

[0063] Anchor molecules for metabolomics are obtained in the same manner. In one specific embodiment, the anchor molecules for metabolomics include (KEGG number). C01035 , C00133 , C01017 , C00507 , C01606 , C14704 , C19285 .

[0064] In practice, the first-level network is obtained by expanding the anchor molecules.

[0065] Specifically, a first-level network for each omics is constructed based on anchor molecules for that omics, including:

[0066] For each type of omics, the anchor molecule of that omics is used as the basic node;

[0067] The direct neighbor molecules of the anchor molecule are extracted from the interaction database of this omics as supplementary nodes;

[0068] The basic nodes, supplementary nodes, and edges between nodes constitute the first-level network of this type of omics.

[0069] In practice, the STRING database can be used as the proteomics interaction database. Downloading is done by batch submitting UniProt IDs, setting the species as *Homo sapiens*, and setting the minimum interaction confidence threshold to 0.7. Only interaction edges with a total evidence score greater than or equal to 0.7 in the STRING database are retained; edges below the threshold are removed.

[0070] In practice, the anchor molecules of proteomics serve as the basic nodes of the primary network of proteomics. For each protein anchor molecule, all its directly adjacent nodes are extracted as supplementary nodes to form the primary network.

[0071] A first-order network is denoted as an undirected weighted graph. ,in This represents the set of vertices in a first-level network, including anchor protein molecules and their neighboring protein molecules. This represents the set of edges in the first-level network. To achieve personalized feature expression, proteins not detected in patient omics sequencing are removed during network generation, retaining only the truly detected proteins to ensure consistency between the network and individual sample data.

[0072] In practice, the KEGG database can be used as the interaction database for metabolomics. The KEGG Compound IDs corresponding to anchor metabolites are retrieved, and all upstream and downstream metabolites that directly react with that metabolite, along with their corresponding enzymatic reaction nodes, are extracted as supplementary nodes to construct the nodes of the primary metabolic network. The primary network of metabolomics is also defined as an undirected weighted graph. ,in Represents the set of metabolites and reaction nodes. This represents the edge set. Similarly, to maintain individualized characteristics, only metabolites actually detected in patient metabolomics sequencing are retained, and undetected nodes are removed.

[0073] In implementation, the characteristics of each node in the primary network include the node's attribute characteristics and topological structure characteristics;

[0074] The attribute characteristics of a node include the expression level of the corresponding molecule in the patient and the structural characteristics of the molecule;

[0075] The topological characteristics of a node include its degree centrality, betweenness centrality, and eigenvector centrality.

[0076] In implementation, for proteomics, each node in the primary network corresponds to the expression level of a molecule in the patient, that is, the standardized expression level of the protein corresponding to the node in the patient's omics sequencing. The structural features of the molecule are the sequence representation of the protein itself, which can be encoded using a fixed-length vector based on a deep pre-trained language model ESM to capture the structural and functional information in the protein sequence.

[0077] In implementation, for metabolomics, the expression level of the corresponding molecule in the patient is represented by the standardized abundance of the metabolite at each node in the primary network. The structural features of the molecule are the digital representation of the molecular structure of the metabolite corresponding to the node. In implementation, Morgan's cyclic molecular fingerprint can be used to generate a fixed-length vector as the digital representation of the molecular structure, comprehensively reflecting the structural information and potential response characteristics of the metabolite.

[0078] Through the above process, each first-level network is eventually transformed into a comprehensive feature space, which contains both the dynamic expression information of the individual patient and the attribute information of the molecules themselves, while retaining the topological structure features, thus laying a solid foundation for subsequent expansion of second-level, third-level, and other networks.

[0079] It should be noted that the nodes of the primary network in different patients correspond to the same molecules, only the feature values ​​of the nodes are different.

[0080] For the second-level network and each subsequent level network of the patient to be predicted, it is necessary to extend the previous level network based on molecular interaction relationships.

[0081] Specifically, based on molecular interaction relationships, the previous level network is extended to obtain the second level network and each subsequent level network, including:

[0082] S221. Take any one of the second-level networks and each subsequent level network as the current level network;

[0083] S222. Based on the nodes of the previous level network, the nodes of the current level network are obtained according to the molecular interaction relationship;

[0084] S233. Determine the edges of the current level network based on the nodes of the current level network and the trained graph autoencoder corresponding to the current level; thus obtaining the current level network.

[0085] Specifically, nodes in the current level network are derived from nodes in the previous level network based on molecular interactions, including:

[0086] S2221. Use the nodes of the previous level network as the basic nodes of the current level network;

[0087] S2222: Select the external neighbor molecules of the molecules corresponding to the nodes of the previous level network from the interaction database, and use them as candidate nodes of the current level network.

[0088] S2223. Select nodes with high relevance to the basic nodes from the candidate nodes as the extension nodes of the current level network; the basic nodes and extension nodes constitute the nodes of the current level network.

[0089] When constructing the second level and each subsequent level of the network, the network at that level will be used as the current network. The following explanation uses the second-level network as an example, with the second-level network as the current network.

[0090] A secondary network must contain nodes from the primary network; therefore, nodes from the primary network are first used as the foundational nodes for the secondary network.

[0091] Then, the external neighboring molecules of the molecules corresponding to the nodes of the previous level network are selected from the interaction database as candidate nodes for the current level network.

[0092] For example, in proteomics, candidate nodes are obtained by extracting all external neighboring molecules of the protein molecule corresponding to a node in the first-level network from the STRING database. External neighboring molecules of a protein molecule are those that are not in the previous level network.

[0093] Furthermore, nodes with high relevance to the base nodes are selected from the candidate nodes to serve as extension nodes for the current level of the network. Specifically, this includes:

[0094] S22231. Calculate the comprehensive relevance score of each candidate node based on topological affinity, molecular expression consistency, and functional semantic relevance.

[0095] S22232. Select the first number of candidate nodes with the highest comprehensive scores as the expansion nodes of the current level network.

[0096] Specifically, the formula for calculating the overall relevance score is as follows:

[0097] ;

[0098] in, Indicates candidate nodes Topological affinity, Indicates candidate nodes Molecular expression consistency Indicates candidate nodes Functional semantic relevance, , and This represents the weighting coefficient.

[0099] This indicates the degree of closeness between candidate nodes and primary network nodes in the network topology. In practice, a random walk model with restart can be applied to the graph structure formed by the primary network nodes and candidate nodes to obtain the steady-state probability of each candidate node as the topological affinity of the candidate node. The numerical range is uniformly normalized to [0,1].

[0100] This is used to reflect whether the expression pattern of a candidate node in a patient sample is consistent with the average expression pattern of the parent network node. In practice, the average feature of the structural features of the parent network node can be calculated, and then the similarity between the candidate node and the average feature can be calculated using cosine similarity to obtain the molecular expression consistency of the candidate node.

[0101] This method describes the similarity between the functional semantics of candidate nodes and the functional set of the parent network, relying on GO annotations and KEGG pathway information. In implementation, for proteomics, the textual semantic features of the GO gazes of each node in the parent network are extracted and an average feature vector is calculated. The textual semantic features of the GO gazes of each candidate node are also extracted, and their similarity to the average feature vector is calculated using Jaccard algorithm to obtain the functional semantic relevance of the candidate node. In implementation, for metabolomics, the textual semantic features of the KEGG pathway information of each node in the parent network are extracted and an average feature vector is calculated. The textual semantic features of the KEGG pathway information of each candidate node are also extracted, and their similarity to the average feature vector is calculated using Jaccard algorithm to obtain the functional semantic relevance of the candidate node.

[0102] It should be noted that the weight coefficients vary for different network levels. , and Different values ​​can be set. For example, for a two-level network, the weight coefficients are set to 0.25, 0.35, and 0.4, respectively. For a three-level network, the weights can be adjusted. The adjustment of the weights is based on the fact that the three-level expansion has a wider coverage and sparser connections between nodes compared to the two-level expansion. As the topological radius increases, relying solely on the network structure will strengthen the bias towards highly connected nodes and weaken the orientation towards actual functions. Therefore, the topological factor is reduced to 10%. At the same time, while expression consistency is still important over a larger range, it is affected by noise and individual differences, so it is set to 30%. Functional semantics can maintain the strongest biological explanatory power and mechanistic rationality in cross-path and cross-path expansion, so it is increased to 60%. That is, the weight coefficients are set to 0.1, 0.3, and 0.6, respectively.

[0103] The candidate nodes are ranked from highest to lowest based on their overall relevance score. During implementation, the first number of candidate nodes with the highest scores are selected and, together with the base nodes, constitute the nodes of the current level network. This first number is, for example, 100. If the number of candidate nodes is less than this first number, all are included. This ensures the controllability of the secondary network's scale and the balance among the anchor points.

[0104] It should be noted that the characteristics of each extended node are the same as those of the nodes in the aforementioned primary network, including the node's attribute characteristics and topological structure characteristics.

[0105] After determining the nodes of the current level network, the edges of the current level network are determined based on the trained graph autoencoder corresponding to the current level.

[0106] Specifically, the graph autoencoder corresponding to the trained current-level network is obtained using the following method:

[0107] A graph autoencoder is constructed, comprising an encoding module and a decoding module; the encoding module is used to obtain the embedding representation of each node using a multi-layer stacked graph convolutional network; the decoding module is used to predict the probability that there is an edge between node pairs based on the node embedding representation.

[0108] The graph autoencoder is trained based on the constructed training sample set to obtain the graph autoencoder corresponding to the current level network.

[0109] In practice, the training sample set constructed can be the same as the training sample set constructed in step S211 above.

[0110] In practice, the encoding module of the graph autoencoder can employ a multi-layered stacked graph convolutional network.

[0111] The first layer of the graph convolutional network in the encoding module takes as input the features of each node in the current layer, and the last layer outputs the embedded representation of each node. This embedding space maps the topological structure and attribute information of the nodes into a continuous space, thus facilitating subsequent relationship reconstruction.

[0112] The feature propagation formula for each layer of the graph convolutional network in the encoding module is:

[0113] ;

[0114] in, , for The degree matrix, where A represents the adjacency matrix of the current-level network nodes. Indicates the first Parameters of layer graph convolutional networks, Indicates the first Input features of layered graph convolutional networks Indicates the first The output features of a graph convolutional network. For the first layer of a graph convolutional network, its input features are the features of each node. For each subsequent layer of a graph convolutional network, its input features are the features output by the previous layer.

[0115] The decoding module of the graph autoencoder calculates the inner product of the embedding representations of node pairs to predict the probability that an edge exists between the node pairs. For any two nodes, it calculates... and The inner product of the two edges is mapped to the [0,1] interval using the sigmoid function to obtain the confidence score of the predicted edge. . Let represent the embedding representation of the i-th node. This represents the embedding representation of the j-th node. Let represent the probability that there is an edge between the i-th node and the j-th node predicted by the graph autoencoder. Thus, the entire predicted adjacency matrix... This represents the model's judgment on whether all node pairs are connected.

[0116] By comparing the actual adjacency matrix A between the current-level network nodes with the predicted adjacency matrix... The parameters of the graph autoencoder can be updated.

[0117] Taking the trained graph autoencoder corresponding to the second-level network as an example. For each sample in the training sample set, its first-level network is obtained using the aforementioned method, and the nodes of the second-level network are obtained based on the nodes of the first-level network. Then, based on the second-level network nodes of each sample, the graph autoencoder corresponding to the second-level network is trained, and the parameters of the graph autoencoder are updated to obtain the trained graph autoencoder corresponding to the second-level network. This process is repeated to obtain the trained graph autoencoder corresponding to the third-level network, which will not be elaborated further here.

[0118] Specifically, the training loss of the graph autoencoder is calculated using the following formula:

[0119] ;

[0120] in, Indicates the reconstruction loss. This indicates the loss due to external biological constraints.

[0121] During implementation, the reconstruction error can be calculated using the following formula:

[0122] ;

[0123] Where N represents the number of samples in the current training batch, This represents the number of nodes in the current level network for the k-th sample. Let represent the probability predicted by the graph autoencoder that there is an edge between the i-th node and the j-th node. This indicates whether there is an edge between the i-th node and the j-th node in the true adjacency matrix.

[0124] In practice, to further improve the accuracy of the model, this application adds an external biological constraint loss to the training loss, using external biological evidence to constrain the edge prediction of the model.

[0125] It's important to note that the number of nodes in the current level of the network may vary for each sample, but the feature dimensions of each node are the same. For graph convolutional networks, it's not necessary for the number of nodes to be the same for each training sample; it's sufficient to ensure that the feature dimensions of each node are consistent.

[0126] Specifically, the loss due to external biological constraints is calculated using the following formula:

[0127] ;

[0128] Where N represents the number of samples in the current training batch, This represents the set of pathways or functional categories to which the nodes of the current-level network of the k-th sample belong. Indicates the pathway or functional category in the set. Let represent the set of nodes in the current-level network of the k-th sample that belong to path or functional type p. Represents a set The embedding representation of the i-th node in the array. Represents a set The mean of the embedding representations of all nodes in the array. Represents a set Number of nodes This represents the 2-norm of a matrix.

[0129] It should be noted that, regarding proteomics, This represents the set of functional categories (e.g., GO functional categories) to which the nodes of the current-level network belong for the k-th sample. For metabolomics, This represents the set of pathways (e.g., KEGG pathways) to which the nodes of the current-level network of the k-th sample belong.

[0130] During implementation, when two nodes appear in the same GO function category or KEGG pathway, Regularization will push their embedding vectors closer together to increase the confidence of the edge in the prediction results. Conversely, if two nodes lack any functional annotations or omics commonalities, regularization will penalize their tendency to connect, reducing the likelihood that the edge will be preserved.

[0131] After the model training converges, the trained graph autoencoder corresponding to the current level is obtained.

[0132] For the patient to be predicted, the features of the nodes of the current level network are input into the trained graph autoencoder corresponding to the current level to determine the edges between the nodes of the current level network, thereby obtaining the current level network.

[0133] After the secondary network is constructed, it can be further extended to tertiary, quaternary, and so on, to capture more global network features. The construction process is the same as that of the secondary network and will not be repeated here. This yields a multi-level molecular interaction network for each omics, composed of multiple layers. After obtaining the multi-level molecular interaction network for the patient to be predicted, a trained deep learning model is used to extract high-dimensional biological features based on this network. Then, based on these high-dimensional biological features, multi-omics data, and clinical data, the pain type of the patient is predicted.

[0134] With the development of artificial intelligence technology, the application of deep learning and reinforcement learning techniques for pattern recognition, pattern classification, and image recognition in the medical field is becoming increasingly widespread.

[0135] Specifically, the deep learning model includes a graph feature extraction module and a classifier; the graph feature extraction module is used to extract high-dimensional biological features of the patient to be predicted based on a multi-level molecular interaction network; the classifier is used to predict the patient's pain type based on the patient's high-dimensional biological features, multi-omics data and clinical data.

[0136] In implementation, the graph feature extraction module can employ a graph convolutional network structure with an attention mechanism, with each level of the network corresponding to each type of omics. The classifier can use existing classification heads.

[0137] In implementation, the classifier uses several fully connected layers for non-linear mapping, with ReLU as the activation function in the hidden layers. Batch normalization and Dropout layers are introduced between layers to prevent overfitting. The final output layer is a one-dimensional Sigmoid unit, outputting the predicted value. This represents the probability of the patient experiencing moderate to severe pain after surgery.

[0138] In practice, the deep learning model is trained using the aforementioned training sample set (provided that a multi-layer, multi-level molecular interaction network corresponding to each sample is obtained). The training loss can be the cross-entropy loss between the prediction result and the pain type label.

[0139] Each level of the network uses deep learning methods for feature propagation and representation learning, thereby obtaining high-dimensional feature representations of anchor molecules at different network scales.

[0140] The node features of each level of the network for the patient to be predicted are input into the corresponding graph convolutional network with attention mechanism to obtain multi-layer network features centered on anchor molecules. This refers to high-dimensional biological characteristics, which are then used to input high-microbiological characteristics, multi-omics data, and clinical data into a deep learning model's classifier to predict the patient's pain type.

[0141] During implementation, the model output can be divided into "high risk" and "low risk" categories based on thresholds, and stratified risk alerts can be provided when necessary, such as indicating that the patient belongs to the "very high risk group", to help clinicians formulate differentiated analgesia strategies.

[0142] In clinical practice, doctors can obtain omics samples through blood collection during routine preoperative examinations and complete testing and data input within a short time. The system automatically performs feature extraction and model calculations, generating a risk prediction report within minutes. This report serves as an important reference for anesthesiologists and pain management specialists in developing individualized analgesia plans, thereby enabling preoperative early warning, dynamic intraoperative adjustments, and individualized postoperative intervention.

[0143] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0144] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting postoperative acute pain, characterized by, Includes the following steps: Acquire multi-omics and clinical data of patients to be predicted; A first-level network for each omics is constructed based on the anchor molecules of each omics; the second-level network and each subsequent network are obtained by extending the previous-level network based on molecular interaction relationships; each level of network is used as a multi-level molecular interaction network for the patient to be predicted. A trained deep learning model is used to extract high-dimensional biological features of the patient to be predicted based on the multi-level molecular interaction network; and the pain type of the patient to be predicted is predicted based on the high-dimensional biological features, multi-omics data and clinical data. Based on molecular interaction relationships, the previous level network is extended to obtain the second level network and each subsequent level network, including: Take any one of the second-level networks and each subsequent level network as the current-level network; The nodes of the current level network are obtained based on the molecular interaction relationships of the nodes of the previous level network. The edges of the current-level network are determined based on the nodes of the current-level network and the trained graph autoencoder corresponding to the current level; thus, the current-level network is obtained. The nodes of the current level network are derived from the nodes of the previous level network based on the molecular interaction relationships, including: Use the nodes of the higher-level network as the base nodes of the current-level network; The external neighbor molecules of the molecules corresponding to the nodes of the previous level network are selected from the interaction database and used as candidate nodes for the current level network. Nodes with high relevance to the base nodes are selected from the candidate nodes and used as extension nodes of the current level network; the base nodes and extension nodes constitute the nodes of the current level network. Nodes with high relevance to the base nodes are selected from the candidate nodes to serve as extension nodes for the current level of the network, including: The overall relevance score of each candidate node is calculated based on topological affinity, molecular expression consistency, and functional semantic relevance. The first number of candidate nodes with the highest overall scores are selected as the expansion nodes for the current level of the network.

2. The method for predicting acute postoperative pain according to claim 1, characterized in that, The graph autoencoder corresponding to the trained current-level network is obtained in the following way: A graph autoencoder is constructed, comprising an encoding module and a decoding module; the encoding module is used to obtain the embedding representation of each node using a multi-layer stacked graph convolutional network; the decoding module is used to predict the probability that there is an edge between node pairs based on the node embedding representation. The graph autoencoder is trained based on the constructed training sample set to obtain the graph autoencoder corresponding to the current level network.

3. The method for predicting acute postoperative pain according to claim 2, characterized in that, The training loss of the graph autoencoder corresponding to the current level is calculated as follows: ； in, Indicates the reconstruction loss. Indicates loss due to external biological constraints. This represents the weighting parameter.

4. The method for predicting acute postoperative pain according to claim 3, characterized in that, The loss due to external biological constraints is calculated using the following formula: ； Where N represents the number of samples in the current training batch, This represents the set of pathways or functional categories to which the nodes of the current-level network of the k-th sample belong. Indicates the pathway or functional category in the set. Let represent the set of nodes in the current-level network of the k-th sample that belong to path or functional type p. Represents a set The embedding representation of the i-th node in the array. Represents a set The mean of the embedding representations of all nodes in the array. Represents a set Number of nodes This represents the 2-norm of a matrix.

5. The method for predicting acute postoperative pain according to claim 3, characterized in that, The reconstruction error is calculated using the following formula: ； Where N represents the number of samples in the current training batch, This represents the number of nodes in the current level network for the k-th sample. Let represent the probability predicted by the graph autoencoder that there is an edge between the i-th node and the j-th node. This indicates whether there is an edge between the i-th node and the j-th node in the true adjacency matrix.

6. The method for predicting acute postoperative pain according to claim 1, characterized in that, The characteristics of each node in each level of the network include the node's attribute characteristics and topological characteristics; The attribute characteristics of a node include the expression level of the corresponding molecule in the patient and the structural characteristics of the molecule; The topological characteristics of a node include its degree centrality, betweenness centrality, and eigenvector centrality.

7. The method for predicting acute postoperative pain according to claim 1, characterized in that, The anchor molecules for each omics were obtained using the following method: Data from multiple postoperative patients were collected to construct a training sample set; the postoperative patient data included multi-omics data and pain types. The samples in the training sample set were divided into two groups according to the type of pain. For each molecule in each omics, differential expression analysis was performed between groups to identify anchor molecules for that omics based on the false discovery rate.

Images