A method, device and product for predicting the risk of postoperative complications

The method addresses the challenges of capturing causal relationships and confounding factors in deep learning models by using a structural causal graph and backdoor-adjusted inference, enhancing the accuracy of postoperative complication predictions in coronary artery bypass grafting.

CN118609822BActive Publication Date: 2024-12-06TIANJIN CHEST HOSPITAL +1
View PDF 6 Cites 0 Cited by

Patent Information

Application Number
CN202411080462.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-08
Publication Date
2024-12-06
Estimated Expiration
2044-08-08

AI Technical Summary

Technical Problem

Existing deep learning models for predicting postoperative complications in coronary artery bypass grafting face challenges in capturing true causal relationships, are prone to interpretability issues, and are affected by confounding factors due to limited and unevenly distributed medical data, leading to inaccurate predictions.

Method used

A method using a structural causal graph and backdoor-adjusted causal inference framework to construct a risk prediction model, combined with GCN and LSTM networks for feature extraction, and a multi-head attention module to decouple causal and non-causal features, along with a small sample learning approach to optimize the model.

Benefits of technology

The method effectively identifies and utilizes real causal features while eliminating confounding factors, improving the accuracy of postoperative complication predictions even with limited and uneven data distribution.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118609822B_ABST
    Figure CN118609822B_ABST
Patent Text Reader

Abstract

The present invention discloses a method, device and product for predicting the risk of postoperative complications, and relates to the field of postoperative complication risk prediction. The postoperative complication risk prediction method comprises constructing a structural causal graph according to the causal relationship between each variable in the electronic health record data; according to the structural causal graph, using the causal inference method of backdoor adjustment to construct a risk prediction model based on the causal inference framework of backdoor adjustment; using the small sample learning method, using samples in the existing data set to pre-train the risk prediction model; constructing multiple support sets according to samples of the target category; and optimizing the pre-trained risk prediction model using multiple support sets; and using the optimized risk prediction model to predict the risk of postoperative complications. The present invention can improve the accuracy of the prediction of complications of coronary artery bypass grafting.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of postoperative complication risk prediction, and in particular to a postoperative complication risk prediction method, device and product. Background Art

[0002] Coronary Artery Bypass Grafting (CABG) is a common treatment for cardiovascular diseases. Its surgical results are affected by many factors, resulting in postoperative complications and mortality risks that cannot be ignored. Therefore, accurately predicting the patient's postoperative health status is of great significance for improving the quality of medical services and reducing surgical risks. With the popularization of electronic health records (EHR), a large amount of clinical data has become an available resource, but traditional manual analysis methods are inefficient and easily affected by subjective factors. Therefore, the development of an automated surgical risk prediction model based on EHR data can not only optimize the clinical decision-making process, but also effectively improve the efficiency of medical resource utilization.

[0003] In recent years, a variety of complex models such as EuroSCOREII and STSPROM have been developed to guide clinical decision-making. They are based on a large amount of patient data and logistic regression to predict postoperative outcomes. They have been widely verified and regarded as the gold standard for risk prediction. However, these models have limitations in dealing with complex nonlinear relationships, require expert knowledge to build, are time-consuming and labor-intensive. Nowadays, machine learning has made major breakthroughs in both theory and technology, and it may be more capable of predicting postoperative outcomes than traditional statistical analysis. Machine learning models can capture higher-order, nonlinear relationships between predictors more deeply, thereby providing more robust predictions and can automatically learn and extract features from data, reducing manual involvement. There are currently many methods based on machine learning or deep learning that predict patients' health status after surgery based on EHR data, and have achieved good results.

[0004] However, deep learning-based methods face the following challenges in the medical field. First, due to their complexity, deep learning models often appear as "black boxes", resulting in a lack of interpretability of the models. Second, these methods focus mainly on capturing the correlation of data rather than true causal relationships, which may lead to erroneous inferences based on misleading correlations. In the medical field, an effective surgical risk prediction model should be constructed based on reasonable causal inferences. In addition, since medical data involves patient privacy issues, high-quality coronary artery bypass grafting data is difficult to obtain, which may cause existing deep learning models to overfit on limited data.

[0005] In recent years, some research works have achieved some results in addressing the problems of difficulty in capturing the true causal relationship between information and limited data based on medical data inference. Li et al. proposed a causal Markov model based on variational autoencoder (VAE) to distinguish hidden variables that are causally related to the disease and those that are not directly related to the disease, thereby avoiding learning false associations from the data and providing a robust method for early prediction of irreversible diseases. Nie et al. used the instrumental variable learning method to combine chest X-ray (CXR) images with EHR data to classify CXR images, giving instrumental variables medical semantics, thereby effectively eliminating pseudo-correlations and further improving the classification accuracy and explanatory power of the model. Nie et al. also effectively classified chest X-ray images through structured causal models and backdoor adjustment methods, while eliminating pseudo-correlations and enhancing the explanatory power of the model. In response to the problem of insufficient medical data, Shang et al. used graph neural networks to understand the relationship between medical codes, and used this information to pre-train the model on the patient's first visit data, improving the model's understanding ability by simulating prediction tasks in actual scenarios, and fine-tuning the model in subsequent patient visits to improve the accuracy of drug recommendations. Zhang et al. proposed a meta-learning-based framework MetaPred, which trains a meta-learner on multiple related risk prediction tasks, then uses this pre-trained model for target risk prediction and further fine-tunes it using limited samples in the target domain, thereby performing clinical risk prediction in limited longitudinal medical data.

[0006] However, how to construct the corresponding structural causal diagram based on the data in the EHR, capture the true causal relationship between clinical indicators and predicted results, and avoid the interference of confounding factors on the prediction as much as possible; how to effectively control all confounding factors on the backdoor path and accurately express the predicted results; and how to make full use of the information in the existing data set to complete accurate predictions in a limited and unevenly distributed data set are still urgent problems to be solved. Summary of the invention

[0007] The purpose of the present invention is to provide a method, device and product for predicting the risk of postoperative complications, which can improve the accuracy of predicting complications of coronary artery bypass grafting.

[0008] To achieve the above object, the present invention provides the following solutions:

[0009] A method for predicting the risk of postoperative complications, comprising:

[0010] Constructing a structural causal graph according to the causal relationship between each variable in the electronic health record data; the variables in the electronic health record data include: patient information S, features R extracted from S, information O of the patient's surgical stage, causal features X of the patient's surgical stage, non-causal features C of the patient's surgical stage, and predicted results Y;

[0011] According to the structural causal graph, a backdoor-adjusted causal inference method is used to construct a risk prediction model based on a backdoor-adjusted causal inference framework; the risk prediction model is used to predict the risk of postoperative complications based on mixed causal characteristics;

[0012] Using the small sample learning method, samples from the existing dataset are used to pre-train the risk prediction model based on the causal inference framework of backdoor adjustment;

[0013] Construct multiple support sets based on samples of the target category; and use multiple support sets to optimize the pre-trained risk prediction model;

[0014] The optimized risk prediction model was used to predict the risk of postoperative complications;

[0015] The prediction process of the risk prediction model is:

[0016] The patient's own information S is classified, constructed into a graph network containing two types of nodes: patient nodes and disease nodes, and the corresponding adjacency matrix is ​​generated;

[0017] According to the adjacency matrix, the GCN network is used to extract feature R from the patient's own information S;

[0018] According to the information O of the patient’s surgical stage, the LSTM network is used for feature extraction;

[0019] According to the extracted features, a multi-head attention module is used to extract the causal features X of the patient's surgical stage and the non-causal features C of the patient's surgical stage from the information O of the patient's surgical stage;

[0020] The extracted causal feature X of the patient's surgical stage and the non-causal feature C of the patient's surgical stage are hierarchically integrated using a structured method to obtain a mixed causal feature;

[0021] Prediction of postoperative complication risk based on mixed causal characteristics.

[0022] Optionally, the structure of the structural causal graph is:

[0023] S→R→{X, Y};

[0024] X←O→C;

[0025] {R, X, C}→Y;

[0026] Among them, → indicates the transition from cause to effect.

[0027] Optionally, the method of using a backdoor-adjusted causal inference method according to the structural causal graph to construct a risk prediction model based on a backdoor-adjusted causal inference framework specifically includes:

[0028] Cut off the connection between the causal feature X of the patient's surgical stage and the feature R extracted from S and the information O of the patient's surgical stage;

[0029] pass To express the true causal relationship from the causal feature X of the patient's surgical stage to the predicted outcome Y; Indicates intervention; then:

[0030] ;

[0031] in, ,..., }, ,..., }, and All are stratified for the i-th specific confounder; represents the conditional probability of Y given X, R, and C; is the conditional probability of r given S; is given The conditional probability of c under the condition of ; P(r) and P(c) are the marginal probability distributions of R and C respectively; represents the conditional probability of Y given X and R; represents the conditional probability of Y given X and C; Represents a set of features extracted from the patient's own information; Represents a set of non-causal features extracted from information about the patient's surgical stage.

[0032] Optionally, constructing multiple support sets according to samples of the target category and optimizing the pre-trained risk prediction model using the multiple support sets specifically includes:

[0033] Select specific target categories to construct the support set of M-way P-shot and the query set consisting of individual samples; where M-way P-shot means that each support set contains M categories and each category has P samples; then each support set includes M×P samples;

[0034] The pre-trained risk prediction model is used to calculate the mixed causal features of each category in the support set;

[0035] The average of the mixed causal features of each category in the support set is taken as the centroid;

[0036] Determine the similarity score between the mixed causal features and the centroids of the samples in the query set;

[0037] The pre-trained risk prediction model is optimized based on the similarity score.

[0038] Optionally, the loss function of the pre-training process of the risk prediction model is:

[0039] ;

[0040] in, is the loss function of the pre-training process; and are the weights for balancing decoupling and causal inference adjustment, respectively; is the causal loss function, which is used to measure the consistency between the predicted probability and the true label; is the supervised classification loss function; is the loss function of the non-causal feature C, which is used to keep the contribution of non-causal features to all categories equal.

[0041] A computer device comprises: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the postoperative complication risk prediction method.

[0042] Optionally, the memory is a computer-readable storage medium.

[0043] A computer program product comprises a computer program, wherein when the computer program is executed by a processor, the postoperative complication risk prediction method is implemented.

[0044] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0045] The present invention provides a method, device and product for predicting the risk of postoperative complications. A causal inference framework based on backdoor adjustment is proposed. While accurately identifying and utilizing real causal features, it effectively eliminates the impact of non-causal features (confounding factors) on model predictions; it uses a GCN network to process the patient's own information, and uses an LSTM model to process the patient's surgical stage information, and uses a multi-head attention module to decouple the information during the operation; it adopts a small sample learning method to select other category samples from the existing data to pre-train the model; and then constructs several support sets according to the target category to fine-tune the pre-trained risk prediction model. The present invention accurately identifies and utilizes real and stable causal features, while effectively eliminating the impact of confounding factors on model predictions; it incorporates a small sample learning method, so that the model can show strong performance even when the samples are limited and unevenly distributed, ensuring the ability to accurately identify key features; therefore, the present invention can improve the accuracy of predicting postoperative complications of coronary artery bypass grafting. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0047] Figure 1 A schematic diagram of a process for predicting the risk of postoperative complications provided by the present invention;

[0048] Figure 2 A structural causal diagram constructed based on clinical data of patients undergoing coronary artery bypass grafting;

[0049] Figure 3 Schematic diagram of the network structure for predicting the risk of coronary artery bypass grafting based on causal inference;

[0050] Figure 4 Schematic diagram of the small sample learning stage of the model;

[0051] Figure 5 ROC curves of various models for predicting various types of postoperative complications. DETAILED DESCRIPTION

[0052] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0053] The purpose of the present invention is to provide a method, device and product for predicting the risk of postoperative complications, which can improve the accuracy of predicting complications of coronary artery bypass grafting.

[0054] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0055] like Figure 1 As shown, a method for predicting the risk of postoperative complications provided by the present invention comprises:

[0056] S101, constructing a structural causal graph according to the causal relationship between each variable in the electronic health record data. The variables in the electronic health record data include: patient information S, feature R extracted from S, information O of the patient's surgical stage, causal feature X of the patient's surgical stage, non-causal feature C of the patient's surgical stage, and predicted result Y.

[0057] like Figure 2 As shown, the structure of the structural causal graph is:

[0058] S→R→{X, Y}; the current relationship indicates that the preoperative self-related information may affect both the clinical operation in the surgical stage and the postoperative results, so R points to the causal feature X and the predicted result Y in the surgical stage at the same time;

[0059] X←O→C; the current relationship indicates that C and X coexist in O. These two paths can be regarded as the process of feature extraction; X can represent the causal features that are crucial to diagnosis in the surgical stage information, while C can represent the non-causal features that may mislead the diagnosis;

[0060] {R, X, C}→Y; the current relationship indicates that the predicted result Y is mainly determined by the causal feature X of the surgical stage, but is also inevitably affected by the feature R extracted from S and the non-causal feature C; among them, → represents cause to effect.

[0061] S102, based on the structural causal graph, adopt the causal inference method of backdoor adjustment to construct a risk prediction model based on the causal inference framework of backdoor adjustment, such as Figure 3 shown.

[0062] The risk prediction model is used to predict the risk of postoperative complications based on mixed causal features. The present invention cuts off the connection between the causal feature and all its non-descendant nodes, so that the remaining connections in the structural causal graph can also express the prediction probability, eliminating false correlations.

[0063] The prediction process of the risk prediction model is:

[0064] S201, classify the patient's own information S, construct a graph network including two types of nodes: patient nodes and disease nodes, and generate a corresponding adjacency matrix.

[0065] The patient's own information S is classified according to the numerical information of clinical indicators and the classification information of whether there is a specific underlying disease, and a graph network containing two types of nodes: patient nodes and disease nodes is constructed.

[0066] Among them, the characteristics of the patient node are the patient's own numerical information, and each disease node , where m is the disease node index, Represents a collection of disease types, each disease node has a -, one-hot encoding , with the mth bit as 1 as its unique identifier, Represents the number of elements in the set, that is, the total number of diseases.

[0067] If the classification information indicates that the patient has a certain disease, a connection is established between the patient node and the corresponding disease node. Through these connections, a corresponding adjacency matrix is ​​generated while constructing a graph network containing patient and disease nodes. In this adjacency matrix, each element represents whether there is a connection between a specific patient and disease node, thereby reflecting the relationship between the patient and the disease.

[0068] S202, according to the adjacency matrix, the GCN network is used to extract the feature R from the patient's own information S.

[0069] ;

[0070] in, It is the representation of feature R extracted from the patient's own information. MLP It is a multi-layer perceptron used to further process and refine the features output by the graph convolutional network (GCN). A Represents the adjacency matrix between patient and disease nodes in the graph. It is an encoder based on the GCN network.

[0071] S203, using an LSTM network to extract features based on the information O of the patient's surgical stage.

[0072] The characteristics of the patient's surgical stage information O can be expressed as ,in, The feature matrix representing the information of the surgical stage of the i-th patient, represents the number of patients in the i-th surgery A vector of time points, represents the number of time steps in the operation phase, Represents the dimension of the feature, Representation Matrix is a dimensional real matrix, where represents the number of time points, and is the dimension of the feature vector at each time point. LSTM is used as the feature extraction network for the information O of the patient's surgical stage:

[0073] ;

[0074] in, Encapsulates the temporal dynamics of LSTM encoding, It is an LSTM based encoder.

[0075] S204, extracting the causal feature X of the patient's surgical stage and the non-causal feature C of the patient's surgical stage from the information O of the patient's surgical stage based on the extracted features using a multi-head attention module;

[0076] Introduce a multi-head attention module for decoupling. In the multi-head attention module, each layer l Update the query vector of the previous layer , thereby generating a new query representation and :

[0077] ;

[0078] ;

[0079] ;

[0080] ;

[0081] in, is the query vector from the previous layer ( K ) and value ( V ) vector is the composite feature value output by LSTM, defined as K = V = , d express K and VThe dimension of , the softmax function is used to calculate the self-attention weight, 1-softmax(·) is used to extract non-causal features that are opposite to causal features, and a zero vector is used. As the first layer of the multi-head attention mechanism, it is used to start the query process for features, where: Q represents the number of query vectors, represents the causal feature X extracted from the information O of the patient’s surgical stage, Represents the non-causal feature C of the patient's surgical stage extracted from the information O of the patient's surgical stage.

[0082] The goal of the causal feature extraction stage is to accurately estimate the causal features corresponding to the true labels. Therefore, the supervised classification loss The definition is as follows:

[0083] ;

[0084] ;

[0085] in, represents a data set, Indicates that the model responds to a specific input The predicted probability output of the category, () represents the classifier, represents the true label of sample d.

[0086] S205, the extracted causal feature X of the patient's surgical stage and the non-causal feature C of the patient's surgical stage are hierarchically integrated using a structured method to obtain a mixed causal feature and .

[0087] Following the backdoor adjustment theory, a structured approach was adopted to address confounding factors by hierarchically integrating confounding factors and causal features. The extracted R and C were stratified and randomly combined with the causal feature X waiting to be classified to obtain the mixed causal feature and And make predictions, the causal loss function is defined as follows:

[0088] ;

[0089] ;

[0090] ;

[0091] in, and are derived from the mixed information R and C through a hierarchical process, and is by taking the causal characteristics Non-causal features with stratification and formed by combining and represents the predicted probability obtained from the classifier. is a causal loss function that measures the consistency between the predicted probability and the true label.

[0092] S206, Prediction of postoperative complication risk based on mixed causal characteristics.

[0093] The embodiment of the present invention uses a backdoor adjustment method to control all non-descendant confounding factors on the path in the structural causal graph, and cuts off the connection between X and the feature R of the patient's own information and various information O during the operation:

[0094] Since the backdoor paths X←R→Y and X←O→C→Y, the traditional correlation The interference of R and C cannot capture the true causal relationship between X and Y; therefore, the present invention uses causal intervention to cut off the backdoor path and uses replace , giving the model the ability to exploit causal features, where indicates intervention;

[0095] Use the remaining connections in the causal diagram to express , assuming that confounding factors can be observed and stratified, i.e. ,..., }and ,..., },in, and All are stratified for the i-th specific confounder.

[0096] On this basis, the specific causal relationship formula is:

[0097] ;

[0098] in, ,..., }, ,..., },in, and All are stratified for the i-th specific confounder; represents the conditional probability of Y given X, R, and C, is the conditional probability of r given S, is given The conditional probability of c under the condition of , P(r) and P(c) are the marginal probability distributions of R and C respectively, represents the conditional probability of Y given X and R, represents the conditional probability of Y given X and C, Represents a set of features extracted from the patient's own information. Represents a set of non-causal features extracted from information about the patient's surgical stage.

[0099] S103, using the small sample learning method, using samples in the existing data set to pre-train the risk prediction model, such as Figure 4 shown.

[0100] The loss function of the pre-training process is:

[0101] ;

[0102] in, is the loss function for pre-training; and They are all hyperparameters, which are the weights for balancing decoupling and causal reasoning adjustment, thereby controlling the impact of the patient's own information and non-causal characteristics on the prediction during the calculation process. is a causal loss function that measures the consistency between the predicted probability and the true label. is the supervised classification loss function. is the loss function of C, which is used to keep the contribution of non-causal features to all categories equal.

[0103] S104, constructing multiple support sets according to samples of the target category, and optimizing the pre-trained risk prediction model using the multiple support sets.

[0104] S104 specifically includes:

[0105] Select specific target categories to construct the support set of M-way P-shot and the query set consisting of individual samples; where M-way P-shot means that each support set contains M categories and each category has P samples; then each support set includes M×P samples;

[0106] Calculate the mixed causal features of each category in the support set using the pre-trained risk prediction model and ;

[0107] Mixed causal features for each category in the support set and The average value of is taken as the centroid;

[0108] Determine the similarity score between the mixed causal features and the centroids of the samples in the query set and ;

[0109] Based on similarity score and The correspondence between the mixed causal features of the query set samples and the average causal features of each category in the support set is evaluated, and the pre-trained risk prediction model is optimized and adjusted.

[0110] Classification loss function based on similarity score The definition is as follows:

[0111] ;

[0112] in, and Respectively and The feature is the true category of the i-th sample The predicted probability of is a hyperparameter used to balance the weights of the two loss terms.

[0113] S105, use the optimized risk prediction model to predict the risk of postoperative complications.

[0114] The feasibility of the method of the present invention is verified below with reference to specific examples.

[0115] The dataset used for feasibility verification is an EHR-format dataset provided by Tianjin University Chest Hospital, which records the clinical information of 4,506 patients undergoing CABG surgery. It contains demographic information, medical diagnosis, laboratory results, vital signs, and detailed information related to the surgery, including preoperative patient information, detailed surgical stage data, and various postoperative complications.

[0116] In terms of data processing, indicators and patient samples with more than 30% missing values ​​were excluded, and the numerical information was normalized to ensure the consistency of data scale. The KNN method was used to estimate the remaining missing values ​​of numerical variables. For categorical information, it was assumed that missing entries indicated absence, so missing values ​​in categorical data were considered zero.

[0117] Given that there were significantly fewer patients with postoperative complications than without in this dataset, and to counteract the tendency of the model to predict the more general class, the dataset was adjusted and filtered to ensure a relative balance between patients with and without complications. This step was done before training and testing the various models to keep the dataset balanced.

[0118] The training is set to 200 epochs with a batch size of 128 patient samples. and are set to 1, and The value of the training set was set to 0.5. The Adam optimizer was used during training, with an initial learning rate of 1e-3, which was reduced to 20% of the previous value every 10 epochs. The patient data were randomly divided into a training set and a test set in a ratio of 8:2.

[0119] During the pre-training phase, we focus on the category with a larger sample size: patients with postoperative pneumonia, so that the model can classify this category and facilitate the extraction and learning of basic features.

[0120] In the fine-tuning stage, low cardiac output, new-onset atrial fibrillation, perioperative myocardial infarction, cardiac arrest, or ventricular fibrillation were selected as target categories to construct the support set for model fine-tuning. Four separate binary classification predictions were made for each target category, each time building a support set containing the target category and the post-operative health category to achieve accurate classification predictions.

[0121] The evaluation indicators used are Accuracy, Precision, Recall, F1 score and FPR, which are calculated as follows:

[0122] ;

[0123] ;

[0124] ;

[0125] ;

[0126] ;

[0127] Among them, TP, FP, TN and FN respectively represent the number of true positive, false positive, true negative and false negative samples in the prediction results. Accuracy is the most intuitive performance indicator, which indicates the proportion of correct predictions of the model, that is, the proportion of correctly predicted positive and negative examples in all predictions; Precision indicates the proportion of samples predicted as positive that are actually positive, reflecting the accuracy of the model in predicting positive classes; Recall indicates the proportion of samples correctly predicted as positive among all actual positive classes, reflecting the ability of the model to capture positive classes; F1 score is the harmonic mean of precision and recall, which is used to measure the balance between the two; FPR indicates the proportion of samples incorrectly predicted as positive among all actual negative classes, reflecting the proportion of negative classes incorrectly marked as positive when the model predicts positive classes.

[0128] Table 1, Table 2, Table 3, and Table 4 respectively show the experimental results of the method of the present invention in predicting whether CABG patients have complications such as low cardiac output, new atrial fibrillation, perioperative myocardial infarction, cardiac arrest or ventricular fibrillation. Compared with the experimental results of previous studies, the method of the present invention has certain advantages. The ROC curves of each model for predicting various postoperative complications are shown in Figure 1. Figure 5 As shown, the method of the present invention also has certain advantages in the ROC curve results.

[0129] Table 1 Comparison of experimental results for low cardiac output prediction

[0130]

[0131] Table 2 Comparison of experimental results for prediction of new-onset atrial fibrillation

[0132]

[0133] Table 3 Comparison of experimental results for the prediction of perioperative myocardial infarction

[0134]

[0135] Table 4 Comparison of experimental results for prediction of cardiac arrest or ventricular fibrillation

[0136]

[0137] In summary, the method of the present invention combines modules such as causal inference, backdoor adjustment, and small sample learning to establish a risk prediction model based on EHR data. It can effectively eliminate the influence of confounding factors on model prediction while accurately identifying and utilizing real and stable causal features. Even when the samples are limited and unevenly distributed, the model can still demonstrate strong performance, ensuring the ability to accurately identify key features.

[0138] In some embodiments, the present invention further provides a computer device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the postoperative complication risk prediction method. The memory provided by the present invention is a computer-readable storage medium.

[0139] In some embodiments, the present invention also provides a computer program product, including a computer program, which implements the postoperative complication risk prediction method when executed by a processor.

[0140] It should be noted that the object information (including but not limited to object device information, object personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in the present invention are all information and data authorized by the object or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant laws, regulations and standards of relevant countries and regions.

[0141] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided by the present invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided by the present invention may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided by the present invention may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but are not limited to this.

[0142] The technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0143] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A method for predicting the risk of postoperative complications, characterized in that: include: Construct a structural causal diagram based on the causal relationships between variables in the electronic health record data; The variables in the electronic health record data include: patient information S, features R extracted from S, information O of the patient's surgical stage, causal features X of the patient's surgical stage, non-causal features C of the patient's surgical stage, and predicted results Y; According to the structural causal graph, a backdoor-adjusted causal inference method is used to construct a risk prediction model based on a backdoor-adjusted causal inference framework; the risk prediction model is used to predict the risk of postoperative complications based on mixed causal characteristics; Using the small sample learning method, samples from the existing dataset are used to pre-train the risk prediction model based on the causal inference framework of backdoor adjustment; Construct multiple support sets based on samples of the target category; and use multiple support sets to optimize the pre-trained risk prediction model; The optimized risk prediction model was used to predict the risk of postoperative complications; The prediction process of the risk prediction model is: The patient's own information S is classified, constructed into a graph network containing two types of nodes: patient nodes and disease nodes, and the corresponding adjacency matrix is ​​generated; According to the adjacency matrix, the graph convolutional network is used to extract the feature R from the patient's own information S; According to the information O of the patient’s surgical stage, the LSTM network is used for feature extraction; According to the extracted features, a multi-head attention module is used to extract the causal features X of the patient's surgical stage and the non-causal features C of the patient's surgical stage from the information O of the patient's surgical stage; The extracted causal feature X of the patient's surgical stage and the non-causal feature C of the patient's surgical stage are hierarchically integrated using a structured method to obtain a mixed causal feature; Prediction of postoperative complication risk based on mixed causal characteristics; The structure of the structural causal graph is: S→R→{X, Y}; X←O→C; {R, X, C}→Y; Among them, → indicates the transition from cause to effect; According to the structural causal diagram, a causal inference method of backdoor adjustment is adopted to construct a risk prediction model based on a causal inference framework of backdoor adjustment, which specifically includes: Cut off the connection between the causal feature X of the patient's surgical stage and the feature R extracted from S and the information O of the patient's surgical stage; pass To express the true causal relationship from the causal feature X of the patient's surgical stage to the predicted outcome Y; Indicates intervention; then: ; Where C={c1, c2,..., c n }, R={r1, r2,..., r n }, c i and r i All are stratified for the i-th specific confounder; represents the conditional probability of Y given X, R, and C; is the conditional probability of r given S; is the conditional probability of c given O; P(r) and P(c) are the marginal probability distributions of R and C respectively; represents the conditional probability of Y given X and R; represents the conditional probability of Y given X and C; Represents the set of features extracted from the patient's own information S; represents the set of non-causal features extracted from the information O of the patient's surgical stage.

2. A method for predicting postoperative complication risk according to claim 1, characterized in that: The method of constructing multiple support sets based on samples of the target category and optimizing the pre-trained risk prediction model using the multiple support sets specifically includes: Select specific target categories to construct the support set of M-way P-shot and the query set consisting of individual samples; where M-way P-shot means that each support set contains M categories and each category has P samples; then each support set includes M×P samples; The pre-trained risk prediction model is used to calculate the mixed causal features of each category in the support set; The average of the mixed causal features of each category in the support set is taken as the centroid; Determine the similarity score between the mixed causal features and the centroids of the samples in the query set; The pre-trained risk prediction model is optimized based on the similarity score.

3. A method for predicting postoperative complication risk according to claim 1, characterized in that: The loss function of the pre-training process of the risk prediction model is: ; in, is the loss function of the pre-training process; and are the weights for balancing decoupling and causal inference adjustment, respectively; is the causal loss function, which is used to measure the consistency between the predicted probability and the true label; is the supervised classification loss function; is the loss function of the non-causal feature C, which is used to keep the contribution of non-causal features to all categories equal.

4. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the postoperative complication risk prediction method according to any one of claims 1 to 3.

5. A computer device according to claim 4, characterized in that: The memory is a computer-readable storage medium.

6. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the postoperative complication risk prediction method according to any one of claims 1 to 3 is implemented.

Citation Information

Patent Citations

  • Competitive risk survival analysis method based on causal inference

    CN114418420A

  • Method and system for improving prediction precision of postoperative complications of esophagus cancer

    CN115206530A

  • Disease prediction and early warning system based on causal network uncertainty reasoning

    CN115862869A

  • Birth defect risk factor assessment method based on causal learning

    CN116978556A

  • Disease risk assessment method and system based on clinical scale and causal discovery

    CN117711624A