System and method for lung health management
Through multimodal data manifold learning and alignment, personalized distribution modeling and time series deep learning model based on self-attention mechanism, combined with the dual optimization module, the problems of multimodal data fusion, personalized modeling and global optimization in lung health management are solved, and high-precision lung disease risk prediction and personalized health assessment are achieved.
Patent Information
- Application Number
- CN202510044382.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-11
- Publication Date
- 2025-05-27
AI Technical Summary
In the lung health management, the prior art has problems such as insufficient multimodal data fusion expression ability, inaccurate dynamic modeling of personalized features, insufficient depth of time series data analysis, and poor global optimization effect of the system.
Multimodal data manifold learning and alignment technology are used to generate shared potential feature space; through a personalized distribution modeling module, based on energy function and variational inference method, the patient's personalized health characteristics are captured; combined with a time series deep learning model based on the self-attention mechanism, comprehensive feature fusion and disease risk prediction are carried out; using dual optimization modules, the parameters of each module are jointly optimized to ensure the convergence of the system's global parameters.
It has achieved effective integration of multimodal data, accurately captured personalized health characteristics, improved the accuracy of lung disease risk prediction and systematic global optimization effect, and improved the in-depth insight and personalized evaluation capabilities of lung health management.
Smart Images
Figure CN120048481A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of health management, and specifically to a system and method for pulmonary health management. Background Art
[0002] In the field of pulmonary health management, although certain progress has been made in the existing technologies, there are still many deficiencies in dealing with complex multi-modal data, personalized differential modeling, and global optimization. These technical defects directly affect the early identification of pulmonary diseases, the accuracy of risk prediction, and the effectiveness of individualized health management. The following analyzes the deficiencies in the existing technologies from specific perspectives.
[0003] In the existing technologies, the processing of multi-modal data of patients often adopts simple splicing or linear combination methods, ignoring the non-linear distribution characteristics of each modal data in the high-dimensional space. This processing method is difficult to reflect the complex associations between time series features, static features, and environmental features. The incoordination of multi-modal data has become one of the main obstacles to accurately evaluating the pulmonary health status.
[0004] Most of the existing pulmonary health management technologies rely on population statistical models or risk assessment methods based on standard templates. These methods usually ignore the individual differential characteristics of patients and can only describe population laws, making it difficult to accurately reflect the personalized health status of patients. Especially when patients have special medical histories or are significantly different from population characteristics, the existing technologies cannot effectively capture the uniqueness of patients, resulting in large deviations in the evaluation results.
[0005] The changes in pulmonary health status often have significant time dependence and dynamic characteristics, but the existing technologies have weak analysis capabilities for time series features. Most existing methods can only capture short-term trends, ignoring the long-term dependence and global patterns of time series features. This deficiency directly limits the in-depth insight into the health status of patients, especially the early identification ability of acute disease risks. Summary of the Invention
[0006] Aiming at the deficiencies of the existing technologies, the present invention provides a system and method for pulmonary health management, which solves the problems of insufficient multi-modal data fusion and expression ability, inaccurate dynamic modeling of personalized features, insufficient depth of time series data analysis, and poor global optimization effect in the existing technologies.
[0007] To achieve the above objectives, the present invention is realized through the following technical solutions: A system for pulmonary health management includes the following steps; A multi-modal data manifold learning and alignment module, which is used to receive the multi-modal data of patients, extract the low-dimensional manifold embedding representations of each modal data, and align the low-dimensional manifold embedding representations of different modalities to generate a shared latent feature space; The personalized distribution modeling module is used to generate the personalized distribution of a patient in the shared latent feature space based on the patient's personalized features; The disease risk prediction module is used to predict the risk of lung diseases of a patient according to the time series features, static features and environmental features of the patient; The dual optimization module is used to jointly optimize the parameters of each module and dynamically adjust the global optimization objective of the system.
[0008] Preferably, the multimodal data manifold learning and alignment module includes; The data input unit is used to receive the multimodal data of the patient, and the multimodal data includes time series data, static feature data and environmental data; The manifold dimensionality reduction unit is used to perform dimensionality reduction processing on the time series data, static feature data and environmental data through a non-linear manifold learning algorithm to obtain corresponding low-dimensional manifold representations; The manifold alignment unit is used to align the low-dimensional manifold representations of different modalities to generate a shared latent feature space.
[0009] Preferably, the manifold dimensionality reduction unit constructs low-dimensional embedded representations of each modality data through a non-linear dimensionality reduction method, and the non-linear dimensionality reduction method includes a manifold learning algorithm based on neighborhood structure. Among them, the low-dimensional manifold representation reflects the non-linear distribution structure of the patient's multimodal data in the high-dimensional space.
[0010] Preferably, the personalized distribution modeling module includes; The distribution generation unit is used to generate the latent distribution of the patient based on the patient's personalized features. This latent distribution is defined by an energy function and reflects the personalized characteristics of the patient in the latent feature space; The distribution optimization unit is used to optimize the latent distribution of the patient based on the variational inference method to align it with the feature distribution of the global model.
[0011] Preferably, the distribution optimization unit is optimized by minimizing the difference between the personalized distribution and the global model distribution, where the personalized distribution is generated based on the individual characteristics of the patient, and the global model distribution is generated by the population data in the multimodal feature space.
[0012] Preferably, the disease risk prediction module includes; The time series feature processing unit is used to perform dynamic feature extraction on the time series data; The comprehensive feature fusion unit is used to fuse the time series features, static features and environmental features to generate disease risk prediction inputs; The risk scoring unit is used to generate the disease risk score of the patient according to the comprehensive features.
[0013] Preferably, the risk scoring unit calculates the risk probability of the patient's lung disease through a time series model, where the time series model includes a deep learning model based on the self-attention mechanism.
[0014] Preferably, the dual optimization module jointly optimizes the personalized distribution modeling objective, the manifold alignment objective, and the disease risk prediction objective of the patient to ensure the convergence of the global system parameters. Among them, the objectives of the joint optimization include the dynamic adjustment of the patient's personalized distribution, the alignment accuracy of the modal feature space, and the accuracy of risk prediction.
[0015] Preferably, the dual optimization module updates the parameters through an alternating optimization algorithm, which includes fixing the parameters of one module and optimizing the parameters of the other module, and finally realizes the minimization of the global loss function.
[0016] A method for lung health management includes the following steps; S1. Receive the time series data, static feature data, and environmental data of the patient, and generate low-dimensional manifold representations of each modality through a non-linear manifold learning algorithm; S2. Align the low-dimensional manifold representations of each modality to generate a shared latent feature space; S3. Generate the latent distribution of the patient based on the patient's personalized features, and optimize the latent distribution through the variational inference method; S4. In the shared latent feature space, predict the risk of lung disease according to the time series features, static features, and environmental features of the patient; S5. Jointly optimize the parameters of the multi-modal data alignment, patient personalized distribution modeling, and disease risk prediction modules through the dual optimization method to ensure the global convergence and prediction accuracy of the system.
[0017] The present invention provides a system and method for lung health management. It has the following beneficial effects: 1. By adopting the technical solution of multi-modal data manifold learning and alignment, the present invention realizes the unified expression of time series data, static feature data, and environmental feature data, achieving the technical effect of effective fusion between different data modalities. Compared with the existing solutions that only perform simple splicing or linear processing on multi-modal data, the problems of insufficient information correlation between modalities and inconsistent feature expressions are solved.
[0018] 2. Through the personalized distribution modeling module, the present invention adopts the distribution generation and optimization method based on the energy function to accurately capture the personalized health characteristics of the patient, achieving the technical effect of dynamically adapting to the changes in the patient's characteristics. Compared with the existing solutions that rely on the group feature distribution model, the problems of insufficient capture of individual differential features and difficulty in achieving accurate health assessment are solved.
[0019] 3. By combining a time - series deep - learning model based on self - attention mechanism and a comprehensive feature fusion mechanism, the present invention accurately predicts the risk of lung diseases in patients, achieving the technical effect of real - time dynamic assessment of health status. Compared with the existing models that only rely on single - modality data or lack the ability to capture dynamic features, it overcomes the problems of insufficient sensitivity to time - series features and low prediction accuracy.
[0020] 4. Through the dual - optimization module, the present invention adopts a technical solution of joint optimization and dynamic weight adjustment, realizing the global parameter convergence of multiple modules of the system and achieving the technical effect of improving the overall performance and stability of the system. Compared with the existing solutions in which each module is optimized independently, resulting in the dispersion of global performance, it solves the technical bottlenecks of poor system coordination and insufficient global optimization effect. BRIEF DESCRIPTION OF THE DRAWINGS
[0021] Figure 1 is the system framework diagram of the present invention; Figure 2 is the schematic diagram of the multi - modality data manifold learning and alignment module of the present invention; Figure 3 is the schematic diagram of the personalized distribution modeling module of the present invention; Figure 4 is the schematic diagram of the disease risk prediction module of the present invention; Figure 5 is the flowchart of the method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0022] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.
[0023] Please refer to the attached Figure 1 - attached Figure 4 , the embodiments of the present invention provide a system for lung health management, including the following steps; A multi - modality data manifold learning and alignment module, which is used to receive the multi - modality data of patients, extract the low - dimensional manifold embedding representation of each modality data, and align the low - dimensional manifold embedding representations of different modalities to generate a shared latent feature space; Specifically, in this embodiment, the key low-dimensional feature representations are extracted from the high-dimensional features through a non-linear manifold learning method. Subsequently, through a modality alignment algorithm, the time series data, static feature data, and environmental data are aligned to a shared latent feature space. This latent feature space can better preserve the key relationships between modalities and the non-linear characteristics in the high-dimensional space; In this embodiment, the multi-modal data comes from a variety of data collection devices, including but not limited to wearable devices, electronic medical record systems, and environmental monitoring devices. The data often has different feature dimensions and data types in its original form. The time series data contains dynamic features that change over time, while the static features and environmental features are mainly attribute data with fixed dimensions. In data processing, it is necessary to first preprocess the data of different modalities; Time series data represents the set of dynamic features of the patient within the time range, where is the feature dimension, and the changes in blood oxygen saturation and respiratory rate can both be expressed as time series features. The static features include the basic information of the patient, including age, gender, genetic information, and past medical history. The environmental features include environmental variable information such as the air quality index (AQI) and PM2.5 concentration in the area where the patient is located; This module first standardizes the above multi-modal data to ensure that the feature values between modalities are within the same range, and uses the sliding window technique to segment the time series data to extract short-term dynamic features; The manifold learning algorithm can embed high-dimensional features into a low-dimensional manifold space in a neighborhood-preserving manner. For time series data, a non-linear manifold learning algorithm based on the neighborhood structure can be used; For time series data , the objective function of manifold dimensionality reduction can be defined as;
[0024] where: is the dimensionality reduction function of the time series data; is the weight matrix, representing the neighborhood relationship between points and in the original high-dimensional space; is the dimension of the time series features after dimensionality reduction, usually satisfying For the static features and environmental features , the dimensionality reduction functions are and , the goal is to map the global characteristics in the high-dimensional space to a lower-dimensional latent space; After completing the dimensionality reduction operations for each modality, the system needs to align the low-dimensional manifold representations of different modalities. The manifold representations of different modalities are distributed in independent manifold spaces, and these low-dimensional manifolds are embedded into a shared latent space , and modality alignment can be achieved by minimizing the alignment error between modalities. The objective function is defined as follows;
[0025] where: is the alignment mapping function; are the time series features, static features, and environmental features after dimensionality reduction, respectively; is the number of samples; During the alignment process, weight parameters can also be introduced to adjust the importance of features in each modality. For patients with large changes in the air quality index, higher weights can be assigned to environmental features to better capture the relationship with the patient's health status;
[0026] where: are the weight parameters of the time series features, static features, and environmental features, respectively; is the energy mapping function of features in different modalities; The energy function can be constrained by introducing a regularization term. To avoid the patient's features deviating too much from the population feature distribution, an influence term of the population features can be added to the energy function;
[0027] where: is the regularization coefficient; is the mean of the global feature distribution; Through the implementation of this module, the latent features of multi-modal data are effectively extracted and integrated.
[0028] The personalized distribution modeling module is used to generate the personalized distribution of the patient in the shared latent feature space based on the patient's personalized features; Specifically, in this embodiment, it is used to generate the unique personalized distribution of the patient according to the low-dimensional embedded features of the patient in the shared latent feature space. This module relies on the unified latent feature space provided by the aforementioned multi-modal data manifold learning and alignment module , through dynamic distribution modeling, the differences in the individual health characteristics of patients can be reflected, the unique health status of patients can be captured, and the problem that the existing system generally relies on group models and ignores individual differences can be solved; In this embodiment, the health status of the patient is in the latent feature space The distribution has a certain degree of non-linearity and dynamics. By generating the probability distribution unique to the patient , which represents the feature distribution of the patient in the latent space, and by optimizing and adjusting this distribution to make it more in line with the personalized characteristics of the patient; The personalized distribution of the patient is defined as a distribution form based on the energy function;
[0029] Where: represents the latent features of the patient; is the distribution parameter; is the energy function of the patient's personalized distribution, which reflects the characteristic deviation of the patient in the latent feature space; The denominator part is the normalization term, ensuring that is the probability distribution; The energy function can be modeled by a neural network and learned from the feature input of the patient. For the input feature , the energy function can be defined as;
[0030] Where: is a learnable mapping function for generating the energy value of the latent space according to the input features; Specifically, the design of the energy function can be optimized by considering the feature weights of the patient's specific modality. Time series data may play a dominant role, while static features have less influence. At this time, the energy function can be defined as;
[0031] Where: is the regularization coefficient; is the mean of the global feature distribution; Distribution optimization In order to optimize the personalized distribution of the patient , this module adopts the variational inference method to minimize the difference between the personalized distribution and the true distribution , and the optimization goal can be expressed as the following loss function;
[0032] Wherein: KL is the KL divergence between the personalized distribution and the global distribution; is the log-likelihood term, reflecting the ability of the personalized distribution to generate patient characteristics; Dynamically update the personalized distribution parameters through an online learning method , for the newly input patient characteristics, the system first infers the initial distribution in its latent feature space, and then updates the parameters through gradient descent to make the personalized distribution more conform to the current state of the patient; The optimized personalized distribution , will be used as one of the core inputs of the disease risk prediction module, providing the patient-specific distribution characteristics for the risk prediction model.
[0033] The disease risk prediction module is used to predict the pulmonary disease risk of a patient based on the patient's time series characteristics, static characteristics, and environmental characteristics; Specifically, in this embodiment, based on the shared latent feature space generated by the multi-modal data manifold learning and alignment module, and the patient-specific distribution generated by the personalized distribution modeling module, the accurate prediction of the pulmonary disease risk is realized. By combining the dynamic features of the time series data, the static features, and the global influence of the environmental data, a risk score that can reflect the current health state of the patient is generated. The prediction result of the model is directly provided to the clinician or the patient to support the early intervention and personalized treatment of the pulmonary disease; In this embodiment, the input of the module includes three parts; The shared low-dimensional latent feature space obtained from the multi-modal data manifold learning and alignment module ; The patient-specific distribution obtained from the personalized distribution modeling module ; The comprehensive feature representation of the dynamic time series features , static features and environmental features ; Through these inputs, the module uses a deep learning-based risk prediction model to generate the disease risk score of the patient; The time series feature processing unit is used to extract features from the input dynamic feature data; The time series features represent the low-dimensional dynamic features of the patient within the time range, including feature dimensions. To capture the short-term fluctuations and long-term trends of the patient’s health status, this unit uses a time series model based on the self-attention mechanism. The self-attention mechanism can effectively extract the global dependency features of the time series by calculating the relationship weights between each time point in the time series. The calculation process includes the following steps:
[0034] in: denote query, key, and value matrices respectively; is the dimension of the key vector; is the length of the time series; Through multiple multi-head attention mechanisms and feed-forward network layers, global and local features of the time series are extracted to generate high-order representations of the time series. ,in It is a high-order feature dimension; The combined feature fusion unit is used to combine time series features , static features and environmental characteristics Perform feature fusion; Feature fusion can be achieved through a fully connected network, whose input is the feature vectors of different modalities and the output is a comprehensive feature representation , the fusion process can be expressed as;
[0035] in: are the weight matrices of time series, static features, and environmental features respectively; is the bias vector; is the activation function; The risk score generation unit is based on comprehensive features Generate a lung disease risk score for your patient ,Specifically, the rating generation process can be mapped to a scalar output through a fully ,connected layer;
[0036] in: is the risk score weight matrix; Generate bias vector for scoring; Output is a probability value in the range [0,1], indicating the probability that the patient is at risk of acute exacerbation or worsening of lung disease; The risk score generation unit can be extended by introducing a regularization term based on the personalized distribution to enhance the adaptability of the model to individual characteristics. Specifically, during the score generation process, the following optimization objective can be introduced;
[0037] where: is the true disease risk label of the -th patient; is the predicted disease risk score; is the number of patient samples; is the regularization coefficient; The prediction accuracy can be improved through transfer learning techniques. For patient groups with similar characteristics, a global model can be first trained on large-scale patient data, and then the model can be adapted to the personalized distribution of new patients through fine-tuning .
[0038] The dual optimization module is used to jointly optimize the parameters of each module and dynamically adjust the global optimization objective of the system Specifically, in this embodiment, it is used to jointly optimize the parameters of the multi-modal data manifold learning and alignment module, the personalized distribution modeling module, and the disease risk prediction module to ensure the optimality and stability of the global system performance. By introducing dual variables, this module combines the optimization objectives of different sub-modules into a unified optimization framework to achieve dynamic adjustment and joint optimization of the parameters; In this embodiment, the manifold alignment module may give priority to the alignment accuracy of multi-modal features while ignoring the adaptability to the specific distribution of patients, and the disease risk prediction module may be more inclined to minimize the prediction error but may ignore the integrity of global features. Therefore, this module combines these objectives into an overall optimization problem through the dual optimization method; The total optimization objective of the system can be expressed as;
[0039] where: is the loss function of manifold alignment, which is used to measure the alignment accuracy of different modal features in the shared latent space; is the loss function of personalized distribution modeling, which is used to optimize the matching between the patient-specific distribution and the global distribution; is the loss function of disease risk prediction, which reflects the error between the predicted output and the true risk; and are weight coefficients used to balance the optimization objectives of different modules; Specifically, the manifold alignment loss function can be defined as;
[0040] where: are the low-dimensional embedded representations of time series, static features, and environmental features respectively; is the alignment mapping function between modalities; The objective of the personalized distribution modeling loss function is to minimize the difference between the personalized distribution and the global distribution, and its expression is;
[0041] where: KL represents the KL divergence between the personalized distribution and the global distribution; represents the log-likelihood term; The loss function of the disease risk prediction module directly measures the error between the predicted value and the true value, and its definition is;
[0042] where; is the true risk label of the th patient; is the predicted risk score; To introduce higher optimization flexibility, this module transforms the global objective of the system into a dual problem through Lagrange multipliers, and its dual form can be expressed as;
[0043] where: are dual variables used to control the accuracy of manifold alignment; is the allowed alignment error threshold; and are the parameters of the personalized distribution modeling module and the risk prediction module respectively; This module updates the parameters in an alternating optimization manner, and the optimization process includes the following steps; First, fix and , and update the dual variable , and its update rule is;
[0044] Where: is the learning rate; indicates whether the current alignment error exceeds the threshold; Subsequently, fix , and optimize and , to minimize the global loss .
[0045] Through such alternating updates, it can be ensured that the global loss function gradually converges to the optimal solution; is used to adjust the weight coefficient according to the specific characteristics of the patient and , for patients dominated by time series data, the weight of can be appropriately increased, thereby enhancing the accuracy of manifold alignment; In this embodiment, the module accepts the manifold alignment loss, personalized distribution loss, and prediction error generated by the foregoing module as inputs, and coordinates the objectives of each module through a joint optimization method. Finally, by optimizing the output, it is ensured that the performance of the entire system reaches the global optimal state.
[0046] The method for pulmonary health management described below can be correspondingly referred to the system for pulmonary health management described above; Please refer to the appendix Figure 1 , the method for pulmonary health management, includes the following steps; S1. Receive the time series data, static feature data, and environmental data of the patient, and generate low-dimensional manifold representations of each modality through a non-linear manifold learning algorithm; Specifically, the multi-modal data of the patient comes from multiple devices and systems. The time series data is usually obtained through wearable devices or medical monitoring devices, and the static feature data can be extracted from the patient's electronic health record or medical history record. The environmental feature data comes from external environmental monitoring devices or public environmental data interfaces, such as the data of regional air quality monitoring stations; First, it is necessary to perform standardization processing on its data format. For time series data, usually unify its sampling frequency to ensure that the data from different devices can be aligned and comparable. The data with different sampling frequencies can be aligned on the time axis through an interpolation algorithm; For static feature data, the processing usually includes filling in missing values and normalization. Missing values can be processed by various methods, such as interpolation, mean filling, or statistical filling methods based on historical data. Normalization ensures that the numerical ranges of different features are consistent to avoid some features having a disproportionate impact on the overall result during subsequent analysis; Time series data is organized as a multi-dimensional array, where one dimension represents the variation of a specific feature over time. Static feature data is organized in the form of a fixed-length vector, and environmental feature data is integrated as part of the time series data for synchronous analysis with the patient's dynamic physiological features.
[0047] S2. Align the low-dimensional manifold representations of each modality to generate a shared latent feature space; Specifically, the multi-modal data from step S1 includes time series data, static feature data, and environmental feature data. The characteristics and distributions of these data may have significant differences in the original high-dimensional space. Therefore, it is necessary to reduce the dimension of each modality of data separately through non-linear manifold learning methods; For time series data, it is embedded into a low-dimensional space through a manifold learning method, thereby retaining the dynamic change pattern and local geometric relationship of the time series data. While reducing the data dimension, the relationship between neighborhood points in the original high-dimensional data can be maintained. The time series data after dimension reduction is output in the form of a low-dimensional embedding representation, providing feature input for subsequent modules; Static feature data is also processed through a manifold learning method to further extract potential feature representations. In this way, the non-linear relationship between static features can be better mined, improving its applicability in subsequent analysis. Through the non-linear manifold learning method, environmental features can be mapped to a low-dimensional space to extract their key features and conduct correlation analysis with the patient's physiological features; The process of modality alignment is achieved by designing an alignment mapping function, so that time series features, static features, and environmental features have a consistent representation form in the shared feature space. The alignment method minimizes the alignment error between modalities, enabling the distribution of data from different modalities in the shared space to better reflect their inherent correlation.
[0048] S3. Generate the patient's latent distribution based on the patient's personalized features and optimize this latent distribution through variational inference methods; Specifically, the health features of patients have strong personalized characteristics. The distributions of different patients in the latent feature space may vary significantly. By constructing the patient's personalized distribution, these differential features are captured, and this distribution is dynamically optimized and adjusted based on the patient's historical data and real-time data; The personalized distribution of patients is represented in the form of a probability distribution, which is used to describe the health characteristics of patients in the latent feature space. The model can reflect the feature distribution density of patients in the shared latent space. The distribution model is defined in a parametric way so that it can be dynamically adjusted according to the specific characteristics of patients. When generating the personalized distribution, the system combines multi-modal data such as the time series features, static features, and environmental features of patients, and performs alignment and dimensionality reduction processing through the shared latent feature space generated in step S2. Therefore, it can be directly used as input in this step. Specifically, the personalized distribution model of patients uses these input features to generate the parameters of the probability distribution, so as to describe the distribution shape of the health characteristics of patients in the latent space. The method based on variational inference is adopted. The goal of variational inference is to minimize the difference between the patient-specific distribution and the true distribution, so as to ensure that the generated personalized distribution can accurately reflect the actual feature state of the patient. The observed data of the patient is used as a constraint on the distribution model during the optimization process, so as to improve the adaptability of the model.
[0049] S4. In the shared latent feature space, predict the risk of lung diseases according to the time series features, static features, and environmental features of patients. Specifically, use a risk prediction model based on deep learning to comprehensively analyze the input features and output the risk score of the patient. The risk prediction model first extracts dynamic feature patterns from the time series features. The time series features are important inputs for disease risk prediction, and their change trends and short-term fluctuations directly reflect the physiological state of the patient. The model extracts global and local patterns of the time series features by introducing a deep learning architecture based on the self-attention mechanism. The self-attention mechanism can capture the correlations between different time points in the time series, so as to enhance the model's understanding of long-term trends and short-term changes. The static feature data is input into the model as global background information in this embodiment, and the environmental feature data is used as dynamic input to reflect the impact of the current external environment where the patient is located on lung health. By combining these static and dynamic features, the model can comprehensively evaluate the health status of the patient. Optimize by introducing the patient's personalized distribution information generated in step S3. The personalized distribution reflects the uniqueness of the patient and the feature distribution state in the latent space. By using the personalized distribution as a regularization constraint term of the prediction model, the system can more accurately capture the patient-specific risk factors. For patients with specific medical histories, their distribution deviations can guide the model to assign higher weights in specific feature dimensions. The final output of the disease risk prediction model is a risk score, which usually ranges from 0 to 1. The higher the score, the greater the risk that the patient is in an acute exacerbation or deterioration of the lung disease. The generation of the risk score not only depends on the input features and the patient's personalized distribution, but also through the dynamic weight adjustment of the time series features to ensure that the model can adapt to the short-term fluctuations of the patient's health status; The system can generate accurate and personalized lung disease risk scores, providing a basis for subsequent intervention measures.
[0050] S5. Jointly optimize the parameters of the multi-modal data alignment, patient personalized distribution modeling, and disease risk prediction modules through the dual optimization method to ensure the global convergence and prediction accuracy of the system.
[0051] Specifically, in this step, by introducing a dual optimization framework, the objective loss functions of each module are combined into an overall optimization problem. Through the introduction of dual variables, the system can dynamically balance the optimization objectives among the modules and achieve the joint training of parameters through an alternating optimization strategy; The overall optimization objective includes three core loss functions: the manifold alignment loss, the personalized distribution modeling loss, and the disease risk prediction error. The manifold alignment loss is used to measure the alignment accuracy of different modal data in the shared latent feature space. The personalized distribution modeling loss is used to adjust the matching degree between the patient-specific distribution and the real data. The disease risk prediction error directly reflects the accuracy of the prediction result; The dual variables dynamically adjust the objective weights of each module to ensure the objective coordination in the overall optimization process. When the manifold alignment accuracy is low, the dual variables will increase the weight of the manifold alignment loss, thus guiding the system to preferentially optimize the parameters of the relevant modules. When the disease risk prediction error is high, the dual variables will correspondingly adjust the priority of this objective; The joint optimization adopts an alternating optimization strategy. First, fix the parameters of some modules and optimize the parameters of other modules. Subsequently, after the parameters are updated, readjust the dual variables to balance the optimization objectives of each module. The process continues to iterate until the global loss function of the system converges; First, fix the parameters of the manifold alignment module and the personalized distribution modeling module, and only optimize the parameters of the disease risk prediction module to minimize the prediction error. After the parameters of the prediction module are updated, the system readsjusts the values of the dual variables according to the current manifold alignment accuracy and the matching degree of the personalized distribution. The alternating optimization strategy can effectively avoid the conflicts between the optimization objectives of each module and ensure the global convergence of the system; It realizes the global joint optimization of multiple modules, not only improving the independent performance of each module, but also significantly improving the coordination and adaptability between modules.
[0052] Although embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.
Claims
1. A system for lung health management, characterized in that, The steps include: A multimodal data manifold learning and alignment module is used to receive multimodal data of patients, extract low-dimensional manifold embedding representations of each modality data, and align the low-dimensional manifold embedding representations of different modalities to generate a shared latent feature space; A personalized distribution modeling module, used for generating a personalized distribution of the patient in a shared latent feature space based on the patient's personalized features; A disease risk prediction module is used to predict the patient's lung disease risk based on the patient's time series characteristics, static characteristics and environmental characteristics; The dual optimization module is used to jointly optimize the parameters of each module and dynamically adjust the global optimization goal of the system.
2. The system for lung health management according to claim 1, characterized in that: The multimodal data manifold learning and alignment module includes: A data input unit, used to receive multimodal data of a patient, wherein the multimodal data includes time series data, static feature data, and environmental data; A manifold dimension reduction unit is used to perform dimension reduction processing on the time series data, static feature data and environmental data through a nonlinear manifold learning algorithm to obtain a corresponding low-dimensional manifold representation; Manifold alignment unit, which is used to align low-dimensional manifold representations of different modalities to generate a shared latent feature space.
3. The system for lung health management according to claim 2, characterized in that: The manifold dimensionality reduction unit constructs a low-dimensional embedding representation of each modality data through a nonlinear dimensionality reduction method, and the nonlinear dimensionality reduction method includes a manifold learning algorithm based on a neighborhood structure, wherein the low-dimensional manifold representation reflects the nonlinear distribution structure of the patient's multimodal data in a high-dimensional space.
4. The system for lung health management according to claim 1, characterized in that: The personalized distribution modeling module includes: A distribution generating unit, used for generating a latent distribution of the patient based on the personalized features of the patient, where the latent distribution is defined by an energy function and reflects the personalized characteristics of the patient in the latent feature space; The distribution optimization unit is used to optimize the latent distribution of patients based on the variational inference method to align it with the feature distribution of the global model.
5. The system for lung health management according to claim 4, characterized in that: The distribution optimization unit performs optimization by minimizing the difference between the personalized distribution and the global model distribution, wherein the personalized distribution is generated based on the individual characteristics of the patient, and the global model distribution is generated by group data in a multimodal feature space.
6. The system for lung health management according to claim 1, characterized in that: The disease risk prediction module includes: A time series feature processing unit, used for dynamic feature extraction of time series data; A comprehensive feature fusion unit, which is used to fuse time series features, static features, and environmental features to generate disease risk prediction input; The risk scoring unit is used to generate a disease risk score for a patient based on the comprehensive features.
7. The system for lung health management according to claim 6, characterized in that: The risk scoring unit calculates the patient's lung disease risk probability through a time series model, wherein the time series model includes a deep learning model based on a self-attention mechanism.
8. The system for lung health management according to claim 1, characterized in that: The dual optimization module ensures the convergence of global system parameters by jointly optimizing the patient's personalized distribution modeling objectives, manifold alignment objectives and disease risk prediction objectives, wherein the joint optimization objectives include dynamic adjustment of the patient's personalized distribution, alignment accuracy of the modal feature space and accuracy of risk prediction.
9. The system for lung health management according to claim 1, characterized in that: The dual optimization module updates parameters through an alternating optimization algorithm, which includes fixing the parameters of one module and optimizing the parameters of another module, ultimately minimizing the global loss function.
10. A method for lung health management, according to the system for lung health management according to claims 1-9, characterized in that: The steps include: S1, receiving the patient's time series data, static feature data and environmental data, and generating a low-dimensional manifold representation of each modality through a nonlinear manifold learning algorithm; S2, aligning the low-dimensional manifold representations of each modality to generate a shared latent feature space; S3, generating a potential distribution of patients based on their personalized characteristics, and optimizing the potential distribution through a variational inference method; S4. Predict lung disease risk based on the patient's time series characteristics, static characteristics, and environmental characteristics in the shared latent feature space; S5. The parameters of multimodal data alignment, patient personalized distribution modeling and disease risk prediction modules are jointly optimized through the dual optimization method to ensure the global convergence and prediction accuracy of the system.