Medical adverse event risk prediction, prevention and control method based on causal inference

Abstract

The invention discloses a medical adverse event risk prediction, prevention and control method based on causal inference. The method comprises the following steps: S1, collecting and preprocessing multi-modal medical data; s2, constructing a causal network through a time slicing technology; s3, introducing ethical factors to form an ethical causal network; s4, generating a causal enhancement vector; s5, constructing a dynamic risk prediction model, and outputting an evaluation result; s6, analyzing a path contribution degree, and screening high-risk paths and nodes; s7, generating a personalized prevention and control scheme; and S8, generating a transparent report, and displaying the transparent report to the medical personnel and the patient. According to the method, the dynamic causal relationship among the medical variables is accurately modeled through the dynamic causal network, the multi-modal data fusion and the ethical correction technology, the personalized prevention and control scheme is generated in combination with the optimization algorithm, comprehensive support is provided for medical decision making, and implementation of precise medical treatment is promoted.

Description

Technical Field

[0001] The present invention relates to the field of medical technology, and in particular to a method for predicting and preventing medical adverse event risks based on causal inference. Background Art

[0002] In the field of modern medicine, the prediction and prevention of adverse medical events have always been an important topic in clinical practice and medical management. With the continuous accumulation of medical data and the rapid development of artificial intelligence technology, methods based on big data and artificial intelligence have been widely used in medical risk assessment. However, in the existing technology, most methods mainly rely on traditional machine learning or deep learning models. Although these models can mine potential patterns in the data, they lack causal inference capabilities and are difficult to accurately reveal the causal relationship behind adverse medical events. This limitation makes the model often lack of explanatory power when facing complex medical scenarios, making it difficult to apply the prediction results to clinical decision-making, especially when dealing with multimodal data and dynamically changing risk factors.

[0003] Traditional medical risk prediction technologies usually rely on static statistical methods or experience-based rule engines. These methods are highly dependent on data and have limited ability to handle complex causal relationships. For example, although regression models and Bayesian networks can analyze the dependencies between variables to a certain extent, most of them assume that the relationships between variables are linear or fixed. This assumption is often not true in medical scenarios because medical data often has highly nonlinear and dynamically changing characteristics. In addition, traditional methods find it difficult to fully utilize multimodal data, such as patients' electronic medical records (EMRs), monitoring data, laboratory test results, and medical images, which further limits the accuracy and comprehensiveness of risk prediction.

[0004] Although existing deep learning methods can automatically extract features from complex data, they also face many challenges. On the one hand, deep learning models are data-driven, and their optimization goal is often to improve the accuracy of predictions rather than to explore causal relationships, so the results of the model lack interpretability. On the other hand, deep learning models lack a systematic framework when dealing with temporal dynamics and multimodal data fusion, such as the changing patterns of time series data and the interactive relationships between different data modalities. In addition, existing deep learning technologies find it difficult to fully consider ethical factors when predicting adverse medical events, such as the fairness of resource allocation and the fragility of patients. This neglect may cause the prediction model to deviate from medical ethical principles in practical applications, further limiting its promotion in highly sensitive scenarios.

[0005] Dynamic causal inference technology has attracted attention in the medical field in recent years. It can model the dynamic influence between medical variables with causal relationships as the core. However, existing causal inference methods still face many limitations in practice. For example, when facing changes in causal relationships in the time dimension, existing causal inference methods find it difficult to build a flexible dynamic causal network. At the same time, most causal inference technologies rely on single-modal data, making it difficult to uniformly model multimodal medical data. In addition, causal inference technology also shows obvious deficiencies in incorporating ethical constraints. For example, how to quantify the impact of ethical factors on causal paths and how to balance ethical corrections and predictive performance have not been effectively solved in existing technologies.

[0006] Intervention strategy optimization methods based on reinforcement learning or optimization algorithms have been applied in many fields, but there are still applicability issues in medical scenarios. Existing intervention strategy optimization methods mainly focus on short-term risk reduction and lack a global perspective on long-term intervention effects. At the same time, these methods often ignore personalized needs in the design of intervention strategies, making it difficult to tailor intervention plans according to the specific circumstances of patients. In addition, most existing optimization algorithms fail to fully utilize the information of dynamic causal networks, nor do they introduce ethical revisions to constrain strategy optimization, which leads to deficiencies in fairness and interpretability of optimization results.

[0007] Therefore, how to provide a method for risk prediction and prevention of adverse medical events based on causal inference is an urgent problem that technicians in this field need to solve. Summary of the invention

[0008] One purpose of the present invention is to propose a method for risk prediction and prevention of adverse medical events based on causal inference. The present invention makes full use of dynamic causal inference technology, multimodal data fusion technology and particle swarm optimization algorithm to construct a causal network model with ethical correction function, and describes in detail the implementation steps in high-risk path screening, key node annotation and intervention strategy optimization. The present invention has the following advantages: the accuracy of risk prediction and the interpretability of causal relationships are improved through dynamic causal networks; multimodal data fusion comprehensively integrates information such as patient electronic medical records, monitoring data and laboratory test results; the fairness of prediction and intervention strategies is ensured by introducing ethical correction factors; and personalized and long-term precision medical prevention and control is achieved through particle swarm optimization algorithm and reinforcement learning dynamic optimization of intervention strategies. The present invention has significant safety, reliability and scalability in the field of adverse medical event prediction and prevention and control.

[0009] A method for predicting and preventing adverse medical event risks based on causal inference according to an embodiment of the present invention comprises the following steps:

[0010] S1. Collect multimodal medical data, preprocess the data, and generate cleaned multimodal medical data;

[0011] S2. Based on the cleaned multimodal medical data, the time dimension is divided through time slicing technology to construct a causal network;

[0012] S3. Based on the constructed causal network, define the ethical factor indicators, construct the ethical weight matrix, quantify the weight influence of ethical factors on the causal path, and embed the causal network to form an ethical causal network;

[0013] S4. Based on the ethical causal network and cleaned multimodal medical data, a causal feature embedding module is designed, a graph neural network is used to process node features, and multimodal features are integrated to generate a causal enhancement vector;

[0014] S5. Use causal enhancement vectors to build a dynamic risk prediction model, combine the time dimension with recursive neural networks, capture the dynamic pattern of adverse events, and output risk assessment results;

[0015] S6. Based on the risk assessment results and ethical causal network, analyze the contribution of causal paths, screen high-risk paths and key nodes based on ethical weights, and generate intervention recommendations;

[0016] S7. Use the screened high-risk pathways and key nodes, combined with risk assessment results, to dynamically adjust intervention recommendations and generate personalized prevention and control plans;

[0017] S8. Generate dynamic transparency reports based on personalized prevention and control plans and show them to medical staff and patients.

[0018] Optionally, the S2 specifically includes:

[0019] S21. Based on the cleaned multimodal medical data, the data is divided into time dimensions using time slicing technology, and each time slice T is defined. i It is a subset of medical data within a fixed time interval;

[0020] S22, in each time slice T i In the above example, we extract the causal variable X j , where X j Represents potential causal factors that affect adverse medical events, including patient characteristics, treatment methods, and external environmental variables;

[0021] S23, based on each time slice T i and the extracted causal variable X j , construct the time-slice subnetwork of the causal network, and calculate the conditional causal influence between variables through the extended causal probability formula:

[0022]

[0023] in, For variables In time slice T i The probability distribution in for right The conditional probability of For variables The parent node set of , m is the total number of causal variables, k is the causal variable index; For variables The conditional probability at time t is, is the causal variable at time t, serving as the “antecedent” in the causal relationship, is the causal variable at time t, serving as the “consequence” in the causal relationship;

[0024] S24. Use a dynamic Bayesian network to connect each time slice sub-network to form a causal network including the time dimension:

[0025]

[0026] Among them, P(X t |X t-1 ) is the state of the causal network at time t is determined by the state at t-1;

[0027] S25. Extract the causal paths and key nodes in the causal network, and mark the paths in the causal paths that have a significant impact on the risk of adverse events.

[0028] Optionally, the S3 specifically includes:

[0029] S31. Define the ethical factor index E based on the constructed causal network k , where E k Including fairness, resource priority and patient vulnerability indicators, quantifying ethical modifications to path weights in causal networks;

[0030] S32. Construct the ethical weight matrix W ethical ,element Indicates that at time t, the variable Relative to variable Ethical impact weight:

[0031]

[0032] Among them, E k is the ethical factor, α k is the weight ratio of ethical factors, p is the total number of ethical factors, k is the index of ethical factors, i and j are the indexes of variables in the causal network;

[0033] S33, based on the ethical weight matrix W ethical, adjust the weights of causal paths in the causal network Added corrections to path adaptation factor and node influence factor:

[0034]

[0035] in, is the ethical correction weight of the path, is the nth adjacent node pair node The influence value of , β is the path adaptation factor, N is the node The number of adjacent nodes of and is the node, n is the node index;

[0036] S34, calculate the path score S(P) by using the path score formula based on multi-dimensional indicators k ), and sort by score, and select the first q highly sensitive paths as the highly ethically sensitive paths P sensitive :

[0037]

[0038] Among them, P k is the kth path, is the ethical sensitivity, T is the total length of the time dimension, t is the index of the time slice, δ t is the time weight;

[0039] S35, based on high ethical sensitivity path P sensitive , generate ethical causal network G ethical , the network contains the node relationships and path weights corrected by ethical factors:

[0040]

[0041] Among them, θ is the path weight threshold.

[0042] Optionally, the S4 specifically includes:

[0043] S41, based on ethical causal network G ethical and cleaned multimodal medical data, extracting causal enhancement features, including node features, path features, and network structure features;

[0044] S42, build a causal feature embedding module to embed the node features in the causal network and the ethical modification weight of the path Mapping to low-dimensional feature space:

[0045]

[0046] Among them, Zembed is the embedded causal enhancement feature vector, X is the node feature matrix, is the relationship function between node pairs, σ is the activation function, W is the trainable parameter matrix, i and j are the indices of the variables in the causal network, and is a node;

[0047] S43, based on the embedded causal enhancement feature vector Z embed , the graph neural network is used to perform multi-layer updates on the embedded features, and the update rule of the node features is:

[0048]

[0049] in, is the feature vector of node i in the lth layer, with an initial value of Z embed , W (l) is the trainable weight matrix of the lth layer, N(i) is the set of neighbor nodes of node i, is the attention weight of the lth layer, is the path feature aggregation function, is the feature vector of node j in the lth layer, σ is the activation function, is the feature vector of node i in the l+1th layer;

[0050] S44. Updated node features and ethical modification weight Introducing attention mechanism for path weight To optimize:

[0051]

[0052] in, is the node correlation score function, exp is the exponential function, and k is the node index;

[0053] S45, Path Weight Based on Optimization and the updated node feature H (L) , combined with the features F extracted from multimodal medical data modal , generate the causal enhancement vector F enhance :

[0054]

[0055] Among them, V is the node set, i is the index of the node currently being processed, and concat is the concatenation operation. It is the feature vector extracted from multimodal medical data.

[0056] Optionally, the S5 specifically includes:

[0057] S51, based on causal enhancement vector F enhance , construct a dynamic risk prediction model, the input of the dynamic risk prediction model is the causal enhancement vector F enhance and the corresponding target variable Y t , where Y t represents the risk label of adverse events at time t;

[0058] S52. Define the time series state transition rules of the risk prediction model, and build a dynamic risk prediction model based on the recursive neural network. The state transition formula is:

[0059] h t =σ(W h ·h t-1 +W x ·F enhance +b h );

[0060] Among them, h t is the hidden state vector, σ is the activation function, W h and W x are the weight matrices of hidden state and input features, respectively, and b h is the bias term, h t-1 To provide a memory of the risks of the previous moment;

[0061] S53. Combined with hidden state vector h t , by introducing attention mechanism, time-dependent weight and global causal information into the risk prediction in the time dimension to output the risk value

[0062]

[0063] Among them, softmax is the activation function, W y is the output layer weight matrix, b y is the bias term of the output layer, T is the length of the historical time step, and h k is the hidden state at time k, α k is the temporal attention weight, ψ(W global ,tk) is the global causal correction function, W global is the global causal weight feature, the current moment in the t time dimension;

[0064] S54. Introducing multiple maintenance correction mechanisms to improve risk value Adjustments are made to combine ethical corrections with path weights, time dependencies, and characteristics of highly sensitive nodes:

[0065]

[0066] Among them, λ1 ,λ 2 ,λ 3 is the modified weight coefficient, P sensitive A collection of highly ethically sensitive paths. is the optimized path weight, P is the set of highly ethically sensitive paths P sensitive The total number of is the risk assessment result after multiple repairs, i and j are node indexes;

[0067] S55. Output risk assessment results Combined with the causal enhancement vector F enhance Provides causal path impact analysis.

[0068] Optionally, the S6 specifically includes:

[0069] S61. Based on causal network and risk assessment results Calculate the contribution C(P) of each causal path:

[0070]

[0071] in, Node feature The amount of change, is the optimized path weight, P is the causal path, i and j are the node indexes, and is a node;

[0072] S62. Using the ethical weight matrix W ethical , modify the path contribution and screen high-risk paths P risk :

[0073] P risk = {P|C(P)·W ethical (P)≥θ 1};

[0074] Among them, W ethical (P) is the comprehensive ethical weight of path P, θ 1 Screening thresholds for risk;

[0075] S63. For the selected high-risk paths P risk , calculate the path priority R(P), taking into account the path length, ethical sensitivity and node characteristics:

[0076]

[0077] in, For Node To Node The ethical correction path weight, L(P) is the total length of path P, For Node ethical sensitivity, For Node ethical sensitivity;

[0078] S64, in the high-risk path set P risk In the example, we label the key node set N key , the key node marking rules are:

[0079]

[0080] Among them, θ 2 Filter thresholds for node sensitivity;

[0081] S65, based on the priority sorting R(P) and the key node set N key , generate intervention recommendations G intervene , specifically:

[0082] G intervene ={(P,N key )|P∈P risk ,rank(R(P))≤G};

[0083] Where G is the number of high-priority paths set, and rank(R(P)) is the ranking of path P in the priority sorting.

[0084] Optionally, the S7 specifically includes:

[0085] S71. From the high-risk path set P risk and the key node set N key Let’s start by defining the optimization objective function to balance risk reduction and ethical constraints:

[0086]

[0087] in, To optimize the objective function, C(P) is the risk contribution of path P, R(P) is the priority score of path P, For Node ethical sensitivity, α, β are weight coefficients;

[0088] S72. According to the high-risk path set P risk and the key node set N key , define the intervention action set A:

[0089]

[0090] Among them, a i For path P and node The i-th intervention action implemented, ΔC(P,a i ) is the intervention action a i The risk contribution reduction value of path P, θ a is the minimum effect threshold of the intervention action;

[0091] S73. Combine high-risk path set P risk and the key node set N key , by quantifying the risk reduction effect and implementation cost of each intervention action, and evaluating i The combined effect:

[0092]

[0093] Among them, R t is the combined effect of the intervention, γ a is the cost penalty coefficient, w j For key nodes The importance weight of For intervention action a i For Node The impact value of

[0094] S74, using particle swarm optimization algorithm, based on the high-risk path set P risk , intervention action set A and intervention comprehensive effect R t , optimizing the intervention strategy through dynamic adjustment of particle position and velocity:

[0095]

[0096] Among them, π * (S t ) is the optimal intervention strategy, π is the candidate intervention strategy, S t is the state space, ΔC(P,π) is the effect of reducing the contribution of intervention strategy π to path P, γ·Cost(π) is the implementation cost of intervention strategy π, combined with the cost penalty coefficient γ, ω is the inertia weight, v t is the particle velocity, argmax π To find the intervention strategy π that maximizes the optimization objective function * ;

[0097] S75. Combining risk assessment results with optimized intervention strategies * (S t ), generate personalized prevention and control plan G plan :

[0098] G plan ={(P,N key ,a i )|P∈Prisk ,a i =π * (S t )};

[0099] Among them, N key is the key node set, P risk is the set of high-risk paths, a i is the intervention action and P is the path.

[0100] The beneficial effects of the present invention are:

[0101] The present invention has achieved remarkable beneficial effects in the field of adverse medical event prediction and prevention and control by introducing dynamic causal networks and ethical correction factors, combining multimodal data fusion and particle swarm optimization algorithm.

[0102] First, the present invention uses dynamic causal networks to dynamically model complex causal relationships, which can accurately capture the causal paths between medical variables that change over time, providing strong theoretical support and higher interpretability for the risk prediction of adverse medical events.

[0103] Secondly, the present invention uses multimodal data fusion technology to unify modeling and analysis of multi-source medical data such as electronic medical records, monitoring data, and laboratory test results, which greatly improves the comprehensiveness and accuracy of risk prediction and effectively avoids information loss caused by data isolation in existing technologies. At the same time, the present invention introduces ethical factors to quantitatively correct the contribution of causal paths, ensuring the balance of risk prediction and intervention strategies in ethical dimensions such as fairness, patient vulnerability, and resource priority, and solving the application limitations of existing technologies in ethically sensitive scenarios.

[0104] In addition, the present invention dynamically optimizes the intervention strategy by combining the particle swarm optimization algorithm and reinforcement learning, and can formulate personalized prevention and control plans based on high-risk paths and key nodes, which not only improves the accuracy of intervention measures, but also enhances the effect of long-term intervention and the efficiency of resource utilization. In short, while improving the accuracy of medical risk prediction, the present invention takes into account ethical fairness and the operability of intervention strategies, provides innovative solutions for medical decision support and patient safety management, and has wide application value and significant practical significance. BRIEF DESCRIPTION OF THE DRAWINGS

[0105] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0106] Figure 1 A flowchart of a method for predicting and preventing adverse medical event risks based on causal inference proposed by the present invention;

[0107] Figure 2 This is a schematic diagram of the causal network construction and ethical weight embedding of a medical adverse event risk prediction and prevention method based on causal inference proposed in the present invention. DETAILED DESCRIPTION

[0108] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0109] refer to Figure 1-2 , a method for risk prediction and prevention of adverse medical events based on causal inference, comprising the following steps:

[0110] S1. Collect multimodal medical data, preprocess the data, and generate cleaned multimodal medical data;

[0111] S2. Based on the cleaned multimodal medical data, the time dimension is divided through time slicing technology to construct a causal network;

[0112] S3. Based on the constructed causal network, define the ethical factor indicators, construct the ethical weight matrix, quantify the weight influence of ethical factors on the causal path, and embed the causal network to form an ethical causal network;

[0113] S4. Based on the ethical causal network and cleaned multimodal medical data, a causal feature embedding module is designed, a graph neural network is used to process node features, and multimodal features are integrated to generate a causal enhancement vector;

[0114] S5. Use causal enhancement vectors to build a dynamic risk prediction model, combine the time dimension with recursive neural networks, capture the dynamic pattern of adverse events, and output risk assessment results;

[0115] S6. Based on the risk assessment results and ethical causal network, analyze the contribution of causal paths, screen high-risk paths and key nodes based on ethical weights, and generate intervention recommendations;

[0116] S7. Use the screened high-risk pathways and key nodes, combined with risk assessment results, to dynamically adjust intervention recommendations and generate personalized prevention and control plans;

[0117] S8. Generate dynamic transparency reports based on personalized prevention and control plans and show them to medical staff and patients.

[0118] In this implementation, S2 specifically includes:

[0119] S21. Based on the cleaned multimodal medical data, the data is divided into time dimensions using time slicing technology, and each time slice T is defined. iIt is a subset of medical data within a fixed time interval;

[0120] S22, in each time slice T i In the above example, we extract the causal variable X j , where X j Represents potential causal factors that affect adverse medical events, including patient characteristics, treatment methods, and external environmental variables;

[0121] S23, based on each time slice T i and the extracted causal variable X j , construct the time-slice subnetwork of the causal network, and calculate the conditional causal influence between variables through the extended causal probability formula:

[0122]

[0123] in, For variables In time slice T i The probability distribution in for right The conditional probability of For variables The parent node set of , m is the total number of causal variables, k is the causal variable index; For variables The conditional probability at time t is, is the causal variable at time t, serving as the “antecedent” in the causal relationship, is the causal variable at time t, serving as the “consequence” in the causal relationship;

[0124] S24. Use a dynamic Bayesian network to connect each time slice sub-network to form a causal network including the time dimension:

[0125]

[0126] Among them, P(X t |X t-1 ) is the state of the causal network at time t is determined by the state at t-1;

[0127] S25. Extract the causal paths and key nodes in the causal network, and mark the paths in the causal paths that have a significant impact on the risk of adverse events.

[0128] In this implementation, S3 specifically includes:

[0129] S31. Define the ethical factor index E based on the constructed causal network k , where E kIncluding fairness, resource priority and patient vulnerability indicators, quantifying ethical modifications to path weights in causal networks;

[0130] S32. Construct the ethical weight matrix W ethical ,element Indicates that at time t, the variable Relative to variable Ethical impact weight:

[0131]

[0132] Among them, E k is the ethical factor, α k is the weight ratio of ethical factors, p is the total number of ethical factors, k is the index of ethical factors, i and j are the indexes of variables in the causal network;

[0133] S33, based on the ethical weight matrix W ethical , adjust the weights of causal paths in the causal network Added corrections to path adaptation factor and node influence factor:

[0134]

[0135] in, is the ethical correction weight of the path, is the nth adjacent node pair node The influence value of , β is the path adaptation factor, N is the node The number of adjacent nodes of and is the node, n is the node index;

[0136] S34, calculate the path score S(P) by using the path score formula based on multi-dimensional indicators k ), and sort by score, and select the first q highly sensitive paths as the highly ethically sensitive paths P sensitive :

[0137]

[0138] Among them, P k is the kth path, is the ethical sensitivity, T is the total length of the time dimension, t is the index of the time slice, δ t is the time weight;

[0139] S35, based on high ethical sensitivity path P sensitive , generate ethical causal network G ethical , the network contains the node relationships and path weights corrected by ethical factors:

[0140]

[0141] Among them, θ is the path weight threshold.

[0142] In this implementation, S4 specifically includes:

[0143] S41, based on ethical causal network G ethical and cleaned multimodal medical data, extracting causal enhancement features, including node features, path features, and network structure features;

[0144] S42, build a causal feature embedding module to embed the node features in the causal network and the ethical modification weight of the path Mapping to low-dimensional feature space:

[0145]

[0146] Among them, Z embed is the embedded causal enhancement feature vector, X is the node feature matrix, is the relationship function between node pairs, σ is the activation function, W is the trainable parameter matrix, i and j are the indices of the variables in the causal network, and is a node;

[0147] S43, based on the embedded causal enhancement feature vector Z embed , the graph neural network is used to perform multi-layer updates on the embedded features, and the update rule of the node features is:

[0148]

[0149] in, is the feature vector of node i in the lth layer, with an initial value of Z embed , W (l) is the trainable weight matrix of the lth layer, N(i) is the set of neighbor nodes of node i, is the attention weight of the lth layer, is the path feature aggregation function, is the feature vector of node j in the lth layer, σ is the activation function, is the feature vector of node i in the l+1th layer;

[0150] S44. Updated node features and ethical modification weight Introducing attention mechanism for path weight To optimize:

[0151]

[0152] in, is the node correlation score function, exp is the exponential function, and k is the node index;

[0153] S45, Path Weight Based on Optimization and the updated node feature H( L ), combined with the features F extracted from multimodal medical data modal , generate the causal enhancement vector F enhance :

[0154]

[0155] Among them, V is the node set, i is the index of the node currently being processed, and concat is the concatenation operation. It is the feature vector extracted from multimodal medical data.

[0156] In this implementation manner, S5 specifically includes:

[0157] S51, based on causal enhancement vector F enhance , construct a dynamic risk prediction model, the input of the dynamic risk prediction model is the causal enhancement vector F enhance and the corresponding target variable Y t , where Y t represents the risk label of adverse events at time t;

[0158] S52. Define the time series state transition rules of the risk prediction model, and build a dynamic risk prediction model based on the recursive neural network. The state transition formula is:

[0159] h t =σ(W h ·h t-1 +W x ·F enhance +b h );

[0160] Among them, h t is the hidden state vector, σ is the activation function, W h and W x are the weight matrices of hidden state and input features, respectively, and b h is the bias term, h t-1 To provide a memory of the risks of the previous moment;

[0161] S53. Combined with hidden state vector h t , by introducing attention mechanism, time-dependent weight and global causal information into the risk prediction in the time dimension to output the risk value

[0162]

[0163] Among them, softmax is the activation function, W y is the output layer weight matrix, b y is the bias term of the output layer, T is the length of the historical time step, and h k is the hidden state at time k, α k is the temporal attention weight, ψ(W global ,tk) is the global causal correction function, W global is the global causal weight feature, the current moment in the t time dimension;

[0164] S54. Introducing multiple maintenance correction mechanisms to improve risk value Adjustments are made to combine ethical corrections with path weights, time dependencies, and characteristics of highly sensitive nodes:

[0165]

[0166] Among them, λ 1 ,λ 2 ,λ 3 is the modified weight coefficient, P sensitive A collection of highly ethically sensitive paths. is the optimized path weight, P is the set of highly ethically sensitive paths P sensitive The total number of is the risk assessment result after multiple repairs, i and j are node indexes;

[0167] S55. Output risk assessment results Combined with the causal enhancement vector F enhance Provides causal path impact analysis.

[0168] In this implementation manner, S6 specifically includes:

[0169] S61. Based on causal network and risk assessment results Calculate the contribution C(P) of each causal path:

[0170]

[0171] in, Node feature The amount of change, is the optimized path weight, P is the causal path, i and j are the node indexes, and is a node;

[0172] S62. Using the ethical weight matrix W ethical , modify the path contribution and screen high-risk paths P risk :

[0173] P risk = {P|C(P)·W ethical (P)≥θ 1};

[0174] Among them, W ethical (P) is the comprehensive ethical weight of path P, θ 1 Screening thresholds for risk;

[0175] S63. For the selected high-risk paths P risk , calculate the path priority R(P), taking into account the path length, ethical sensitivity and node characteristics:

[0176]

[0177] in, For Node To Node The ethical correction path weight, L(P) is the total length of path P, For Node ethical sensitivity, For Node ethical sensitivity;

[0178] S64, in the high-risk path set P risk In the example, we label the key node set N key , the key node marking rules are:

[0179]

[0180] Among them, θ 2 Filter thresholds for node sensitivity;

[0181] S65, based on the priority sorting R(P) and the key node set N key , generate intervention recommendations G intervene , specifically:

[0182] G intervene ={(P,N key )|P∈P risk ,rank(R(P))≤G};

[0183] Where G is the number of high-priority paths set, and rank(R(P)) is the ranking of path P in the priority sorting.

[0184] In this implementation manner, the S7 specifically includes:

[0185] S71. From the high-risk path set P risk and the key node set N keyLet’s start by defining the optimization objective function to balance risk reduction and ethical constraints:

[0186]

[0187] in, To optimize the objective function, C(P) is the risk contribution of path P, R(P) is the priority score of path P, For Node ethical sensitivity, α, β are weight coefficients;

[0188] S72. According to the high-risk path set P risk and the key node set N key , define the intervention action set A:

[0189]

[0190] Among them, a i is the path P and the node The i-th intervention action implemented, ΔC(P,a i ) is the intervention action a i The risk contribution reduction value of path P, θ a is the minimum effect threshold of the intervention action;

[0191] S73. Combine high-risk path set P risk and the key node set N key , by quantifying the risk reduction effect and implementation cost of each intervention action, and evaluating i The comprehensive effect:

[0192]

[0193] Among them, R t is the combined effect of the intervention, γ a is the cost penalty coefficient, w j For key nodes The importance weight of For intervention action a i For Node The impact value of

[0194] S74, using particle swarm optimization algorithm, based on the high-risk path set P risk , intervention action set A and intervention comprehensive effect R t , optimizing the intervention strategy through dynamic adjustment of particle position and velocity:

[0195]

[0196] Among them, π * (St ) is the optimal intervention strategy, π is the candidate intervention strategy, S t is the state space, ΔC(P,π) is the effect of reducing the contribution of intervention strategy π to path P, γ·Cost(π) is the implementation cost of intervention strategy π, combined with the cost penalty coefficient γ, ω is the inertia weight, v t is the particle velocity, argmax π To find the intervention strategy π that maximizes the optimization objective function * ;

[0197] S75. Combining risk assessment results with optimized intervention strategies * (S t ), generate personalized prevention and control plan G plan :

[0198] G plan ={(P,N key ,a i )|P∈P risk ,a i =π * (S t )};

[0199] Among them, N key is the key node set, P risk is the set of high-risk paths, a i is the intervention action and P is the path.

[0200] Embodiment 1:

[0201] In order to verify the feasibility of the present invention in implementation, the present invention is applied to the intensive care unit of a tertiary hospital. In this scenario, the patient's condition is complicated and the multimodal data is complex, including electronic medical records, monitoring data, laboratory test results, etc. At the same time, medical resources are limited, and doctors need to make decisions quickly, which puts higher requirements on the prediction and prevention of adverse medical events. However, existing medical risk assessment methods are mostly based on static statistics or single-modality data analysis, which cannot dynamically model changes in patients' conditions, and it is difficult to consider complex factors such as ethical fairness, resulting in inaccurate prediction results and limited effectiveness of intervention measures.

[0202] To solve the above problems, the present invention was applied and verified in the ICU scenario of the hospital. First, the multimodal medical data of 100 ICU patients were collected and cleaned, including the patients' electronic medical records, real-time monitoring data (heart rate, blood pressure, blood oxygen saturation, etc.) and laboratory test data (white blood cell count, C-reactive protein, coagulation index, etc.). Subsequently, the dynamic causal network between the patient's condition variables was constructed through the dynamic causal network technology proposed in the present invention. During the construction process, the causal path is dynamically updated by combining time slicing technology and ethical correction factors, and the contribution of the path to the risk of adverse events is quantified.

[0203] In specific applications, the present invention focused on the prediction and prevention of adverse drug reactions. During the data analysis process, the system identified 50 high-risk causal paths. These paths were weighted by ethical correction factors, and 15 high-priority paths were further screened, including some key paths involving vulnerable patients (such as elderly patients and patients with multiple comorbidities).

[0204] Based on the high-risk path screening results, the present invention uses a particle swarm optimization algorithm to dynamically optimize the intervention strategy. For example, for an elderly patient, the causal network analysis shows that "antibiotic dose adjustment->abnormal liver function->infection risk" is the main adverse event path. After dynamic optimization, the system recommends that doctors reduce the antibiotic dose, increase the frequency of liver function monitoring, and prioritize the allocation of anti-infection drug resources to patients. Experimental data show that through the above intervention, the patient's liver function indicators significantly improved within 48 hours after the intervention (alanine aminotransferase decreased from 150U / L to 45U / L), and the risk of infection decreased by 30%.

[0205] In order to verify the overall effect of the present invention, the differences between the present invention and the prior art in terms of risk prediction accuracy, intervention effect and ethical fairness were compared and tested. In the application of 100 patients, the risk prediction accuracy of the present invention reached 92.5%, which was significantly higher than the 78.6% of the prior art; in terms of the intervention effect on adverse events, the present invention reduced the incidence of adverse events in patients from 35% to 20%, while the prior art only reduced it to 28%. In addition, in the patient grouping fairness test, the proportion of intervention measures for vulnerable patients in the present invention reached 45%, which was much higher than the 25% of the prior art, reflecting significant ethical fairness.

[0206] Table 1 Comparative data between the present invention and the prior art in the prediction and prevention of adverse medical events

[0207]

[0208] The present invention is significantly superior to the existing technology in the prediction and prevention of adverse medical events. In terms of risk prediction accuracy, the present invention reaches 92.5%, which is higher than the 78.6% of the existing technology, showing the ability of dynamic causal networks to accurately model complex causal relationships. In high-risk path screening, the screening accuracy of the present invention is 89.3%, which is significantly better than the 72.1% of the existing technology, proving the effectiveness of ethical correction factors and causal analysis.

[0209] For vulnerable patients, the intervention coverage of the present invention is 45.0%, which is much higher than the 25.0% of the existing technology, highlighting its advantage in fair allocation of resources. At the same time, the incidence of adverse events was reduced from 35.0% to 20.0%, and the improvement rate of the high-risk patient group reached 65.0%, which is significantly better than the existing technology, indicating that the present invention has excellent performance in precise intervention and resource optimization.

[0210] In addition, the multimodal data utilization rate of the present invention reaches 95.0%, and the priority intervention rate is 85.0%, which are far higher than the 70.0% and 60.0% of the existing methods respectively. Through the particle swarm optimization algorithm, the present invention shortens the intervention optimization time to 12 minutes, greatly improving the efficiency. These data fully prove that the present invention has comprehensive advantages in accuracy, fairness and efficiency, and provides an efficient and reliable solution for medical scenarios.

[0211] In summary, the present invention not only significantly improves the accuracy of risk prediction and the effectiveness of intervention measures, but also achieves comprehensive transcendence in ethical fairness and efficiency, providing an innovative solution for the prediction and prevention of adverse medical events.

[0212] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for risk prediction and prevention of adverse medical events based on causal inference, characterized in that: The steps include: S1. Collect multimodal medical data, preprocess the data, and generate cleaned multimodal medical data; S2. Based on the cleaned multimodal medical data, the time dimension is divided through time slicing technology to construct a causal network; S3. Based on the constructed causal network, define the ethical factor indicators, construct the ethical weight matrix, quantify the weight influence of ethical factors on the causal path, and embed the causal network to form an ethical causal network; S4. Based on the ethical causal network and cleaned multimodal medical data, a causal feature embedding module is designed, a graph neural network is used to process node features, and multimodal features are integrated to generate a causal enhancement vector; S5. Use causal enhancement vectors to build a dynamic risk prediction model, combine the time dimension with recursive neural networks, capture the dynamic pattern of adverse events, and output risk assessment results; S6. Based on the risk assessment results and ethical causal network, analyze the contribution of causal paths, screen high-risk paths and key nodes based on ethical weights, and generate intervention recommendations; S7. Use the screened high-risk pathways and key nodes, combined with risk assessment results, to dynamically adjust intervention recommendations and generate personalized prevention and control plans; S8. Generate dynamic transparency reports based on personalized prevention and control plans and show them to medical staff and patients.

2. A method for predicting and preventing medical adverse events based on causal inference according to claim 1, characterized in that: The S2 specifically includes: S21. Based on the cleaned multimodal medical data, the data is divided into time dimensions using time slicing technology, and each time slice T is defined. i It is a subset of medical data within a fixed time interval; S22, in each time slice T i In the above example, we extract the causal variable X j , where X j Represents potential causal factors that affect adverse medical events, including patient characteristics, treatment methods, and external environmental variables; S23, based on each time slice T i and the extracted causal variable X j , construct the time-slice subnetwork of the causal network, and calculate the conditional causal influence between variables through the extended causal probability formula: in, For variables In time slice T i The probability distribution in for right The conditional probability of For variables The parent node set of , m is the total number of causal variables, k is the causal variable index; For variables The conditional probability at time t is, is the causal variable at time t, serving as the "antecedent" in the causal relationship, is the causal variable at time t, serving as the "consequence" in the causal relationship; S24. Use a dynamic Bayesian network to connect each time slice sub-network to form a causal network including the time dimension: Among them, P(X t |X t-1 ) is the state of the causal network at time t is determined by the state at t-1; S25. Extract the causal paths and key nodes in the causal network, and mark the paths in the causal paths that have a significant impact on the risk of adverse events.

3. A method for predicting and preventing medical adverse events based on causal inference according to claim 1, characterized in that: The S3 specifically includes: S31. Define the ethical factor index E based on the constructed causal network k , where E k Including fairness, resource priority and patient vulnerability indicators, quantifying ethical modifications to path weights in causal networks; S32. Construct the ethical weight matrix W ethical ,element Indicates that at time t, the variable Relative to variable Ethical impact weight: Among them, E k is the ethical factor, α k is the weight ratio of ethical factors, p is the total number of ethical factors, k is the index of ethical factors, i and j are the indexes of variables in the causal network; S33, based on the ethical weight matrix W ethical , adjust the weights of causal paths in the causal network Added corrections to path adaptation factor and node influence factor: in, is the ethical correction weight of the path, is the nth adjacent node pair node The influence value of , β is the path adaptation factor, N is the node The number of adjacent nodes of and is the node, n is the node index; S34, calculate the path score S(P) by using the path score formula based on multi-dimensional indicators k ), and sort by score, and select the first q highly sensitive paths as the highly ethically sensitive paths P sensitive : Among them, P k is the kth path, is the ethical sensitivity, T is the total length of the time dimension, t is the index of the time slice, δ t is the time weight; S35, based on high ethical sensitivity path P sensitive , generate ethical causal network G ethical , the network contains the node relationships and path weights corrected by ethical factors: Among them, θ is the path weight threshold.

4. A method for predicting and preventing medical adverse events based on causal inference according to claim 1, characterized in that: The S4 specifically includes: S41, based on ethical causal network G ethical and cleaned multimodal medical data, extracting causal enhancement features, including node features, path features, and network structure features; S42, build a causal feature embedding module to embed the node features in the causal network and the ethical modification weight of the path Mapping to low-dimensional feature space: Among them, Z embed is the embedded causal enhancement feature vector, X is the node feature matrix, is the relationship function between node pairs, σ is the activation function, W is the trainable parameter matrix, i and j are the indices of the variables in the causal network, and is a node; S43, based on the embedded causal enhancement feature vector Z embed , the graph neural network is used to perform multi-layer updates on the embedded features, and the update rule of the node features is: in, is the feature vector of node i in the lth layer, with an initial value of Z embed , W (l) is the trainable weight matrix of the lth layer, N(i) is the set of neighbor nodes of node i, is the attention weight of the lth layer, is the path feature aggregation function, is the feature vector of node j in the lth layer, σ is the activation function, is the feature vector of node i in the l+1th layer; S44. Updated node features and ethical modification weight Introducing attention mechanism for path weight To optimize: in, is the node correlation score function, exp is the exponential function, and k is the node index; S45, Path Weight Based on Optimization and the updated node feature H (L) , combined with the features F extracted from multimodal medical data modal , generate the causal enhancement vector F enhance : Among them, V is the node set, i is the index of the node currently being processed, and concat is the concatenation operation. It is the feature vector extracted from multimodal medical data.

5. The method for predicting and preventing adverse medical event risks based on causal inference according to claim 1, characterized in that: The S5 specifically includes: S51, based on causal enhancement vector F enhance , construct a dynamic risk prediction model, the input of the dynamic risk prediction model is the causal enhancement vector F enhance and the corresponding target variable Y t , where Y t represents the risk label of adverse events at time t; S52. Define the time series state transition rules of the risk prediction model, and build a dynamic risk prediction model based on the recursive neural network. The state transition formula is: h t =σ(W h ·h t-1 +W x ·F enhance +b h ); Among them, h t is the hidden state vector, σ is the activation function, W h and W x are the weight matrices of hidden state and input features, respectively, and b h is the bias term, h t-1 To provide a memory of the risks of the previous moment; S53. Combined with hidden state vector h t , by introducing attention mechanism, time-dependent weight and global causal information into the risk prediction in the time dimension to output the risk value Among them, softmax is the activation function, W y is the output layer weight matrix, b y is the bias term of the output layer, T is the length of the historical time step, and h k is the hidden state at time k, α k is the temporal attention weight, ψ(W global ,tk) is the global causal correction function, W global is the global causal weight feature, the current moment in the t time dimension; S54. Introducing multiple maintenance correction mechanisms to improve risk value Adjustments are made to combine ethical corrections with path weights, time dependencies, and characteristics of highly sensitive nodes: Among them, λ1, λ2, λ3 are the modified weight coefficients, P sensitive A collection of highly ethically sensitive paths. is the optimized path weight, P is the set of highly ethically sensitive paths P sensitive The total number of is the risk assessment result after multiple repairs, i and j are node indexes; S55. Output risk assessment results Combined with the causal enhancement vector F enhance Provides causal path impact analysis.

6. A method for predicting and preventing adverse medical event risks based on causal inference according to claim 1, characterized in that: The S6 specifically includes: S61. Based on causal network and risk assessment results Calculate the contribution C(P) of each causal path: in, Node feature The amount of change, is the optimized path weight, P is the causal path, i and j are the node indexes, and is a node; S62. Using the ethical weight matrix W ethical , modify the path contribution and screen high-risk paths P risk : P risk ={P∣C(P)·W ethical (P)≥θ1}; Among them, W ethical (P) is the comprehensive ethical weight of path P, θ1 is the risk screening threshold; S63. For the selected high-risk paths P risk , calculate the path priority R(P), taking into account the path length, ethical sensitivity and node characteristics: in, For Node To Node The ethical correction path weight, L(P) is the total length of path P, For Node ethical sensitivity, For Node ethical sensitivity; S64, in the high-risk path set P risk In the example, we label the key node set N key , the key node marking rules are: Among them, θ2 is the node sensitivity screening threshold; S65, based on the priority sorting R(P) and the key node set N key , generate intervention recommendations G intervene , specifically: G intervene ={(P,N key )∣P∈P risk ,rank(R(P))≤G}; Where G is the number of high-priority paths set, and rank(R(P)) is the ranking of path P in the priority sorting.

7. A method for predicting and preventing medical adverse events based on causal inference according to claim 1, characterized in that: The S7 specifically includes: S71. From the high-risk path set P risk and the key node set N key Let’s start by defining the optimization objective function to balance risk reduction and ethical constraints: in, To optimize the objective function, C(P) is the risk contribution of path P, R(P) is the priority score of path P, For Node ethical sensitivity, α, β are weight coefficients; S72. According to the high-risk path set P risk and the key node set N key , define the intervention action set A: Among them, a i For path P and node The i-th intervention action implemented, ΔC(P,a i ) is the intervention action a i The risk contribution reduction value of path P, θ a is the minimum effect threshold of the intervention action; S73. Combine high-risk path set P risk and the key node set N key , by quantifying the risk reduction effect and implementation cost of each intervention action, and evaluating i The comprehensive effect: Among them, R t is the combined effect of the intervention, γ a is the cost penalty coefficient, w j For key nodes The importance weight of For intervention action a i For Node The impact value of S74, using particle swarm optimization algorithm, based on the high-risk path set P risk , intervention action set A and intervention comprehensive effect R t , optimizing the intervention strategy through dynamic adjustment of particle position and velocity: Among them, π * (S t ) is the optimal intervention strategy, π is the candidate intervention strategy, S t is the state space, ΔC(P,π) is the effect of reducing the contribution of intervention strategy π to path P, γ·Cost(π) is the implementation cost of intervention strategy π, combined with the cost penalty coefficient γ, ω is the inertia weight, v t is the particle velocity, argmax π To find the intervention strategy π that maximizes the optimization objective function * ; S75. Combining risk assessment results with optimized intervention strategies * (S t ), generate personalized prevention and control plan G plan : G plan ={(P,N key ,a i )∣P∈P risk ,a i =π * (S t )}; Among them, N key is the key node set, P risk is the set of high-risk paths, a i is the intervention action and P is the path.

Images