Multi-modal ICU patient death risk prediction method fusing causal effect estimation

Abstract

The invention discloses a multi-mode ICU patient death risk prediction method fusing causal effect estimation. The method comprises the steps that S1, original data samples in the patient treatment process are collected; s2, constructing a causal structure diagram; s3, based on a causal-driven anti-fact data generation method, performing minority class data enhancement to obtain a causal structure diagram of data balance; s4, forming multi-modal patient information representation based on the causal structure diagram; and S5, inputting the multi-modal patient information representation into the death risk prediction model, fusing ATE estimation, and outputting a death risk prediction probability. The ICU death rate prediction with high prediction performance and high interpretability is realized by combining the causal graph and the patient information for feature processing; in the data enhancement process, a BKA algorithm is introduced, intervention feature sampling and intermediate variable migration are jointly optimized, the causal consistency, representativeness and diversity of generated samples are effectively balanced, and the generalization ability and robustness of the model in key minority class event prediction tasks are improved.

Description

Technical Field

[0001] The present invention belongs to the field of medical big data technology, and specifically relates to a multimodal ICU patient mortality risk prediction method integrating causal effect estimation. Background Art

[0002] Mortality prediction is particularly critical in the intensive care unit (ICU), as ICU staff must rapidly make decisions and implement appropriate interventions to save lives based on mortality risk assessments. Furthermore, mortality prediction holds significant promise for improving healthcare management efficiency and reducing unnecessary mortality. Due to its widespread demand, a growing number of studies have explored mortality prediction from various perspectives, including disease diagnosis-based mortality risk prediction, clinical time series modeling, and the integration of domain knowledge into mortality prediction models. Currently, various traditional scoring systems are used in clinical research to predict mortality in various patient populations, with moderate accuracy (AUROC values ​​of approximately 0.65-0.85). These scoring systems include the APACHE II / III score, the SAPS II / III score, and the SOFA score. However, these scoring tools have several inherent limitations. First, their predictive accuracy declines over time with increasing life expectancy, changing public health conditions, and the emergence of new diseases. For example, systems such as APACHE, SOFA, and SAPS may experience calibration bias as patient populations evolve and diagnostic and treatment technologies develop. Second, some scoring systems use a static assessment model, relying primarily on data collected on the patient's first day in the ICU, forcing clinicians to rely on subjective judgments that are prone to bias. Third, cross-national validation studies have shown significant heterogeneity in the performance of these scoring systems, suggesting that their lack of adaptability to diverse patient populations may be an important factor leading to performance differences.

[0003] The limitations of traditional scoring systems have led to a surge in the use of machine learning (ML) and deep learning (DL) technologies in the fields of medical diagnosis and prognosis prediction. With the popularization of electronic health record (EHR) systems, the acquisition of massive amounts of clinical data has enabled models to be continuously updated and optimized based on real-time clinical information. Early DL applications for mortality prediction mainly used simple feedforward neural network architectures, and their performance was comparable to that of traditional scoring systems. Recent research advances have led to the emergence of more advanced DL models, which can be divided into three categories based on the input data type: (1) time series-based models: using clinical time series data such as vital signs and laboratory results to predict patient outcomes; (2) text-based models: extracting prognostic information from clinical text containing key semantic information; and (3) hybrid models: integrating multimodal data to provide a more comprehensive mortality prediction solution. Recent studies on ICU mortality prediction based on the public MIMIC-III database have shown that the area under the receiver operating characteristic curve (AUROC) of the model is between 0.8 and 0.9, and the area under the precision-recall curve (AUPRC) is approximately 0.3-0.7. However, despite achieving breakthroughs in prediction accuracy, these methods have the following key limitations: (1) they are not compatible with traditional clinically recognized physical diagnosis and treatment processes, limiting the application of medical knowledge; (2) the models are poorly interpretable, and their "black box" nature restricts their practical clinical application; (3) they use a fixed feature set, and their stability and robustness remain to be verified. Although deep learning models primarily focus on improving prediction performance, the medical field places greater emphasis on the generalization ability of models and the interpretability of domain-specific knowledge. There is an urgent need to enhance the model's ability to extract underlying causal structures to support clinical decision-making.

[0004] Furthermore, the development of prognostic models faces the problem of data imbalance, a common challenge in machine learning classification tasks. This is especially true when the number of samples in one class far exceeds that of other classes, causing the model to heavily favor the majority class and neglect the minority class. Data imbalance is particularly prominent in machine learning applications in healthcare. This imbalanced distribution can lead predictive models to misclassify high-risk patients as negative, potentially delaying critical interventions. Summary of the Invention

[0005] In response to the above-mentioned deficiencies in the prior art, the multimodal ICU patient mortality risk prediction method that integrates causal effect estimation provided by the present invention solves the problem that traditional deep learning models have difficulty in capturing causal effects, resulting in poor generalization ability of prediction results between different data distributions, and the problem that traditional models use unbalanced data distribution, which in turn affects the accuracy of model prediction results.

[0006] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is: a multimodal ICU patient mortality risk prediction method integrating causal effect estimation, comprising the following steps: S1. Collect raw data samples during the patient's treatment process and perform preprocessing; Among them, the original data samples include patients' diagnosis, procedures, medications, demographic information, and ICU observation indicators; S2. Construct a causal structure diagram using diagnosis as the intervention variable, death as the outcome variable, and operation and medication as mediating variables between diagnosis and death; S3. Based on the causal-driven counterfactual data generation method, minority class data enhancement is performed on the diagnosis, operation, and drug in the causal structure graph to obtain a data-balanced causal structure graph. The variables in the causal structure graph are converted into encoding features in a unified format through the Transformer encoder. S4. Concatenate the coded features in a unified format, and process the concatenated features, demographic information, and ICU observation indicators through a multi-layer perceptron. Concatenate the outputs of the multi-layer perceptron according to a causal structure diagram to form a multimodal patient information representation. S5. Input the multimodal patient information representation into the mortality risk prediction model and integrate the causal effect ATE estimation to output the final mortality risk prediction probability.

[0007] Furthermore, the preprocessing in step S1 includes detecting and correcting outliers in the original data samples, performing maximum-minimum normalization on continuous data, performing label encoding on data types, and processing missing data values.

[0008] Furthermore, in step S2, based on the inter-group causal relationship and intra-group causal relationship of the diagnosis, death, operation and medication variables, diagnosis is used as an intervention variable, death is used as an outcome variable, and operation and medication are used as mediating variables between diagnosis and death to construct a causal structure diagram; Among them, for the inter-group causal relationship of variables, the inter-group causal relationship that conforms to medical common sense is obtained based on the entity medical relationship between the variables; for the intra-group causal relationship of variables, the causal discovery method based on reinforcement learning is used to explore the relationship between the features within the variables, combined with the causal discovery paradigm based on variable order and the pre-training method, the intra-group causal relationship is obtained by searching the space of variable order and collective variable selection.

[0009] Furthermore, in step S3, the method for performing minority class data enhancement includes: S31. Dataset that will form the causal structure diagram Divide into majority class subsets and minority class subset ; in, represents the feature vector of the i-th sample, represents the label of the i-th sample, Represents the total number of samples; in the feature vector In the model, the diagnosis variable group serves as the intervention variable group T, death serves as the outcome variable, and operation and medication serve as the mediating variables M between diagnosis and death; represents the intervention variable group of the i-th sample; S32. For minority class subsets For each intervention variable group in the majority class subset Use the nearest neighbor 1-NN to match and find the majority class sample with the closest propensity score , forming a counterfactual sample pair, which will not be matched successfully to the majority class samples as a counterfactual generation target; S33, from the current minority class subset In the BKA algorithm, the intervention variable group is sampled and a new sample is generated. ; S34, based on the causal path in the causal structure diagram, the generated new sample Most of the samples that are not successfully paired Use the nearest neighbor 1-NN to match, and select the majority class samples that are successfully matched Migrate the mediating variable M in the sample, generate new minority class counterfactual instances, and copy the migrated mediating variable to the new sample In the final counterfactual minority class sample, construct ; S35. Based on the constructed counterfactual minority samples and majority samples, a data-balanced causal structure diagram is formed.

[0010] Furthermore, the step S33 includes the following sub-steps: S33-1. Use kernel density estimation to estimate the probability density of the intervention variable group and then generate the initial population; S33-2, using the BKA algorithm to globally optimize the bandwidth and sampling distribution weight of the estimated probability density; S33-3. Based on the optimized bandwidth and sampling distribution weight, after repeated sampling for each intervention variable group several times, the confidence interval of the mean and standard deviation of the intervention characteristic in the intervention variable group is calculated; S33-4. Use the confidence interval as the parameter search boundary of the BKA algorithm, and use the mutation strategy of the particle swarm optimization algorithm to perform global search and local search until the set number of iterations is reached and the multi-objective fitness function of the BKA algorithm converges.

[0011] Furthermore, the multi-objective fitness function of the BKA algorithm is: Where, represents the weight parameter, represents the diversity index, Represents representative indicators, represents the causal consistency index, represents the eigenvalue of the generated sample, represents the feature mean, represents the inverse probability weight of each sample calculated from the propensity score estimate, represents the propensity score for generating samples, represents the propensity score of the true unmatched majority class sample, represents the average treatment effect of the generated samples, represents the average treatment effect of the real sample, and n represents the number of samples.

[0012] Furthermore, in step S3, converting the variables in the causal structure diagram into coding features in a unified format includes: Representing the diagnosis, operation, and medication as categorical feature representation and text feature representation, respectively, and encoding the text feature representation through a Transformer encoder to obtain corresponding encoding features; Among them, for categorical feature representation, if the patient has a diagnosis, operation or medication record, the corresponding categorical feature representation is recorded as 1, otherwise it is recorded as 0; For text feature representation, we used Clinical BERT to segment the diagnosis and operation text data to obtain the corresponding text feature representation. We also used the dual-view molecular pre-training model to segment the SMILES strings for the medication text data to obtain the corresponding text feature representation. During encoding through the Transformer encoder, the embedding information is determined based on the causal path in the causal structure graph, where the text feature representation of the diagnosis is used as the intervention variable, mortality is used as the outcome variable, and operations and medications are located in the path from diagnosis to structure.

[0013] Furthermore, the step S4 includes: Processing the diagnostic coding features through a first multi-layer perceptron to obtain a first output; The diagnosis coding feature is concatenated with the medication coding feature and the operation coding feature on its causal path, and processed by a second multi-layer perceptron to obtain a second output; The diagnostic coding feature is concatenated with the operational coding feature on its causal path, and processed by a third multi-layer perceptron to obtain a third output; Processing the demographic information through a fourth multi-layer perceptron to obtain a fourth output; Processing the ICU observation indicators through a fifth multi-layer perceptron to obtain a fifth output; The first to fifth outputs are spliced ​​together according to the causal structure diagram to form a multimodal patient information representation.

[0014] Furthermore, in step S5, the death risk prediction model includes a multi-head self-attention layer, a fully connected layer, a GEGLU activation function, a fully connected layer, a linear layer and a Sigmoid activation layer connected in sequence.

[0015] Furthermore, the loss function of the death risk prediction model is for: Where, represents the total loss, represents the hyperparameter weight, represents the sigmoid function, y represents the true label, and y' represents the predicted probability. Indicates treatment status.

[0016] The beneficial effects of the present invention are: (1) The method of the present invention effectively integrates stable causal relationships and treatment effect information in the process of mortality risk prediction by integrating the causal structure diagram, showing stronger stability under different environments, and solving the problem that traditional deep learning models have difficulty in capturing causal effects, resulting in poor generalization ability of prediction results between different data distributions; further, clinicians and medical practitioners can use domain expertise to modify the causal structure diagram to better guide model learning.

[0017] (2) Unlike the traditional method of establishing causal relationships based on a single feature, the present invention constructs causal relationships based on feature categories in the process of constructing a causal structure diagram, reducing the dependence on specific features in the category, making it easier to adapt to different environments. This generalization ability greatly improves the practicality of the model in actual medical applications.

[0018] (3) The clinical data used in the method of the present invention come from various sources, including structured tabular data (such as diagnostic codes, surgical procedures, medication records, vital signs, etc.) and a large amount of unstructured text data (such as descriptions of drug molecular structures, etc.). In view of the complexity of multimodal data, the present invention adopts a Transformer-based encoding framework to effectively integrate tabular category features with text sequence features, realize deep interaction and semantic integration of different modal information, and enhance the model's comprehensive understanding of the patient's status and prediction of mortality risk.

[0019] (4) The method of the present invention introduces causal effect estimation ATE to promote the model's understanding and utilization of causal relationships, thereby providing support for clinical personalized treatment decisions.

[0020] (5) To address the problem of uneven distribution of original data samples, the method of the present invention proposes to enhance minority class data through a causal-driven counterfactual data generation method, introduces a causal structure graph in the data enhancement process, and uses the BKA algorithm to optimize intervention feature sampling, adjust the upsampling distribution parameters of intervention features, and increase sample diversity; further, optimize the migration of mediating variables to ensure causal path consistency; by jointly optimizing sampling and migration, comprehensively optimizing the generation of intervention features and mediating variables, and retaining high-quality counterfactual samples, the diversity and causal consistency performance are balanced.

[0021] (6) In the process of minority class data enhancement, the method of the present invention adopts a causal reasoning framework based on propensity score matching. The generated samples strictly remain on the real data manifold and focus on matching the intervention variable group, maximizing the generation effect while maintaining interpretability. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 Flowchart of the multimodal ICU patient mortality risk prediction method integrating causal effect estimation provided by the present invention.

[0023] Figure 2 Schematic diagram of the ICU patient mortality risk prediction framework provided by the present invention.

[0024] Figure 3 Schematic diagram of the minority class data enhancement process provided by the present invention. DETAILED DESCRIPTION

[0025] The specific embodiments of the present invention are described below to facilitate understanding of the present invention by those skilled in the art. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations utilizing the concepts of the present invention are protected.

[0026] A multimodal ICU patient mortality risk prediction method integrating causal effect estimation, such as Figure 1-2 As shown, the following steps are included: S1. Collect raw data samples during the patient's treatment process and perform preprocessing; Among them, the original data samples include patients' diagnosis, procedures, medications, demographic information, and ICU observation indicators; S2. Construct a causal structure diagram using diagnosis as the intervention variable, death as the outcome variable, and operation and medication as mediating variables between diagnosis and death; S3. Based on the causal-driven counterfactual data generation method, minority class data enhancement is performed on the diagnosis, operation, and drug in the causal structure graph to obtain a data-balanced causal structure graph. The variables in the causal structure graph are converted into encoding features in a unified format through the Transformer encoder. S4. Concatenate the coded features in a unified format, and process the concatenated features, demographic information, and ICU observation indicators through a multi-layer perceptron. Concatenate the outputs of the multi-layer perceptron according to a causal structure diagram to form a multimodal patient information representation. S5. Input the multimodal patient information representation into the mortality risk prediction model and integrate the causal effect ATE estimation to output the final mortality risk prediction probability.

[0027] In the embodiment of the present invention, based on the implementation process of the above steps S1 to S5, the following is constructed: Figure 2 The multimodal ICU patient mortality risk prediction framework (Causality-Oriented prediction with Temporal and Treatment-Aware Augmentation, CoTTA) is shown.

[0028] In step S1 of the embodiment of the present invention, for the collected original data samples, diagnosis refers to the doctor's disease identification information of the patient, including anemia, chronic kidney disease and hypotension, etc., operation refers to the intervention or surgery performed on the patient during the ICU, medication refers to the drug treatment of the patient after discharge from the ICU, demographic information includes the patient's gender, age and admission type, etc., and ICU observation indicators are the patient's vital signs and laboratory tests within 48 hours before admission, including the maximum and minimum values ​​of indicators such as anion gap, albumin, body temperature and calcium concentration.

[0029] In step S1 of the embodiment of the present invention, the preprocessing of the above-mentioned raw data samples includes outlier detection and correction of the raw data samples, maximum-minimum normalization of continuous data, label encoding of data types, and processing of missing data values; specifically, outlier detection and correction mainly involves identifying abnormal entry errors based on medical common sense and implementing corrective measures; for continuous data, sklearn.preprocessing.MinMaxScaler provided by scikit-learn is used for minimum-maximum normalization; for data types, sklearn.preprocessing.LabelEncoder provided by scikit-learn is used for label encoding; in missing value processing, time series data is forward filled using the last observation value, and other missing values ​​are interpolated using the median.

[0030] In step S2 of the embodiment of the present invention, based on the inter-group causal relationship and intra-group causal relationship of the diagnosis, death, operation and medication variables, diagnosis is used as an intervention variable, death is used as an outcome variable, and operation and medication are used as mediating variables between diagnosis and death to construct a causal structure diagram; Among them, for the inter-group causal relationship of variables, the inter-group causal relationship that conforms to medical common sense is obtained based on the entity medical relationship between the variables; for the intra-group causal relationship of variables, the causal discovery method based on reinforcement learning is used to explore the relationship between the features within the variables, combined with the causal discovery paradigm based on variable order and the pre-training method, the intra-group causal relationship is obtained by searching the space of variable order and collective variable selection.

[0031] In the causal structure diagram of this embodiment, the diagnosis variable group serves as the intervention variable group (Treatment, T), the death outcome serves as the outcome variable (Y), and the operation and medication serve as the mediating variables (Mediator, M) between the diagnosis and the outcome.

[0032] In step S3 of this embodiment of the present invention, to address the imbalanced distribution of raw data samples, this embodiment proposes a treatment-aware augmentation (TAA) method that integrates a causal graph structure, propensity score matching, and the Black Kite Algorithm (BKA) to generate causally consistent, representative, and diverse counterfactual instances for minority class samples, thereby alleviating the class imbalance problem in clinical event prediction. This method uses causal graph modeling to clarify the causal paths between variables, combines propensity score estimation to achieve sample matching, and introduces BKA as a core optimization mechanism to guide the joint optimization of intervention variable sampling and mediating variable migration strategies. BKA simulates the hunting behavior of black kites, balancing the three goals of causal consistency, diversity, and representativeness in the process of exploring the globally optimal counterfactual samples, thereby improving the quality of generated samples and the performance of the prediction model on minority classes.

[0033] Specifically, in step S3 of this embodiment, Figure 3 As shown in Figure 2, the methods for enhancing minority class data include: S31. Dataset that will form the causal structure diagram Divide into majority class subsets and minority class subset ; in, represents the feature vector of the i-th sample, represents the label of the i-th sample, Indicates the total number of samples; in the feature vector In the model, the diagnosis variable group serves as the intervention variable group T, death serves as the outcome variable, and operation and medication serve as the mediating variables M between diagnosis and death; Represents the intervention variable group of the i-th sample; in the causal structure diagram, it contains the causal link from the intervention variable group T (for the intervention variable group of the label) to the label Y; S32. For minority class subsets For each intervention variable group in the majority class subset Use the nearest neighbor 1-NN to match and find the majority class sample with the closest propensity score , forming a counterfactual sample pair, which will not be matched successfully to the majority class samples as a counterfactual generation target; Among them, the propensity score is used to measure the probability of an individual having the target label under the condition of controlling the covariate B (demographic information). The propensity score is estimated using Logistic regression. The propensity score of the i-th sample is The calculation formula is: Where, represents the intervention variable of the i-th sample, represents the control covariate of the i-th sample.

[0034] The nearest neighbor 1-NN calculation formula is as follows: For each intervention variable group in each minority class sample, a nearest majority class sample is found, which is counted as a successful pairing. The successful match constitutes a counterfactual sample pair, and the majority class samples that fail to match will enter the next stage as the target of counterfactual generation; S33, from the current minority class subset In the BKA algorithm, the intervention variable group is sampled and a new sample is generated. ; S34, based on the causal path in the causal structure diagram, the generated new sample Most of the samples that are not successfully paired Use the nearest neighbor 1-NN to match, and select the majority class samples that are successfully matched Migrate the mediating variable M in the sample, generate new minority class counterfactual instances, and copy the migrated mediating variable to the new sample In the final counterfactual minority class sample, construct ; S35. Based on the constructed counterfactual minority samples and majority samples, a data-balanced causal structure diagram is formed.

[0035] In step S33 of this embodiment, the intervention feature sampling is optimized through the Black Kite Optimization Algorithm (BKA), and the upsampling distribution parameters of the intervention variable group T are adjusted to increase sample diversity. The mediating variable migration is optimized to ensure causal path consistency. The sampling and migration are jointly optimized to comprehensively optimize the generation of the intervention variable group T and the mediating variable M, as well as the high-quality counterfactual sample retention strategy, to balance diversity and causal consistency performance.

[0036] The BKA algorithm is a nature-inspired swarm intelligence optimization algorithm that imitates the hunting behavior of black kites and has the following core mechanisms: initialization, generating a set of candidate solutions to represent optimization parameters; fitness evaluation, evaluated through fitness function; global search, exploring the solution space through random steps, simulating the wide-area search of black kites, and looking for potential optimal solutions; local search, fine-tuning near high-fitness solutions, simulating precise hunting, and optimizing local solutions; group collaboration, referring to the optimal solution and local excellent solution in the population, updating the direction of candidate solutions, accelerating convergence, and dynamic adjustment: balancing global and local searches according to the iterative progress; termination condition, reaching the maximum number of iterations or fitness convergence, and outputting the optimal solution.

[0037] Based on this, step S33 of this embodiment includes the following sub-steps: S33-1. Use kernel density estimation to estimate the probability density of the intervention variable group and then generate the initial population; S33-2, using the BKA algorithm to globally optimize the bandwidth and sampling distribution weight of the estimated probability density; S33-3. Based on the optimized bandwidth and sampling distribution weight, after repeated sampling for each intervention variable group several times, the confidence interval of the mean and standard deviation of the intervention characteristic in the intervention variable group is calculated; S33-4. Use the confidence interval as the parameter search boundary of the BKA algorithm, and use the mutation strategy of the particle swarm optimization algorithm to perform global search and local search until the set number of iterations is reached and the multi-objective fitness function of the BKA algorithm converges.

[0038] In the above process, the BKA algorithm adopts a multi-objective optimization strategy and decomposes the optimization goal into the following sub-goals: (1) diversity goal, which calculates the T feature variance of the generated new samples to ensure that a wide range of minority class distribution is covered; (2) representative goal, which minimizes the propensity score matching difference between the new samples and the real unmatched majority class samples; (3) causal consistency goal, which retains high-quality samples and evaluates whether the new samples conform to the path of T → M → Y based on the causal graph, and calculates the causal effect ATE of the new samples and difference.

[0039] In this embodiment, a weighted optimal solution is used to evaluate the quality of generated samples. The goal is to maximize this value to optimize the diversity, representativeness, and causal consistency of the generated samples. Therefore, the multi-objective fitness function of the BKA algorithm is: Where, represents the weight parameter, represents the diversity index, Represents representative indicators, represents the causal consistency index, represents the eigenvalue of the generated sample, represents the feature mean, represents the inverse probability weight of each sample calculated from the propensity score estimate, represents the propensity score for generating samples, represents the propensity score of the true unmatched majority class sample, represents the average treatment effect of the generated samples, represents the average treatment effect of the real sample, and n represents the number of samples.

[0040] in, It is a diversity indicator (T feature variance), which represents the diversity of samples after standardization. The goal is to maximize this value to cover the minority class distribution; Representative indicators (propensity score distance), where The inverse probability weight of each sample is calculated based on the propensity score: , the inverse probability weight as sample weight enhances the sample representativeness; The causal consistency index represents the difference between the average treatment effect (ATE) of the newly generated sample and the ATE of the real sample (the smaller the difference, the more consistent it is with the original causal structure).

[0041] In step S3 of the embodiment of the present invention, converting the variables in the causal structure diagram into coding features in a unified format includes: Representing the diagnosis, operation, and medication as categorical feature representation and text feature representation, respectively, and encoding the text feature representation through a Transformer encoder to obtain corresponding encoding features; Among them, for categorical feature representation, if the patient has a diagnosis, operation or medication record, the corresponding categorical feature representation is recorded as 1, otherwise it is recorded as 0; For text feature representation, we used Clinical BERT to segment the diagnosis and operation text data to obtain the corresponding text feature representation. We also used the dual-view molecular pre-training model to segment the SMILES strings for the medication text data to obtain the corresponding text feature representation. During encoding through the Transformer encoder, the embedded information is determined according to the causal path in the causal structure diagram, where the text feature representation corresponding to the diagnosis is used as the intervention variable, the mortality rate is used as the outcome variable, and the operation and medication are located in the path from diagnosis to structure. Specifically, the initial representation of each group of text is obtained through the Transformer encoder, and the mediating variables are hierarchically processed along the causal path. The operation variable group and the medication variable group are spliced ​​and fused with their respective ancestor nodes, and finally the fused mediating variables are used as the encoding features.

[0042] In step S4 of the embodiment of the present invention, based on the coding features obtained by the above-mentioned Transformer encoder, including diagnosis coding features, operation coding features, and medication coding features, the operation coding features and medication coding features are respectively spliced ​​with the nodes on their causal paths and processed by independent multi-layer perceptrons; at the same time, demographic information and ICU observation indicators are also processed by the multi-layer perceptron, and the outputs obtained by the multi-layer perceptron processing are spliced ​​as the final patient information representation. Specifically, step S4 of this embodiment includes: Processing the diagnostic coding features through a first multi-layer perceptron to obtain a first output; The diagnosis coding feature is concatenated with the medication coding feature and the operation coding feature on its causal path, and processed by a second multi-layer perceptron to obtain a second output; The diagnostic coding feature is concatenated with the operational coding feature on its causal path, and processed by a third multi-layer perceptron to obtain a third output; Processing the demographic information through a fourth multi-layer perceptron to obtain a fourth output; Processing the ICU observation indicators through a fifth multi-layer perceptron to obtain a fifth output; The first to fifth outputs are spliced ​​together according to the causal structure diagram to form a multimodal patient information representation.

[0043] In step S5 of the embodiment of the present invention, as Figure 2 As shown in Figure 1, the death risk prediction model includes a multi-head self-attention layer, a fully connected layer, a GEGLU activation function, a fully connected layer, a linear layer, and a Sigmoid activation layer connected in sequence; Specifically, the input sequence x representing multimodal patient information first passes through a multi-head self-attention mechanism, mapping it into tensors consisting of query Q, key K, and value V. The self-attention layer is followed by a feedforward neural network (FFN) consisting of two fully connected layers. The first layer maps the input to a higher-dimensional intermediate space. This layer then performs a nonlinear transformation using a gated linear unit (GLU) activation function (GEGLU). The output then passes through a series of linear layers with ReLU activation and dropout regularization, ultimately achieving a binary classification task and predicting the patient's mortality risk.

[0044] In this example, the average treatment effect (ATE) is used to evaluate the overall average impact of the intervention on the entire population for the entire mortality risk prediction process (CoTTA). Each patient has two potential outcomes: survival or death. If the outcomes of both states of the same patient can be observed simultaneously, causal inference is more direct, but this is not possible in reality. During training, the predictions are divided into positive and negative samples based on the true labels. The ATE of the positive sample is estimated to be , where d represents the treatment status. CoTTA introduces ATE into the loss function. Previous research has explored incorporating causal inference into deep learning loss design, improving model performance by penalizing the difference between the treatment group and the control group.

[0045] In this embodiment, binary cross entropy loss and ATE loss are used to jointly constrain the model to accurately predict the difference in results between counterfactual groups. This strategy gives greater weight to samples whose true results are close to the decision boundary and whose predicted probabilities differ significantly.

[0046] Based on this, the loss function of the death risk prediction model in the embodiment of the present invention is for: Where, represents the total loss, represents the hyperparameter weight, represents the sigmoid function, y represents the true label, and y' represents the predicted probability. Represents the treatment status. By adjusting α on the validation set, we can balance the mortality prediction accuracy and the ability to capture causal relationships, thereby improving the overall performance of the model.

[0047] In an embodiment of the present invention, an experimental test example of the above-mentioned death risk prediction method on an actual medical public dataset is provided.

[0048] In this example, the MIMIC-III (version 1.4) dataset (ICU, a widely recognized benchmark dataset in the field of mortality prediction) was used. MIMIC-III contains desensitized clinical data of patients in the emergency department and ICU of Beth Israel Deaconess Medical Center in Boston, USA, as well as the homologous external validation MIMIC-IV (version 3.1) dataset and an elderly patient dataset from West China Hospital of Sichuan University (WCHSU) in China for validation.

[0049] Based on the above dataset, the method of the present invention (CoTTA) was compared with the existing models that can predict mortality risk. The comparison results are shown in Table 1. Table 1: Comparison of model prediction results In Table 1, SAPS-II is a scoring system designed to predict the mortality of critically ill patients. It takes into account 12 physiological characteristics within the first 24 hours after ICU admission. The total score is calculated by summing the scores of each feature, and then the total score is converted into the probability of death; GRU-D (deep learning model based on gated recurrent unit) is a deep learning model that uses recurrent neural networks to analyze multivariate time series data; IPNET (interpolation-prediction network) is a new deep learning architecture based on semi-parametric interpolation network; MC (multi-task channel-level LSTM) is an enhancement to the LSTM network, which uses a single neural model at the channel-level LSTM to simultaneously predict multiple tasks; MTRNN is a multi-task recurrent neural network. RNN, combined with attention mechanism, is specifically used to predict hospital mortality; HM-AS is a deep learning model using natural language processing (NLP) technology, used to predict in-hospital mortality; MultiModal-1DCNN is a deep neural network architecture that uses recurrent neural networks to process time series data and convolutional neural networks to analyze clinical text; Vital+EntityEmb is a multimodal neural network that jointly trains time series signals and unstructured clinical text representations to predict the risk of in-hospital mortality in ICU patients; DECAF is a general deep cascade framework for predicting the potential risks of all physiological functions in each clinical stage; GAN (c-med GAN) is a variant of conditional medical generative adversarial network (c-med GAN) used to predict the mortality of ICU patients; MMDL (multimodal deep learning) is an integration of multimodal deep learning models consisting of feedforward neural networks and gated recurrent unit networks.

[0050] As shown in Table 1, the proposed CoTTA framework demonstrates transferability under a given causal structure, demonstrating comparable performance across all three datasets, compared to traditional scoring methods (SAPS-II), time-series-based prediction models (GRU-D, IPNET, MC, MTRNN), text-based prediction models (IHM-AS), and multimodal data-based prediction models (MultiModal-1DCNN, Viral+EntityEmb, DECAF, c-medGAN, MMDL). By integrating causal structures, CoTTA demonstrates enhanced stability across diverse environments. Clinicians and healthcare practitioners can leverage their domain expertise to modify the causal structure graph to better guide model learning. Unlike traditional methods that establish causal relationships based on individual features, the proposed method constructs relationships based on feature categories, reducing reliance on specific features within categories and making it more adaptable to diverse environments. This generalization significantly enhances the model's practicality in real-world medical applications, demonstrating the potential of causal machine learning in clinical settings.

[0051] In this embodiment, for the data enhancement part of the present invention, multiple groups of control experiments are provided in this embodiment to demonstrate the advantages of the data enhancement method of the present invention in the evaluation of category imbalanced datasets.

[0052] In this embodiment, two core indicators are used: AUPRC (area under the precision-recall curve) and MCC (Matthews correlation coefficient). AUPRC measures the model's ability to distinguish different categories by comprehensively considering precision and recall, and is particularly sensitive to the model's recognition efficiency for minority samples. MCC provides a balanced assessment of classification accuracy by integrating four types of discrimination results: true positives, true negatives, false positives, and false negatives.

[0053] The comparative experimental results are shown in Table 2; Table 2: Comparison results of different data balancing methods on three datasets Experimental results show that compared with widely used interpolation-based methods, the proposed method exhibits superior performance on all three datasets. The proposed method improves on the following aspects: 1. Similarity: Maintaining a certain degree of similarity with the target instance to ensure similarity of counterfactual pairs; 2. Sparsity: Minimizing feature differences compared to the target sample, replacing feature difference matching with matching of the intervention variable group facilitates intuitive interpretation; 3. Rationality: The modified features come from the intervention variable group, and their values ​​are derived from the true values; 4. Accessibility: The proposed method can cover most instances of the majority class, ensuring the wide applicability of the interpretation; 5. Diversity: The proposed method can present multi-angle interpretations through different feature combinations, or provide a choice of multiple sets of alternative solutions involving different feature differences.

[0054] Specific embodiments are used in the present invention to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core ideas. At the same time, for those skilled in the art, according to the ideas of the present invention, there may be changes in the specific implementation methods and application scopes. In summary, the contents of this specification should not be understood as limiting the present invention.

[0055] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and it should be understood that the scope of protection of the present invention is not limited to such specific descriptions and embodiments. Those skilled in the art can make various other specific variations and combinations based on the technical teachings disclosed in the present invention without departing from the essence of the present invention, and such variations and combinations are still within the scope of protection of the present invention.

Claims

1. A multimodal ICU patient mortality risk prediction method integrating causal effect estimation, characterized by: The following steps are involved: S1. Collect raw data samples during the patient's treatment process and perform preprocessing; Among them, the original data samples include patients' diagnosis, procedures, medications, demographic information, and ICU observation indicators; S2. Construct a causal structure diagram using diagnosis as the intervention variable, death as the outcome variable, and operation and medication as mediating variables between diagnosis and death; S3. Based on the causal-driven counterfactual data generation method, minority class data enhancement is performed on the diagnosis, operation, and drug in the causal structure graph to obtain a data-balanced causal structure graph. The variables in the causal structure graph are converted into encoding features in a unified format through the Transformer encoder. S4. Concatenate the coded features in a unified format, and process the concatenated features, demographic information, and ICU observation indicators through a multi-layer perceptron. Concatenate the outputs of the multi-layer perceptron according to a causal structure diagram to form a multimodal patient information representation. S5. Input the multimodal patient information representation into the mortality risk prediction model and integrate the causal effect ATE estimation to output the final mortality risk prediction probability.

2. The multimodal ICU patient mortality risk prediction method integrating causal effect estimation according to claim 1 is characterized in that: The preprocessing in step S1 includes detecting and correcting outliers in the original data samples, performing maximum-minimum normalization on continuous data, label encoding on data types, and processing missing data values.

3. The multimodal ICU patient mortality risk prediction method integrating causal effect estimation according to claim 1 is characterized in that: In step S2, based on the inter-group causal relationship and intra-group causal relationship of the diagnosis, death, operation and medication variables, diagnosis is used as the intervention variable, death is used as the outcome variable, and operation and medication are used as mediating variables between diagnosis and death to construct a causal structure diagram; Among them, for the inter-group causal relationship of variables, the inter-group causal relationship that conforms to medical common sense is obtained based on the entity medical relationship between the variables; for the intra-group causal relationship of variables, the causal discovery method based on reinforcement learning is used to explore the relationship between the features within the variables, combined with the causal discovery paradigm based on variable order and the pre-training method, the intra-group causal relationship is obtained by searching the space of variable order and collective variable selection.

4. The multimodal ICU patient mortality risk prediction method integrating causal effect estimation according to claim 1 is characterized in that: In step S3, the method for enhancing minority class data includes: S31. Dataset that will form the causal structure diagram Divide into majority class subsets and minority class subset ; in, represents the feature vector of the i-th sample, represents the label of the i-th sample, Indicates the total number of samples; in the feature vector In the model, the diagnosis variable group serves as the intervention variable group T, death serves as the outcome variable, and operation and medication serve as the mediating variables M between diagnosis and death; represents the intervention variable group of the i-th sample; S32. For minority class subsets For each intervention variable group in the majority class subset Use the nearest neighbor 1-NN to match and find the majority class sample with the closest propensity score , forming a counterfactual sample pair, which will not be matched successfully to the majority class samples as a counterfactual generation target; S33, from the current minority class subset In the BKA algorithm, the intervention variable group is sampled and a new sample is generated. ; S34, based on the causal path in the causal structure diagram, the generated new sample Most of the samples that are not successfully paired Use the nearest neighbor 1-NN to match, and select the majority class samples that are successfully matched Migrate the mediating variable M in the sample, generate new minority class counterfactual instances, and copy the migrated mediating variable to the new sample In the final counterfactual minority class sample, construct ; S35. Based on the constructed counterfactual minority samples and majority samples, a data-balanced causal structure diagram is formed.

5. The multimodal ICU patient mortality risk prediction method integrating causal effect estimation according to claim 4 is characterized in that: The step S33 includes the following sub-steps: S33-1. Use kernel density estimation to estimate the probability density of the intervention variable group and then generate the initial population; S33-2, using the BKA algorithm to globally optimize the bandwidth and sampling distribution weight of the estimated probability density; S33-3. Based on the optimized bandwidth and sampling distribution weight, after repeated sampling for each intervention variable group several times, the confidence interval of the mean and standard deviation of the intervention characteristic in the intervention variable group is calculated; S33-4. Use the confidence interval as the parameter search boundary of the BKA algorithm, and use the mutation strategy of the particle swarm optimization algorithm to perform global search and local search until the set number of iterations is reached and the multi-objective fitness function of the BKA algorithm converges.

6. The multimodal ICU patient mortality risk prediction method integrating causal effect estimation according to claim 5 is characterized in that: The multi-objective fitness function of the BKA algorithm is: Where, represents the weight parameter, represents the diversity index, Represents representative indicators, represents the causal consistency index, represents the eigenvalue of the generated sample, represents the feature mean, represents the inverse probability weight of each sample calculated from the propensity score estimate, represents the propensity score for generating samples, represents the propensity score of the true unmatched majority class sample, represents the average treatment effect of the generated samples, represents the average treatment effect of the real sample, and n represents the number of samples.

7. The multimodal ICU patient mortality risk prediction method integrating causal effect estimation according to claim 1 is characterized in that: In step S3, converting the variables in the causal structure diagram into coding features in a unified format includes: Representing the diagnosis, operation, and medication as categorical feature representation and text feature representation, respectively, and encoding the text feature representation through a Transformer encoder to obtain corresponding encoding features; Among them, for categorical feature representation, if the patient has a diagnosis, operation or medication record, the corresponding categorical feature representation is recorded as 1, otherwise it is recorded as 0; For text feature representation, we used Clinical BERT to segment the diagnosis and operation text data to obtain the corresponding text feature representation. We also used the dual-view molecular pre-training model to segment the SMILES strings for the medication text data to obtain the corresponding text feature representation. During encoding through the Transformer encoder, the embedding information is determined based on the causal path in the causal structure graph, where the text feature representation of the diagnosis is used as the intervention variable, mortality is used as the outcome variable, and operations and medications are located in the path from diagnosis to structure.

8. The multimodal ICU patient mortality risk prediction method integrating causal effect estimation according to claim 7 is characterized in that: The step S4 comprises: Processing the diagnostic coding features through a first multi-layer perceptron to obtain a first output; The diagnosis coding feature is concatenated with the medication coding feature and the operation coding feature on its causal path, and processed by a second multi-layer perceptron to obtain a second output; The diagnostic coding feature is concatenated with the operational coding feature on its causal path, and processed by a third multi-layer perceptron to obtain a third output; Processing the demographic information through a fourth multi-layer perceptron to obtain a fourth output; Processing the ICU observation indicators through a fifth multi-layer perceptron to obtain a fifth output; The first to fifth outputs are spliced ​​together according to the causal structure diagram to form a multimodal patient information representation.

9. The multimodal ICU patient mortality risk prediction method integrating causal effect estimation according to claim 1 is characterized in that: In step S5, the death risk prediction model includes a multi-head self-attention layer, a fully connected layer, a GEGLU activation function, a fully connected layer, a linear layer and a Sigmoid activation layer connected in sequence.

10. The multimodal ICU patient mortality risk prediction method integrating causal effect estimation according to claim 1 is characterized in that: The loss function of the death risk prediction model for: Where, represents the total loss, represents the hyperparameter weight, represents the sigmoid function, y represents the true label, and y' represents the predicted probability. Indicates treatment status.

Images