Multiple sclerosis intervention recommendation generation system based on individual immune profile

Abstract

This invention discloses a multiple sclerosis intervention suggestion generation system based on individual immune characteristics, comprising: a multi-source immune data acquisition and temporal standardization processing layer, a multi-scale immune feature extraction and immune status index calculation layer, and a non-therapeutic intervention scheme multi-objective optimization and generation layer. This invention has the following advantages and effects: the multi-source immune data acquisition and temporal standardization processing layer achieves temporal alignment, dimensional normalization, and missing value imputation of the original dataset of multi-dimensional immune features; the multi-scale immune feature extraction and immune status index calculation layer extracts features from the time domain, frequency domain, and nonlinear dynamics in parallel and fuses them to calculate the immune status index; and the non-therapeutic intervention scheme multi-objective optimization and generation layer generates non-therapeutic intervention schemes, thereby effectively solving the problems of difficulty in integrating multi-source heterogeneous immune data, incomplete immune status assessment, and lack of personalization and practicality in intervention suggestions.

Description

Technical Field

[0001] This invention relates to the field of medical and health management technology, and in particular to a system for generating intervention recommendations for multiple sclerosis based on individual immune characteristics. Background Technology

[0002] In the field of personalized health management, particularly in generating non-therapeutic intervention recommendations for immune-related conditions, existing technologies face a core problem: the inability to effectively integrate multi-source heterogeneous immune data and extract comprehensive features to accurately assess individual immune status, resulting in intervention recommendations lacking personalization and practicality. Existing technologies typically rely on single-type or limited-dimensional immune data, lacking the ability to standardize multi-source data over time, leading to difficulties in data integration and significant information loss. Furthermore, existing methods are largely limited to time-domain statistical feature analysis, failing to capture the complex dynamic behavior of the immune system from multiple scales, such as frequency domain and nonlinear dynamics, resulting in incomplete and inaccurate immune status assessments. Simultaneously, the lack of a multi-objective optimization mechanism makes it impossible to simultaneously balance individual tolerability, feasibility, and expected improvement in immune status when generating intervention recommendations, often resulting in recommendations that are detached from individual conditions and have poor executability. Therefore, there is an urgent need for a multiple sclerosis intervention recommendation generation system based on individual immune characteristics that can overcome the above core problems, achieving multi-scale feature fusion and quantitative assessment of immune status based on multi-source immune data, and generating personalized and actionable non-therapeutic intervention recommendations through multi-objective optimization. Summary of the Invention

[0003] The purpose of this invention is to provide a multiple sclerosis intervention suggestion generation system based on individual immune characteristics to solve the problems mentioned in the background art.

[0004] The above-mentioned technical objective of this invention is achieved through the following technical solution: a multiple sclerosis intervention suggestion generation system based on individual immune characteristics, comprising:

[0005] Multi-source immune data acquisition and temporal standardization processing layer: Collects individual genomic data, proteomics data, and dynamic immune cell count data to form a multi-dimensional immune feature raw dataset; performs temporal alignment, dimensional normalization, and missing value imputation on the multi-dimensional immune feature raw dataset to generate a standardized individual immune feature temporal matrix.

[0006] Multi-scale immune feature extraction and immune status index calculation layer: Receives the individual immune feature time-series matrix as input, and processes it in parallel through a time-domain feature extraction unit, a frequency-domain feature decomposition unit, and a nonlinear dynamics feature parsing unit to generate a multi-scale immune feature set containing time-domain feature subsets, frequency-domain feature subsets, and nonlinear dynamics feature subsets; based on the multi-scale immune feature set, feature fusion is achieved through feature semantic encoding and cross-modal feature interaction, and the immune status index is calculated through a pre-trained immune status assessment model;

[0007] Multi-objective optimization and generation layer for non-therapeutic intervention programs: Based on the immune status index, an optimization model is constructed with the objectives of individual tolerability, feasibility, and expected improvement in immune status as the goals; a multi-objective evolutionary algorithm is used to solve the optimization model to obtain a set of Pareto optimal non-therapeutic intervention programs; according to preset decision rules, the final program is selected from the Pareto optimal solution set and encoded into a structured, executable set of personalized health management recommendations.

[0008] By adopting the above technical solutions, the integration and standardization of genomics, proteomics, and dynamic immune cell counting data are achieved through a multi-source immune data acquisition and temporal standardization processing layer. This solves the integration difficulties caused by data heterogeneity and temporal inconsistency in existing technologies, ensuring data integrity and comparability. The multi-scale immune feature extraction and immune status index calculation layer comprehensively captures the dynamic behavior and complex patterns of the immune system through parallel processing of time domain, frequency domain, and nonlinear dynamic features, significantly improving the accuracy and comprehensiveness of immune status assessment and overcoming the limitations of traditional methods that rely solely on time domain features. The multi-objective optimization and generation layer for non-therapeutic intervention programs constructs an optimization model with individual tolerability, feasibility, and improvement in immune status as objectives, and uses a multi-objective evolutionary algorithm to generate Pareto optimal solutions. This ensures a balance among multiple key objectives in the intervention program, making the final health management recommendations more personalized, executable, and scientifically reliable, effectively solving the problem of existing technologies where recommended programs are divorced from individual actual conditions.

[0009] A further setting includes, within the multi-source immune data acquisition and time-series standardization processing layer, the steps for acquiring individual genomic data, proteomics data, and dynamic immune cell count data include:

[0010] Multi-source immune characteristic data are collected through a distributed biosensor node network, which includes three types of data acquisition devices: gene sequencing nodes, mass spectrometry analysis nodes, and flow cytometry counting nodes, to acquire genomics data, proteomics data, and dynamic immune cell count data, respectively. The genomics data includes single nucleotide polymorphism (SNP) site genotyping data, gene expression abundance time-series data, and dynamic changes in DNA methylation levels. The proteomics data includes cytokine concentration fluctuation data, autoantibody titer time-series data, and complement protein activity level data. The dynamic immune cell count data includes absolute T lymphocyte subset count data, B lymphocyte activation status distribution data, and natural killer cell cytotoxicity activity index data.

[0011] By adopting the above technical solution, multi-source immune data is collected through a distributed biosensor node network, including gene sequencing nodes, mass spectrometry analysis nodes, and flow cytometry counting nodes. This ensures the comprehensiveness and professionalism of data collection, covering multiple immune characteristics from the genetic to the cellular level, and providing a rich and high-quality data foundation. The specifically collected genomics data, proteomics data, and dynamic immune cell count data enable in-depth analysis of individual immune status from multiple dimensions, enhancing the accuracy and reliability of subsequent feature extraction and status assessment, and providing data support for generating highly personalized intervention recommendations.

[0012] A further setting is that, in the multi-scale immune feature extraction and immune status index calculation layer, the steps of parallel processing by the time-domain feature extraction unit, the frequency-domain feature decomposition unit, and the nonlinear dynamic feature analysis unit include:

[0013] The temporal feature extraction unit traverses the individual immune feature time-series matrix through a sliding time window, and calculates the statistical features and trend characteristics of the genomic data sequence, proteomics data sequence, and immune cell dynamic count data sequence in each window; the statistical features include mean, variance, skewness, and kurtosis, and the trend characteristics include linear fitting slope and cumulative change.

[0014] The frequency domain feature decomposition unit performs a fast Fourier transform on the individual immune feature time-series matrix, transforming the genomics data sequence, proteomics data sequence, and immune cell dynamic count data sequence from the time domain to the frequency domain, and extracting the frequency domain energy distribution features. The frequency domain energy distribution features include the main frequency band energy ratio, spectral entropy, and frequency band energy ratio. At the same time, wavelet transform is used to extract the time-frequency joint distribution features of the genomics data sequence, proteomics data sequence, and immune cell dynamic count data sequence.

[0015] The nonlinear dynamics feature analysis unit employs recursive quantitative analysis and sample entropy algorithms to quantify the complexity and determinism of genomics data sequences, proteomics data sequences, and immune cell dynamic counting data sequences, respectively, in order to calculate the corresponding recursive quantitative analysis features and sample entropy features. The recursive quantitative analysis features include the density of recursion points and the determinism coefficient in the recursion graph.

[0016] By adopting the above technical solutions, the time-domain feature extraction unit calculates statistical features and trend characteristics through a sliding time window, effectively capturing the fluctuation patterns and long-term trends of immune data in the time dimension, providing basic time-series insights for immune status assessment; the frequency-domain feature decomposition unit extracts frequency-domain energy distribution features and time-frequency joint distribution features through fast Fourier transform and wavelet transform, effectively revealing the periodic behavior and energy distribution patterns of immune data in the frequency domain, making up for the frequency-domain dynamic information that traditional time-domain analysis cannot capture; the nonlinear dynamics feature analysis unit uses recursive quantitative analysis and sample entropy algorithms to quantify data complexity and determinism, enabling the identification of chaotic and nonlinear behaviors in the immune system, thereby more comprehensively reflecting the intrinsic dynamic characteristics of the immune system. The parallel processing of these three units can deeply analyze immune data at multiple scales, significantly improving the dimensionality and accuracy of feature extraction, and providing richer and more discriminative feature inputs for subsequent immune status index calculation.

[0017] A further setting is that, in the multi-scale immune feature extraction and immune status index calculation layer, the step of feature fusion based on the multi-scale immune feature set through feature semantic encoding and cross-modal feature interaction includes:

[0018] The feature semantic coding network maps the statistical features and trend features in the time-domain feature subset to time-domain semantic representation vectors in a unified semantic space; maps the frequency-domain energy distribution features and time-frequency joint distribution features in the frequency-domain feature subset to frequency-domain semantic representation vectors in a unified semantic space; and maps the recursive quantitative analysis features and sample entropy features in the nonlinear dynamics feature subset to nonlinear dynamics semantic representation vectors in a unified semantic space. The feature semantic coding network adopts a multilayer perceptron structure and realizes the mapping from feature space to semantic space through nonlinear transformation.

[0019] Based on the temporal semantic representation vector, the frequency domain semantic representation vector, and the nonlinear dynamics semantic representation vector, a multidimensional feature association map is constructed. Based on the multidimensional feature association map, a graph attention network is used to achieve cross-domain feature fusion, outputting enhanced temporal semantic representation vector, enhanced frequency domain semantic representation vector, and enhanced nonlinear dynamics semantic representation vector. Then, after cross-modal feature interaction is achieved through a feature cross-network, a cross-modal interaction feature vector is generated.

[0020] Based on the cross-modal interaction feature vector, the contribution score of each feature dimension is calculated using a feature importance evaluation algorithm based on random forest; a dynamic feature contribution threshold is set, and the feature contribution threshold is automatically adjusted according to the total number of feature dimensions and fusion quality requirements; feature dimensions with contribution scores higher than the feature contribution threshold are selected to form an optimized fusion immune feature vector.

[0021] By adopting the above technical solutions, time-domain, frequency-domain, and nonlinear dynamic features are mapped to a unified semantic space through a feature semantic coding network, eliminating the semantic gap between different feature modalities and achieving consistency and comparability of feature representations. The cross-domain feature fusion mechanism based on multi-dimensional feature association graphs and graph attention networks can automatically learn deep associations and dependencies between features, enhancing the contextual information of feature representations and generating more robust and representative fused features. Through feature importance assessment and dynamic threshold filtering, feature dimensions with high contribution can be preferentially retained, reducing interference from noise and redundant features, and optimizing the quality and information density of feature vectors. Quality assessment of information entropy and variance contribution rate of the fused feature vectors ensures the scientific rigor and reliability of the feature fusion process. Finally, the immune status index is accurately calculated through an immune status assessment model, providing a quantitative and reliable basis for intervention plan generation and ultimately improving decision-making quality.

[0022] A further configuration involves the following: in the multi-scale immune feature extraction and immune status index calculation layer, the nodes in the constructed multi-dimensional feature association map include time-domain feature nodes, frequency-domain feature nodes, and nonlinear dynamic feature nodes, with each node corresponding to a semantic representation vector. By calculating the cosine similarity of the semantic representation vectors between nodes, association edges between nodes are established. When the cosine similarity exceeds a preset similarity threshold, a weighted connection edge is established between the corresponding nodes. The weight value of the weighted connection edge is proportional to the cosine similarity.

[0023] By adopting the above technical solutions, a multidimensional feature association graph is constructed, which nodes the time-domain, frequency-domain, and nonlinear dynamic features. Weighted connection edges are established through cosine similarity, making the semantic relationships between features visible and quantifiable. This graph structure provides a clear topological framework for subsequent feature fusion. Dynamically establishing connection edges based on similarity thresholds ensures that only semantically highly related features are connected, avoiding interference from irrelevant features and improving the accuracy and efficiency of feature association. The weights of the weighted connection edges are proportional to the cosine similarity, enabling the graph attention network to more precisely utilize the correlation strength between features for information aggregation, thereby enhancing the targeting and effectiveness of feature fusion and providing a foundation for generating high-quality fused immune feature vectors.

[0024] A further setting involves, in the multi-scale immune feature extraction and immune state index calculation layer, the step of outputting the enhanced temporal semantic representation vector, the enhanced frequency domain semantic representation vector, and the enhanced nonlinear dynamic semantic representation vector, including:

[0025] The graph attention network calculates the attention coefficients of each node and its neighboring nodes through a multi-head attention mechanism, and performs weighted aggregation on the semantic representation vectors of the neighboring nodes based on the attention coefficients. After multi-layer graph attention propagation, each node obtains an enhanced semantic representation vector that incorporates neighborhood information. The enhanced temporal semantic representation vector, the enhanced frequency domain semantic representation vector, and the enhanced nonlinear dynamic semantic representation vector are output respectively.

[0026] By adopting the above technical solution, the graph attention network calculates the attention coefficients between nodes through a multi-head attention mechanism, which can adaptively focus on important neighboring nodes and achieve dynamic allocation of weights for different features. By weighted aggregation of the semantic representation vectors of neighboring nodes, each node can fuse local and global contextual information to obtain an enhanced semantic representation vector. This not only improves the representational power of individual features but also captures the synergistic effect between features. Through multi-layer graph attention propagation, information is efficiently diffused in the graph, making the final output enhanced semantic representation vector more comprehensively reflect the multi-scale characteristics of the immune system. This provides richer and more accurate feature inputs for subsequent cross-modal feature interaction and immune status index calculation, significantly improving the overall analysis performance.

[0027] A further setting is that, in the multi-scale immune feature extraction and immune status index calculation layer, the step of calculating the immune status index using a pre-trained immune status assessment model includes:

[0028] The fused immune feature vector is input into a pre-trained immune status assessment model. The immune status assessment model adopts a multi-layer feedforward neural network structure, which includes an input layer, multiple hidden layers, and an output layer. The number of neurons in the input layer is consistent with the dimension of the fused immune feature vector. The hidden layer uses a non-linear activation function to realize the non-linear transformation of features. The output layer uses a sigmoid activation function to map the calculation results to a preset standardized numerical range to generate an immune status index.

[0029] By adopting the above technical solution, a pre-trained multi-layer feedforward neural network is used as the immune status assessment model. Leveraging its powerful nonlinear fitting capability, it can efficiently process high-dimensional fused immune feature vectors and uncover the complex mapping relationship between features and immune status. The hidden layers use nonlinear activation functions to enhance the model's feature transformation capabilities, enabling it to capture deep patterns in the data. The output layer uses a sigmoid activation function to standardize the calculation results to a preset range, resulting in an immune status index with clear numerical meaning and comparability, facilitating direct use in subsequent model optimization. This structure ensures the accuracy and stability of the immune status index calculation, providing a reliable and quantitative decision-making basis for intervention plan generation, thus improving practicality and credibility.

[0030] A further setting is that the processing steps of the multi-objective optimization and generation layer of the non-therapeutic intervention program include the following steps:

[0031] An optimization model with three core optimization objectives is constructed based on the immune status index. These objectives are: individual tolerance optimization, which quantifies the physiological and psychological adaptability of the intervention program to an individual through a tolerance assessment function; feasibility optimization, which measures the feasibility of the intervention program under realistic conditions through a feasibility assessment function; and immune status improvement optimization, which constructs an improvement assessment function based on the gap between the current and target immune status indices. This improvement assessment function is used to evaluate the expected ability of non-therapeutic intervention programs to narrow this gap. All three core optimization objectives use the immune status index as a key input parameter, and the weighting coefficients are adjusted in a coordinated manner through the numerical range of the immune status index.

[0032] A non-dominated sorting genetic algorithm is used to solve the optimization model, generating a Pareto optimal solution set. During the solution process, constraints are verified for each candidate solution. These constraints include individual tolerance constraints and feasibility constraints. Individual tolerance constraints quantify the adaptability of the intervention plan to the individual's physiological and psychological state through a tolerance evaluation function, while feasibility constraints measure the executability of the intervention plan under realistic conditions through a feasibility evaluation function. For candidate solutions that do not meet the constraints, a penalty function method is used to process them, quantifying the degree of constraint violation as a penalty term for the core optimization objective.

[0033] The optimal implementation scheme is selected from the Pareto optimal solution set, compiled into a structured set of personalized health management suggestions, and then output.

[0034] By adopting the above technical solutions and constructing an optimization model with individual tolerability, feasibility, and improvement in immune status as core objectives, multiple key aspects of the intervention plan can be comprehensively considered, ensuring that the generated recommendations are both scientific and humane. The non-dominated sorting genetic algorithm is used to solve the optimization model, efficiently exploring the solution space and generating a Pareto optimal solution set. These solutions represent the optimal solutions in terms of trade-offs between different objectives, providing users with multiple optional strategies. Constraint verification and penalty function methods are used to handle invalid solutions, enhancing the robustness of the optimization process and ensuring that the final solution meets the requirements of practical feasibility and individual tolerability. Encoding the optimal solution into a structured set of health management recommendations makes the output clear, operable, and easy for users to implement directly, effectively solving the problem of recommendations being detached from reality in existing technologies and improving the acceptability and effectiveness of intervention recommendations.

[0035] A further setting involves the following step in the multi-objective optimization and generation layer of the non-therapeutic intervention program: all three core optimization objectives use the immune status index as a key input parameter, and the weight coefficients are adjusted in a coordinated manner through the numerical range of the immune status index. This includes:

[0036] A preset threshold for the immune status index range is used to divide the range of immune status index values ​​into a low immune status index range, a medium immune status index range, and a high immune status index range. The low immune status index range corresponds to an immune status index that is lower than the first threshold, the medium immune status index range corresponds to an immune status index that is between the first threshold and the second threshold, and the high immune status index range corresponds to an immune status index that is higher than the second threshold.

[0037] Establish a weight coefficient mapping: When the immune status index is in the low immune status index range, generate a weight configuration scheme dominated by the goal of improving immune status; when the immune status index is in the medium immune status index range, generate a weight configuration scheme dominated by the goal of feasibility optimization; when the immune status index is in the high immune status index range, generate a weight configuration scheme dominated by the goal of optimizing individual tolerance.

[0038] A smooth transition zone for weighting coefficients is set at the boundary of adjacent immune status index intervals. An S-shaped function is used to achieve a continuous and gradual change in the weighting coefficients, avoiding abrupt changes in the weighting coefficients at the boundary.

[0039] By adopting the above technical solution and dynamically adjusting the weight coefficients according to the numerical range of the immune status index, different optimization objectives can be adaptively prioritized. For example, the focus is on improvement in low immune status, feasibility in medium immune status, and tolerability in high immune status. This intelligent weight allocation mechanism makes the intervention plan more in line with the actual needs of the individual's current immune status. Setting a smooth transition zone for the weight coefficients and using an S-shaped function to achieve continuous gradual change avoids optimization instability caused by abrupt weight changes, ensuring the smoothness and convergence of the optimization process. This linkage adjustment mechanism not only improves the flexibility and adaptability of the optimization model, but also makes the generated intervention plan more personalized and practical, and can better meet the diverse needs of individuals with different immune statuses.

[0040] A further setting involves using a method based on approximating ideal solutions to select the optimal implementation scheme from the Pareto optimal solution set within the multi-objective optimization and generation layer of the non-therapeutic intervention scheme. Specifically:

[0041] Based on the numerical distribution of the individual tolerance optimization objective function, the feasibility optimization objective function, and the immune status improvement optimization objective function in the Pareto solution set, the optimal reference value and the worst reference value of each objective function are determined.

[0042] Calculate the combined score of each Pareto solution relative to the optimal reference value and the worst reference value, and select the Pareto solution with the highest combined score as the optimal implementation solution.

[0043] By adopting the above technical solution, the optimal implementation plan is selected from the Pareto optimal solution set using a ranking method based on approximation of ideal solutions. By calculating the comprehensive score of each solution relative to the best and worst reference values, the overall performance of each plan on multiple objectives can be objectively and quantitatively evaluated, avoiding subjective bias. This method ensures that the final selected plan achieves the best balance in terms of tolerability, feasibility, and improvement in immune status, taking into account both the theoretical effects of the plan and the actual implementation conditions. Through this scientific decision-making mechanism, highly optimized and executable personalized health management recommendations can be output, significantly improving the quality of intervention plans and user satisfaction, and effectively solving the trade-off problem in multi-objective decision-making.

[0044] In summary, the present invention has the following beneficial effects: by integrating multi-source immune data, extracting features at multiple scales, and optimizing for multiple objectives, it achieves accurate assessment of individual immune status and generates highly personalized non-therapeutic intervention plans, significantly improving the accuracy and practicality of health management. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of an embodiment;

[0046] Figure 2This is a schematic diagram of the workflow of the multi-source immune data acquisition and time-series standardization processing layer in the embodiment;

[0047] Figure 3 This is a schematic diagram of the workflow of the multi-scale immune feature extraction and immune status index calculation layer in the embodiment;

[0048] Figure 4 This is a schematic diagram of the workflow of the multi-objective optimization and generation layer of the non-therapeutic intervention scheme in the embodiment. Detailed Implementation

[0049] The present invention will be further described in detail below with reference to the accompanying drawings.

[0050] like Figures 1-4 As shown;

[0051] This embodiment discloses a multiple sclerosis intervention suggestion generation system based on individual immune characteristics, including:

[0052] Multi-source immune data acquisition and temporal standardization processing layer: Collect individual genomic data, proteomics data, and dynamic immune cell count data to form a multi-dimensional immune feature raw dataset; perform temporal alignment, dimensional normalization, and missing value imputation on the multi-dimensional immune feature raw dataset to generate a standardized individual immune feature temporal matrix.

[0053] The specific implementation process is as follows: A distributed biosensor node network is deployed to collect raw datasets of multi-dimensional immune characteristics of individuals in real time. The distributed biosensor node network consists of three types of dedicated data acquisition devices: gene sequencing nodes, mass spectrometry analysis nodes, and flow cytometry counting nodes. Gene sequencing nodes utilize high-throughput sequencing technology to periodically acquire genomic data, specifically including single nucleotide polymorphism (SNP) genotyping data, gene expression abundance time-series data, and dynamic changes in DNA methylation levels. Among these, SNP genotyping data extracts variant sites through genome-wide association analysis; gene expression abundance time-series data quantifies transcript levels using RNA sequencing technology; and dynamic changes in DNA methylation levels are captured by bisulfite sequencing to capture fluctuations in epigenetic modifications. The mass spectrometry analysis node uses mass spectrometry to detect proteomics data, including cytokine concentration fluctuations, autoantibody titer time series data, and complement protein activity levels. Cytokine concentration fluctuation data is quantified using liquid chromatography-mass spectrometry (LC-MS / MS) to measure changes in inflammatory factors such as IL-6 and TNF-α. Autoantibody titer time series data is input after calibration using enzyme-linked immunosorbent assay (ELISA), and complement protein activity levels are quantified using functional assays. The flow cytometry counting node uses flow cytometry to collect dynamic immune cell counting data, including absolute T lymphocyte subset counts, B lymphocyte activation status distribution data, and natural killer cell cytotoxicity indicators. Absolute T lymphocyte subset counts are obtained after staining with CD4+ / CD8+ marker fluorescent antibodies. B lymphocyte activation status distribution data is based on CD19 and CD86 expression levels to classify activation levels. Natural killer cell cytotoxicity indicators are obtained by detecting target cell lysis rates using a chromium release assay or flow cytometry. All data acquisition devices are integrated with timestamp recording function to ensure that each data point is accompanied by the accurate acquisition time, and transmit the raw data to the central processing unit through wireless or wired communication protocols to form a multi-dimensional immune feature raw dataset.

[0054] Next, the original dataset of the aforementioned multi-dimensional immune features is subjected to time-series standardization to eliminate issues such as asynchronous time, inconsistent dimensions, and missing data. Time-series alignment first unifies the time axis of the multi-source data by setting a baseline time series (equally spaced time points in hours or days). For non-uniformly sampled data points, linear interpolation or spline interpolation methods are used to map them to the baseline time points, ensuring that data from all dimensions are aligned at the same timestamp. For example, gene expression abundance time-series data may be collected daily, while cytokine concentration data may be collected every 6 hours. Interpolation algorithms are used to align the cytokine concentration data to daily time points, forming a consistent time series. Dimensional normalization aims to eliminate unit differences between different data dimensions. Each data dimension (such as gene expression abundance, cytokine concentration, and immune cell count) is normalized separately, mapping its numerical range uniformly to the [0,1] interval. For skewed distribution data (such as cytokine concentration), a logarithmic transformation is performed beforehand to improve the distribution pattern before normalization. Missing value imputation addresses potential missing data during data collection using a multiple imputation method based on time series and feature correlation: First, missing locations are identified. For continuous time series missing values, an autoregressive integral moving average model is used to predict the missing values. For random missing values, the k-nearest neighbor algorithm is used to imputate based on complete data from similar individuals. After imputation, data consistency is verified to ensure that the imputed values ​​do not introduce significant bias. Finally, all processed data are integrated into a standardized individual immune feature time series matrix, with time as the row and immune feature as the column, including genomic data sequences, proteomics data sequences, and immune cell dynamic count data sequences.

[0055] Multi-scale immune feature extraction and immune status index calculation layer: Receives the individual immune feature time-series matrix as input, and processes it in parallel through a time-domain feature extraction unit, a frequency-domain feature decomposition unit, and a nonlinear dynamics feature analysis unit to generate a multi-scale immune feature set containing time-domain feature subsets, frequency-domain feature subsets, and nonlinear dynamics feature subsets. Based on the multi-scale immune feature set, feature fusion is achieved through feature semantic encoding and cross-modal feature interaction, and the immune status index is calculated through a pre-trained immune status assessment model.

[0056] The specific implementation process is as follows: The time-domain feature extraction unit uses a sliding time window mechanism to traverse the entire time-series matrix. The window size is dynamically set according to the data sampling frequency and feature stability requirements. For example, for data collected daily, the window length can be set to 7 days to cover the immune fluctuation cycle within a week. Within each sliding window, the unit calculates statistical features and trend characteristics for genomics data sequences, proteomics data sequences, and immune cell dynamic count data sequences, respectively. Statistical features include mean, variance, skewness, and kurtosis. The mean reflects the average level of the data within the window, variance characterizes the degree of data fluctuation, skewness describes the asymmetry of the distribution, and kurtosis measures the sharpness of the distribution. The trend characteristics are calculated by linear regression analysis to calculate the slope of the linear fit, quantifying the upward or downward trend of the data sequence, and the cumulative change is calculated as an indicator of the total change in the data within the window. In specific calculations, for time-series gene expression abundance data in genomics sequences, the mean is calculated as the arithmetic mean of expression abundance at all time points within the window, the variance is the sum of squares of the deviations of each point from the mean, and skewness and kurtosis are calculated based on the third and fourth central moments, respectively. The slope of the linear fit is obtained by fitting a straight line using the least squares method, and the cumulative change is the difference between the values ​​at the end and the beginning of the window. The processing of proteomics data sequences and immune cell dynamic counting data sequences is similar, but parameters need to be adjusted according to the data characteristics. For example, cytokine concentration data may exhibit a log-normal distribution, so a logarithmic transformation is performed before calculating the statistics. All calculated feature values ​​are output in window order, forming a time-domain feature subset.

[0057] The frequency domain feature decomposition unit performs frequency domain analysis on the time-series matrix of individual immune features in parallel. First, it uses Fast Fourier Transform (FFT) to convert genomic, proteomics, and immune cell dynamics count data sequences from the time domain to the frequency domain, obtaining the spectral information of each sequence. The FFT employs a radix-2 algorithm to decompose the time-domain signal into multiple frequency components and calculates the amplitude and phase of each component. Based on the spectral results, the unit extracts frequency domain energy distribution features, including the dominant frequency band energy proportion, spectral entropy, and frequency band energy ratio. The dominant frequency band energy proportion is calculated by identifying the highest-energy frequency band in the spectrum (usually corresponding to the main cycle of the immune rhythm) and calculating the proportion of energy in that band to the total energy. Spectral entropy, based on the principle of information entropy, quantifies the uniformity of spectral energy distribution; a higher value indicates a more dispersed spectrum. The frequency band energy ratio is calculated by dividing specific frequency bands (e.g., low-frequency bands correspond to long-term trends, and high-frequency bands correspond to short-term fluctuations) to reveal immune dynamics at different time scales. Simultaneously, the unit extracts joint time-frequency distribution features through wavelet transform. The wavelet transform uses Morlet wavelets as basis functions and generates time-frequency spectra through scaling and translation operations, thereby capturing the local frequency characteristics of non-stationary signals. For genomic data sequences, wavelet transform can identify periodic fluctuations in gene expression; for proteomics data sequences, it helps detect sudden changes in cytokine concentrations; and for dynamic immune cell counting data sequences, it can reveal multi-scale oscillation patterns in cell subpopulation counting. All frequency domain feature values ​​are integrated to form a frequency domain feature subset.

[0058] The nonlinear dynamics feature analysis unit focuses on quantifying the complexity and determinism of immune data sequences. It employs recursive quantitative analysis and sample entropy algorithms in parallel to process genomics, proteomics, and immune cell dynamic counting data sequences. Recursive quantitative analysis first maps each data sequence to a high-dimensional space through phase space reconstruction, calculating a recursion graph to visualize the recursive behavior of the sequence. The recursion point density in the recursion graph represents the frequency of similar states occurring in the sequence, while the determinism coefficient quantifies the predictability of the sequence; high determinism indicates a regular pattern in the sequence. The sample entropy algorithm is used to assess the complexity of the sequence, quantifying irregularity by calculating the probability of pattern repetition in the sequence; a higher value indicates a more complex sequence. Specifically, for genomics data sequences, recursive quantitative analysis reveals the dynamic stability of gene expression, while sample entropy reflects the uncertainty of expression fluctuations; for proteomics data sequences, recursion point density indicates recurring patterns in cytokine concentrations, while sample entropy measures the randomness of concentration changes; for immune cell dynamic counting data sequences, the determinism coefficient assesses the predictability of cell counts, while sample entropy quantifies the degree of chaos in the counts. All nonlinear dynamic eigenvalues, including recursion point density, deterministic coefficients, and sample entropy, are aggregated to form a subset of nonlinear dynamic features. The processing results of these three units—the time-domain feature subset, the frequency-domain feature subset, and the nonlinear dynamic feature subset—together constitute a multi-scale immune feature set, providing a foundation for subsequent feature fusion and immune status index calculation.

[0059] The feature semantic encoding network maps the statistical features and trend features in the time-domain feature subset to time-domain semantic representation vectors in a unified semantic space. Simultaneously, it maps the frequency-domain energy distribution features and time-frequency joint distribution features in the frequency-domain feature subset to frequency-domain semantic representation vectors in a unified semantic space. Furthermore, it maps the recursive quantitative analysis features and sample entropy features in the nonlinear dynamics feature subset to nonlinear dynamics semantic representation vectors in a unified semantic space. The feature semantic encoding network employs a multilayer perceptron structure, achieving the mapping from the feature space to the semantic space through nonlinear transformation. Specifically, the multilayer perceptron contains three hidden layers, each using the ReLU activation function for nonlinear transformation, and the output dimension of each semantic representation vector is set to 32 dimensions. During the mapping process, for the time-domain feature subset, six dimensions of input statistical features and trend features are mapped to a 32-dimensional time-domain semantic representation vector using a multilayer perceptron. For the frequency-domain feature subset, five dimensions of input frequency-domain energy distribution features and time-frequency joint distribution features are mapped to a 32-dimensional frequency-domain semantic representation vector. For the nonlinear dynamics feature subset, three dimensions of input recursive quantitative analysis features and sample entropy features are also mapped to a 32-dimensional nonlinear dynamics semantic representation vector. All mapping processes are optimized using the stochastic gradient descent algorithm with mean squared error as the loss function to minimize reconstruction error and ensure that the semantic representation retains the key information of the original features.

[0060] Based on the temporal semantic representation vector, frequency domain semantic representation vector, and nonlinear dynamic semantic representation vector, a multidimensional feature association graph is constructed. The nodes in the multidimensional feature association graph include temporal feature nodes, frequency domain feature nodes, and nonlinear dynamic feature nodes, each node corresponding to a semantic representation vector. Association edges between nodes are established by calculating the cosine similarity of their semantic representation vectors; specifically, for any two nodes, the cosine similarity of their semantic representation vectors is calculated. When the cosine similarity exceeds a preset similarity threshold (0.7), a weighted connection edge is established between the corresponding nodes; the weight of the weighted connection edge is proportional to the cosine similarity. After the graph is constructed, a graph attention network is used to achieve cross-domain feature fusion; the graph attention network calculates the attention coefficient between each node and its neighboring nodes through a multi-head attention mechanism, and performs weighted aggregation of the semantic representation vectors of neighboring nodes based on the attention coefficients. Specifically, for each node, its attention coefficient is calculated through a query-key-value mechanism: first, the node's semantic representation vector is linearly transformed into a query vector, a key vector, and a value vector; then, the dot product of the query vector and the key vectors of all neighboring nodes is calculated, and the attention coefficient is obtained by normalization using the softmax function. To control model complexity, the graph attention network uses four attention heads, each independently calculating the attention coefficient and generating intermediate representations. Finally, the outputs of all heads are concatenated and linearly transformed to obtain the enhanced semantic representation vector. After multiple layers of graph attention propagation (e.g., two layers), each node obtains an enhanced semantic representation vector that incorporates neighborhood information; it outputs enhanced temporal semantic representation vectors, enhanced frequency domain semantic representation vectors, and enhanced nonlinear dynamic semantic representation vectors, maintaining a 32-dimensional dimension.

[0061] Next, cross-modal feature interaction is achieved through a feature crossover network. This network first concatenates the enhanced temporal semantic representation vector, the enhanced frequency-domain semantic representation vector, and the enhanced nonlinear dynamics semantic representation vector to form a 96-dimensional combined feature vector. Then, a feature crossover method based on factorization machines is used to perform pairwise crossovers of all feature dimensions in the combined feature vector, calculating the interaction strength of each feature crossover pair. The interaction strength is calculated using the inner product of latent feature vectors, assigning a latent feature vector (10-dimensional) to each feature dimension, and obtaining the weight of the feature crossover by calculating the inner product of the latent feature vectors. Simultaneously, a multi-head attention mechanism is introduced to assign different attention weights to each feature crossover term, highlighting important feature interaction combinations. Specifically, two attention heads are used to improve computational efficiency; each head calculates the attention score of the crossover term, and a higher-order feature interaction information is generated through weighted summation. Significant feature interaction terms are extracted from the feature cross-network to generate high-order feature interaction information. The high-order feature interaction information is then weighted and concatenated with the enhanced temporal semantic representation vector, the enhanced frequency semantic representation vector, and the enhanced nonlinear dynamic semantic representation vector. The weights are dynamically determined by the dynamic feature importance evaluation module based on the predictive contribution of each feature to the final immune status index. The dynamic feature importance evaluation module uses a gradient boosting decision tree model, which is trained based on historical data to obtain feature importance scores and adjusts the weight allocation in real time. Finally, a compact and information-rich cross-modal interaction feature vector is generated, with its dimension controlled to around 64 dimensions through weighted concatenation and selective retention.

[0062] Based on the cross-modal interaction feature vector, the contribution score of each feature dimension is calculated using a feature importance evaluation algorithm based on random forest. The random forest model contains 100 decision trees, and the contribution score is quantified by calculating the number of splits and purity improvement of the feature in the tree. A dynamic feature contribution threshold is set, which is automatically adjusted according to the total number of feature dimensions and fusion quality requirements. Specifically, the threshold is initially set to the median of the contribution score. If the information entropy is lower than a set value (0.5), the threshold is lowered to retain more features; if the variance contribution rate is higher than 80%, the threshold is increased to filter key features. Feature dimensions with contribution scores higher than the feature contribution threshold are selected to form an optimized fused immune feature vector, which is then input into the pre-trained immune status evaluation model.

[0063] The optimized fused immune feature vector is input into a pre-trained immune status assessment model. This model employs a multi-layer feedforward neural network structure, comprising an input layer, multiple hidden layers, and an output layer. The number of neurons in the input layer is exactly the same as the dimension of the optimized fused immune feature vector. For example, if the optimized fused immune feature vector is 64-dimensional, the input layer has 64 neurons, each receiving a feature dimension value from the fused immune feature vector, ensuring that the input data matches the network structure. The hidden layers use a non-linear activation function to achieve non-linear transformation of the features, specifically the ReLU activation function, to introduce non-linear mapping capabilities and enhance the model's ability to capture and abstract complex immune feature patterns. The number of hidden layers can be set according to actual needs. For example, three hidden layers can be set, with the number of neurons decreasing sequentially: the first hidden layer contains 128 neurons, the second contains 64 neurons, and the third contains 32 neurons. This layer-by-layer compression method gradually extracts high-level abstract features and reduces the risk of overfitting. The output layer employs the Sigmoid activation function to map the output of the hidden layers to a pre-defined standardized numerical range [0, 1], generating an immune status index. The immune status assessment model is pre-trained using historical immune feature data and corresponding expert-annotated immune status labels. When calculating the immune status index, the optimized fused immune feature vector propagates forward through the input layer, hidden layer, and output layer. The input layer passes the feature vector to the first hidden layer, which then performs a non-linear transformation using the ReLU activation function before outputting it to the second hidden layer, and so on. Finally, the output layer generates the immune status index using the Sigmoid function. This index, as a standardized scalar value, reflects an individual's overall state level under current immune characteristics and serves as a core input parameter for subsequent multi-objective optimization and generation layers of non-therapeutic intervention programs, guiding the generation of personalized health management recommendations. The entire calculation process ensures high efficiency and accuracy, enabling the real-time output of reliable immune status assessment results under dynamic immune data.

[0064] Multi-objective optimization and generation layer for non-therapeutic intervention programs: Based on the immune status index, an optimization model is constructed with the objectives of individual tolerability, feasibility, and expected improvement in immune status as the goals; a multi-objective evolutionary algorithm is used to solve the optimization model to obtain a set of Pareto optimal non-therapeutic intervention programs; according to preset decision rules, the final program is selected from the Pareto optimal solution set and encoded into a structured, executable set of personalized health management recommendations.

[0065] The specific implementation process is as follows: Based on the immune status index, an optimization model is constructed that includes three core optimization objectives: individual tolerance optimization objective, feasibility optimization objective, and immune status improvement optimization objective. The individual tolerance optimization objective quantifies the adaptability of the intervention program to the individual's physiological and psychological state through a tolerance assessment function. This function comprehensively considers the individual's historical health data, subjective tolerance feedback, and physiological indicators (such as fatigue level, sleep quality, etc.), and its mathematical expression is:

[0066] ;

[0067] in, Indicating intervention plan Individual tolerance score; For this is the first Individual indicators in the intervention program The specific values ​​to be taken (e.g., pain score, psychological stress index, etc.); For the first The weight of each indicator reflects its importance in overall tolerance; The total number of indicators used to assess an individual's tolerance score; This is the physiological response coefficient; , and To adjust parameters, the feasibility optimization objective measures the feasibility of the intervention plan under realistic conditions through a feasibility assessment function. This function evaluates the plan's time cost, resource requirements, and environmental fit, and the formula is:

[0068] ;

[0069] in, Indicating intervention plan Feasibility score; Indicates resource availability score; The resource threshold represents the minimum resource level required to execute the plan. and These are the available time and the required time, respectively. and These are the weighting coefficients for the first and second items in the feasibility assessment, respectively. The scaling factor is used. The optimization objective for improving immune status is based on the difference between the current immune status index and the target immune status index, constructing an improvement assessment function. This function predicts the potential for improved immune status after the implementation of the plan, and its expression is:

[0070] ;

[0071] in, Indicating intervention plan The score for improvement in immune status; and These are the target immune status index and the current immune status index, respectively. It is the total number of intervention measures; For the first The expected influencing factors of the intervention on immune status (such as nutritional adjustment and exercise intensity). and This is the scaling factor. All three core optimization objectives use the immune status index as a key input parameter, and the weighting coefficients are adjusted in tandem through the numerical range of the immune status index. Specifically, preset boundary thresholds for the immune status index range (e.g., the first boundary threshold is set to 0.3, and the second boundary threshold is set to 0.7) divide the range of immune status index values ​​into a low immune status index range (immune status index below 0.3), a medium immune status index range (immune status index between 0.3 and 0.7), and a high immune status index range (immune status index above 0.7). When the immune status index is in the low immune status index range, a weight configuration scheme is generated primarily based on the goal of improving immune status, for example, assigning a weight of 0.6 to immune status improvement, 0.2 to individual tolerance, and 0.2 to feasibility. When the immune status index is in the medium immune status index range, a weight configuration scheme is generated primarily based on the goal of optimizing feasibility, for example, assigning a weight of 0.5 to feasibility, 0.3 to immune status improvement, and 0.2 to individual tolerance. When the immune status index is in the high immune status index range, a weight configuration scheme is generated primarily based on the goal of optimizing individual tolerance, for example, assigning a weight of 0.6 to individual tolerance, 0.3 to feasibility, and 0.1 to immune status improvement. At the boundaries of adjacent immune status index ranges, a smooth transition zone for the weight coefficients is set, using the Sigmoid function to achieve a continuous and gradual change in the weight coefficients, avoiding abrupt changes in the weight coefficients at the boundaries. The formula is as follows:

[0072] ;

[0073] in, For transition weights; This represents the current immune status index; This represents the threshold value for the immune status index interval. The parameter controls the steepness of the transition; the larger the value, the more abrupt the transition.

[0074] After the optimization model is constructed, a non-dominated sorting genetic algorithm is used to solve the model and generate a Pareto optimal solution set. The non-dominated sorting genetic algorithm first initializes the population, with each individual representing a candidate intervention plan, encoded as multi-dimensional decision variables (such as diet plan, exercise frequency, and psychological intervention intensity). The population is iteratively evolved through crossover, mutation, and selection operations. Non-dominated sorting stratifies individuals based on their performance on the three optimization objectives, and crowding calculation ensures the diversity of the solution set. During the solution process, each candidate solution is constrained, including individual tolerance constraints and feasibility constraints. Individual tolerance constraints are quantified using a tolerance evaluation function, requiring the solution score to be no lower than a preset tolerance threshold (e.g., 0.5); feasibility constraints are measured using a feasibility evaluation function, requiring the solution score to be higher than a feasibility threshold (e.g., 0.6). For candidate solutions that do not meet the constraints, a penalty function method is used to process them, quantifying the degree of constraint violation as a penalty term for the core optimization objective. For example, if a tolerance constraint is violated, the penalty term is subtracted from the corresponding core optimization objective. ,in, The value of the penalty item; The penalty coefficient is... The tolerance threshold; Indicating intervention plan Individual tolerance score.

[0075] When selecting the optimal implementation scheme from the Pareto optimal solution set, a ranking method based on approximation of the ideal solution is adopted. First, based on the numerical distribution of the individual tolerance optimization objective function, feasibility optimization objective function, and immune status improvement optimization objective function in the Pareto solution set, the optimal and worst reference values ​​for each objective function are determined: the optimal reference value is the maximum (for improvement) or minimum (for cost-related indicators of tolerance and feasibility) value of each objective function in the solution set, and the worst reference value is the opposite extreme value. Then, a comprehensive score for each Pareto solution relative to the optimal and worst reference values ​​is calculated, using the following formula: ;in, Indicates the first The overall score of each Pareto solution (candidate intervention); To solve Euclidean distance to the optimal reference value; To solve The Euclidean distance to the worst reference value was calculated. The Pareto solution with the highest overall score was selected as the optimal implementation plan. Finally, the optimal implementation plan was compiled into a structured set of personalized health management recommendations, including specific interventions, implementation schedules, resource allocation instructions, and expected outcome assessments. This was then output to users or healthcare professionals through a visual interface or document format to ensure the recommendations were actionable and personalized. Throughout the process, the immune status index and optimization model were periodically updated to adapt to the dynamic changes in individual immune status, achieving continuous optimization.

[0076] This specific embodiment is merely an explanation of the present invention and is not intended to limit the invention. After reading this specification, those skilled in the art can make modifications to this embodiment without contributing any inventive step, but such modifications are protected by patent law as long as they are within the scope of the claims of the present invention.

Claims

1. A multiple sclerosis intervention recommendation generation system based on individual immune characteristics, characterized in that, include: Multi-source immune data acquisition and time-series standardization processing layer: Collects individual genomic data, proteomics data, and dynamic immune cell count data to form a multi-dimensional raw dataset of immune features; The original dataset of multidimensional immune features is subjected to temporal alignment, dimensional normalization, and missing value imputation to generate a standardized temporal matrix of individual immune features. Multi-scale immune feature extraction and immune status index calculation layer: Receives the individual immune feature time-series matrix as input, and processes it in parallel through a time-domain feature extraction unit, a frequency-domain feature decomposition unit, and a nonlinear dynamics feature analysis unit to generate a multi-scale immune feature set containing time-domain feature subsets, frequency-domain feature subsets, and nonlinear dynamics feature subsets; based on the multi-scale immune feature set, feature fusion is achieved through feature semantic encoding and cross-modal feature interaction, and the immune status index is calculated through a pre-trained immune status assessment model; Multi-objective optimization and generation layer for non-therapeutic intervention programs: Based on the immune status index, an optimization model is constructed with the objectives of individual tolerability, feasibility, and expected improvement in immune status as the goals; a multi-objective evolutionary algorithm is used to solve the optimization model to obtain a set of Pareto optimal non-therapeutic intervention programs; according to preset decision rules, the final program is selected from the Pareto optimal solution set and encoded into a structured, executable set of personalized health management recommendations.

2. The multiple sclerosis intervention suggestion generation system based on individual immune characteristics according to claim 1, characterized in that: In the multi-source immune data acquisition and time-series normalization processing layer, the steps for acquiring individual genomic data, proteomics data, and dynamic immune cell count data include: Multi-source immune characteristic data are collected through a distributed biosensor node network, which includes three types of data acquisition devices: gene sequencing nodes, mass spectrometry analysis nodes, and flow cytometry counting nodes, to acquire genomics data, proteomics data, and dynamic immune cell count data, respectively. The genomics data includes single nucleotide polymorphism (SNP) site genotyping data, gene expression abundance time-series data, and dynamic changes in DNA methylation levels. The proteomics data includes cytokine concentration fluctuation data, autoantibody titer time-series data, and complement protein activity level data. The dynamic immune cell count data includes absolute T lymphocyte subset count data, B lymphocyte activation status distribution data, and natural killer cell cytotoxicity activity index data.

3. The multiple sclerosis intervention suggestion generation system based on individual immune characteristics according to claim 1, characterized in that: In the multi-scale immune feature extraction and immune status index calculation layer, the steps of parallel processing by the time-domain feature extraction unit, the frequency-domain feature decomposition unit, and the nonlinear dynamic feature analysis unit include: The temporal feature extraction unit traverses the individual immune feature time-series matrix through a sliding time window, and calculates the statistical features and trend characteristics of the genomic data sequence, proteomics data sequence, and immune cell dynamic count data sequence in each window; the statistical features include mean, variance, skewness, and kurtosis, and the trend characteristics include linear fitting slope and cumulative change. The frequency domain feature decomposition unit performs a fast Fourier transform on the individual immune feature time-series matrix, transforming the genomics data sequence, proteomics data sequence, and immune cell dynamic count data sequence from the time domain to the frequency domain, and extracting the frequency domain energy distribution features. The frequency domain energy distribution features include the main frequency band energy ratio, spectral entropy, and frequency band energy ratio. At the same time, wavelet transform is used to extract the time-frequency joint distribution features of the genomics data sequence, proteomics data sequence, and immune cell dynamic count data sequence. The nonlinear dynamics feature analysis unit employs recursive quantitative analysis and sample entropy algorithms to quantify the complexity and determinism of genomics data sequences, proteomics data sequences, and immune cell dynamic counting data sequences, respectively, in order to calculate the corresponding recursive quantitative analysis features and sample entropy features. The recursive quantitative analysis features include the density of recursion points and the determinism coefficient in the recursion graph.

4. The multiple sclerosis intervention suggestion generation system based on individual immune characteristics according to claim 3, characterized in that: In the multi-scale immune feature extraction and immune status index calculation layer, the step of feature fusion based on the multi-scale immune feature set through feature semantic encoding and cross-modal feature interaction includes: The feature semantic coding network maps the statistical features and trend features in the time-domain feature subset to time-domain semantic representation vectors in a unified semantic space; maps the frequency-domain energy distribution features and time-frequency joint distribution features in the frequency-domain feature subset to frequency-domain semantic representation vectors in a unified semantic space; and maps the recursive quantitative analysis features and sample entropy features in the nonlinear dynamics feature subset to nonlinear dynamics semantic representation vectors in a unified semantic space. The feature semantic coding network adopts a multilayer perceptron structure and realizes the mapping from feature space to semantic space through nonlinear transformation. Based on the temporal semantic representation vector, the frequency domain semantic representation vector, and the nonlinear dynamics semantic representation vector, a multidimensional feature association map is constructed. Based on the multidimensional feature association map, a graph attention network is used to achieve cross-domain feature fusion, outputting enhanced temporal semantic representation vector, enhanced frequency domain semantic representation vector, and enhanced nonlinear dynamics semantic representation vector. Then, after cross-modal feature interaction is achieved through a feature cross-network, a cross-modal interaction feature vector is generated. Based on the cross-modal interaction feature vector, the contribution score of each feature dimension is calculated using a feature importance evaluation algorithm based on random forest; a dynamic feature contribution threshold is set, and the feature contribution threshold is automatically adjusted according to the total number of feature dimensions and fusion quality requirements; feature dimensions with contribution scores higher than the feature contribution threshold are selected to form an optimized fusion immune feature vector.

5. The multiple sclerosis intervention suggestion generation system based on individual immune characteristics according to claim 4, characterized in that: In the multi-scale immune feature extraction and immune status index calculation layer, the nodes in the constructed multi-dimensional feature association map include time-domain feature nodes, frequency-domain feature nodes, and nonlinear dynamic feature nodes, each node corresponding to a semantic representation vector; by calculating the cosine similarity of the semantic representation vectors between nodes, association edges between nodes are established; when the cosine similarity exceeds a preset similarity threshold, a weighted connection edge is established between the corresponding nodes; the weight value of the weighted connection edge is proportional to the cosine similarity.

6. The multiple sclerosis intervention suggestion generation system based on individual immune characteristics according to claim 5, characterized in that: In the multi-scale immune feature extraction and immune status index calculation layer, the steps of outputting the enhanced temporal semantic representation vector, the enhanced frequency domain semantic representation vector, and the enhanced nonlinear dynamic semantic representation vector include: The graph attention network calculates the attention coefficients of each node and its neighboring nodes through a multi-head attention mechanism, and performs weighted aggregation on the semantic representation vectors of the neighboring nodes based on the attention coefficients. After multi-layer graph attention propagation, each node obtains an enhanced semantic representation vector that incorporates neighborhood information. The enhanced temporal semantic representation vector, the enhanced frequency domain semantic representation vector, and the enhanced nonlinear dynamic semantic representation vector are output respectively.

7. The multiple sclerosis intervention suggestion generation system based on individual immune characteristics according to claim 6, characterized in that: In the multi-scale immune feature extraction and immune status index calculation layer, the step of calculating the immune status index using a pre-trained immune status assessment model includes: The fused immune feature vector is input into a pre-trained immune status assessment model. The immune status assessment model adopts a multi-layer feedforward neural network structure, which includes an input layer, multiple hidden layers, and an output layer. The number of neurons in the input layer is consistent with the dimension of the fused immune feature vector. The hidden layer uses a non-linear activation function to realize the non-linear transformation of features. The output layer uses a sigmoid activation function to map the calculation results to a preset standardized numerical range to generate an immune status index.

8. The multiple sclerosis intervention suggestion generation system based on individual immune characteristics according to claim 1, characterized in that: The process of multi-objective optimization and generation layer processing of the non-therapeutic intervention scheme includes the following steps: An optimization model with three core optimization objectives is constructed based on the immune status index. The three core optimization objectives are: individual tolerance optimization objective, which quantifies the adaptability of the intervention to an individual's physiological and psychological state through a tolerance assessment function; feasibility optimization objective, which measures the feasibility of the intervention under realistic conditions through a feasibility assessment function; and immune status improvement optimization objective, which constructs an improvement assessment function based on the gap between the current immune status index and the target immune status index. The improvement assessment function is used to evaluate the expected ability of non-therapeutic intervention to narrow this gap. All three core optimization objectives use the immune status index as a key input parameter, and the weight coefficients are adjusted in a coordinated manner through the numerical range of the immune status index. A non-dominated sorting genetic algorithm is used to solve the optimization model, generating a Pareto optimal solution set. During the solution process, constraints are verified for each candidate solution. These constraints include individual tolerance constraints and feasibility constraints. Individual tolerance constraints quantify the adaptability of the intervention plan to the individual's physiological and psychological state through a tolerance evaluation function, while feasibility constraints measure the executability of the intervention plan under realistic conditions through a feasibility evaluation function. For candidate solutions that do not meet the constraints, a penalty function method is used to process them, quantifying the degree of constraint violation as a penalty term for the core optimization objective. The optimal implementation scheme is selected from the Pareto optimal solution set, compiled into a structured set of personalized health management suggestions, and then output.

9. The multiple sclerosis intervention suggestion generation system based on individual immune characteristics according to claim 8, characterized in that: In the multi-objective optimization and generation layer of the non-therapeutic intervention program, all three core optimization objectives use the immune status index as a key input parameter. The step of adjusting the weight coefficients in a coordinated manner through the numerical range of the immune status index includes: A preset threshold for the immune status index range is used to divide the range of immune status index values ​​into a low immune status index range, a medium immune status index range, and a high immune status index range. The low immune status index range corresponds to an immune status index that is lower than the first threshold, the medium immune status index range corresponds to an immune status index that is between the first threshold and the second threshold, and the high immune status index range corresponds to an immune status index that is higher than the second threshold. Establish a weight coefficient mapping: When the immune status index is in the low immune status index range, generate a weight configuration scheme dominated by the goal of improving immune status; when the immune status index is in the medium immune status index range, generate a weight configuration scheme dominated by the goal of feasibility optimization; when the immune status index is in the high immune status index range, generate a weight configuration scheme dominated by the goal of optimizing individual tolerance. A smooth transition zone for weighting coefficients is set at the boundary of adjacent immune status index intervals. An S-shaped function is used to achieve a continuous and gradual change in the weighting coefficients, avoiding abrupt changes in the weighting coefficients at the boundary.

10. The multiple sclerosis intervention suggestion generation system based on individual immune characteristics according to claim 8, characterized in that: In the multi-objective optimization and generation layer of the non-therapeutic intervention scheme, the optimal implementation scheme is selected from the Pareto optimal solution set using a ranking method based on approximation of ideal solutions. Specifically: Based on the numerical distribution of the individual tolerance optimization objective function, the feasibility optimization objective function, and the immune status improvement optimization objective function in the Pareto solution set, the optimal reference value and the worst reference value of each objective function are determined. Calculate the combined score of each Pareto solution relative to the optimal reference value and the worst reference value, and select the Pareto solution with the highest combined score as the optimal implementation solution.

Images