Rau risk generation and prediction method and system based on causal bootstrapped diffusion model

Abstract

The application discloses a RAU risk generation and prediction method and system based on a causal guided diffusion model, belongs to the cross technical field of medical information technology and artificial intelligence, and does not directly predict a risk label, but learns and simulates complete evolution distribution from a current health state to multiple possible health states in the future. Through Monte Carlo sampling, the application can generate a large number of 'future selves' conforming to individual characteristics, and statistically obtain risk probabilities from the 'future selves'. The method can better capture the uncertainty of disease development, provide a probability distribution instead of a single point estimate, and the prediction result is more in line with the randomness nature of a biological process.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of medical information technology and artificial intelligence, and in particular to a method and system for generating and predicting RAU risk based on a causal guided diffusion model. Background Technology

[0002] Recurrent aphthous ulcers (RAU), one of the most common types of oral mucosal diseases, cause long-term distress to hundreds of millions of patients worldwide due to their recurrent attacks, severe pain, and impact on quality of life. The pathological characteristics of RAU are round or oval ulcers on the oral mucosa that are red, yellow, concave, and painful. Although self-limiting, its recurrence cycle is unpredictable, severely affecting patients' eating, speech function, and mental health.

[0003] The medical community generally considers Recurrent Antihypertensive Ulcer (RAU) to be a complex, multifactorial, immune-mediated disease. Its potential risk factors are broad, encompassing immune system dysregulation (such as T-cell subset disorders and cytokine network abnormalities), systemic diseases (such as gastrointestinal diseases and hematological disorders), micronutrient deficiencies (such as iron, zinc, and B vitamins), oral microbiota imbalance, and even psychological stress, lifestyle habits, and dietary habits. These factors are not simply linearly additive but constitute a highly nonlinear, dynamically changing, and complex causal network.

[0004] Current methods for assessing the risk of recurrent acute urinary tract infections (RAUs) rely primarily on clinicians' experience-based medical history taking and limited laboratory testing. This model has fundamental limitations: 1. Lack of predictive ability and passivity: Assessment often occurs "after" an ulcer has already occurred, rather than "before" it occurs, failing to provide proactive risk warnings and leaving patients in a state of passive response. 2. Insufficient quantification and standardization: The assessment criteria rely heavily on doctors' subjective experience and lack a unified, objective, and quantifiable risk scale, making it difficult to conduct large-scale, standardized risk screening and management; 3. Inadequate modeling of complex relationships: Traditional statistical models (such as logistic regression) or conventional discriminative machine learning models (such as support vector machines and random forests) struggle to capture the deep, nonlinear interactions and dynamic evolution patterns behind multimodal and heterogeneous data. They typically answer "whether" a relapse will occur, but cannot describe "how" a relapse occurs. 4. The "black box" problem and lack of interpretability: Even some deep learning models operate like a "black box" in their prediction process, unable to clearly explain to doctors and patients "why" a high-risk assessment is made, or what specific factors played a decisive role in the current situation. This greatly limits their application value in guiding personalized and precise interventions.

[0005] The limitations of existing technologies in predicting RAU risk stem from their adherence to a "discriminative" technological paradigm. This paradigm aims to assign a static risk label to individuals, neglecting the fundamental characteristic that RAU risk evolves dynamically over time and has multiple possible future paths. To achieve truly forward-looking predictions, the technological paradigm must shift from "discrimination" to "generative," that is, from "judging the current state" to "generating a set of possible future evolutionary paths." Generative artificial intelligence models, especially diffusion models (DDPMs), learn the inherent probability distribution of data and sample new samples from it, which aligns perfectly with the need to simulate multiple future possibilities in terms of technical principles. However, translating this general principle into a specific technical solution applicable to RAU risk generation, prediction, and interpretable attribution remains a gap in current technology. Therefore, to overcome the significant shortcomings of existing technologies in terms of accuracy, forward-looking perspective, personalization, and interpretability, a novel method and system based on a generative paradigm is urgently needed. Summary of the Invention

[0006] The purpose of this invention is to address the technical deficiencies in the existing technology by providing a method for generating and predicting RAU risk based on a causal guided diffusion model.

[0007] Another object of the present invention is to provide the above-described system.

[0008] The technical solution adopted to achieve the purpose of this invention is: A method for generating and predicting RAU risk based on a causal guided diffusion model includes the following steps: Step 1: Obtain the user's multimodal health data, preprocess the multimodal health data, and then extract features. Step 2: Construct a Causal Guided-Diffusion Model (CG-DM), train it, and then generate a set of target vectors for future health states using the trained CG-DM. Step 3: Calculate the user's RAU relapse probability based on the set of target vectors obtained in Step 2. And identify and quantify key risk drivers, and rank the quantified key risk drivers; Step 4, integrate RAU recurrence probability The key risk drivers, after being sorted, are visualized.

[0009] In the above technical solution, the causal guided diffusion model includes a multimodal conditional encoder and a causal guided inverse denoising network. The system includes a risk reasoning and attribution module; the multimodal conditional encoder processes and encodes the multimodal health data to form a unified conditional embedding vector; and a causal-guided inverse denoising network gradually adds Gaussian noise to a noise-free target vector to generate a noisy target vector. Guided by the conditional embedding vector, the noisy target vector is iterated multiple times in reverse to remove noise, thereby training the causal-guided inverse denoising network to obtain a target vector for the future health state. The risk reasoning and attribution module is used to drive the causal-guided inverse denoising network to perform multiple Monte Carlo samplings during the reasoning phase to generate the distribution of future states and calculate the risk probability from them. At the same time, it analyzes the causal cross-attention module inside the causal-guided inverse denoising network to identify key risk factors.

[0010] In the above technical solution, the multimodal health data includes static data. Time series data Image data and questionnaire data The multimodal conditional encoder extracts features from multimodal health data through multiple parallel sub-networks, and then concatenates or fuses the features extracted from each sub-network to obtain a conditional embedding vector c, as shown in the following formula: ; ; ; ; ; in, It is a multilayer perceptron. For gated loop unit, It is a convolutional neural network. For splicing or attention fusion operations, These are the feature vectors output by each sub-network.

[0011] In the above technical solution, the target vector of the future health state It is a K-dimensional vector whose dimensions include key biomarkers and clinical characteristics after a certain point in the future.

[0012] In the above technical solution, the causal guided inverse denoising network For the U-Net architecture with embedded causal cross-attention modules, at each time step of the reverse process... The causal-guided inverse denoising network receives the noisy target vector through a causal cross-attention module. Encoding vector and conditional embedding vector As the original input features, to predict the injected noise Output attention weight matrix Its attention weight matrix The calculation process is as follows: ; ; ; ; ; Attention ;

[0013] in, Noisy target vector Intermediate representation in causal-guided inverse denoising networks; For time step The corresponding encoding vector; Linear ( ) is a learnable linear projection layer; To represent features after incorporating time information, They are respectively by and conditional embedding vector The generated queries, keys, and values; The projection matrix is ​​learnable; The dimension of the key vector; The attention weight matrix reveals the causal contribution of current risk factors to the generation of future states. This attention mechanism dynamically focuses on the conditional embedding vector at each step of the denoising process. The part most relevant to the current denoising task.

[0014] In the above technical solution, the steps for the risk reasoning and attribution module to calculate the recurrence probability of RAU are as follows: Step a, in the conditional embedding vector Guided by this, for an initial pure noise vector Using the trained causal-guided inverse denoising network ,from arrive Iterate and gradually denoise to generate a target vector for the future health state. ; Step b, Repeat step a several times to generate a set of target vectors for future health states. ; Step c: Count the number of target vectors for future health states that meet the criteria for RAU relapse in the set obtained in step b. ; Step d: Calculate the recurrence probability of RAU. .

[0015] In the above technical solution, step a, specifically the process of using the trained causal-guided inverse denoising network, is as follows: the causal-guided inverse denoising network... The internal causal cross-attention module will use the attention weight matrix AND value vector Multiplication yields weighted features AttentionOutput The weighted features are passed through an output projection layer and then added to the input features of the attention module through residual connections to form an updated feature representation. This representation is then fed into the next layer of U-Net for subsequent calculations.

[0016] In the above technical solution, the risk reasoning and attribution module identifies key risk factors by analyzing the attention weight matrix generated during the generation process and recorded independently of the denoising process. The specific steps are as follows: Step A, Attention Weight Recording: During the reverse denoising process of step a once or multiple times, record each or several preset time steps. At that time, the attention weight matrix calculated by the causal cross-attention module Attention weight matrix The dimension is ,in It is the sequence length of the intermediate features in the network. It is a conditional embedding vector Dimensions.

[0017] Step B, Attention Score Aggregation: For each time step The attention weight matrix, along its row dimensions Summarize to get a length of a one-dimensional vector One-dimensional vector of all time steps Weighted summation is performed to obtain a final aggregated attention score vector. ,in, For time step The weight of the corresponding one-dimensional vector.

[0018] Step C, Attribution Mapping and Recognition: Aggregate the attention score vector Mapping back to the original input health features, based on the construction of the multimodal conditional encoder, the known conditional embedding vectors... The correspondence between each dimension interval and the original input health features will be used to aggregate the attention score vector. The scores of dimensions corresponding to the same original health feature are accumulated to obtain the total attention score of that original health feature. The higher the total attention score, the stronger the guiding role of that original health feature factor in the evolution from the current state to a future high-risk state, i.e., it is a key risk driver. Another aspect of the invention also includes a risk generation and prediction system for recurrent aphthous ulcers (RAU) based on a causal guided diffusion model. The system is used to implement the method and includes: a data interface module, a preprocessing module, a risk prediction core module, and a visualization and interpretation module. The data interface module is used to acquire users' multimodal health data from multiple data sources; the preprocessing module is used to clean and standardize the acquired multimodal health data and construct model inputs; the risk prediction core module includes a causal guided diffusion model (CG-DM), which, guided by the user's current multimodal health data, generates a set of target vectors for future health status and the user's RAU relapse probability. And identify key risk drivers; the visualization and interpretation module is used to show users or doctors the probability of RAU recurrence risk in the form of charts and highlight the key risk drivers identified.

[0019] Compared with the prior art, the beneficial effects of the present invention are: This invention represents a paradigm shift from "discriminative" to "generative" approaches, resulting in more probabilistic and realistic predictions. Instead of directly predicting a risk label, it learns and simulates the complete evolutionary distribution from the current health state to various possible future health states. Through Monte Carlo sampling, this invention can generate a large number of "future selves" that conform to individual characteristics and statistically derive the risk probability from them. This method better captures the uncertainty of disease development, providing a probability distribution rather than a single point estimate, and the prediction results are more consistent with the stochastic nature of biological processes. It incorporates a clear causal attribution mechanism, completely resolving the "black box" problem. The core causal-guided inverse noise reduction network, through its internal cross-attention mechanism, must "look back" at current health indicators at each step in generating the future state to obtain guidance. These attention weights naturally and dynamically reveal which current factors contribute most to high-risk future outcomes. This provides doctors with direct, quantifiable scientific evidence for developing personalized interventions (such as recommending vitamin supplementation, adjusting sleep patterns, or managing stress). By deeply integrating multimodal heterogeneous data, a truly panoramic and personalized assessment is achieved. The multimodal conditional encoder of this invention can uniformly process and encode data from vastly different sources such as genes, imaging, time-series monitoring, and subjective questionnaires, using it as unified conditional information to guide the entire generation process. This ensures that predictions are based on the most comprehensive health profile of each individual, achieving a high degree of personalization. It possesses the potential for data augmentation and counterfactual analysis. A well-trained model can not only predict risk but also serve as a high-quality digital twin data generator, producing large amounts of realistic, time-evolving synthetic patient data to alleviate the scarcity of medical data. More valuablely, by modifying certain values ​​in the conditional embedding vector c (e.g., simulating "if vitamin B12 levels are increased by 20%), counterfactual analysis can be performed to observe changes in future risk probabilities, thereby quantifying the potential effects of specific interventions. Attached Figure Description

[0020] Figure 1 This is a schematic flowchart of the method of the present invention. Detailed Implementation

[0021] The present invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0022] Example 1 like Figure 1 As shown, a method for generating and predicting RAU risk based on a causal guided diffusion model includes the following steps: Step 1: Obtain the user's multimodal health data, preprocess the multimodal health data, and then extract features; wherein, the multimodal health data includes: static data. Includes gender, age, family history of RAU, systemic medical history (stent history, diabetes), and smoking history (none, past smoker, current smoker). Time series data. This includes records of RAU attacks over the past year (date, duration, and number of ulcers for each attack), regular hematological test results (white blood cell count, folic acid, and vitamin B12 levels), and continuous data recorded via wearable devices or apps (such as daily steps, sleep duration, and heart rate variability (HRV)). Imaging data. This includes standardized oral endoscopic images taken by professional physicians or by users themselves to assess mucosal health; and an oral microbiome OTU / ASV abundance table obtained from 16S rRNA sequencing. Questionnaire data. This includes a regularly completed psychological and lifestyle questionnaire, a stress scale rating system (PSS10 stress scale), a food questionnaire (milk, eggs, seafood, nuts, tomatoes, pineapple, peppers), and the amount of exposure to each item (none, small amount, large amount).

[0023] Step 2: Construct a Causal Guided-Diffusion Model (CG-DM), train it to obtain the trained CG-DM model, and then generate a set of target vectors for future health states under given conditional information. The CG-DM model includes a multimodal conditional encoder, a forward diffusion process, and a causal guided inverse denoising network. And a risk reasoning and attribution module; specifically including the following steps: Step 2.1: Collect target vectors from the longitudinal cohort study data. The target vector is used to depict a snapshot of the health status over a one-month prediction window. Preferably, the... ,in Indicates whether an ulcer has occurred. This indicates the average diameter of the ulcer. This represents the average pain score. This indicates the predicted serum ferritin level.

[0024] Step 2.2: Input the multimodal health data extracted in Step 1 into the multimodal conditional encoder for encoding, forming a conditional embedding vector c. Preferably, the multimodal conditional encoder includes multiple parallel sub-networks, namely... It is a multilayer perceptron. For gated loop unit, For convolutional neural networks, a pre-trained ResNet-18 is used as... For processing image data using convolutional neural networks (Oral cavity image), extract 256-dimensional features A two-level gated cyclic unit (GRU) is used as a timing encoder to process timing data. Extract 128-dimensional features A multilayer perceptron (MLP) with an entity embedding layer processes static data. and questionnaire data Extract 128-dimensional features and The features extracted from each of the above sub-networks are concatenated or fused and passed through a linear layer to form a 512-dimensional conditional embedding vector c. The formula is as follows: ; ; ; ; ; in, It is a multilayer perceptron. For gated loop unit, It is a convolutional neural network. For splicing or attention fusion operations, These are the feature vectors output by each sub-network.

[0025] Step 2.3, guiding the inverse denoising network During training, a forward diffusion process is required, which is based on a pre-defined variance scheduling table. (For example, from) linearly increase to Total steps ), at each time step in the variance scheduling table Within, all are directed towards a noise-free target vector. Add Gaussian noise The process involves adding T steps at a time until a noisy target vector is obtained. The formula used is as follows: in, and ,in, From the start (time step 1) to the current time step All information proportions The cumulative product. It represents the sum of the original, noise-free image. Beginning, then After adding noise, the original image information What percentage was retained in total? It is a time step At that time, for the previous image The proportion of information retained. It is determined by variance. Decide, ,because It's the proportion of noise added, so It refers to the proportion of original information retained.

[0026] Step 2.4, in the guided reverse denoising network The U-Net architecture, which embeds causal cross-attention modules in multiple resolution levels of the encoder and decoder, integrates intermediate features at each resolution level. The conditional embedding vector c obtained in step 2.2 will be used as the key and value for the noisy target vector obtained in step 2.3. After multiple inverse iterations to remove noise, a causal-guided inverse denoising network is trained to obtain the target vector of the future health state. Therefore, when deciding how to denoise, the user's current health condition must be taken into account. For example, if the conditional embedding vector c encodes features such as "recent extreme stress" and "low vitamin B12," the causal cross-attention module will assign higher weights to these features, thereby guiding the denoising process to generate a future state that is more likely to contain ulcers. Preferably, the loss function used during training is: ,in, , where represents the expected value, and represents the average value across different training data. Different time steps Different noises , : indicates from the training dataset In the process, a clean, noise-free "real" sample is randomly selected. .

[0027] Step 2.5, in the conditional embedding vector Guided by this, for an initial pure noise vector Using the trained causal-guided inverse denoising network ,from arrive Iterate and gradually denoise to generate a target vector for the future health state. Repeat the above process 1000 times to generate a set of target vectors for future health states. = ;in, ;(in ),in, : i.e., time step ; The currently known input, i.e., the time step. A noisy vector; The trained causal-guided inverse denoising network is used as input to the current noisy vector. Conditional embedding vectors and time step Then predict in The noise contained therein; This is a scaling factor. This factor is needed to adjust the noise amplitude to match the current step. The part in parentheses is... One estimate is made using the estimated noise. From Subtract noise from the middle, and then pass through Scale adjustments are made to obtain the desired result. A preliminary estimate; A controlled random noise term is injected into each step of the reverse denoising process to ensure the randomness of the generation process.

[0028] Step 2.6: Count the number of target vectors in the set obtained in Step 2.5 that meet the criteria for RAU recurrence. Calculate the recurrence probability of RAU For example, count the number of target vectors representing future health status with a dimension value greater than 0.5, indicating whether an ulcer has occurred. The user's RAU recurrence probability .

[0029] Step 3: Identify key risk drivers using the risk reasoning and attribution module. The specific steps are as follows: During one or more iterations of denoising using the trained causal-guided inverse denoising network, record and weightedly sum all time steps in the iterative denoising process. Attention weight matrix The weighted attention weight matrix is ​​then summarized according to the feature dimensions of the conditional embedding vector c to obtain the total attention score for each original input feature. A higher total attention score indicates a stronger guiding role for the factor in the evolution from the current state to a future high-risk state, thus identifying it as a key risk driver. For example, for the conditional embedding vector... The attention weight matrix is ​​obtained by summing and weighting the 512-dimensional features. Dimensions 10 to 20 correspond to "psychological stress," and dimensions 21 to 30 correspond to "vitamin B levels." By summing the total attention weights across these dimensions, the "contribution score" for each macro-factor can be obtained. For example, if the total score for "psychological stress" is the highest, the system determines it to be the current key risk driver.

[0030] Preferably, its attention weight matrix The calculation process is as follows: ; ; ; ; ; Attention ;

[0031] in, Noisy target vector Intermediate representation in U-Net residual blocks; For time step The corresponding encoding vector; Linear ( ) is a learnable linear projection layer; It is a feature representation that incorporates time information. They are respectively by and conditional embedding vector The generated queries, keys, and values; The projection matrix is ​​learnable; It is the dimension of the key vector; This forms the final attention weight matrix, used to reveal the "causal" contribution of current risk factors to the generation of future states. This attention mechanism dynamically focuses on the conditional embedding vector at each step of the denoising process. The part most relevant to the current denoising task.

[0032] Step 4: Result Visualization and Interpretation. Visualize and interpret the RAU recurrence probability calculated in Step 3. (e.g., "78%)" and a bar chart showing key risk drivers are presented to users / doctors, and personalized recommendations are generated based on these key risk drivers. The bar chart clearly lists the top five key risk drivers and their relative contribution, such as: "Recent mental stress (45%)", "Insufficient sleep (25%)", "Spicy diet (15%)", "Oral flora - abnormal abundance of Streptococcus (10%)", and "Genetic susceptibility (5%)".

[0033] Example 2 This embodiment provides a risk generation and prediction system for recurrent aphthous ulcers (RAU) based on a causal guided diffusion model. The system can be deployed on a cloud server and provides services to end users and medical institutions via Web API or a mobile app. The system includes: a data interface module, a preprocessing module, a risk prediction core module, and a visualization and interpretation module.

[0034] The data interface module is responsible for secure data exchange with hospital HIS / LIS systems, user mobile apps, and gene testing service providers to acquire users' multimodal health data. The preprocessing module cleans and standardizes the acquired multimodal health data and constructs model inputs. The risk prediction core module includes a causal guided diffusion model (CG-DM), which, guided by the user's current multimodal health data, generates a set of target vectors for future health status and the user's RAU relapse probability. And identify key risk drivers; the visualization and interpretation module is used to show users or doctors the probability of RAU recurrence risk in the form of charts and highlight the key risk drivers identified.

[0035] The visualization and report generation module generates an interactive web report. The report includes: (1) a prominent risk dashboard displaying the final risk probability and the corresponding risk level (low / medium / high). (2) a dynamic, interactive risk attribution graph (such as a sunburst or bar chart) that users can click to view the specific values ​​and historical trends of each risk factor. (3) a user-friendly suggestion text library automatically generated based on the attribution results. For example, if "low vitamin B12" is the main factor, the system will automatically prompt: "Your vitamin B12 level may pose a major risk to your recent oral health. It is recommended that you consult a doctor and consider increasing your intake of foods rich in vitamin B12, such as animal liver, red meat, eggs, and dairy products."

[0036] Example 3 Application of counterfactual reasoning: In this embodiment, the system of the present invention is used to evaluate the effectiveness of potential interventions. A user predicted by the system to have an 85% risk of RAU recurrence within the next month had their primary risk drivers attributed to "extremely high PSS stress scale score" and "extremely low sleep quality index." The physician can operate within the system's "intervention simulation" interface: 1. Create "counterfactual" conditions: The physician embeds the user's conditions into a vector. Make a copy and manually modify the corresponding "PSS score" and "sleep duration" values ​​to an "ideal" range (e.g., simulating the state after psychological counseling and increased sleep duration). This yields a new "counterfactual" conditional embedding vector. 2. Perform "counterfactual" predictions: The system uses this modified... As a guide, the risk reasoning module is run again to generate a new future state distribution and calculate a new risk probability. 3. Evaluate the intervention effect: Assuming calculations show... The risk of RAU relapse can be reduced to 25%. The system can then show doctors and patients: "Through effective stress management and improved sleep, your risk of RAU relapse is expected to be significantly reduced from 85% to 25%." This quantitative feedback greatly enhances patient adherence to medical advice and provides data support for doctors to select the most effective intervention pathway.

[0037] This invention utilizes a novel generative AI paradigm to achieve accurate prediction and deep causal attribution of the risk of relapse in recurrent acute respiratory distress syndrome (RAU). This approach not only improves predictive accuracy but, more importantly, opens the door to truly personalized, explainable, and interventionable precision health management, providing a revolutionary technological path to conquering this stubborn and common disease.

[0038] The above description is only a preferred embodiment of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for generating and predicting RAU risk based on a causal guided diffusion model, characterized in that, Includes the following steps: Step 1: Obtain the user's multimodal health data, preprocess the multimodal health data, and then extract features. Step 2: Construct a causal guided diffusion model CG-DM, train it, and generate a set of target vectors for future health states through the trained causal guided diffusion model. The causal guided diffusion model CG-DM includes a multimodal conditional encoder, a causal guided inverse denoising network, and a risk reasoning and attribution module. Step 3: Calculate the user's RAU relapse probability based on the set of target vectors obtained in Step 2. And identify and quantify key risk drivers, and rank the quantified key risk drivers; Step 4, integrate RAU recurrence probability Visualize the key risk drivers after ranking; The multimodal conditional encoder processes and encodes the multimodal health data to form a unified conditional embedding vector c. The causal-guided inverse denoising network progressively adds Gaussian noise to a noise-free target vector to generate a noisy target vector. Guided by the conditional embedding vector, it removes noise from the noisy target vector through multiple inverse iterations to train the causal-guided inverse denoising network and obtain a target vector for the future health state. The risk reasoning and attribution module drives the causal-guided inverse denoising network to perform multiple Monte Carlo samplings during the reasoning phase to generate a distribution of future states and calculate the risk probability. Simultaneously, it analyzes the causal cross-attention module within the causal-guided inverse denoising network to identify key risk factors.

2. The RAU risk generation and prediction method according to claim 1, characterized in that, The multimodal health data includes static data. Time series data Image data and questionnaire data The multimodal conditional encoder extracts features from multimodal health data through multiple parallel sub-networks, and then concatenates or fuses the features extracted from each sub-network to obtain a conditional embedding vector c, as shown in the following formula: ； ； ； ； ； in, It is a multilayer perceptron. For gated loop unit, It is a convolutional neural network. For splicing or attention fusion operations, These are the feature vectors output by each subnetwork.

3. The RAU risk generation and prediction method according to claim 1, characterized in that, The target vector of future health status It is a K-dimensional vector whose dimensions include key biomarkers and clinical characteristics after a certain point in the future.

4. The RAU risk generation and prediction method according to claim 1, characterized in that, The causal-guided inverse denoising network is a U-Net architecture embedded with a causal cross-attention module. At each time step of the inverse process... The causal-guided inverse denoising network receives the noisy target vector, encoding vector, and conditional embedding vector through a causal cross-attention module. As the original input features, to predict the injected noise Output attention weight matrix Its attention weight matrix The calculation process is as follows: ； ； ； ； ； Attention ; in, Noisy target vector Intermediate representation in causal-guided inverse denoising networks; For time steps The corresponding encoding vector; Linear ( () is a learnable linear projection layer; To represent the features after incorporating time information, They are respectively by and conditional embedding vector The generated queries, keys, and values; The projection matrix is ​​learnable; The dimension of the key vector; This is the attention weight matrix.

5. The RAU risk generation and prediction method according to claim 1, characterized in that, The steps for calculating the relapse probability of RAU in the Risk Reasoning and Attribution module are as follows: Step a, in the conditional embedding vector Guided by this, for an initial pure noise vector Using the trained causal-guided inverse denoising network ,from arrive Iterate and gradually denoise to generate a target vector for the future health state. ; Step b: Repeat step a N times to generate a set of target vectors for future health states. ; Step c: Count the number of target vectors for future health status that satisfy RAU relapse in the set obtained in step b. ; Step d: Calculate the recurrence probability of RAU. .

6. The RAU risk generation and prediction method according to claim 5, characterized in that, In step a, the specific process of using the trained causal-guided inverse denoising network is as follows: the causal cross-attention module inside the causal-guided inverse denoising network assigns the attention weight matrix... AND value vector Multiplication yields weighted features AttentionOutput The weighted features are passed through an output projection layer and then added to the input features of the attention module through residual connections to form an updated feature representation. This representation is then fed into the next layer of U-Net for subsequent calculations.

7. The RAU risk generation and prediction method according to claim 4, characterized in that, The risk reasoning and attribution module identifies key risk factors by analyzing the attention weight matrix generated during the generation process and recorded independently of the denoising process. The specific steps are as follows: Step A, Attention Weight Recording: During the reverse denoising process of step a once or multiple times, record each or several preset time steps. At that time, the attention weight matrix calculated by the causal cross-attention module Attention weight matrix The dimension is ,in It is the sequence length of the intermediate features in the network. It is a conditional embedding vector The dimension; Step B, Attention Score Aggregation: For each time step The attention weight matrix, along its row dimensions Summarize to get a length of a one-dimensional vector One-dimensional vector of all time steps Weighted summation is performed to obtain a final aggregated attention score vector. ,in, For time steps The weights of the corresponding one-dimensional vector; Step C, Attribution Mapping and Recognition: Aggregate the attention score vector Mapping back to the original input health features, based on the construction of the multimodal conditional encoder, the known conditional embedding vectors... The correspondence between each dimension interval and the original input health feature is determined by summing the dimension scores corresponding to the same original health feature in the aggregated attention score vector to obtain the total attention score of the original health feature.

8. A RAU risk generation and prediction system based on a causal guided diffusion model, characterized in that, The system is used to implement the RAU risk generation and prediction method as described in any one of claims 1 to 7, and the system includes: a data interface module, a preprocessing module, a risk prediction core module, and a visualization and interpretation module; The data interface module is used to acquire users' multimodal health data from multiple data sources; the preprocessing module is used to clean and standardize the acquired multimodal health data and construct model inputs; the risk prediction core module includes a causal guided diffusion model, which, guided by the user's current multimodal health data, generates a set of target vectors for future health status and the user's RAU relapse probability. And identify key risk drivers; the visualization and interpretation module is used to show users or doctors the probability of RAU recurrence risk in the form of charts and highlight the key risk drivers identified.

Images