A health analysis method and apparatus

Abstract

The application discloses a health analysis method and device. The method comprises the following steps: obtaining historical health observation data of a plurality of individuals, including time series data of health characteristic indexes, health intervention measures, health risk results and individual inherent health data; constructing a causal inference model comprising an encoder and a decoder, wherein the encoder adopts a variational gated recurrent neural network and is used for extracting noise-robust latent variables from the time series data; training the causal inference model based on a preset loss function comprising a regularization term based on enhanced inverse propensity weighting, obtaining a health analysis model, and analyzing to-be-analyzed health observation data of a target object based on the health analysis model. By using the technical scheme, the calculation complexity of model training can be reduced, the stability of the throughput and response speed of online inference service can be improved, and the memory occupation during model inference can be reduced, so that the model can be deployed on a resource-limited device.

Description

Technical Field

[0001] This application relates to the fields of computer systems and intelligent health technology, specifically to a health analysis method and device based on deep learning and causal inference. Background Technology

[0002] Currently, the analysis of time-series data such as physiological electrical signals and activity sequences in the field of intelligent health services often relies on recurrent neural networks (RNNs) and long short-term memory (LSTM) networks. However, when these models are deployed on resource-constrained edge computing devices, they often face dual bottlenecks from hardware deployment and model training robustness, which restricts the practical application of related intelligent health services.

[0003] In terms of hardware deployment, traditional RNNs / LSTMs, due to their inherent sequential dependency characteristics, are difficult to fully parallelize, resulting in high computational latency during inference and failing to meet the stringent real-time feedback requirements of scenarios such as health monitoring. Simultaneously, the large number of model parameters and state memory units leads to significant memory consumption, making efficient operation difficult in resource-constrained environments. Furthermore, continuous gating operations and state updates result in high power consumption, impacting device battery life and the feasibility of long-term health monitoring.

[0004] In model training, the presence of time-varying noise in the training data, coupled with the sensitivity of RNNs / LSTMs to noisy, mixed, and non-stationary time-series healthy data, frequently leads to vanishing or exploding gradients, resulting in unstable training and difficulty in convergence. Consequently, developers often rely on extensive experimentation to perform tedious hyperparameter tuning of learning rate, network depth, initialization methods, etc., severely wasting computational time and storage resources, making the training process inefficient and costly. Summary of the Invention

[0005] This application proposes a health analysis method and device to address numerous technical problems associated with existing RNN / LSTM systems, including high computational latency, large memory footprint, and high power consumption in hardware deployment. Furthermore, it addresses the instability during training with noisy time-series data, requiring tedious hyperparameter tuning and wasting computational resources. By employing a variational gated recurrent network, the hidden states are transformed from deterministic vectors into low-dimensional distributed parameters, thereby optimizing memory access and enhancing computational stability. Additionally, by introducing a regularization term based on enhanced inverse tendency probability weighting into the preset loss function, the gradient descent direction is guided, reducing the number of model training iterations and improving the system reliability after deployment. This method is particularly suitable for real-time health data analysis in scenarios with limited computational resources or high service stability requirements.

[0006] This application provides a health analysis method, the method comprising:

[0007] Obtain a historical health observation dataset; wherein the historical health dataset includes historical health observation data of several individuals, and the historical health observation data includes time-series data of health characteristic indicators, health intervention measures, health risk outcomes, and individual inherent health data;

[0008] A causal inference model is constructed, comprising an encoder and a decoder. The encoder employs a variational gated recurrent neural network to extract noise-resistant latent variables from time-series data, and the decoder is used to predict the causal effect of health interventions on health risk outcomes based on the latent variables.

[0009] Based on a preset loss function, the causal inference model is trained using the historical health observation dataset to obtain a health analysis model, which is then used to analyze the health observation data of the target object; wherein, the preset loss function includes a regularization term based on the probability weighting of enhanced adverse tendency.

[0010] Furthermore, the encoder includes a latent variable learning network, a first intervention prediction network, and a first outcome prediction network. The latent variable learning network includes a variationally gated recurrent unit based on variational information bottleneck, used to determine latent variables based on individual inherent health data, health characteristic indicators, and health intervention measures. The first intervention prediction network is used to predict the health intervention measures at the current moment based on the latent variables. The first outcome prediction network is used to reconstruct the health risk outcome at the next moment based on the latent variables.

[0011] The decoder includes a latent variable transit network, a second intervention prediction network, and a second outcome prediction network. The latent variable transit network is used to recursively update latent variables during counterfactual prediction. The second intervention prediction network is used to predict health intervention measures at corresponding times based on the latent variables. The second outcome prediction network is used to predict health risk outcomes at future times based on the updated latent variables and the selection results of health intervention measures, and to determine the causal effect of health intervention measures on health risk outcomes.

[0012] Furthermore, latent variables are identified based on individual-specific health data, health characteristic indicators, and health interventions, including:

[0013] For each time step in the time series data, the hidden state is updated using a variational gated cyclic unit based on the variational information bottleneck;

[0014] The update rule for the variational gated recurrent unit based on the variational information bottleneck is as follows:

[0015] Based on the hidden state and health intervention measures of the previous moment, and the health characteristic indicators of the current moment, calculate the mean of the reset gate, update gate, candidate hidden states, and the logarithmic variance of the candidate hidden states at the current moment.

[0016] Random noise is sampled from a standard normal distribution, and candidate hidden states are determined based on the random noise, the mean of the candidate hidden states, and the log-variance of the candidate hidden states.

[0017] The hidden state at the current moment is determined based on the hidden state at the previous moment, the update gate at the current moment, and the candidate hidden states;

[0018] The distribution of latent variables is determined based on the last hidden state of the time series data;

[0019] The latent variables are identified by sampling from the distribution of the latent variables using reparameterization techniques.

[0020] Furthermore, determine the causal effect of health interventions on health risk outcomes, including:

[0021] The historical health observation dataset was sampled to obtain covariate samples;

[0022] The covariate samples are input into the encoder to obtain the latent variable distribution corresponding to the covariate samples, and the latent variable distribution is sampled to obtain latent variable samples;

[0023] The latent variable samples and the selection results of health intervention measures are input into the decoder to obtain the latent outcome distribution, and the latent outcome distribution is sampled to obtain the counterfactual outcome samples;

[0024] Based on the counterfactual outcome sample, determine the causal effect of health interventions on health risk outcomes.

[0025] Furthermore, based on the counterfactual outcome sample, the causal effect of the health intervention on the health risk outcome is determined, including: using formula (2) to determine the individual causal effect of the health intervention on the health risk outcome, and using formula (3) to determine the average causal effect of the health intervention on the health risk outcome.

[0026] Furthermore, based on a preset loss function, the causal inference model is trained using the historical health observation dataset to obtain a health analysis model, including:

[0027] The historical health observation dataset is divided into several mini-batches as training data for each iteration. For each individual in each training batch:

[0028] The individual's historical health observation data is input into the causal inference model to obtain predictions of health intervention measures and health risk outcomes;

[0029] Based on the enhanced prediction function, the enhanced prediction value is determined according to the prediction of the health intervention measures and the prediction of the health risk outcomes;

[0030] Based on the preset loss function, the loss function value containing a regularization term weighted by the enhanced adverse tendency probability is determined according to the enhanced predicted value and the observed value of the health risk result.

[0031] Based on the loss function value, the model parameters of the causal inference model are adjusted;

[0032] The health analysis model is obtained when the model converges or the number of iterations reaches the preset number.

[0033] Furthermore, the loss function also includes a regularization term based on the probability weighting of the enhanced adverse tendency, as shown in formula (1).

[0034] Furthermore, the preset loss function also includes a reconstruction loss term and a regularization term based on variational information bottleneck;

[0035] The reconstruction loss term is used to constrain the ability of the latent variables to reconstruct observed health interventions and health risk outcomes; the regularization term based on variational information bottleneck is used to constrain the distribution of the latent variables to be close to a preset prior distribution.

[0036] Furthermore, the sufficiency assumptions of the causal inference model include:

[0037] Hypothesis 1: The potential outcomes of a health intervention are equal to the observed outcomes;

[0038] Assumption 2: For all individuals in the historical health observation dataset, the probability of selecting any health intervention is greater than zero, regardless of health characteristic indicators;

[0039] Hypothesis 3: For all individuals in the historical health observation dataset, health interventions are independent of the potential outcomes of health interventions.

[0040] This application also provides a computing device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the health analysis method described above.

[0041] The embodiments of this application adopt the following technical solution: acquiring historical health observation data of several individuals, including time-series data of health characteristic indicators, health intervention measures, health risk outcomes, and individual inherent health data; constructing a causal inference model including an encoder and a decoder, wherein the encoder adopts a variational gated recurrent neural network to extract noise-resistant latent variables from the time-series data; training the causal inference model based on a preset loss function to obtain a health analysis model; and analyzing the health observation data of the target object based on the health analysis model.

[0042] The above-described technical solutions adopted in the embodiments of this application can achieve the following beneficial effects:

[0043] This solution proposes a health analysis scheme that reduces the computational complexity of model training, thereby reducing training time while achieving the same prediction accuracy; it also reduces memory usage during model inference, enabling deployment on resource-constrained embedded devices; it improves the numerical stability of the training process, reducing the frequency of gradient anomalies and minimizing resource waste caused by training interruptions; and it enhances the throughput and response speed stability of the online inference service. Attached Figure Description

[0044] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0045] Figure 1 A schematic diagram of the causal inference model for estimating causal effects using static observation data provided in this application;

[0046] Figure 2 A flowchart illustrating the process of estimating causal effects using RMSNs provided in this application;

[0047] Figure 3 A schematic flowchart of the health analysis method provided in Embodiment 1 of this application;

[0048] Figure 4 This is a schematic diagram of a causal inference model for estimating causal effects using time-series observation data, provided in Embodiment 1 of this application.

[0049] Figure 5 This is an architectural diagram of the causal inference model provided in Embodiment 2 of this application;

[0050] Figure 6 This is a schematic diagram of the structure of the variational gated loop unit based on the variational information bottleneck provided in Embodiment 2 of this application;

[0051] Figure 7This is a schematic diagram of the structure of a computing device provided in Embodiment 3 of this application. Detailed Implementation

[0052] Causal inference is a key research topic in many fields, including statistics, computer science, and economics. The goal of causal inference is to estimate the degree of change in outcomes when a decision is intervened in; that is, to calculate how much a variable (i.e., the observed outcome) changes when the value of a specific variable (e.g., an intervention) changes. Unlike traditional correlation analysis, it calculates causal relationships between things, reflecting the internal causal connections and guiding efficient and accurate decision-making or analyzing the development logic of things. For example, in the medical field, we can understand the acquisition and development process of diseases by estimating the causal relationship between diseases and pathogenic factors, thereby better preventing and intervening in diseases. Causal inference algorithms have broad applications in the life insurance industry. They can be used to evaluate the effectiveness of marketing strategies; assist marketers in making personalized product recommendations to customers; and have particularly high application value in customer service, such as helping customers provide health management plans and making personalized product recommendations based on customer circumstances.

[0053] The difficulty in estimating causal effects lies not only in the intricate causal relationships, but more importantly, in the fact that only the true outcome—the actual situation—can be observed. Counterfactual outcomes, those contrary to known facts, are forever unobservable. Since the causal effect of an intervention can be determined by subtracting the factual outcome from the counterfactual outcome, estimating the counterfactual outcome is key to estimating causal effects. The most reliable method for assessing causal effects is randomized controlled trials (RCTs), also known as double-blind trials. However, they have several problems: first, RCTs are very costly; second, they cannot be conducted in some scenarios; third, RCTs only focus on the sample mean, not the individuals; and fourth, they cannot provide causal explanations. In the era of big data, using historical observational data for causal inference is a more economical and practical method.

[0054] For causal inference models that estimate causal effects using static observation data, such as Figure 1 As shown in the observational data model, an individual's health characteristic indicator X influences latent variable Z, which in turn influences health intervention A. Both latent variable Z and health intervention A jointly influence health risk outcome Y. For example, older adults tend to avoid vaccination, resulting in the observed health intervention being non-vaccination. In the inferential model, the influence of latent variable Z on health intervention A is removed to simulate a randomized controlled experiment and obtain a pure causal effect.

[0055] For causal inference models that estimate causal effects using dynamic observation data, current technologies propose two main approaches in the field of causal inference based on historical observation data: statistical causal inference algorithms and machine learning-based causal inference algorithms. Statistical methods include: causal inference methods based on regression-based intervention allocation and covariate fitting; the inverse propensity score-weighted method (IPW) based on propensity score (representing an individual's preference for intervention), which adjusts for selection bias in observation data; inverse propensity score-weighted algorithms based on confounding balance; and causal inference algorithms based on the dual robust estimation algorithm (DRL) and enhanced inverse propensity score-weighted (AIPW). Machine learning-based methods include: decision tree-based methods include causal inference models based on classification and regression trees (CART) [9]; causal inference algorithms based on Bayesian additive regression trees (BART); causal inference algorithms based on decision trees and random forests; causal multi-task Gaussian processes (CMGP) based methods; causal inference models based on generative adversarial networks (GANs); multi-task causal models based on biased score weights; edge structure models based on recurrent neural networks (RMSNs), which can process time-series observation data; and G-Net proposed a recurrent neural network model based on the G-computation formula.

[0056] Taking RMSNs as an example, Figure 2 This is a schematic diagram of the structure of an RMSNs provided in this application. Figure 2 As shown, its encoder, propensity score predictor, and decoder are typically constructed using recurrent neural networks (RNNs). Time-dependent covariates are used as input to the decoder, encoded by the RNN to generate state vectors at each time step. The decoder then predicts the intervention outcome, and finally, the propensity score is used for inverse propensity probability weighting. The causal effect of the intervention is estimated based on the factual outcome and the predicted counterfactual outcome.

[0057] The existing algorithms mentioned above all have certain limitations. Statistical causal inference algorithms are inefficient and have low accuracy when processing large amounts of historical observation data, and the optimization process is also difficult. While existing machine learning-based causal inference algorithms can handle large amounts of observation data, their interpretability is low, meaning that the model output has low weighting for the contribution of input features, resulting in unreliable causal results. Therefore, this application proposes a new causal inference algorithm that can simultaneously achieve both interpretability and accuracy in causal analysis.

[0058] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0059] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0060] Example 1

[0061] Figure 3 This is a schematic flowchart of the health analysis method provided in Embodiment 1 of this application. Figure 3 As shown, the method includes:

[0062] S11, Obtain historical health observation dataset; wherein, the historical health dataset includes historical health observation data of several individuals, and the historical health observation data includes time series data of health characteristic indicators, health intervention measures, health risk outcomes, and individual inherent health data.

[0063] Historical health observation datasets refer to retrospective data sets from the real world used for model training and validation; they are health observation data that actually occurred under natural conditions.

[0064] An individual refers to a single, independent unit of observation within a dataset, namely a single patient or healthy participant. Each individual has a unique, longitudinal, time-series data record.

[0065] Health characteristics indicators are dynamic variables that are measured or assessed at a specific point in time to reflect an individual's health status. Examples of health characteristics indicators include age, weight, blood pressure, blood glucose levels, blood lipid levels, BMI, electrocardiogram characteristics, and quantitative indicators in medical imaging reports.

[0066] Health interventions refer to proactive behaviors or treatments applied to an individual. Examples of health interventions include taking specific medications, undergoing surgery, following specific diet or exercise programs, and receiving vaccinations.

[0067] Health risk outcomes refer to the outcome variables used to measure health status observed at the next time point after the implementation of a health intervention. They directly reflect the potential impact of the health intervention. For example, health risk outcomes may include the risk of disease, whether a disease recurs, whether specific complications occur, whether physiological indicators are met, readmission rate, survival status, and changes in quality of life scores.

[0068] Individual-specific health data refers to observational data about an individual that does not change over time, used to characterize their underlying health background. Examples of individual-specific health data include gender, date of birth, ethnicity, genetic markers, allergy history, and family medical history.

[0069] This approach can obtain historical health observation datasets from publicly available clinical data sources or those provided by individuals. Based on this, data from different data sources can be correlated and merged using the individual's unique identifier; all timestamps can be unified to the same time zone and sorted by individual and timestamp to form a longitudinal time-series record for each individual; variables can be defined and coded for health characteristic indicators, health interventions, health risk outcomes, and individual-specific health data; and missing and outlier values ​​in the historical health observation data can be preprocessed.

[0070] Specifically, given a historical health observation dataset ,in Indicators representing an individual's health characteristics as they change over time; This represents an individual's inherent health data that does not change over time. and It is a covariate; These are health interventions chosen by the individual; It refers to the outcome in the period following an individual's intervention, i.e., the health risk outcome. It is also a covariate. Part of it.

[0071] S12, Construct a causal inference model; wherein, the causal inference model includes an encoder and a decoder, the encoder employs a variational gated recurrent neural network to extract noise-robust latent variables from time-series data, and the decoder is used to predict the causal effect of health interventions on health risk outcomes based on the latent variables.

[0072] Variational modeling refers to outputting the parameters of a probability distribution instead of a definite hidden state value. It uses reparameterization techniques to sample from this distribution, enabling the model to learn and express the uncertainty in the data.

[0073] Variational gated recurrent neural networks (GRNNs) are hardware-friendly, lightweight sequence modeling units. They achieve this by changing the hidden states from deterministic vectors to low-dimensional distributed parameters (mean, variance).

[0074] Memory access optimization: During backpropagation, the caching requirement for intermediate activation values ​​can be reduced, thereby reducing the pressure on video memory bandwidth.

[0075] Enhanced computational stability: Distributed representations are inherently insensitive to input noise, reducing drastic changes in internal activation values ​​caused by fluctuations in input data quality (profanity), thereby suppressing gradient anomalies, improving the utilization of processor (especially GPU stream processor) computing units, and avoiding a large amount of invalid computation caused by gradient problems.

[0076] In causal inference, health interventions can be simplified into a binary classification problem, i.e. For example, suppose This is the control group, meaning no medication was taken. It is the experimental group, indicating that taking the drug is beneficial to the individual. ,That may be or Individuals in Health risk outcomes at that time With individuals Health risk outcomes at that time Subtracting the two yields the individual causal effect. However, for individuals in the control group... Because no medication was taken Therefore, the result It can be observed; it is a factual result. And for... It can never be achieved, so Unobservable results are counterfactual. Therefore, causal inference models need to infer counterfactual results from historical health observation data and then calculate the causal effect.

[0077] Specifically, definable Health intervention measures Predicted values ​​of health risk outcomes at that time yes The model considers historical records of individual health characteristics, inherent health data, and health interventions at any given time. Therefore, the goal of the constructed causal inference model is:

[0078] ;

[0079] That is, estimate After a moment until The causal effect within, in the formula, .

[0080] Understandably, causal relationships in time series data are more complex than those in static data. Figure 4 This is a schematic diagram of causal relationships under time-series data provided in Embodiment 1 of this application.

[0081] like Figure 4As shown, this relates to the observation data model, which is the actual data generation process observed in the real world. In time... Latent variables at time Influences an individual at the present time Health characteristics indicators at that time Next time Latent variables at time Health characteristics indicators Current time Health characteristics indicators at that time Influences the individual in the next moment Latent variables at time Health characteristics indicators and health interventions Current time Health intervention measures Influences the individual in the next moment Latent variables at time Health characteristics indicators and health interventions The next time Health characteristics indicators at that time Influences the individual in the next moment Health intervention measures The next time Latent variables at time Influences the individual in the next moment Health characteristics indicators at that time and health interventions .

[0082] For intervention models, that is, idealized theoretical models used for causal inference. In time... Latent variables at time Influences an individual at the present time Health characteristics indicators at that time Next time Latent variables at time Health characteristics indicators ,time Health characteristics indicators at that time Influences the individual in the next moment Latent variables at time and health characteristic indicators ,time Health intervention measures Influences the individual in the next moment Health characteristics indicators at that time Latent variables and health interventions The next time Latent variables at time Influences the individual in the next moment Health characteristics indicators at that time and health interventions .

[0083] Therefore, this application proposes an intervention impact prediction algorithm based on variational gated recurrent neural networks, which uses time-series observation data to train a supervised causal inference model.

[0084] The causal inference model constructed in this application can be broadly divided into an encoder and a decoder, which consist of a variational gated recurrent neural network and a fully connected neural network. The variational gated recurrent neural network uses tanh and sigmoid activation functions, the fully connected layers use ELU (Exponential Linear Unit) as the activation function, and the intervention selection prediction classification network uses the softmax activation function. The encoder and decoder networks use Dropout, enabling the model to also provide uncertainty intervals for counterfactual results, thus giving the model's predictions higher confidence.

[0085] In some embodiments, optionally, the sufficiency assumptions of the causal inference model include:

[0086] Hypothesis 1: The potential outcomes of a health intervention are equal to the observed outcomes;

[0087] Assumption 2: For all individuals in the historical health observation dataset, the probability of selecting any health intervention is greater than zero, regardless of health characteristic indicators;

[0088] Hypothesis 3: For all individuals in the historical health observation dataset, health interventions are independent of the potential outcomes of health interventions.

[0089] In this application, counterfactual outcomes are predicted based on the above three sufficiency assumptions to determine causal effects from time-series observation data. Among these,

[0090] Assumption 1 is the assumption of stable individual intervention outcomes. That is, if the intervention plan is selected... So, intervention The potential outcome equals the observed outcome. Mathematical representation: ;

[0091] Assumption 2, for all individuals in the dataset, The probability of choosing any intervention is greater than zero, and the probability of choosing any covariate is greater than zero. Irrelevant. Mathematical representation: ;

[0092] Assumption 3 is an unconfusion-free assumption. That is, given the covariates... Potential results Intervention with choice Independent. Mathematical representation: .

[0093] Therefore, counterfactual results can be estimated using deep learning models. Specifically, this application uses... Calculate the potential outcomes of health risk results and learn the conditional distribution. The encoder; using Predictive intervention, learning condition distribution .in, In the model, use This represents the extracted confounding variables.

[0094] S13, Based on a preset loss function, the causal inference model is trained using the historical health observation dataset to obtain a health analysis model, which is then used to analyze the health observation data of the target object; wherein, the preset loss function includes a regularization term based on the probability weighting of enhanced adverse tendency.

[0095] A predefined loss function is a target function that is defined in advance before model training to quantify the performance of a causal inference model.

[0096] The regularization term based on the probability weighting of the enhanced inverse tendency is a "system scheduling strategy to guide the training process to converge quickly".

[0097] A health analysis model is a model obtained by training a causal inference model using historical health observation datasets. It can learn the causal laws of health status evolution from historical data.

[0098] This scheme uses historical health observation datasets as input. The encoder receives time-series data of health characteristics, health interventions, and individual-specific health data to predict health risk outcomes at a single time step and extract confounding variables. The intermediate states of the variational gated recurrent neural network serve as latent variables. The decoder receives pre-defined health interventions, predicted health risk outcomes, and individual-specific health data. The latent variables output by the encoder are used as the initial states of the decoder's variational gated recurrent neural network to predict counterfactual outcomes and select appropriate health interventions.

[0099] The training objective of the encoder and decoder is to minimize the loss function of the outcome prediction network and the intervention prediction network, and to train the model by minimizing the loss function.

[0100] In this embodiment, optionally, the regularization term based on the enhanced inverse tendency probability weighting is:

[0101] ; (1)

[0102] In the formula, Indicates time Health risk outcomes, To enhance the prediction function, ;

[0103] This represents the time predicted by health interventions and latent variables based on time t. Health risk outcomes, This represents the health intervention predicted at time t based on latent variables at time t. , , and Indicates model parameters.

[0104] It is understandable that in observational time-series data, the allocation of interventions is not random. For example, doctors may be more inclined to prescribe a certain drug to patients with more severe conditions. This phenomenon of intervention allocation being related to individual characteristics is called selection bias.

[0105] To avoid estimation bias caused by selection bias, this application introduces a regularization term in the loss function based on the enhanced inverse tendency probability weighting.

[0106] Specifically, an enhanced prediction function can be defined. To estimate health risk outcomes and health interventions .in, For the newly introduced model parameters, This is the inverse probability weighted term.

[0107] Based on enhanced prediction function Define regularization terms For the results of actual observation Compared with the prediction results after weighting adjustment by enhanced adverse propensity The difference between them is expressed in this application as mean square error.

[0108] Finally, a regularization term based on the probability weighting of the enhanced adverse tendency is introduced into the original preset loss function to form a new optimization objective:

[0109] ;

[0110] In the formula, is a hyperparameter used to control the strength of regularization. The regularization process involves solving the following formula:

[0111]

[0112] By solving the above formula, the entire training process satisfies the augmented inverse probability weighted estimation, that is, it satisfies the nonparametric estimation equation.

[0113] In this scheme, the regularization term introduces a dynamic weight adjustment mechanism into the loss function, which has the following effects: 1) Guiding the optimization direction: In the early stages of training, it acts like a "navigation system," guiding gradient descent to avoid "erroneous optimization paths" caused by proliferative biases that are difficult for the model to correct later. 2) Reducing the number of iterations: By avoiding lingering on erroneous paths, the model can reach a robust convergence point with fewer training iterations, thereby directly saving CPU / GPU computing time and energy consumption. 3) Improving system reliability: After deployment, the obtained model exhibits stability to small perturbations (proliferation) in the input data, reducing the frequency of recalculation or alarms triggered by large fluctuations in prediction results in the online inference service, and improving the overall reliability of the service system.

[0114] In this embodiment, optionally, the preset loss function further includes a reconstruction loss term and a regularization term based on variational information bottleneck;

[0115] The reconstruction loss term is used to constrain the ability of the latent variables to reconstruct health intervention measures and health risk outcomes; the regularization term based on variational information bottleneck is used to constrain the distribution of the latent variables to be close to a preset prior distribution.

[0116] Selection bias in historical data may lead to some decision choices being correlated with certain individual characteristics; for example, older adults are more likely to avoid vaccination, resulting in incorrect conclusions. Without adjusting for selection bias caused by time dependence, the model's counterfactual estimates will be unreliable. Therefore, this application uses neural networks. Distribution of learning conditions The hidden confounding variables are regularized through variational gated recurrent neural units, which are regularization terms based on variational information bottlenecks. By leveraging the idea of ​​variational information bottlenecks, the model is forced to extract as much important information as possible and overcome selection bias.

[0117] Specifically, the preset loss function can be expressed as: ;in, Indicates the preset loss function. Represents the reconstruction loss term. This represents the regularization term based on the variational information bottleneck. Among them,

[0118] ;

[0119] ;

[0120] In the formula, This represents the number of individuals in the historical health observation dataset. This indicates the number of time steps in the time series data. This indicates the number of Monte Carlo samplings performed from the latent variable distribution. Represents an individual In time Health intervention measures, Represents an individual In time Health risk outcomes, Represents an individual In time Health characteristic indicators, Indicates time Potential variables, This represents the conditional distribution of the health risk outcomes predicted by the decoder. This represents the conditional distribution of health interventions predicted by the decoder. Represents the relative entropy between two distributions. This represents the posterior distribution of the latent variables in the encoder output. This represents the prior distribution of the latent variables. This is a hyperparameter.

[0121] In this scheme, the model is reconstructed to ensure that it accurately fits the historical patterns in the historical health observation dataset by reconstructing the loss term. At the same time, the variational information bottleneck regularization term is used to actively compress potential representations and filter time-varying mixed information, thereby driving the model to learn a more generalizable state representation.

[0122] In this embodiment, optionally, the causal inference model is trained using the historical health observation dataset based on a preset loss function to obtain a health analysis model, including:

[0123] The historical health observation dataset is divided into several mini-batches as training data for each iteration. For each individual in each training batch:

[0124] The individual's historical health observation data is input into the causal inference model to obtain predictions of health intervention measures and health risk outcomes;

[0125] Based on the enhanced prediction function, the enhanced prediction value is determined according to the prediction of the health intervention measures and the prediction of the health risk outcomes;

[0126] Based on the preset loss function, the loss function value containing a regularization term weighted by the enhanced adverse tendency probability is determined according to the enhanced predicted value and the observed value of the health risk result.

[0127] Based on the loss function value, the model parameters of the causal inference model are adjusted;

[0128] The health analysis model is obtained when the model converges or the number of iterations reaches the preset number.

[0129] In this scheme, an iterative optimization algorithm is used to train the constructed causal inference model. Specifically, the historical health observation dataset is divided into several batches of small-batch training data. For each training iteration, the following operations are performed: For the historical health observation data of each individual in the current small-batch data, it is input into the causal inference model to obtain the model output health intervention propensity score prediction and health risk outcome prediction; Based on the health intervention propensity score prediction and the observed actual health interventions, the inverse propensity score weight is calculated; The health risk outcome prediction is combined with the inverse propensity score weight and a learnable correction parameter to calculate the enhanced prediction value; According to the enhanced prediction value and the observed real health risk outcome, the mean squared error is calculated as a double robust loss term; The double robust loss term, the reconstruction loss term, and the regularization term based on the variational information bottleneck are combined according to preset weights to form a complete loss function value; The gradient of the loss function value with respect to all model parameters is calculated using the backpropagation algorithm; According to the gradient, the model parameters are updated using stochastic gradient descent or its variant optimization algorithm; The above iterative process is repeated until the model's performance on the validation set converges or the preset maximum number of training iterations is reached, finally obtaining the trained health analysis model.

[0130] The technical solution provided in this embodiment employs a variational gated recurrent network to transform the hidden state from a deterministic vector into a low-dimensional distributed parameter, thereby optimizing memory access and enhancing computational stability. Furthermore, by introducing a regularization term based on enhanced inverse tendency probability weighting into the preset loss function, the gradient descent direction is guided, reducing the number of model training iterations and improving the system reliability after deployment. This reduces the computational complexity of model training, decreasing training time while achieving the same prediction accuracy; reduces memory usage during model inference, enabling deployment on resource-constrained embedded devices; improves the numerical stability of the training process, reducing the frequency of gradient anomalies and minimizing resource waste due to training interruptions; and enhances the throughput and response speed stability of the online inference service.

[0131] Example 2

[0132] To enable those skilled in the art to more clearly understand the technical solution provided in this embodiment, a preferred implementation method is also provided as a specific solution. It should be noted that this embodiment is merely an example and does not limit the specific scope of protection of this solution.

[0133] To address the problems of low counterfactual prediction accuracy and poor interpretability in existing causal inference models, this invention provides a causal inference model that specifically solves the following problems:

[0134] Existing statistical causal inference models suffer from low data processing efficiency and low analysis accuracy, and the optimization process is also quite difficult. By introducing a variational information bottleneck mechanism, the efficiency and accuracy of data analysis results can be improved.

[0135] Existing machine learning-based causal inference models suffer from selection bias and poor interpretability. By improving the structure of the encoder and decoder, as well as interpretable latent variables, the credibility of data analysis results can be enhanced.

[0136] In the technical solution of this invention, the encoder includes a latent variable learning network, a first intervention prediction network, and a first outcome prediction network. The latent variable learning network includes a variationally gated recurrent unit based on variational information bottlenecks, used to determine latent variables based on individual inherent health data, health characteristic indicators, and health intervention measures. The first intervention prediction network is used to predict the health intervention measures at the current moment based on the latent variables, and the first outcome prediction network is used to reconstruct the health risk outcome at the next moment based on the latent variables. The decoder includes a latent variable transit network, a second intervention prediction network, and a second outcome prediction network. The latent variable transit network is used to recursively update latent variables during counterfactual prediction. The second intervention prediction network is used to predict the health intervention measures at the corresponding moment based on the latent variables. The second outcome prediction network is used to predict the health risk outcome at future moments based on the updated latent variables and the selection results of health intervention measures, and to determine the causal effect of health intervention measures on health risk outcomes.

[0137] By designing an encoder-decoder architecture, the two tasks of "learning historical patterns from observation data" and "simulating counterfactual outcomes under different future interventions" are structurally separated and synergistically optimized, thereby resolving the inherent contradiction of simultaneously pursuing prediction accuracy and unbiased causal inference in time series data. Figure 5 This is an architectural diagram of the causal inference model provided in Embodiment 2 of this application, as shown below. Figure 5 As shown, the architecture of the causal inference model includes the encoder and the decoder, wherein,

[0138] The encoder includes a latent variable learning network, a first intervention prediction network, and a first outcome prediction network; wherein...

[0139] The latent variable learning network includes a variationally gated recurrent unit based on variational information bottleneck, used to determine latent variables based on individual inherent health data, health characteristic indicators, and health intervention measures;

[0140] The first intervention prediction network is used to predict health interventions at the current moment based on latent variables;

[0141] The first result prediction network is used to reconstruct the health risk outcome for the next time step based on latent variables;

[0142] The decoder includes a latent variable transit network, a second intervention prediction network, and a second outcome prediction network; wherein the latent variable transit network is used to recursively update latent variables during the counterfactual prediction process;

[0143] The second intervention prediction network is used to predict health interventions at corresponding times based on latent variables;

[0144] The second outcome prediction network is used to predict future health risk outcomes based on updated latent variables and the selection results of health interventions.

[0145] In the encoder, the input is time-series data of health characteristic indicators. Time series data of health interventions and individual-specific health data The data at each time step is processed using a variationally gated recurrent neural unit based on the variational information bottleneck. This unit receives the individual's inherent health data, health characteristic indicators for the current time step, and the latent variable distribution and health interventions from the previous time step, outputting the latent variable distribution for the current time step. At the final time step, the encoder outputs the distribution parameters of the latent variables. And by employing reparameterization techniques, samples are obtained from it to obtain latent variables. .

[0146] Latent variables generated at the final time step It will be used for two prediction tasks. One is to input the first intervention prediction network to predict the health intervention based on latent variables and pre-set health interventions. Predicting health interventions at the current time step The other is an input-first outcome prediction network, which, based on latent variables and pre-set health interventions, predicts the outcome. Reconstructing health risk outcomes at the next time step .

[0147] In the decoder, the latent state is the output of the encoder. As its initial hidden state, at each subsequent time step, a variationally gated recurrent neural unit based on a variational information bottleneck processes the input data, which includes the individual's inherent health data. Pre-set health intervention measures at the current time step and the predicted results of health risk outcomes at the previous time step. Output the distribution parameters of the latent variables. And by employing reparameterization techniques, samples are obtained from it to obtain latent variables. .

[0148] And at each time step, the latent variables and pre-set health interventions Used for two prediction tasks, one of which is input into a second intervention prediction network to predict health interventions at corresponding time steps. The other is an input to a second outcome prediction network to predict health risk outcomes at the next time step. .

[0149] In this embodiment, optionally, latent variables are determined based on individual-specific health data, health characteristic indicators, and health interventions, including:

[0150] For each time step in the time series data, the hidden state is updated using a variational gated cyclic unit based on the variational information bottleneck;

[0151] The update rule of the variational gated recurrent unit based on the variational information bottleneck is as follows: Based on the hidden state and health intervention measures of the previous time step, and the health characteristic indicators of the current time step, calculate the reset gate, update gate, mean of candidate hidden states, and logarithmic variance of candidate hidden states at the current time step; sample random noise from a standard normal distribution, and determine candidate hidden states based on the random noise, the mean of candidate hidden states, and the logarithmic variance of candidate hidden states; determine the hidden state at the current time step based on the hidden state of the previous time step, the update gate of the current time step, and the candidate hidden states; determine the distribution of latent variables based on the last hidden state of the time series data.

[0152] The latent variables are identified by sampling from the distribution of the latent variables using reparameterization techniques.

[0153] Information bottleneck refers to the process of minimizing redundant input information while maximizing the retention of information relevant to the prediction objective.

[0154] The variational gated recurrent unit based on variational information bottleneck provided in this application introduces uncertainty into the gated recurrent unit, models the hidden state inside the neuron as a Gaussian distribution, and incorporates the idea of ​​information bottleneck. Its structure is as follows: Figure 6 As shown.

[0155] Specifically, at each time step, the state transition function is:

[0156] ;

[0157] in, It is a non-linear transformation function. These are parameters inside the RNN.

[0158] The internal state transitions of a variationally gated recurrent neuron are as follows:

[0159] Reset Door: ;

[0160] Update Gate: ;

[0161] Mean of candidate hidden states: ;

[0162] Variance of candidate hidden states: ;

[0163] sampling: ;

[0164] Candidate hidden state: ;

[0165] Hidden status update: ;

[0166] in, , , and Represents the weight matrix. , , and Indicates the bias term. This represents random noise.

[0167] This approach forces neurons to extract only the most important information at each time step, removing time-dependent confounding variables and improving the interpretability of the data.

[0168] In this embodiment, optionally, determining the causal effect of health intervention measures on health risk outcomes includes: sampling the historical health observation dataset to obtain covariate samples; inputting the covariate samples into the encoder to obtain the latent variable distribution corresponding to the covariate samples, and sampling the latent variable distribution to obtain latent variable samples; inputting the latent variable samples and the selection results of health intervention measures into the decoder to obtain a latent outcome distribution, and sampling the latent outcome distribution to obtain a counterfactual outcome sample; and determining the causal effect of health intervention measures on health risk outcomes based on the counterfactual outcome sample.

[0169] Specifically, the do() operator can be applied to obtain the potential results. distributed:

[0170] ;

[0171] Using approximation To estimate potential outcomes; then from To obtain samples, the average causal effect is calculated based on the sampling results. The sampling rules are as follows:

[0172] ;

[0173] ;

[0174] .

[0175] In this embodiment, optionally, determining the causal effect of health interventions on health risk outcomes based on the counterfactual outcome sample includes: determining the individual causal effect of health interventions on health risk outcomes using the following formula:

[0176] ; (2)

[0177] In the formula, Indicates the first Individual causal effects of individuals Indicates the first Health risk outcomes for individuals receiving health interventions Indicates the first Health risk outcomes for individuals who do not receive health interventions;

[0178] Furthermore, the average causal effect of health interventions on health risk outcomes is determined using the following formula:

[0179] ; (3)

[0180] In the formula, Indicates average causal effect. This indicates the number of individuals receiving health interventions. This indicates the number of individuals who did not receive health interventions.

[0181] Individual causal effect refers to the difference between two counterfactual outcomes obtained when two different health interventions are applied to the same potential state of the same individual.

[0182] The average causal effect refers to the average of the individual causal effects of all individuals in a dataset.

[0183] In this approach, the average causal effect can be calculated by taking the counterfactual outcome for each individual under the two interventions, either for all individuals in the dataset or for a representative sample. Then, the average of the counterfactual outcomes for all individuals under the two interventions is taken, and the results are subtracted. Alternatively, the Monte Carlo method can be used to sample the latent states of each individual multiple times and calculate the predictions under the two interventions separately.

[0184] Determining individual causal effects can be used for precision health analysis, such as recommending the most effective treatment for individuals; and determining average causal effects can be used for public health and policy evaluation to assess the overall cost-effectiveness and promotional value of a health intervention in a population.

[0185] Based on the causal model architecture provided in this application, it is trained separately, and the training method is as follows.

[0186] Phase 1, training the encoder:

[0187] Input training set: ;

[0188] Initialize the model: ;

[0189] Model parameters: ;

[0190] Learning rate: .

[0191] training set Divided into several small batches Batch size for each iteration: ;

[0192] Through forward propagation, the inference results are calculated, and the loss function is also calculated:

[0193] ;

[0194] ;

[0195] KL;

[0196] ;

[0197] Optimize the neural network using backpropagation based on the calculated loss function:

[0198] :

[0199] ;

[0200] :

[0201] ;

[0202] :

[0203] .

[0204] The training of the encoder is completed when the model converges or the preset number of iterations is reached.

[0205] The second stage involves training the decoder.

[0206] Input training set: ;

[0207] Initialize the model: ;

[0208] Model parameters: ;

[0209] Learning rate: ;

[0210] By time step Splitting the dataset: ;

[0211] For each individual and each time step Calculate the hidden variables:

[0212] ;

[0213] training set Divided into several small batches Batch size for each iteration: ;

[0214] Through forward propagation, the inference results are calculated, and the loss function is also calculated:

[0215] ;

[0216] ;

[0217] ;

[0218] ;

[0219] Optimize the neural network using backpropagation based on the calculated loss function:

[0220] :

[0221] ;

[0222] :

[0223] ;

[0224] :

[0225] ;

[0226] The training of the decoder is completed when the model converges or the preset number of iterations is reached.

[0227] Finally, for Estimate the counterfactual consequences and calculate the causal effects.

[0228] This solution provides a causal inference model, comprising an encoder and a decoder. The encoder includes a latent variable learning network, a first intervention prediction network, and a first outcome prediction network. The latent variable learning network includes a variationally gated recurrent unit based on variational information bottlenecks, used to determine latent variables based on individual inherent health data, health characteristic indicators, and health intervention measures. The first intervention prediction network is used to predict the health intervention measures at the current moment based on the latent variables, and the first outcome prediction network is used to reconstruct the health risk outcome at the next moment based on the latent variables. The decoder includes a latent variable transit network, a second intervention prediction network, and a second outcome prediction network. The latent variable transit network is used to recursively update latent variables during counterfactual prediction. The second intervention prediction network is used to predict the health intervention measures at the corresponding moment based on the latent variables, and the second outcome prediction network is used to predict the health risk outcome at future moments based on the updated latent variables and the selection results of health intervention measures, and to determine the causal effect of health intervention measures on health risk outcomes.

[0229] The beneficial effects of the technical solution of this invention include, but are not limited to:

[0230] 1) By dividing the work between the encoder and decoder, each network can focus on optimizing a single objective and achieve objective coordination rather than interference through controllable information interaction via latent variables.

[0231] 2) The variational gated recurrent neural network based on the variational information bottleneck actively compresses redundant information in latent variables during training, forcing the network to learn the decongested common core features of health intervention measures and health risk outcomes, thereby improving the accuracy of the model.

[0232] Example 3

[0233] Those skilled in the art will understand that embodiments of this solution can provide methods, systems, or computer program products. Therefore, this solution can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this solution can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0234] Therefore, this application also proposes a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the methods described in any embodiment of this application.

[0235] This solution is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this solution. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0236] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0237] These computer program instructions can also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1The steps of the function specified in one or more boxes.

[0238] Furthermore, Figure 7 This is a schematic diagram of the structure of an electronic device provided in Embodiment 3 of this application. Figure 7 As shown, this application also proposes an electronic device (or computing device) including a processor 11, a memory 12, and a computer program stored on the memory 12 and executable on the processor 11, wherein the processor 11 executes the computer program to implement the method described in any embodiment of this application.

[0239] In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, a network interface, and memory. Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media. Computer-readable media includes both permanent and non-persistent, removable and non-removable media; information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, modules of a program, or other data.

[0240] It should be noted that the computing device also includes a time-series data acquisition interface for receiving health feature indicators uploaded by wearable devices; the processor is configured with a tensor core for accelerating the computation of variational gated recurrent neural networks; the computing device is a health management server, a smart medical terminal, or an insurance risk assessment platform.

[0241] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0242] The above description is merely an embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of this application should be included within the scope of the claims of this application.

Claims

1. A health analysis method applied to a computing device, characterized in that, The method includes: Obtain a historical health observation dataset; wherein the historical health dataset includes historical health observation data of several individuals, and the historical health observation data includes time-series data of health characteristic indicators, health intervention measures, health risk outcomes, and individual inherent health data; A causal inference model is constructed, comprising an encoder and a decoder. The encoder employs a variational gated recurrent neural network to extract noise-resistant latent variables from time-series data, and the decoder is used to predict the causal effect of health interventions on health risk outcomes based on the latent variables. Based on a preset loss function, the causal inference model is trained using the historical health observation dataset to obtain a health analysis model, which is then used to analyze the health observation data of the target object; wherein, the preset loss function includes a regularization term based on the probability weighting of enhanced adverse tendency.

2. The method according to claim 1, characterized in that, The encoder includes a latent variable learning network, a first intervention prediction network, and a first outcome prediction network. The latent variable learning network includes a variationally gated recurrent unit based on variational information bottleneck, used to determine latent variables based on individual inherent health data, health characteristic indicators, and health intervention measures. The first intervention prediction network is used to predict the health intervention measures at the current moment based on the latent variables. The first outcome prediction network is used to reconstruct the health risk outcome at the next moment based on the latent variables. The decoder includes a latent variable transit network, a second intervention prediction network, and a second outcome prediction network. The latent variable transit network is used to recursively update latent variables during counterfactual prediction. The second intervention prediction network is used to predict health intervention measures at corresponding times based on the latent variables. The second outcome prediction network is used to predict health risk outcomes at future times based on the updated latent variables and the selection results of health intervention measures, and to determine the causal effect of health intervention measures on health risk outcomes.

3. The method according to claim 2, characterized in that, Potential variables were identified based on individual-specific health data, health characteristic indicators, and health interventions, including: For each time step in the time series data, the hidden state is updated using a variational gated cyclic unit based on the variational information bottleneck; The update rule for the variational gated recurrent unit based on the variational information bottleneck is as follows: Based on the hidden state and health intervention measures of the previous moment, and the health characteristic indicators of the current moment, calculate the mean of the reset gate, update gate, candidate hidden states, and the logarithmic variance of the candidate hidden states at the current moment. Random noise is sampled from a standard normal distribution, and candidate hidden states are determined based on the random noise, the mean of the candidate hidden states, and the log-variance of the candidate hidden states. The hidden state at the current moment is determined based on the hidden state at the previous moment, the update gate at the current moment, and the candidate hidden states; The distribution of latent variables is determined based on the last hidden state of the time series data; The latent variables are identified by sampling from the distribution of the latent variables using reparameterization techniques.

4. The method according to claim 2, characterized in that, Determining the causal effect of health interventions on health risk outcomes includes: The historical health observation dataset was sampled to obtain covariate samples; The covariate samples are input into the encoder to obtain the latent variable distribution corresponding to the covariate samples, and the latent variable distribution is sampled to obtain latent variable samples; The latent variable samples and the selection results of health intervention measures are input into the decoder to obtain the latent outcome distribution, and the latent outcome distribution is sampled to obtain the counterfactual outcome samples; Based on the counterfactual outcome sample, determine the causal effect of health interventions on health risk outcomes.

5. The method according to claim 4, characterized in that, Based on the counterfactual outcome sample, determine the causal effect of health interventions on health risk outcomes, including: The individual causal effect of health interventions on health risk outcomes is determined using the following formula: ； In the formula, Indicates the first Individual causal effects of individuals Indicates the first Health risk outcomes for individuals receiving health interventions Indicates the first Health risk outcomes for individuals who do not receive health interventions; Furthermore, the average causal effect of health interventions on health risk outcomes is determined using the following formula: ； In the formula, Indicates average causal effect. This indicates the number of individuals receiving health interventions. This indicates the number of individuals who did not receive health interventions.

6. The method according to claim 1, characterized in that, Based on a preset loss function, the causal inference model is trained using the historical health observation dataset to obtain a health analysis model, including: The historical health observation dataset is divided into several mini-batches as training data for each iteration. For each individual in each training batch: The individual's historical health observation data is input into the causal inference model to obtain predictions of health intervention measures and health risk outcomes; Based on the enhanced prediction function, the enhanced prediction value is determined according to the prediction of the health intervention measures and the prediction of the health risk outcomes; Based on the preset loss function, the loss function value containing a regularization term weighted by the enhanced adverse tendency probability is determined according to the enhanced predicted value and the observed value of the health risk result. Based on the loss function value, the model parameters of the causal inference model are adjusted; The health analysis model is obtained when the model converges or the number of iterations reaches the preset number.

7. The method according to claim 1 or 6, characterized in that, The regularization term based on the enhanced reverse tendency probability weighting is: ； In the formula, Indicates time Observations of health risk outcomes, To enhance the predicted values, ; This represents the time predicted by health interventions and latent variables based on time t. Health risk outcomes, This represents the health intervention predicted at time t based on latent variables at time t. and Indicates model parameters.

8. The method according to claim 1 or 6, characterized in that, The preset loss function also includes a reconstruction loss term and a regularization term based on variational information bottleneck; The reconstruction loss term is used to constrain the ability of the latent variables to reconstruct health intervention measures and health risk outcomes; the regularization term based on variational information bottleneck is used to constrain the distribution of the latent variables to be close to a preset prior distribution.

9. The method according to claim 1, characterized in that, The sufficiency assumptions of the causal inference model include: Hypothesis 1: The potential outcomes of a health intervention are equal to the observed outcomes; Assumption 2: For all individuals in the historical health observation dataset, the probability of selecting any health intervention is greater than zero, regardless of health characteristic indicators; Hypothesis 3: For all individuals in the historical health observation dataset, health interventions are independent of the potential outcomes of health interventions.

10. A computing device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1-9.

Images