Clinical intervention effect prediction method and device based on diffusion cross counterfactual regression network
By using a diffusion-cross counterfactual regression network, the problems of confounding factors, selection bias, and sample imbalance in medical data are solved, enabling more accurate prediction of clinical intervention effects and supporting personalized medical decision-making.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- PEKING UNIV
- Filing Date
- 2026-02-06
- Publication Date
- 2026-06-02
AI Technical Summary
Existing technologies struggle to effectively mitigate potential confounding factors, selection bias, and sample imbalance when processing medical data, resulting in insufficient accuracy and reliability in predicting the effects of clinical interventions and impacting the application value of personalized medicine.
We employ a method based on a diffusion-cross counterfactual regression network. This method reduces the influence of confounding factors through a feature cleansing network module, alleviates selection bias by introducing an integral probability metric, and constructs a cross-network to reduce sample imbalance. We build a model that includes multi-layer feature interactions and a dual-channel deep neural network and optimize the training process to improve prediction accuracy.
It significantly improves the accuracy and robustness of predicting the effects of clinical interventions, can distinguish between effective and ineffective patient subgroups, provides data support for personalized medicine, and optimizes the allocation of medical resources.
Smart Images
Figure CN122135929A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of medical and health information processing technology, and relates to information processing technology for predicting the effect of clinical intervention, and particularly to a method and device for predicting the effect of clinical intervention based on a diffusion cross-counterfactual regression network. Background Technology
[0002] Accurately predicting the impact of clinical interventions on patient outcomes is a core task of personalized medicine. Quantifying the effects of interventions and effectively identifying patient groups sensitive to specific interventions can help physicians develop more appropriate treatment plans for patients with different characteristics and optimize resource allocation.
[0003] With the widespread application of electronic medical records in the medical field, intervention effect prediction technology based on causal inference has developed rapidly. The core of this technology lies in quantitatively assessing the impact of different interventions on patients by analyzing patient characteristics, intervention measures and patient outcomes. However, accurate prediction of clinical intervention effects faces the following three core problems: (1) potential confounding factors: there are patient characteristics that simultaneously affect intervention allocation and patient outcomes but are not directly observed; (2) selection bias: there is a systematic bias in the probability distribution of the intervention group and the control group; (3) sample imbalance: there is a significant difference in the sample size between the intervention group and the control group.
[0004] While traditional machine learning methods offer some interpretability when processing structured data, they struggle to capture complex interactions within the data, lack the ability to handle confounding factors, and exhibit poor generalization ability in small-sample clinical data scenarios. Deep learning methods, although improving feature representation capabilities and demonstrating advantages in processing multimodal complex data, perform poorly in addressing imbalanced sample problems, limiting the accuracy and reliability of predictions and impacting their application value in clinical decision support. Summary of the Invention
[0005] To address the shortcomings of the existing technologies, the present invention aims to provide a method and apparatus for predicting the clinical intervention effect based on a diffusion-cross counterfactual regression network, thereby improving the accuracy and reliability of predicting the clinical intervention effect of patients, and distinguishing between patient subgroups with and without effective intervention, thus providing a reference for personalized medical decision-making.
[0006] The core of this invention includes: defining a feature cleansing network based on a diffusion model to effectively reduce the impact of potential confounding factors in medical data; introducing an integral probability metric to mitigate selection bias caused by non-random assignment; and constructing a cross-network to reduce the interference of sample imbalance. By constructing a patient feature dataset containing dimensions such as demographic information, medical history, and clinical indicators, and by constructing and training a clinical intervention effect prediction model based on a diffusion-cross counterfactual regression network, the accuracy and reliability of predicting patient clinical intervention effects are improved, and patient subgroups with effective and ineffective interventions are distinguished.
[0007] For ease of understanding, the terminology used in this invention is explained as follows:
[0008] Electronic medical records: comprehensive documents that record and manage patient information, medical process, examinations and tests, medications and follow-ups in digital form.
[0009] Causal inference: using statistical or experimental methods to estimate the impact of an intervention on an outcome.
[0010] Diffusion model: A generative model based on a progressive noise addition and denoising process.
[0011] The technical solution adopted in this invention is:
[0012] A clinical intervention efficacy prediction method based on diffusion-cross counterfactual regression networks includes the following steps:
[0013] First, construct a patient characteristic dataset.
[0014] Based on references, expert opinions, and clinical practice, a multi-dimensional core patient characteristic dataset was constructed, including demographic information, medical history, and clinical indicators. The raw data was preprocessed to ensure its quality met the requirements for model training.
[0015] Furthermore, the preprocessing includes normalizing or standardizing continuous features, while removing or filling samples that lack core features or key outcomes, thereby improving the effectiveness and consistency of the data.
[0016] Second, we construct a clinical intervention effect prediction model based on a diffusion cross-counterfactual regression network.
[0017] This invention constructs a clinical intervention effect prediction model based on a diffusion-cross counterfactual regression network. The diffusion-cross counterfactual regression network structure in the clinical intervention effect prediction model includes: a feature cleansing network module, a multi-layer feature interaction network module, a dual-channel deep neural network module, and a cross-network module. The feature cleansing network module is used to reduce the influence of potential confounding factors; the multi-layer feature interaction network module is used to learn shared representation functions and uncover complex relationships between patient features; the dual-channel deep neural network module is used to predict the potential outcomes of patients under different intervention scenarios; and the cross-network module is used to reduce the interference of sample imbalance through a balance learning mechanism. Specifically, it includes the following module steps:
[0018] S1. Feature Cleaning Network Module: Using patient feature datasets, based on the input original patient feature data, through forward and reverse diffusion processes of the diffusion model, the module reduces the interference of potential confounding factors in the data on feature representation, and obtains pure feature codes that can truly reflect the clinical intervention effect of patients.
[0019] S2, Multi-layer Feature Interaction Network Module: Based on the clean feature encoding output by the feature purification network module, the shared representation function is learned through a multi-layer neural network to obtain a representation containing information on individual patient differences.
[0020] S3, Dual-channel deep neural network module: Based on the patient's intervention label, the two channels of the dual-channel deep neural network predict the potential outcomes of the patient receiving and not receiving intervention, and calculate the predicted value of the patient's clinical intervention effect.
[0021] S4, Cross-Network Module: Utilizes feature representation to learn the pseudo-intervention effect, and serves as an intermediate variable to connect the two channels of the dual-channel deep neural network, reducing the interference of sample size imbalance between the intervention group and the control group.
[0022] Furthermore, in step S1, the feature purification network module, the forward diffusion process of the diffusion model gradually adds noise to the original features, interfering with and deconstructing them; the reverse diffusion process, based on a neural network, effectively removes potential confounding factors from the noisy data through iterative denoising operations, restoring pure features layer by layer. This aims to improve feature quality, ensure the effectiveness and accuracy of features, and eliminate the influence of potential confounding factors in medical data.
[0023] Furthermore, in step S2, the multi-layer feature interaction network module introduces an integral probability metric. By measuring the distribution differences between the intervention group and the control group in the shared representation space and making balance adjustments, the prediction bias caused by the uneven distribution of data between the two groups is reduced, thereby optimizing the accuracy of clinical intervention effect estimation.
[0024] Third, we optimized and trained the clinical intervention effect prediction model based on the diffusion cross-counterfactual regression network.
[0025] In the diffusion cross-reverse factual regression network, the loss functions of each module are defined using mean squared error, integral probability measure, etc. to optimize the model. The synergy of multiple loss functions reduces the risk of overfitting and enhances the generalization of the model.
[0026] Furthermore, during training, model performance is continuously optimized by adjusting model parameters and hyperparameters, such as learning rate, number of layers, and number of hidden units. Simultaneously, regularization techniques are introduced to enhance the model's resistance to overfitting and improve its reliability in practical applications.
[0027] In specific implementation, the present invention utilizes the above method to realize a clinical intervention effect prediction system (device) based on a diffusion cross-counterfactual regression network.
[0028] The system employs a patient feature dataset construction method as described in the first step of this invention and any optional embodiment thereof, a clinical intervention effect prediction model based on a diffusion-cross counterfactual regression network as described in the second step of this invention and any optional embodiment thereof, and an optimization method for a clinical intervention effect prediction model based on a diffusion-cross counterfactual regression network as described in the third step of this invention and any optional embodiment thereof. The system includes:
[0029] The patient feature dataset construction module is used to collect core patient features from multiple dimensions, including demographic information, medical history, and clinical indicators, based on the method in the first step, and to construct a standardized, high-quality patient feature dataset.
[0030] The clinical intervention effect prediction model construction module is used to build a feature purification network module, a multi-layer feature interaction network module, a dual-channel deep neural network module, and a cross network module according to the model structure described in the second step. It defines the network structure of each module, realizes the collaborative work between the modules, and constructs a complete clinical intervention effect prediction model based on diffusion cross counterfactual regression network.
[0031] The model optimization and training module is used to train the prediction model using the patient feature dataset according to the optimization method in the third step. By configuring multiple loss function combinations, the hyperparameters of the model are continuously adjusted. When the model performance tends to stabilize or reaches the preset target, the training is stopped, and the optimized clinical intervention effect prediction model based on the diffusion cross counterfactual regression network is obtained.
[0032] The clinical intervention effect prediction model application module is used to process the input patient electronic medical records according to the preprocessing standards of the patient feature dataset construction module, input them into the trained prediction model, calculate the potential outcomes of the patient under the scenarios of receiving intervention and not receiving intervention, as well as the corresponding clinical intervention effect prediction values, distinguish between the "effective subgroup" and "ineffective subgroup" of intervention, output the prediction results, and provide data support for doctors to formulate personalized treatment plans.
[0033] The beneficial technical effects of this invention are as follows:
[0034] This invention provides a method and apparatus for predicting the clinical intervention effect based on a diffusion-cross counterfactual regression network. Through innovative model design, it introduces a diffusion model as a feature purification network module, helping to alleviate the problem of severe interference from potentially confounding features in real-world medical scenarios. It introduces an integral probability metric to balance the spatial distribution of the representation and constructs a cross-network for weight compensation, reducing the impact of selection bias and sample imbalance, thus improving the accuracy and robustness of clinical intervention effect prediction. Experiments show that the model outperforms existing methods in terms of prediction accuracy and overall performance. Furthermore, by quantifying the clinical intervention effect on patients, this invention makes progress in distinguishing between "effective subgroups" and "ineffective subgroups," enabling the extraction of more valuable information from electronic medical records, providing doctors with references and assisting in personalized treatment. From a macro perspective, this invention helps optimize doctors' workflows, promotes the rational allocation of medical resources, and provides an intelligent tool for the development of personalized medicine, which has positive significance for improving diagnostic and treatment efficiency, reducing costs, and improving the patient's medical experience. Attached Figure Description
[0035] Figure 1 This is a schematic diagram of the clinical intervention effect prediction model based on the diffusion cross-counter-factual regression network of the present invention;
[0036] Figure 2 This is a structural block diagram of the clinical intervention effect prediction system (device) based on diffusion cross-counter-factual regression network of the present invention;
[0037] Figure 3 This is an overall flowchart of the clinical intervention effect prediction system (device) based on diffusion cross-counterfactual regression network of the present invention. Detailed Implementation
[0038] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. These examples are used to illustrate the present invention but do not limit its scope in any way.
[0039] Example 1
[0040] This invention provides a method for constructing a patient feature dataset, aiming to provide a high-quality data foundation for subsequent model training. The observable patient information samples can be denoted as D={(y, x, w)}. For each patient, y is the clinical intervention outcome of interest, such as the patient's survival status (whether they die), physiological scores, or prognostic assessment values; x is the patient's characteristics, including demographic information (such as age, gender), medical history, and other health characteristics; w represents a binary intervention variable, i.e., whether or not a certain intervention is received. Let y(1) and y(0) represent the potential clinical intervention outcomes for the patient when receiving intervention (w=1) or not receiving intervention (w=0). Further, let sets T and C represent the intervention sample set and the control sample set, respectively. In this case, their union is... This constitutes a complete set of samples, which is the entire sample space.
[0041] Example 2
[0042] like Figure 1 As shown, this embodiment of the invention provides a clinical intervention effect prediction model based on a diffusion-cross counterfactual regression network. The diffusion-cross counterfactual regression network includes the following modules: a feature purification network module, a multi-layer feature interaction network module, a dual-channel deep neural network module, and a cross-network module. First, the feature purification network module reduces the influence of potential confounding factors in the original features to obtain pure feature encodings. Second, the multi-layer feature interaction network module mines complex relationships between features to generate representations. Then, the dual-channel deep neural network module predicts the intervention effect in intervention / control scenarios respectively, and the cross-network module learns pseudo-intervention effects to compensate for information in the dual-channel prediction results. Finally, the individual's clinical intervention effect is calculated by combining the prediction information output by each module.
[0043] S1, Feature Cleanup Network Module
[0044] Based on the input original patient feature data, noise is gradually introduced during the forward diffusion process of the diffusion model, and then the original features are recovered through iterative learning via the reverse diffusion process. This reduces the influence of potential confounding factors in the data and yields pure feature codes that can truly reflect the clinical intervention effects on patients.
[0045] (1) Forward diffusion process
[0046] Let x t Let x represent the noisy state at any given time. t-1 This refers to the patient features at step t-1. The initial input x0 is the original patient feature x. i That is, x0=x i Noise diffusion is achieved using Markov chains.
[0047]
[0048] Where q represents the probability distribution of the forward diffusion process; N represents the normal distribution; I is the identity matrix; β t ∈(0, 1) controls the noise intensity added in step t, α t =1-β t This refers to the proportion of the original feature signal retained. Through reparameterization and the additivity of two independent Gaussian noises, the noise-added state at any given time can be directly obtained from x0.
[0049]
[0050]
[0051] It can be deduced that,
[0052]
[0053] To avoid completely destroying patient characteristics, a linear noise scheduling strategy was designed to adjust the added noise.
[0054]
[0055] The hyperparameter s∈[0, 1] controls the overall noise level, α min <α max Define the upper and lower bounds of the noise.
[0056] (2) Reverse diffusion process
[0057] Patient characteristics after adding noise Starting from this point, the patient's pure patient characteristics are gradually restored, which is represented as:
[0058]
[0059] Where, p θ It is the probability distribution of the reverse diffusion process. and These are the mean and variance of the Gaussian distribution output by a neural network with learnable parameters θ, respectively.
[0060] (3) Reasoning process
[0061] pass Step-by-step positive noise addition process The original, mixed patient features x0 are destroyed, followed by a reverse denoising process. To restore the pure patient characteristics .
[0062] S2, Multi-layer Feature Interaction Network Module
[0063] Based on the clean features output by the feature sanitization network module, a shared representation function is learned through a multi-layer neural network. Generate a representation z that includes information on individual differences. i Then, the intervention outcome is predicted from this representation. Let the distributions of the intervention group and the control group generated by the representation function be denoted as . and In particular, an integral probability metric is introduced to balance the two distributions and alleviate the selection bias problem.
[0064] S3, Dual-channel Deep Neural Network Module
[0065] A dual-channel deep neural network was designed to estimate the intervention response function μ1(·) and the control response function μ0(·), i.e., the potential outcome with and without intervention. Therefore, the patient's intervention effect can be expressed as:
[0066]
[0067] S4, Cross-Network Module
[0068] Learning the effect of pseudo-intervention by using extracted patient feature representations Furthermore, by using it as an intermediate variable, a bidirectional connection is established between the intervention response function and the control response function, which can balance the learning of the two response functions during training.
[0069] Specifically, the cross-network subtracts the predicted pseudo-intervention effect from the output of the intervention response function to obtain the cross-control response. This indicates the outcome when individuals in the intervention group did not receive the intervention.
[0070]
[0071] Simultaneously, the predicted pseudo-intervention effect is added to the output of the control response function to obtain the cross-intervention response. This indicates the outcome if individuals in the control group receive the intervention.
[0072]
[0073] This embodiment constructs a diffusion-cross counterfactual regression network through four modules: a feature sanitization network module to reduce the influence of potential confounding factors in medical data through a diffusion model; a multi-layer feature interaction network module to capture complex relationships between features; a dual-channel deep neural network module to predict intervention effects in different scenarios; a cross-network to learn pseudo-intervention effects; and a bidirectional network connecting intervention and control responses. The collaborative work of these multiple modules effectively alleviates the problems of "potential confounding factors," "selection bias," and "sample imbalance" faced by existing clinical intervention effect prediction models.
[0074] Example 3
[0075] This embodiment provides an optimized training method for a clinical intervention effect prediction model based on a diffusion cross-reverse factual regression network. By adjusting the model parameters and hyperparameters according to the loss function, the model accuracy and generalization are optimized, ensuring that the model can stably output accurate prediction results in real clinical data scenarios.
[0076] (1) Loss function of feature sanitization network module
[0077] To learn the key parameter θ of the backdiffusion process, the optimization objective of the diffusion model is to maximize the Evidence Lower Bound (ELBO) of the original patient features x0. This metric simultaneously constrains the model's performance in both feature reconstruction and denoising matching. By maximizing the ELBO, the model can make the feature distribution generated during the backdiffusion process as close as possible to the true feature distribution, thereby improving the accuracy of feature reconstruction. This process is equivalent to minimizing the difference between the model output and the target features. Therefore, the loss function of the feature sanitization network module can be expressed as a weighted sum of the model's prediction errors, denoted as . .
[0078] (2) Loss function of multi-layer feature interaction network module
[0079] The multi-layer feature interaction network module specifically introduces an integral probability metric to balance the distributions of the intervention and control groups: the integral probability metric (IPM) is a measure of the difference between probability distributions. Due to the difference between the distributions of the intervention and control groups, the integral probability metric IPM... , This directly constitutes the theoretical upper limit for estimating the loss of individual intervention effects; therefore, it can be achieved by minimizing IPM ( , To optimize individual efficacy estimation.
[0080] (3) Loss function of dual-channel deep neural network module
[0081] The objective functions corresponding to the intervention response function and the control function in the dual-channel deep neural network module are as follows: minimizing the difference between the predicted potential result and the true value.
[0082]
[0083]
[0084] loss function of dual-channel deep neural network modules That is, the optimization objective of the intervention response function. The optimization objective of the control response function The weighted sum, where ω1 and ω2 are the weights:
[0085]
[0086] (4) Loss function of cross-network module
[0087] The loss function for predicting the pseudo-intervention effect in the cross-network module is also to minimize the difference between the predicted potential result and the true value.
[0088]
[0089]
[0090] Cross-network module loss function This refers to the cross-cutting intervention response optimization objective. , and the cross-response optimization objective The weighted sum, where ω3 and ω4 are the weights:
[0091]
[0092] Therefore, the complete loss function of the diffusion cross-regression network This can be expressed as,
[0093]
[0094] Where λ1-λ4 are the weights of each loss; The loss function for the feature cleanup network module; This is the integral probability measure between the distributions of the intervention group and the control group, i.e., the loss function of the multi-layer feature interaction network module; is the loss function for the dual-channel deep neural network module; This is the loss function for the cross-network module.
[0095] This embodiment presents an optimized training method for a clinical intervention effect prediction model based on a diffusion-cross counterfactual regression network. Through multi-dimensional optimization strategies, it leverages the diffusion mechanism to reduce the impact of potential confounding factors, mitigates selection bias through distribution balance adjustment, and reduces the interference of sample imbalance by utilizing the cross-network. Combined with hyperparameter tuning and regularization techniques, the model's prediction accuracy and generalization performance in real-world clinical scenarios are enhanced, providing reliable performance assurance for practical applications.
[0096] Example 4
[0097] like Figure 2 and Figure 3As shown, this embodiment of the invention provides a clinical intervention effect prediction system (device) using a diffusion-cross counterfactual regression network. The system employs a patient feature dataset construction method as described in Embodiment 1 and any optional embodiment of this invention, a clinical intervention effect prediction model using a diffusion-cross counterfactual regression network as described in Embodiment 2 and any optional embodiment of this invention, and an optimization method for the clinical intervention effect prediction model using a diffusion-cross counterfactual regression network as described in Embodiment 3 and any optional embodiment of this invention. The system includes:
[0098] The patient feature dataset construction module is used to collect multi-dimensional core features of patients and construct a standardized, high-quality patient feature dataset according to the method described in Example 1.
[0099] The clinical intervention effect prediction model construction module is used to build a feature purification network module, a multi-layer feature interaction network module, a dual-channel deep neural network module, and a cross network module according to the model structure described in Example 2. It defines the network structure of each module, realizes the collaborative work between the modules, and constructs a complete clinical intervention effect prediction model with a diffusion cross counterfactual regression network.
[0100] The model optimization training module is used to train the prediction model using the patient feature dataset according to the optimization method described in Example 3, and continuously adjust the model hyperparameters by configuring multiple loss function combinations. When the model performance tends to be stable or reaches the preset target, the training is stopped, and the optimized diffusion cross counterfactual regression network clinical intervention effect prediction model is obtained.
[0101] The clinical intervention effect prediction model application module is used to process the input patient electronic medical records according to the preprocessing standards of the patient feature dataset construction module, input them into the trained prediction model, obtain the potential results of the patient under the scenarios of receiving intervention and not receiving intervention, calculate the corresponding individual intervention effect prediction value, distinguish between the "effective subgroup" and "ineffective subgroup" corresponding to the intervention, and output the prediction results to provide data support for doctors to formulate personalized treatment plans.
[0102] As can be seen from the above embodiments, the clinical intervention effect prediction method and device based on diffusion-cross counterfactual regression network disclosed in this invention first collects core patient features covering multiple dimensions such as demographic information, medical history, and clinical indicators to construct a standardized, high-quality patient feature dataset. Second, a diffusion-cross counterfactual regression network is constructed. The dataset is sampled, and a feature purification network module eliminates potential confounding factors in the input patient features. A multi-layer feature interaction network module mines complex relationships between features and learns balanced representations. A dual-channel deep neural network module predicts potential outcomes under different intervention scenarios, and a cross-network module learns pseudo-intervention effects to compensate for sample imbalance bias, thereby calculating the predicted value of the patient's clinical intervention effect. Subsequently, a model optimization training module trains and fine-tunes the network based on a multi-loss function joint optimization strategy, significantly improving the model's accuracy and robustness in predicting clinical intervention effects. Finally, through the application module, the patient's electronic medical record data is input into the trained prediction model, outputting the predicted potential outcomes under intervention scenarios, individual intervention effects, and "effective subgroup" identification results, providing scientific data support for doctors to formulate personalized treatment plans.
[0103] The clinical intervention effect prediction method based on diffuse cross-counterfactual regression network proposed in this invention uses the publicly available semi-synthetic dataset IHDP (Infant Health and Development Program) to construct a patient feature dataset. This dataset originates from a randomized trial evaluating the impact of expert home visits on the future cognitive scores of low birth weight infants. To ensure the validity and consistency of the data, samples lacking core patient characteristics and intervention results were removed, and a final sample of 747 eligible patients (including 139 intervention group samples and 608 control group samples) was selected for model training and testing.
[0104] In this embodiment, 25 features were extracted, including infant physiological development indicators (such as birth weight and head circumference) and maternal background characteristics (such as age and education level). A binary intervention label (whether or not an expert home visit was accepted) and a continuous outcome variable (infant cognitive test score) were set to construct and optimize a clinical intervention effect prediction model based on a diffusion-cross counterfactual regression network. Model performance was measured by the absolute error of the average intervention effect (…). ) and the absolute error of the average intervention effect of the intervention group ( A comprehensive evaluation of the metrics was conducted, and the results were compared with existing mainstream causal inference methods on both the training and test sets. The specific comparisons are shown in Table 1 below.
[0105] Table 1. Comparison of evaluation indicators between the present invention and existing causal inference methods
[0106]
[0107] The results in the table show that the examples of this invention demonstrate advantages in all evaluation indicators, fully illustrating the effectiveness and advancement of the proposed model in predicting clinical intervention effects. Through innovative model design, a diffusion model is introduced as a feature purification network module, effectively reducing the interference of potential confounding factors in the observation data. Simultaneously, the multi-layer feature interaction network module, by introducing an integral probability metric, balances the representation space distribution between the intervention group and the control group, effectively mitigating selection bias caused by non-random assignment. Furthermore, the construction of a cross-network to learn pseudo-intervention effects reduces the impact of sample imbalance through a two-way information compensation mechanism, significantly enhancing the model's predictive robustness under complex observation data. Moreover, this invention, by quantifying the individual effects of intervention methods, has made progress in identifying "effective subgroups" and "ineffective subgroups" of interventions, providing doctors with more valuable decision-making references and assisting in personalized diagnosis and treatment. From a macro perspective, this invention is beneficial for optimizing the allocation of medical resources, providing intelligent tools for the development of personalized medicine, and has positive significance for improving diagnostic and treatment efficiency and patient prognosis.
[0108] The methods and systems described in this invention are not limited to the embodiments described in the specific implementation. Other implementation methods derived by those skilled in the art based on the technical solutions of this invention also fall within the scope of technical innovation of this invention.
Claims
1. A method for predicting the effect of clinical intervention based on a diffusion-cross counterfactual regression network, characterized in that, Includes the following steps: 1) Collect multi-dimensional core patient characteristics and construct a patient characteristic dataset; 2) Construct a clinical intervention effect prediction model based on a diffusion-cross counterfactual regression network; The constructed clinical intervention effect prediction model based on diffusion-cross counterfactual regression network includes: a feature purification network module, a multi-layer feature interaction network module, a dual-channel deep neural network module, and a cross-network module; among which, The feature cleansing network module is used to reduce the influence of potential confounding factors in patient feature data. Based on the input original feature data, noise is gradually introduced during the forward diffusion process of the diffusion model, and then the original features are recovered through iterative learning via the reverse diffusion process to obtain a clean feature code that truly reflects the effect of clinical intervention. The multi-layer feature interaction network module is used to learn shared representation functions and explore the correlation between individual patient features. Based on the clean feature encoding output by the feature purification module, the shared representation functions are learned through a multi-layer neural network to generate a representation containing individual difference information. The intervention results are predicted from this representation. An integral probability metric is introduced to balance the distribution of the intervention group and the control group generated by the representation function. The dual-channel deep neural network module is used to predict the potential outcomes for patients under different intervention scenarios. Cross-network modules are used to reduce the interference of imbalanced samples through a balance learning mechanism; 3) The constructed clinical intervention effect prediction model based on diffusion cross counterfactual regression network is optimized and trained using a multi-dimensional optimization strategy to obtain a well-trained clinical intervention effect prediction model. In the diffusion cross-counterfactual regression network, a multi-dimensional loss function is used to collaboratively optimize the clinical intervention effect prediction model; Multidimensional loss functions include mean squared error and integral probability metric loss functions; 4) Input the patient characteristic data to be predicted into the trained clinical intervention effect prediction model to obtain the prediction results, including the potential outcomes of patients under the scenarios of receiving intervention and not receiving intervention, as well as the corresponding individual clinical intervention effect prediction values.
2. The clinical intervention effect prediction method based on diffusion-cross counterfactual regression network as described in claim 1, characterized in that, The forward diffusion process in the feature cleanup network module is as follows: Let x t Let x represent the noisy state at any given time. t-1 This refers to the patient features at step t-1; the initial input x0 is the original patient feature x. i That is, x0=x i ; Noise diffusion using a Markov chain is represented as follows: Where, β t ∈(0, 1) controls the noise intensity added at step t. By reparameterization and the additivity of independent Gaussian noise, the noise-added state x at any time can be obtained at x0. t , is represented as: Design a linear noise scheduling strategy to adjust the added noise, expressed as: Wherein, the hyperparameter s∈[0, 1] controls the overall noise level, α min <α max Define the upper and lower bounds of the noise.
3. The clinical intervention effect prediction method based on diffusion-cross counterfactual regression network as described in claim 2, characterized in that, The reverse diffusion process in the feature cleanup network module is as follows: Patient characteristics after adding noise Starting from this point, the patient gradually recovers pure patient characteristics, represented as: Where, p θ It is the probability distribution of the reverse diffusion process. and These are the mean and variance of the Gaussian distribution output by a neural network with learnable parameters θ, respectively.
4. The clinical intervention effect prediction method based on diffusion-cross counterfactual regression network as described in claim 1, characterized in that, The loss function of the feature cleansing network module is the weighted sum of the model prediction errors, and the optimization objective is to minimize the loss.
5. The clinical intervention effect prediction method based on diffusion-cross counterfactual regression network as described in claim 4, characterized in that, The loss function of the multi-layer feature interaction network module is to minimize the integral probability measure of the intervention group distribution and the control group distribution in the shared representation space.
6. The clinical intervention effect prediction method based on diffusion-cross counterfactual regression network as described in claim 5, characterized in that, The intervention response and control response loss functions of the dual-channel deep neural network module minimize the gap between the predicted potential results and the true values.
7. The clinical intervention effect prediction method based on diffusion-cross counterfactual regression network as described in claim 6, characterized in that, In the cross-network module, the predicted pseudo-intervention effect is subtracted from the output of the intervention response function to obtain the cross-control response, which represents the result when the individual in the intervention group did not receive the intervention; the predicted pseudo-intervention effect is added to the output of the control response function to obtain the cross-intervention response, which represents the result if the individual in the control group received the intervention.
8. The clinical intervention effect prediction method based on diffusion-cross counterfactual regression network as described in claim 7, characterized in that, The complete loss function of a diffusion cross-regression network Represented as: Where λ1-λ4 are the weights of each loss; The loss function for the feature cleanup network module; This is the integral probability measure between the distributions of the intervention group and the control group, i.e., the loss function of the multi-layer feature interaction network module; is the loss function for the dual-channel deep neural network module; This is the loss function for the cross-network module.
9. A clinical intervention effect prediction system based on a diffusion-cross counterfactual regression network implemented using the method of claim 1, characterized in that, The system includes: a patient feature dataset construction module, a clinical intervention effect prediction model construction module, a model optimization and training module, and a clinical intervention effect prediction model application module; wherein... The patient feature dataset construction module is used to collect and extract multi-dimensional core features of patients and build a standardized, high-quality patient feature dataset. The clinical intervention effect prediction model construction module is used to build a feature purification network module, a multi-layer feature interaction network module, a dual-channel deep neural network module, and a cross network module. It defines the network structure of each module, realizes the collaborative work between the modules, and builds a complete clinical intervention effect prediction model based on diffusion cross counterfactual regression network. The model optimization and training module uses the patient feature dataset to train the prediction model and adjusts the model hyperparameters by configuring multiple loss function combinations to train the model and obtain the optimized clinical intervention effect prediction model based on the diffusion cross antifactual regression network. The clinical intervention effect prediction model application module uses the trained prediction model to calculate the potential outcomes for patients under the scenarios of receiving intervention and not receiving intervention, as well as the corresponding predicted values of clinical intervention effects, distinguishes between effective and ineffective subgroups of intervention, and outputs the prediction results.