Alzheimer's disease multi-subtype health management method and system based on five-library linkage

Abstract

The present application relates to the technical field of health management, and more particularly to a kind of Alzheimer's disease multi-subtype health management method and system based on five library linkage.The method obtains multi-dimensional health characteristic data, extracts feature vector and maps to uniform representation space to obtain normalized feature representation, constructs cross-modal feature representation matrix based on time sequence correlation and spatial distribution constraint, matches and identifies Alzheimer's disease subtype category with disease progression mode library, accordingly extracts adaptive strategy combination from health management resource library to generate personalized health management scheme, and obtains strategy effectiveness evaluation result for iterative optimization by correlating real-time feedback data with cross-modal feature representation matrix.The present application effectively improves the personalized matching precision and dynamic adaptability of health management scheme.

Description

Technical Field

[0001] This invention relates to the field of health management technology, and in particular to a health management method and system for multiple subtypes of Alzheimer's disease based on the linkage of five databases. Background Technology

[0002] Current health management methods for Alzheimer's disease generally rely on clinical diagnostic scales (such as MMSE and CDR) and imaging examination results. Physicians determine the disease stage based on experience and formulate uniform routine intervention plans. This approach usually treats patients as a homogeneous group, ignoring the significant heterogeneity of Alzheimer's disease in terms of pathological mechanisms, cognitive decline trajectories, and biomarker distribution, resulting in insufficient matching between management strategies and the actual pathological subtypes of patients.

[0003] While some existing studies have attempted to classify subtypes using single-type data (such as cerebrospinal fluid biomarkers and brain amyloid PET imaging), these methods still have significant limitations: the information correlation between different data sources has not been fully explored, and in particular, there is a lack of ability to model the dynamic evolution of multi-dimensional features (such as cognitive function, biochemical indicators, behavioral scales, neuroimaging, and gene expression) over time, resulting in low robustness of subtype identification results and difficulty in guiding the development of personalized health management plans.

[0004] Furthermore, conventional health management programs are often implemented long-term once established, lacking a real-time response mechanism for changes in the patient's physiology, behavior, and symptoms during implementation. Doctors can only obtain periodic feedback on the effectiveness through regular follow-up visits, and cannot promptly identify and dynamically adjust the effectiveness of the strategy during intervention. This results in a significant lag in optimizing management outcomes, which is not conducive to delaying disease progression. Summary of the Invention

[0005] This invention provides a health management method and system for multiple subtypes of Alzheimer's disease based on the linkage of five databases, which can solve the problems in the prior art.

[0006] A first aspect of this invention provides a health management method for multiple subtypes of Alzheimer's disease based on five-database linkage, comprising: Acquire multidimensional health characteristic data of the target individual; Feature extraction is performed on the multi-dimensional health feature data to obtain a set of feature vectors, and each feature vector in the set of feature vectors is mapped to a unified representation space to obtain a normalized feature representation. Based on the temporal correlation and spatial distribution constraints among the features in the normalized feature representation, a cross-modal feature representation matrix is ​​constructed. The cross-modal feature representation matrix is ​​then matched with a preset disease progression pattern library to identify the Alzheimer's disease subtype category corresponding to the target individual. Based on the Alzheimer's disease subtype, a combination of management strategies that are suitable for the Alzheimer's disease subtype is extracted from the health management resource library to generate a personalized health management plan; Real-time feedback data is collected from the target individual during the execution of the personalized health management plan. The real-time feedback data is then correlated with the cross-modal feature representation matrix to obtain the strategy effectiveness evaluation result. The strategy effectiveness evaluation result is used to iteratively optimize the personalized health management plan.

[0007] Feature extraction is performed on the multi-dimensional health feature data to obtain a feature vector set, and each feature vector in the feature vector set is mapped to a unified representation space to obtain a normalized feature representation, including: The multidimensional health feature data is grouped according to data modality type to obtain multiple modality data subsets. For each modality data subset, prior knowledge features corresponding to the modality data subset are extracted from the disease progression pattern library and the health management resource library, respectively. The prior knowledge features are fused with the modal data subset, and the fused time-domain features and frequency-domain features are extracted to construct the corresponding modal feature vectors. All modal feature vectors are then summarized to obtain the feature vector set. Based on the multi-subtype disease evolution paths in the disease progression pattern library, a semantic association graph between modalities is constructed, and the association weight distribution between feature vectors of different modalities in the feature vector set is determined through the semantic association graph; Based on the correlation weight distribution, each feature vector in the feature vector set is weighted and fused to obtain each weighted and fused feature vector; The weighted and fused feature vectors are projected onto a unified representation space of a preset dimension to obtain aligned feature vectors. The aligned feature vectors are then normalized to obtain the normalized feature representation.

[0008] Based on the multi-subtype disease evolution paths in the disease progression pattern library, a semantic association graph between modalities is constructed. The association weight distribution between feature vectors of different modalities in the feature vector set is determined through this semantic association graph, including: Multiple disease evolution paths corresponding to Alzheimer's disease subtypes are extracted from the disease progression pattern library. Each disease evolution path contains multiple disease stage nodes and the transition relationships between stage nodes. For each disease evolution path, identify the multi-dimensional pathological evolution features involved in the disease stage nodes, and establish a mapping relationship between the disease evolution path and each data modality in the multi-dimensional health feature data; A semantic association graph is constructed based on the mapping relationship. The semantic association graph includes modal nodes and semantic association edges. The modal nodes correspond to each modal feature vector in the feature vector set. The semantic association edges represent the co-occurrence relationship and causal dependency relationship between different modal nodes based on the disease evolution path. Historical intervention case data were extracted from the health management resource database, and the contribution distribution of different modal feature combinations to the intervention effect in the historical intervention case data was statistically analyzed. The semantic association edges in the semantic association graph are weighted according to the contribution distribution, and the association weight distribution is generated based on the corrected semantic association graph.

[0009] Based on the temporal correlation and spatial distribution constraints among the features in the normalized feature representation, a cross-modal feature representation matrix is ​​constructed. Pattern matching is then performed between the cross-modal feature representation matrix and a pre-defined disease progression pattern library to identify the Alzheimer's disease subtype category corresponding to the target individual, including: Extract the temporal correlation features of each feature in the normalized feature representation, and extract the temporal evolution templates corresponding to multiple Alzheimer's disease subtypes from the disease progression pattern library; Alignment calculations are performed between the temporal correlation features and the temporal evolution templates to obtain the temporal matching degree between the normalized feature representation and each temporal evolution template; For each feature in the normalized feature representation, a spatial distribution constraint representing the intermodal collaborative relationship is constructed in the multidimensional feature space. Based on the temporal correlation feature and the spatial distribution constraint, the cross-modal feature representation matrix is ​​constructed. Calculate the structural similarity between the cross-modal feature representation matrix and the standard feature representation matrix templates corresponding to each Alzheimer's disease subtype in the disease progression pattern library; The cross-modal feature representation matrix is ​​correlated with the intervention response features of historical confirmed cases in the health management resource database to obtain the similarity of response patterns between the target individual and historical cases of each subtype. Based on the temporal matching degree, the structural similarity, and the response pattern similarity, the Alzheimer's disease subtype category corresponding to the target individual is determined.

[0010] For each feature in the normalized feature representation, a spatial distribution constraint representing the intermodal collaborative relationship is constructed in the multidimensional feature space. Based on the temporal correlation features and the spatial distribution constraint, the cross-modal feature representation matrix is ​​constructed, including: Each feature in the normalized feature representation is mapped to a multidimensional feature space to obtain the spatial coordinates of each feature in the multidimensional feature space. Multiple modal co-evolutionary patterns corresponding to Alzheimer's disease subtypes are determined from the disease progression pattern library, and the spatial co-evolutionary degree between features in the multidimensional feature space is calculated based on the modal co-evolutionary patterns. Based on the spatial synergy, a spatial distribution constraint is generated that includes spatial distance constraints and synergy direction constraints between features; Multimodal feature evolution data of historical cases are extracted from the health management resource database, and the spatial distribution constraints are corrected based on the multimodal feature evolution data to obtain the corrected spatial distribution constraints; The temporal correlation features are arranged in chronological order to form a temporal dimension sequence, and the features in the normalized feature representation are arranged spatially according to the corrected spatial distribution constraints to form a spatial dimension sequence. The cross-modal feature representation matrix is ​​constructed based on the temporal dimension sequence and the spatial dimension sequence.

[0011] Based on the aforementioned Alzheimer's disease subtypes, a combination of management strategies suitable for each subtype is extracted from a health management resource database to generate a personalized health management plan, including: Extract the set of standard management strategies corresponding to the Alzheimer's disease subtype from the health management resource library, and extract the feature representation of the target individual from the cross-modal feature representation matrix; The characteristic representation of the target individual is matched and calculated with the applicable characteristic conditions of each management strategy in the standard management strategy set to obtain the suitability of each management strategy for the target individual. Extract the disease progression stage classification criteria corresponding to the Alzheimer's disease subtype from the disease progression pattern library, and determine the current disease progression stage of the target individual based on the cross-modal feature representation matrix and the disease progression stage classification criteria; Extract the phased management strategy weight configuration scheme corresponding to the disease progression stage from the health management resource library; The suitability of the target individual is weighted according to the phased management strategy weight configuration scheme, and multiple management strategies are selected from the standard management strategy set based on the weighted suitability to determine the personalized health management plan.

[0012] Real-time feedback data is collected from the target individual during the implementation of the personalized health management plan. This real-time feedback data is then correlated with the cross-modal feature representation matrix to obtain the strategy effectiveness evaluation results, including: Collect real-time feedback data from the target individual during the execution of the personalized health management plan; The baseline feature representation of the target individual is extracted from the cross-modal feature representation matrix, the real-time feedback data is converted into a real-time feature representation of the same dimension as the baseline feature representation, and the feature offset between the real-time feature representation and the baseline feature representation is calculated. The expected feature evolution trajectory corresponding to the Alzheimer's disease subtype category is extracted from the disease progression pattern library, and the feature offset is matched with the expected feature evolution trajectory to obtain the matching consistency result. The typical intervention effect benchmarks of each management strategy in the management strategy combination are extracted from the health management resource library. The feature offset is compared and analyzed with the typical intervention effect benchmarks to determine the actual intervention effect of each management strategy. Based on the matching consistency results and the actual intervention effects, the effectiveness of each management strategy in the management strategy combination is scored to obtain the strategy effectiveness evaluation results.

[0013] A second aspect of this invention provides a health management system for multiple subtypes of Alzheimer's disease based on five databases, comprising: The data acquisition unit is used to acquire multi-dimensional health characteristic data of the target individual. The feature mapping unit is used to extract features from the multi-dimensional health feature data to obtain a set of feature vectors, and to map each feature vector in the set of feature vectors to a unified representation space to obtain a normalized feature representation. The subtype identification unit is used to construct a cross-modal feature representation matrix based on the temporal correlation and spatial distribution constraints between features in the normalized feature representation, and to perform pattern matching between the cross-modal feature representation matrix and a preset disease progression pattern library to identify the Alzheimer's disease subtype category corresponding to the target individual. The solution generation unit is used to extract a combination of management strategies that are compatible with the Alzheimer's disease subtype from the health management resource library based on the Alzheimer's disease subtype category, and generate a personalized health management solution. The feedback analysis unit is used to collect real-time feedback data of the target individual during the execution of the personalized health management plan, and to perform correlation analysis between the real-time feedback data and the cross-modal feature representation matrix to obtain the strategy effectiveness evaluation result. The strategy effectiveness evaluation result is used to iteratively optimize the personalized health management plan.

[0014] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0015] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0016] This invention addresses the individualized health management needs of its users by extracting features from multi-dimensional health characteristic data and mapping them to a unified representation space. This process forms a normalized feature representation, effectively eliminating dimensional and semantic differences between different modalities and improving the robustness and interpretability of feature expression. The cross-modal feature representation matrix integrates temporal correlation and spatial distribution constraints, significantly enhancing the ability to characterize the complex phenotypes of Alzheimer's disease. High-precision matching with a disease progression pattern database enables accurate subtype identification, avoiding misclassification or omission of subtypes caused by traditional single-dimensional classification.

[0017] This invention extracts suitable strategy combinations from a health management resource library based on precise subtype categories to generate highly customized personalized health management plans, maximizing the targeting and synergistic effect of interventions. Real-time feedback data during plan execution is correlated with cross-modal feature representation matrices to dynamically quantify strategy effectiveness, providing objective evidence for iterative optimization and gradually approaching the optimal management path, significantly improving disease progression control and patient quality of life.

[0018] This invention integrates data collection, feature processing, pattern matching, resource scheduling, and feedback loop into a cohesive whole through a five-database linkage mechanism, significantly shortening the cycle from data input to solution generation and optimization. The pattern database continuously accumulates multi-subtype development patterns, the resource database dynamically updates management strategies, and feedback data drives model parameter adjustments, forming a self-evolving healthy management ecosystem. This reduces reliance on human resources and experience bias, making it suitable for long-term, sustainable management of large-scale populations. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the health management method for multiple subtypes of Alzheimer's disease based on the five-database linkage according to an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the process of determining the normalized feature representation according to an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0022] Figure 1 This is a flowchart illustrating a health management method for multiple subtypes of Alzheimer's disease based on five-database linkage, according to an embodiment of the present invention. Figure 1 As shown, this multi-subtype health management method for Alzheimer's disease based on the linkage of five databases includes: Acquire multidimensional health characteristic data of the target individual; Feature extraction is performed on the multi-dimensional health feature data to obtain a set of feature vectors, and each feature vector in the set of feature vectors is mapped to a unified representation space to obtain a normalized feature representation. Based on the temporal correlation and spatial distribution constraints among the features in the normalized feature representation, a cross-modal feature representation matrix is ​​constructed. The cross-modal feature representation matrix is ​​then matched with a preset disease progression pattern library to identify the Alzheimer's disease subtype category corresponding to the target individual. Based on the Alzheimer's disease subtype, a combination of management strategies that are suitable for the Alzheimer's disease subtype is extracted from the health management resource library to generate a personalized health management plan; Real-time feedback data is collected from the target individual during the execution of the personalized health management plan. The real-time feedback data is then correlated with the cross-modal feature representation matrix to obtain the strategy effectiveness evaluation result. The strategy effectiveness evaluation result is used to iteratively optimize the personalized health management plan.

[0023] Figure 2 This is a schematic diagram illustrating the process of determining the normalized feature representation according to an embodiment of the present invention. Figure 2 As shown, feature extraction is performed on the multi-dimensional health feature data to obtain a feature vector set, and each feature vector in the feature vector set is mapped to a unified representation space to obtain a normalized feature representation, including: The multidimensional health feature data is grouped according to data modality type to obtain multiple modality data subsets. For each modality data subset, prior knowledge features corresponding to the modality data subset are extracted from the disease progression pattern library and the health management resource library, respectively. The prior knowledge features are fused with the modal data subset, and the fused time-domain features and frequency-domain features are extracted to construct the corresponding modal feature vectors. All modal feature vectors are then summarized to obtain the feature vector set. Based on the multi-subtype disease evolution paths in the disease progression pattern library, a semantic association graph between modalities is constructed, and the association weight distribution between feature vectors of different modalities in the feature vector set is determined through the semantic association graph; Based on the correlation weight distribution, each feature vector in the feature vector set is weighted and fused to obtain each weighted and fused feature vector; The weighted and fused feature vectors are projected onto a unified representation space of a preset dimension to obtain aligned feature vectors. The aligned feature vectors are then normalized to obtain the normalized feature representation.

[0024] After acquiring multi-dimensional health characteristic data of the target individual, the data is grouped and processed according to its modality. Multi-dimensional health characteristic data typically encompasses various modalities, including neuroimaging data (such as MRI and PET scans), neuroelectrophysiological data (such as EEG signals), cognitive assessment scale data, biomarker detection data (such as Aβ42 and p-tau protein concentrations in cerebrospinal fluid / blood), and daily behavior monitoring data. After modality-based segmentation, each data category forms an independent modality data subset. These subsets differ significantly in data structure, sampling frequency, and information dimensions; therefore, it is necessary to extract corresponding prior knowledge features for each modality data subset.

[0025] For each modality subset, prior knowledge features related to that modality are extracted from both the disease progression pattern database and the health management resource database. The disease progression pattern database stores typical evolutionary patterns of different Alzheimer's disease subtypes across various modalities, such as hippocampal atrophy patterns observed in MRI images and characteristic frequency band power change trends in EEG signals for specific subtypes. The health management resource database contains intervention feedback records related to each modality, reflecting the historical impact of different management strategies on various modality indicators. Integrating these prior knowledge features as auxiliary information with the corresponding modality subsets effectively compensates for the generalization limitations of single data-driven methods in small-sample scenarios.

[0026] After fusing prior knowledge features with a subset of modal data, time-domain and frequency-domain features are extracted from the fusion result. Time-domain feature extraction focuses on the statistical properties of the data along the time axis, including statistics such as mean, variance, kurtosis, and autocorrelation coefficient, as well as dynamic indicators such as the time trend slope reflecting the rate of disease progression. Frequency-domain feature extraction obtains the power spectral density distribution of different frequency bands by performing Fourier transform or wavelet transform on the time-series data. This is particularly important for neurophysiological data such as EEG, where specific frequency bands (e.g.,...) are crucial. , , , The power ratio of each band is of significant importance for classifying Alzheimer's disease subtypes. The time-domain features are concatenated with the frequency-domain features to construct the modality feature vector corresponding to that modality. This process is repeated for all modality subsets, and the resulting set of feature vectors containing all modality information is obtained.

[0027] After obtaining the feature vector set, a semantic association graph between modalities is constructed based on the multi-subtype disease evolution paths stored in the disease progression pattern database. This graph uses the feature vectors of each modality as nodes and the semantic correlation between modalities as edges. The weight of the edges reflects the strength of the co-change between the two modalities during disease progression. Specifically, multi-modal joint evolution records of different subtypes are extracted from the disease progression pattern database, and the co-occurrence frequency and covariation pattern of each modality feature during the subtype evolution process are statistically analyzed to measure the strength of the semantic association between modalities. Let the first... The modality and the first The semantic association strength between modalities is Then the semantic association graph can be represented as an association weight matrix. ,in The Line 1 The column element is Through the analysis of The data is normalized row by row to obtain the correlation weight distribution that reflects the relative importance of each modality feature vector in the fusion process.

[0028] Based on the correlation weight distribution, the feature vectors in the feature vector set are weighted and fused. Let the i-th feature vector be... The feature vectors of each modality are The corresponding normalized association weights are Then the weighted fusion of the first eigenvectors Cross-modal information injection is achieved by weighting other modal feature vectors onto itself. ,in for After normalization, the first Line 1 The column weights are used to assign weights to each feature vector. This weighted fusion operation allows each modality's feature vector to retain its own modal information while incorporating semantic constraints from other related modalities, thereby improving the feature vector's ability to express Alzheimer's disease subtypes. The resulting feature vector sets after weighted fusion still exhibit heterogeneity in dimensionality and distribution among modalities, requiring further spatial alignment processing.

[0029] The weighted and fused feature vectors are projected onto a unified representation space of a preset dimension to achieve spatial alignment of feature vectors from different modalities. The projection operation is implemented through a pre-trained linear or nonlinear mapping function. This mapping function, during training, is based on multi-subtype sample data and aims to maximize the cluster density of samples of the same subtype in the unified representation space while minimizing feature overlap between samples of different subtypes. The selection of the preset dimension must balance information preservation integrity with subsequent computational efficiency, and is typically determined based on the number of subtypes in the disease progression pattern database and the degree of feature differences between subtypes. After obtaining the aligned feature vectors through projection, they are normalized to eliminate differences in units and numerical ranges between different modalities. The normalization process uses a standardization method based on sample distribution, adjusting the aligned feature vectors to a standard distribution with a mean of zero and a variance of unit values, ultimately yielding a normalized feature representation.

[0030] Normalized feature representation integrates temporal and frequency domain information from multimodal data, semantic association constraints between modalities, and prior knowledge of disease progression. It enables consistent description of health characteristics across different modalities within a unified representation space, providing high-quality input representations for subsequent construction of cross-modal feature representation matrices based on temporal correlation and spatial distribution constraints, and for Alzheimer's disease subtype identification. The entire feature extraction and normalization process effectively overcomes the challenges of heterogeneity, high dimensionality, and small sample size in multimodal health data by fully utilizing prior knowledge from disease progression pattern libraries and health management resource libraries, thereby improving the accuracy and robustness of subtype identification.

[0031] Based on the multi-subtype disease evolution paths in the disease progression pattern library, a semantic association graph between modalities is constructed. The association weight distribution between feature vectors of different modalities in the feature vector set is determined through this semantic association graph, including: Multiple disease evolution paths corresponding to Alzheimer's disease subtypes are extracted from the disease progression pattern library. Each disease evolution path contains multiple disease stage nodes and the transition relationships between stage nodes. For each disease evolution path, identify the multi-dimensional pathological evolution features involved in the disease stage nodes, and establish a mapping relationship between the disease evolution path and each data modality in the multi-dimensional health feature data; A semantic association graph is constructed based on the mapping relationship. The semantic association graph includes modal nodes and semantic association edges. The modal nodes correspond to each modal feature vector in the feature vector set. The semantic association edges represent the co-occurrence relationship and causal dependency relationship between different modal nodes based on the disease evolution path. Historical intervention case data were extracted from the health management resource database, and the contribution distribution of different modal feature combinations to the intervention effect in the historical intervention case data was statistically analyzed. The semantic association edges in the semantic association graph are weighted according to the contribution distribution, and the association weight distribution is generated based on the corrected semantic association graph.

[0032] When extracting disease evolution paths corresponding to multiple Alzheimer's disease subtypes from the disease progression pattern database, each evolution path is represented as a directed graph structure. Nodes in the graph correspond to different stages of disease development, and directed edges between nodes correspond to state transitions between adjacent stages. Taking the typical memory-dominant subtype as an example, its evolution path sequentially includes the stages of mild cognitive impairment, episodic memory impairment, executive function impairment, and comprehensive cognitive decline. Each stage node is accompanied by multi-dimensional pathological description information, including biomarker characteristics, neuroimaging manifestations, and cognitive behavior scale score ranges. For other subtypes such as language-dominant, visuospatial-dominant, and frontal lobe-dominant, their respective evolution paths are extracted from the disease progression pattern database in the same manner, ensuring that the extracted set of evolution paths covers all preset Alzheimer's disease subtype categories in the database.

[0033] For each disease evolution path, multidimensional pathological evolution features involved in each disease stage node are identified. These features span multiple data modalities, including neuroimaging modalities (such as structural magnetic resonance imaging volume parameters and PET metabolic distribution), cerebrospinal fluid / blood biochemical modalities (such as Aβ42 / Aβ40 ratio and p-tau217 concentration), cognitive assessment modalities (such as MMSE scores and MoCA sub-scores), and genomic modalities (such as APOE genotype and risk variant sites). When establishing the mapping relationship between the disease evolution path and each data modality, for each disease stage node, the set of modalities with significant pathological significance in that stage is recorded, and the characteristic response direction of that modality in that stage is labeled (increasing, decreasing, or non-monotonic change). This mapping relationship closely links the abstract disease evolution process with specific observable data modalities, providing a structured prior basis for the subsequent construction of semantic association maps.

[0034] Based on the above mapping relationships, a semantic association graph is constructed. Each data modality is represented as a modality node, and the node's attributes contain statistical information on the stage distribution of that modality's appearance across all subtype evolutionary paths. Semantic association edges connect different modality nodes. The existence of an edge is determined by the co-occurrence of two modalities in the same disease stage node (co-occurrence relationship) or by a change in one modality preceding and driving a change in another modality along the evolutionary path (causal dependency relationship). For co-occurrence relationships, let the modality... With mode The number of times they co-occur at the same stage node across all subtype evolutionary paths is Then the co-occurrence intensity Defined as ,in For modality The total number of times it appears in all stage nodes. For modality The total number of occurrences across all stage nodes. For causal dependencies, by analyzing the order of modal characteristic changes in the evolutionary path, a statistical method based on Granger causality testing is used to calculate the modalities. For modes causal dependence strength Finally, the initial weights of semantically related edges. It is obtained by weighted combination of co-occurrence strength and causal dependence strength, i.e. ,in The hyperparameter, used to balance the contribution ratios of the two types of relationships, has a range of values. It can be adjusted according to the characteristics of the specific subtype dataset; the default setting is 0.5.

[0035] When extracting historical intervention case data from the health management resource database, the historical intervention case data includes the patient's baseline characteristic modality combination for each case, the type of intervention strategy used, and post-intervention effect evaluation indicators (such as the magnitude of cognitive function improvement, the duration of maintenance of daily living abilities, and the rate of change in biomarkers). Statistical analysis is performed on the historical intervention case data to calculate the contribution distribution of different modality characteristic combinations to the intervention effect. Specifically, for each modality characteristic combination... (A subset consisting of several modal nodes) The proportion of cases where the intervention effect reaches the preset excellent standard in which the combined features are used to guide intervention decisions is counted as the contribution score of the modal combination. Furthermore, the contribution of the mode combination is decomposed into each pair of mode nodes to obtain the contribution of each pair of modes. Average contribution gain across all combinations covering this mode pair This gain reflects the additional improvement in intervention effect when the two modalities are used together compared to when they are used alone.

[0036] The semantic association edges in the semantic association graph are weighted according to the contribution distribution, and the adjustment strategy is based on contribution gain. For initial edge weights Perform multiplicative adjustments to correct the edge weights. for ,in To adjust the intensity control coefficient, which regulates the impact of historical intervention contributions on the edge weights of the semantic association graph, a default value of 1.0 is used. After adjustment, the edge weights in the semantic association graph simultaneously encode prior knowledge of disease evolution (co-occurrence relationships and causal dependencies) and empirical knowledge of historical interventions (contribution distribution), making the graph more practically valuable for health management of Alzheimer's disease subtypes.

[0037] When generating the association weight distribution based on the corrected semantic association graph, for each modality node Adjust the edge weights between it and all other modal nodes. Normalization is performed to obtain the modes. For the normalized association weight vectors of other modalities, this vector forms a row of the association weight distribution. The normalized association weight vectors of all modal nodes are aggregated to form a complete association weight distribution matrix, which can be directly used in subsequent cross-modal feature fusion steps. In the normalization process, the modified edge weights are transformed using the Softmax function to ensure that the sum of all components of the weight vector corresponding to each modal node is 1, while preserving the relative magnitudes of the weights to avoid fusion bias caused by differences in the original weight dimensions.

[0038] In practical applications, the semantic association graph is not statically fixed, but dynamically updated as historical intervention case data in the health management resource database continuously accumulates. Each time a new batch of historical intervention case data is added, the contribution gain is recalculated. The edge weights of the semantic association graph are incrementally updated according to the above correction process, so that the distribution of association weights can reflect the latest clinical evidence. This dynamic update mechanism ensures the adaptability of the semantic association graph to changes in the characteristics of different patient groups, effectively improving the personalization accuracy and long-term effectiveness of health management programs for multi-subtype Alzheimer's disease.

[0039] Based on the temporal correlation and spatial distribution constraints among the features in the normalized feature representation, a cross-modal feature representation matrix is ​​constructed. Pattern matching is then performed between the cross-modal feature representation matrix and a pre-defined disease progression pattern library to identify the Alzheimer's disease subtype category corresponding to the target individual, including: Extract the temporal correlation features of each feature in the normalized feature representation, and extract the temporal evolution templates corresponding to multiple Alzheimer's disease subtypes from the disease progression pattern library; Alignment calculations are performed between the temporal correlation features and the temporal evolution templates to obtain the temporal matching degree between the normalized feature representation and each temporal evolution template; For each feature in the normalized feature representation, a spatial distribution constraint representing the intermodal collaborative relationship is constructed in the multidimensional feature space. Based on the temporal correlation feature and the spatial distribution constraint, the cross-modal feature representation matrix is ​​constructed. Calculate the structural similarity between the cross-modal feature representation matrix and the standard feature representation matrix templates corresponding to each Alzheimer's disease subtype in the disease progression pattern library; The cross-modal feature representation matrix is ​​correlated with the intervention response features of historical confirmed cases in the health management resource database to obtain the similarity of response patterns between the target individual and historical cases of each subtype. Based on the temporal matching degree, the structural similarity, and the response pattern similarity, the Alzheimer's disease subtype category corresponding to the target individual is determined.

[0040] After obtaining the normalized feature representation, the temporal variation patterns of each feature dimension are extracted. Specifically, for multi-dimensional features such as cognitive assessment scores, neuroimaging indicators, and biomarker concentrations collected from the target individual at different time points, a feature sequence based on a time axis is constructed. A sliding window mechanism is used to capture the changing trends and fluctuations of each feature within adjacent time periods, thereby obtaining temporally correlated features. Simultaneously, temporal evolution templates corresponding to multiple Alzheimer's disease subtypes are extracted from a disease progression pattern library. The disease progression pattern library pre-stores the typical feature evolution trajectories of different subtypes (such as memory-dominant, language-impaired, executive function-declining, and posterior cortical atrophy) at various stages of the disease. Each subtype corresponds to a set of standardized feature change sequences indexed by time points, i.e., temporal evolution templates. These templates are derived from large-scale clinical cohort data, and after expert annotation and statistical modeling, they are written into the disease progression pattern library, possessing high representativeness and discriminative power.

[0041] When aligning the temporal correlation features of the target individual with the temporal evolution templates of each subtype, a dynamic time warping method is used to handle the alignment problem between time series of different lengths. Let the temporal correlation feature sequence of the target individual be... , No. The temporal evolution template sequences of each subtype are as follows ,in The length of the target individual sequence. For the first The length of each subtype template sequence is obtained by constructing the cumulative distance matrix and backtracking the optimal path. and Minimum normalized distance between , and thus Normalization to temporal matching degree , The larger the value, the more likely the target individual's characteristic evolution trajectory is to be related to the first... The closer the subtype templates are, the better. This step effectively solves the problem of sequence length differences caused by inconsistent follow-up frequencies among different individuals, ensuring fairness in cross-individual comparisons.

[0042] Based on the temporal correlation features obtained, spatial distribution constraints are further constructed for each feature in the normalized feature representation in a multidimensional feature space to characterize the cooperative relationship between modalities. The multimodal features at each time point are regarded as a point in a high-dimensional space. By calculating the covariance structure between different modal features, a covariance matrix describing the joint distribution law of each modality feature is obtained. Covariance matrix The Line 1 Column elements Reflects the first The modal feature and the first The degree of linear correlation between modal features is defined by positive values ​​indicating that they change in the same direction and negative values ​​indicating that they change in opposite directions, while the absolute value reflects the strength of the synergy. Based on this, spatial distribution constraints are introduced, requiring that the distribution of each modal feature in the multidimensional feature space satisfy the subtype-specific spatial structure specified by the disease progression pattern library, thereby embedding the synergistic relationship between modalities into the feature representation process.

[0043] Cross-modal feature representation matrix The construction integrates information from two dimensions: temporal correlation features and spatial distribution constraints. Temporal correlation features are used as row indices in the matrix, and the spatial coordination structure of multimodal features is used as column indices. The values ​​of each modal feature at each time point are multiplied by their corresponding spatial constraint weights and then filled into the corresponding positions in the matrix. This results in a cross-modal feature representation matrix that simultaneously reflects the temporal evolution of features and the spatial coordination relationships between modalities. The number of rows in this matrix equals the number of time nodes, and the number of columns equals the number of dimensions of the modal features involved in the representation. Each element in the matrix carries both temporal information and spatial distribution constraint information, giving it strong discriminative ability.

[0044] Calculate the cross-modal feature representation matrix Standard feature representation matrix templates corresponding to each Alzheimer's disease subtype in the disease progression pattern library When assessing structural similarity between components, a matrix distance metric based on the Frobenius norm is used. and If the dimensions are the same (or aligned by interpolation or truncation if they are different), then the structural difference between the two is... Defined as ,in Denotes the Frobenius norm. After normalization, it is converted into structural similarity. , The larger the value, the greater the cross-modal feature representation matrix of the target individual compared to the first modality. The closer the standard templates of each subtype are, the better. Standard feature representation matrix template The average of the cross-modal feature representation matrices of all confirmed cases of this subtype in the disease progression pattern database is obtained, which can represent the typical pattern of this subtype in terms of feature temporal evolution and modal co-structure.

[0045] Cross-modal feature representation matrix Association analysis was conducted with the intervention response characteristics of historical confirmed cases in the health management resource database to provide supplementary evidence for subtype identification from the perspective of intervention effectiveness. The intervention response characteristics of historical confirmed cases record the changes in modal characteristics of different subtype patients after receiving specific intervention programs. These trajectories are also stored in the health management resource database in the form of cross-modal feature representation matrices. By calculating the target individual's... The cosine similarity between the response pattern and the intervention response feature matrix of historical cases of each subtype is used to obtain the response pattern similarity. , The higher the value, the closer the target individual is to the first. The closer the historical cases of a subtype are in their intervention response patterns, the more indirectly the identification and judgment of that subtype is supported.

[0046] Overall time series matching degree Structural similarity Similarity to response pattern The evaluation results from the three dimensions were used to calculate the target individual's classification into the third dimension using a weighted fusion strategy. The composite score of each Alzheimer's disease subtype As shown in the following formula: ; in, , , These are the fusion weights for temporal matching degree, structural similarity, and response pattern similarity, respectively. The sum of these three weights is 1, and all are greater than 0. The values ​​of the fusion weights are adaptively adjusted based on the historical discrimination accuracy of each dimension's evaluation metrics in different subtype recognition tasks, assigning greater weight to dimensions with higher discrimination accuracy. Finally, the comprehensive score is taken. The largest subtype is used as the Alzheimer's disease subtype category corresponding to the target individual and is output to the subsequent personalized health management plan generation stage. When the difference between the highest score and the second highest score is less than a preset threshold, a multi-subtype mixed determination mechanism is triggered, marking the target individual as a mixed subtype, and incorporating management strategies for multiple subtypes into the subsequent management plan to address the actual situation of blurred subtype boundaries in clinical practice.

[0047] For each feature in the normalized feature representation, a spatial distribution constraint representing the intermodal collaborative relationship is constructed in the multidimensional feature space. Based on the temporal correlation features and the spatial distribution constraint, the cross-modal feature representation matrix is ​​constructed, including: Each feature in the normalized feature representation is mapped to a multidimensional feature space to obtain the spatial coordinates of each feature in the multidimensional feature space. Multiple modal co-evolutionary patterns corresponding to Alzheimer's disease subtypes are determined from the disease progression pattern library, and the spatial co-evolutionary degree between features in the multidimensional feature space is calculated based on the modal co-evolutionary patterns. Based on the spatial synergy, a spatial distribution constraint is generated that includes spatial distance constraints and synergy direction constraints between features; Multimodal feature evolution data of historical cases are extracted from the health management resource database, and the spatial distribution constraints are corrected based on the multimodal feature evolution data to obtain the corrected spatial distribution constraints; The temporal correlation features are arranged in chronological order to form a temporal dimension sequence, and the features in the normalized feature representation are arranged spatially according to the corrected spatial distribution constraints to form a spatial dimension sequence. The cross-modal feature representation matrix is ​​constructed based on the temporal dimension sequence and the spatial dimension sequence.

[0048] When mapping each feature in the normalized feature representation to a multidimensional feature space, each feature vector is embedded into a high-dimensional Euclidean space through a nonlinear projection function. This allows features originally distributed across different modalities and dimensions to be compared and analyzed within a unified geometric framework. Specifically, cognitive assessment features, neuroimaging features, biochemical indicator features, genomic features, and behavioral monitoring features, after their respective projection transformations, form corresponding spatial coordinate points. Let the... The spatial coordinates of each feature in the multidimensional feature space are: Its dimension is consistent with the preset dimension of the projection space. This mapping method preserves the semantic information of the original features while eliminating scale differences between different modalities, providing a geometric basis for subsequent calculation of spatial coherence. After mapping, the relative positional relationship of each feature in the multidimensional feature space directly reflects its potential correlation at the disease representation level.

[0049] When identifying modal co-evolutionary patterns corresponding to multiple Alzheimer's disease subtypes from a disease progression pattern database, typical patterns of multimodal feature co-changes at each stage of disease progression are extracted for different subtypes, such as memory-dominant, executive function impairment, and visuospatial impairment. These modal co-evolutionary patterns describe the synchronous, sequential, or mutually restrictive relationships between different features during disease progression. Based on these co-evolutionary patterns, the correlation between any two features in the multidimensional feature space is calculated. and Spatial coordination between The calculation of spatial synergy considers two aspects: first, the Euclidean distance between the two features in the multidimensional feature space; the closer the distance, the closer the two features are in disease representation; second, the frequency with which the two features jointly participate in the same stage of evolution in the co-evolutionary pattern; the higher the frequency, the more stable the synergistic relationship between the two features. After normalizing these two aspects, a weighted sum is obtained. Its value range is The larger the value, the higher the degree of spatial coordination between the two features.

[0050] Based on spatial coordination This generates spatial distribution constraints that include spatial distance constraints between features and cooperative direction constraints. The spatial distance constraint specifies the maximum allowable distance threshold that the spatial coordinates corresponding to features with high cooperativeness should satisfy in the multidimensional feature space. If If the preset coordination threshold is exceeded, then and The Euclidean distance between them must not exceed The cooperative direction constraint stipulates that cooperative feature pairs should evolve along similar directions in the multidimensional feature space, that is, the cosine of the angle between their feature evolution vectors should not be lower than a preset direction consistency threshold. These two types of constraints together constitute a spatial distribution constraint set, which is used to constrain the spatial arrangement of features when constructing a cross-modal feature representation matrix, ensuring that the matrix structure can truly reflect the collaborative relationship between multimodal features, rather than relying solely on temporal information for a single-dimensional description.

[0051] When extracting multimodal feature evolution data from historical cases in a health management resource database, historical cases similar to the target individual in terms of basic demographic characteristics and initial subtype tendency are selected. The spatial coordinate trajectories of these multimodal features at different disease stages are obtained. Based on these historical evolution trajectories, the initially generated spatial distribution constraints are corrected using data-driven methods. Specifically, the correction method involves statistically analyzing the proportion of each feature pair in historical cases that actually satisfy spatial distance and co-directional constraints. If the actual co-directional performance of a feature pair in historical data deviates significantly from the initial constraints, then the feature pair is... and Perform adaptive adjustment. Let the correction coefficient be... Its value is determined by the statistical variance of this feature's contribution to synergistic performance in historical cases; the larger the variance, the larger the correction magnitude, thus enhancing the robustness of the constraint. The corrected spatial distance constraint threshold is denoted as... The corrected directional consistency threshold is denoted as Together, these two constitute the modified spatial distribution constraint, which has stronger individual adaptability and data support compared to the initial constraint.

[0052] When constructing the cross-modal feature representation matrix, the temporally related features are arranged in chronological order to form a time-dimensional sequence. Each column of the time-dimensional sequence corresponds to a time node, and each row corresponds to a feature channel, thus reflecting the temporal dynamics of disease progression in the column direction of the matrix. At the same time, the features in the normalized feature representation are constrained according to the modified spatial distribution. and Arrange spatial dimensions to improve spatial synergy. Higher-order features are arranged adjacently along the row direction of the matrix, thus reflecting the spatial cooperative structure among features in the row direction. The specific arrangement process employs a constraint-based hierarchical clustering method, prioritizing features that satisfy the modified spatial distance constraint into the same cluster, and then finely adjusting the feature order within each cluster according to the cooperative direction constraint, ultimately forming a spatial dimension sequence. Combining the temporal dimension sequence and the spatial dimension sequence according to row-column correspondence yields the cross-modal feature representation matrix. Each element in the matrix simultaneously carries evolutionary information in the temporal dimension and collaborative structural information in the spatial dimension, enabling subsequent pattern matching to be performed at both the temporal and spatial levels, significantly improving the accuracy and stability of subtype identification.

[0053] In practical applications, when the multimodal data of a target individual is partially missing, spatial distribution constraints can also play an auxiliary role in feature completion. Specifically, for missing features, spatial synergy between the missing feature and existing features can be used to complete the feature completion process. and the corrected spatial distance constraints Inferring approximate coordinates of missing features in a multidimensional feature space. This, to some extent, compensates for the impact of missing data on the completeness of the cross-modal feature representation matrix. The inference method is based on the spatial coordinates of known collaborative features, while satisfying... Within the constrained spherical neighborhood, combined with cooperative direction constraints The most probable location of the missing feature is determined, and the coordinates of that location are used as the estimated value of the missing feature and filled into the corresponding position in the matrix. This mechanism makes the construction of cross-modal feature representation matrices highly operable even in real clinical scenarios with incomplete data, enhancing the practicality of overall health management methods.

[0054] Based on the aforementioned Alzheimer's disease subtypes, a combination of management strategies suitable for each subtype is extracted from a health management resource database to generate a personalized health management plan, including: Extract the set of standard management strategies corresponding to the Alzheimer's disease subtype from the health management resource library, and extract the feature representation of the target individual from the cross-modal feature representation matrix; The characteristic representation of the target individual is matched and calculated with the applicable characteristic conditions of each management strategy in the standard management strategy set to obtain the suitability of each management strategy for the target individual. Extract the disease progression stage classification criteria corresponding to the Alzheimer's disease subtype from the disease progression pattern library, and determine the current disease progression stage of the target individual based on the cross-modal feature representation matrix and the disease progression stage classification criteria; Extract the phased management strategy weight configuration scheme corresponding to the disease progression stage from the health management resource library; The suitability of the target individual is weighted according to the phased management strategy weight configuration scheme, and multiple management strategies are selected from the standard management strategy set based on the weighted suitability to determine the personalized health management plan.

[0055] After determining the Alzheimer's disease subtype of the target individual, a combination of management strategies suitable for that subtype needs to be extracted from the health management resource library to generate a personalized health management plan. The health management resource library pre-stores standard management strategy sets for each Alzheimer's disease subtype. Each standard management strategy set consists of multiple independent management strategies, and each strategy is accompanied by a feature condition vector describing its applicability, covering multiple dimensions such as cognitive function score range, neuroimaging indicator thresholds, biomarker concentration ranges, and daily living ability ratings. Based on the identified subtype index, the corresponding standard management strategy set is located and extracted from the health management resource library, while simultaneously using the constructed cross-modal feature representation matrix... Extracting the feature representation vector of the target individual This feature representation vector comprehensively reflects the current health status distribution of the target individual across various modal dimensions.

[0056] The feature representation vector of the target individual In the set of standard management strategies, the first A vector of applicable characteristic conditions for each management strategy Perform matching calculations to obtain the initial fit of the management strategy with the target individual. During matching calculations, normalized Euclidean distance is used to measure the degree of deviation for continuous feature dimensions, while a class consistency indicator function is used to determine the deviation for discrete feature dimensions. The matching results of both types of dimensions are then combined to obtain the final result. The calculation method is as follows: ; in, The total number of feature dimensions. The first vector representing the feature of the target individual The value of dimension, For the first The first management strategy applicable feature condition vector The value of dimension, For the first Importance weights of dimensional features This is the distance metric function for the corresponding dimension. Perform the above calculation sequentially on all management strategies to obtain the complete initial fit vector. ,in This represents the total number of strategies in the standard management strategy set.

[0057] After completing the initial fit calculation, the fit needs to be further refined based on the target individual's current disease progression stage. This involves extracting disease progression stage classification criteria corresponding to the target individual's Alzheimer's disease subtype from a disease progression pattern database. These criteria describe the typical feature distribution range of each stage in the form of a multidimensional feature threshold combination, typically dividing disease progression into several stages such as preclinical, mild cognitive impairment, mild dementia, moderate dementia, and severe dementia. The cross-modal feature representation matrix is ​​then used. The extracted comprehensive characteristics of the target individual are compared one by one with the classification criteria for each stage. The membership degree between the target individual's characteristic distribution and the standard characteristic interval of each stage is calculated. The stage with the highest membership degree is determined as the current disease progression stage of the target individual, and is denoted as stage. .

[0058] Disease progression stage The judgment result directly affects the selection and weighting of subsequent management strategies, and the extraction of stage-specific strategies from the health management resource library. The corresponding phased management strategy weight configuration scheme uses a weight vector. Stored in the form of , where Indicates the stage Next The phase-adaptation weights of each management strategy satisfy... The design of the phased management strategy weighting scheme is based on intervention effect data from a large number of historical cases at different disease progression stages. By statistically analyzing the actual intervention benefits of each management strategy at different stages, the weights are calibrated offline and updated regularly to ensure that the weighting scheme can reflect the inherent relationship between disease progression stages and the effectiveness of management strategies.

[0059] Based on the phased management strategy weight allocation scheme, the initial fit vector of the target individual is determined. Perform weighted processing to obtain the weighted fitness score. : ; Weighted processing results in higher overall scores for management strategies with stronger historical performance at the current stage of disease progression, while management strategies with weaker applicability at this stage, even with high initial fit, will have their overall ranking lowered due to weighting suppression. Weighted fit for all management strategies. Sort in descending order and select the quantity threshold according to a preset strategy. Select the top-ranked management strategies to form a personalized management strategy combination for the target individual.

[0060] After determining the combination of management strategies, the execution parameters of each selected strategy need to be individually adjusted. Each management strategy includes several adjustable execution parameters, such as the daily duration of cognitive training, the intensity level of exercise intervention, the dosage ratio of nutritional supplements, and the frequency of medication administration for drug-assisted management. This is based on the target individual characteristic representation vector. The feature condition vector applicable to this management strategy The deviations across various dimensions are used to adjust the execution parameters through linear interpolation or piecewise mapping, ensuring that the intensity of the management strategy's execution precisely matches the actual state of the target individual. The adjusted execution parameters of each management strategy, together with the strategy content, constitute a complete personalized health management plan text, output in a structured format, including fields such as strategy name, execution parameters, execution cycle, precautions, and expected effect indicators, facilitating understanding and implementation by medical personnel and caregivers of the target individual.

[0061] After a personalized health management plan is generated, the plan content, along with the target individual's subtype, disease progression stage, initial fit vector, and weighted fit vector, are simultaneously written into the historical record table of the health management resource library. This serves as the foundational data for subsequent strategy effectiveness evaluation and plan iteration optimization. When the target individual generates real-time feedback data during plan execution, this feedback data can be correlated with the current plan record to continuously update the historical effectiveness statistics of each management strategy under different subtypes and disease progression stages. This drives the dynamic correction of the phased management strategy weight configuration plan, forming a closed-loop optimization mechanism.

[0062] Real-time feedback data is collected from the target individual during the implementation of the personalized health management plan. This real-time feedback data is then correlated with the cross-modal feature representation matrix to obtain the strategy effectiveness evaluation results, including: Collect real-time feedback data from the target individual during the execution of the personalized health management plan; The baseline feature representation of the target individual is extracted from the cross-modal feature representation matrix, the real-time feedback data is converted into a real-time feature representation of the same dimension as the baseline feature representation, and the feature offset between the real-time feature representation and the baseline feature representation is calculated. The expected feature evolution trajectory corresponding to the Alzheimer's disease subtype category is extracted from the disease progression pattern library, and the feature offset is matched with the expected feature evolution trajectory to obtain the matching consistency result. The typical intervention effect benchmarks of each management strategy in the management strategy combination are extracted from the health management resource library. The feature offset is compared and analyzed with the typical intervention effect benchmarks to determine the actual intervention effect of each management strategy. Based on the matching consistency results and the actual intervention effects, the effectiveness of each management strategy in the management strategy combination is scored to obtain the strategy effectiveness evaluation results.

[0063] After the target individual begins implementing a personalized health management plan, real-time feedback data generated during the plan's execution is continuously collected. This data comes from multiple sources, including physiological signals recorded by wearable devices (such as heart rate variability, sleep structure, and daytime activity levels), task completion and error rates uploaded to a cognitive training platform, behavioral observation scale scores completed by caregivers or family members, and imaging or biochemical test results from regular follow-up examinations. This data is aggregated in batches within fixed time windows to ensure consistency in time granularity and coverage dimensions for each batch of real-time feedback data, thus providing a stable input foundation for subsequent correlation analysis.

[0064] From the cross-modal feature representation matrix Extract the baseline features of the target individual before the program starts. , Dimensions and To maintain consistency with the feature dimensions, and using this as a reference anchor for subsequent comparisons, the collected real-time feedback data undergoes the same preprocessing steps as the initial feature extraction (including denoising, normalization, and modal alignment) to transform it into a reference anchor. Real-time feature representation of the same dimension Based on this, the feature offset between the real-time feature representation and the baseline feature representation is calculated. Its definition is Feature offset Each component corresponds to the magnitude of feature change under different modalities. Positive values ​​indicate that the feature in that dimension has improved or strengthened relative to the baseline, while negative values ​​indicate relative degradation or weakening. To eliminate the interference of differences in the dimensions of features across different modalities on the interpretation of the offset, [the following is a partial translation of the remaining text, which is not possible without further context]. Each component is standardized so that the variation range of each component can be compared horizontally on a unified scale.

[0065] Extract the expected feature evolution trajectory corresponding to the Alzheimer's disease subtype to which the target individual belongs from the disease progression pattern database. , This describes the direction and range of characteristic changes that patients of this subtype should exhibit at different stages under standard intervention conditions. It is derived from statistical fitting of historical case data and is subtype-specific. (Characteristic offset is then described.) and Perform a dimension-by-dimensional matching analysis at the corresponding time points: For each feature dimension, determine... Does the direction of change in this dimension correspond to... The expected direction at this time point is consistent, and the magnitude of change is calculated to see if it falls within the expected range. The consistency of direction and magnitude across all dimensions is combined to obtain the matching consistency result. , The range of values ​​is A higher value indicates that the actual evolution of the target individual's characteristics matches the expected evolutionary trajectory of that subtype more closely, suggesting that the overall scheme guides the individual to change along a normal progression path. Below the preset threshold When the actual evolutionary trajectory deviates, it indicates that further identification of the specific characteristics of the deviation is needed to provide a basis for subsequent adjustments to the scheme.

[0066] Typical intervention efficacy benchmarks corresponding to each management strategy in the management strategy combination were extracted from the health management resource database. ,in For management strategy indexing, The first one is described The typical characteristic change patterns of a management strategy in historical intervention cases are analyzed, including the expected direction of improvement, the mean magnitude of improvement, and the confidence interval for each characteristic dimension. Characteristic offsets are then considered. Corresponding to each strategy Comparative analysis: For each management strategy, calculate... The components related to the main dimensions of this strategy and Cosine similarity between and amplitude deviation rate , Defined as the ratio of the difference between the actual change magnitude and the typical baseline magnitude average to the baseline magnitude average. The degree of alignment between the direction of the intervention effect and the actual situation This reflects the degree of deviation in the intensity of the intervention's effect. (Comprehensive) and Determine the first The actual intervention effect of management strategies :when higher and When the intervention is within a reasonable range, the strategy is deemed effective; when... lower or If the tolerance limit is exceeded, the intervention effect of the strategy is deemed insufficient or there is a directional bias.

[0067] Based on matching consistency results The actual intervention effects of various management strategies Each management strategy in the management strategy portfolio is scored for effectiveness. The calculation takes into account two aspects of information: on the one hand, ... As a global consistency correction factor, it applies a holistic adjustment to the scores of all strategies. When the overall evolutionary trajectory matches well, the upper limit of the score for each strategy is correspondingly increased. On the other hand, with As the core basis for assessing the contribution of individual strategies, it directly reflects the actual effectiveness of the strategy on the target individual. Specifically, Depend on and The weighted combination is determined, that is ,in and These are the fusion weights of the global consistency correction factor and the individual intervention effect, respectively, satisfying... The values ​​for both are adaptively adjusted based on subtype and disease progression stage. Effectiveness scores are assigned to all strategies in the management strategy combination. The results are summarized to form a strategy effectiveness evaluation. The results record the score, scoring basis and main contribution dimensions of each strategy in a structured form for subsequent iterative optimization.

[0068] The results of strategy effectiveness evaluation play a direct driving role in iterative optimization, for Below the validity threshold The management strategy triggers a strategy replacement or parameter adjustment process: a candidate strategy that better matches the current characteristic state of the target individual is retrieved from the health management resource library, or the execution frequency and intensity parameters of the original strategy are modified. For Higher than Furthermore, management strategies that closely align with the expected evolutionary trajectory are retained and their implementation weight in the plan can be appropriately strengthened. After multiple rounds of iterative optimization, the effectiveness scores of each strategy in the personalized health management plan tend to stabilize overall, and the accuracy of the plan's adaptation to the target individual continues to improve, thereby achieving dynamic and precise health management for individuals with different subtypes of Alzheimer's disease.

[0069] A second aspect of this invention provides a health management system for multiple subtypes of Alzheimer's disease based on five databases, comprising: The data acquisition unit is used to acquire multi-dimensional health characteristic data of the target individual. The feature mapping unit is used to extract features from the multi-dimensional health feature data to obtain a set of feature vectors, and to map each feature vector in the set of feature vectors to a unified representation space to obtain a normalized feature representation. The subtype identification unit is used to construct a cross-modal feature representation matrix based on the temporal correlation and spatial distribution constraints between features in the normalized feature representation, and to perform pattern matching between the cross-modal feature representation matrix and a preset disease progression pattern library to identify the Alzheimer's disease subtype category corresponding to the target individual. The solution generation unit is used to extract a combination of management strategies that are compatible with the Alzheimer's disease subtype from the health management resource library based on the Alzheimer's disease subtype category, and generate a personalized health management solution. The feedback analysis unit is used to collect real-time feedback data of the target individual during the execution of the personalized health management plan, and to perform correlation analysis between the real-time feedback data and the cross-modal feature representation matrix to obtain the strategy effectiveness evaluation result. The strategy effectiveness evaluation result is used to iteratively optimize the personalized health management plan.

[0070] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0071] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0072] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0073] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A health management method for multiple subtypes of Alzheimer's disease based on the linkage of five databases, characterized in that, include: Acquire multidimensional health characteristic data of the target individual; Feature extraction is performed on the multi-dimensional health feature data to obtain a set of feature vectors, and each feature vector in the set of feature vectors is mapped to a unified representation space to obtain a normalized feature representation. Based on the temporal correlation and spatial distribution constraints among the features in the normalized feature representation, a cross-modal feature representation matrix is ​​constructed. The cross-modal feature representation matrix is ​​then matched with a preset disease progression pattern library to identify the Alzheimer's disease subtype category corresponding to the target individual. Based on the Alzheimer's disease subtype, a combination of management strategies that are suitable for the Alzheimer's disease subtype is extracted from the health management resource library to generate a personalized health management plan; Real-time feedback data is collected from the target individual during the execution of the personalized health management plan. The real-time feedback data is correlated with the cross-modal feature representation matrix to obtain the strategy effectiveness evaluation result. The strategy effectiveness evaluation result is used to iteratively optimize the personalized health management plan.

2. The method according to claim 1, characterized in that, Feature extraction is performed on the multi-dimensional health feature data to obtain a feature vector set, and each feature vector in the feature vector set is mapped to a unified representation space to obtain a normalized feature representation, including: The multidimensional health feature data is grouped according to data modality type to obtain multiple modality data subsets. For each modality data subset, prior knowledge features corresponding to the modality data subset are extracted from the disease progression pattern library and the health management resource library, respectively. The prior knowledge features are fused with the modal data subset, and the fused time-domain features and frequency-domain features are extracted to construct the corresponding modal feature vectors. All modal feature vectors are then summarized to obtain the feature vector set. Based on the multi-subtype disease evolution paths in the disease progression pattern library, a semantic association graph between modalities is constructed, and the association weight distribution between feature vectors of different modalities in the feature vector set is determined through the semantic association graph; Based on the correlation weight distribution, each feature vector in the feature vector set is weighted and fused to obtain each weighted and fused feature vector; The weighted and fused feature vectors are projected onto a unified representation space of a preset dimension to obtain aligned feature vectors. The aligned feature vectors are then normalized to obtain the normalized feature representation.

3. The method according to claim 2, characterized in that, Based on the multi-subtype disease evolution paths in the disease progression pattern library, a semantic association graph between modalities is constructed. The association weight distribution between feature vectors of different modalities in the feature vector set is determined through this semantic association graph, including: Multiple disease evolution paths corresponding to Alzheimer's disease subtypes are extracted from the disease progression pattern library. Each disease evolution path contains multiple disease stage nodes and the transition relationships between stage nodes. For each disease evolution path, identify the multi-dimensional pathological evolution features involved in the disease stage nodes, and establish a mapping relationship between the disease evolution path and each data modality in the multi-dimensional health feature data; A semantic association graph is constructed based on the mapping relationship. The semantic association graph includes modal nodes and semantic association edges. The modal nodes correspond to each modal feature vector in the feature vector set. The semantic association edges represent the co-occurrence relationship and causal dependency relationship between different modal nodes based on the disease evolution path. Historical intervention case data were extracted from the health management resource database, and the contribution distribution of different modal feature combinations to the intervention effect in the historical intervention case data was statistically analyzed. The semantic association edges in the semantic association graph are weighted according to the contribution distribution, and the association weight distribution is generated based on the corrected semantic association graph.

4. The method according to claim 1, characterized in that, Based on the temporal correlation and spatial distribution constraints among the features in the normalized feature representation, a cross-modal feature representation matrix is ​​constructed. Pattern matching is then performed between the cross-modal feature representation matrix and a pre-defined disease progression pattern library to identify the Alzheimer's disease subtype category corresponding to the target individual, including: Extract the temporal correlation features of each feature in the normalized feature representation, and extract the temporal evolution templates corresponding to multiple Alzheimer's disease subtypes from the disease progression pattern library; Alignment calculations are performed between the temporal correlation features and the temporal evolution templates to obtain the temporal matching degree between the normalized feature representation and each temporal evolution template; For each feature in the normalized feature representation, a spatial distribution constraint representing the intermodal collaborative relationship is constructed in the multidimensional feature space. Based on the temporal correlation feature and the spatial distribution constraint, the cross-modal feature representation matrix is ​​constructed. Calculate the structural similarity between the cross-modal feature representation matrix and the standard feature representation matrix templates corresponding to each Alzheimer's disease subtype in the disease progression pattern library; The cross-modal feature representation matrix is ​​correlated with the intervention response features of historical confirmed cases in the health management resource database to obtain the similarity of response patterns between the target individual and historical cases of each subtype. Based on the temporal matching degree, the structural similarity, and the response pattern similarity, the Alzheimer's disease subtype category corresponding to the target individual is determined.

5. The method according to claim 4, characterized in that, For each feature in the normalized feature representation, a spatial distribution constraint representing the intermodal collaborative relationship is constructed in the multidimensional feature space. Based on the temporal correlation features and the spatial distribution constraint, the cross-modal feature representation matrix is ​​constructed, including: Each feature in the normalized feature representation is mapped to a multidimensional feature space to obtain the spatial coordinates of each feature in the multidimensional feature space. Multiple modal co-evolutionary patterns corresponding to Alzheimer's disease subtypes are determined from the disease progression pattern library, and the spatial co-evolutionary degree between features in the multidimensional feature space is calculated based on the modal co-evolutionary patterns. Based on the spatial synergy, a spatial distribution constraint is generated that includes spatial distance constraints and synergy direction constraints between features; Multimodal feature evolution data of historical cases are extracted from the health management resource database, and the spatial distribution constraints are corrected based on the multimodal feature evolution data to obtain the corrected spatial distribution constraints; The temporal correlation features are arranged in chronological order to form a temporal dimension sequence, and the features in the normalized feature representation are arranged spatially according to the corrected spatial distribution constraints to form a spatial dimension sequence. The cross-modal feature representation matrix is ​​constructed based on the temporal dimension sequence and the spatial dimension sequence.

6. The method according to claim 1, characterized in that, Based on the aforementioned Alzheimer's disease subtypes, a combination of management strategies suitable for each subtype is extracted from a health management resource database to generate a personalized health management plan, including: Extract the set of standard management strategies corresponding to the Alzheimer's disease subtype from the health management resource library, and extract the feature representation of the target individual from the cross-modal feature representation matrix; The characteristic representation of the target individual is matched and calculated with the applicable characteristic conditions of each management strategy in the standard management strategy set to obtain the suitability of each management strategy for the target individual. Extract the disease progression stage classification criteria corresponding to the Alzheimer's disease subtype from the disease progression pattern library, and determine the current disease progression stage of the target individual based on the cross-modal feature representation matrix and the disease progression stage classification criteria; Extract the phased management strategy weight configuration scheme corresponding to the disease progression stage from the health management resource library; The suitability of the target individual is weighted according to the phased management strategy weight configuration scheme, and multiple management strategies are selected from the standard management strategy set based on the weighted suitability to determine the personalized health management plan.

7. The method according to claim 1, characterized in that, Real-time feedback data is collected from the target individual during the execution of the personalized health management plan. This real-time feedback data is then correlated with the cross-modal feature representation matrix to obtain the strategy effectiveness evaluation results, including: Collect real-time feedback data from the target individual during the execution of the personalized health management plan; The baseline feature representation of the target individual is extracted from the cross-modal feature representation matrix, the real-time feedback data is converted into a real-time feature representation of the same dimension as the baseline feature representation, and the feature offset between the real-time feature representation and the baseline feature representation is calculated. The expected feature evolution trajectory corresponding to the Alzheimer's disease subtype category is extracted from the disease progression pattern library, and the feature offset is matched with the expected feature evolution trajectory to obtain the matching consistency result. The typical intervention effect benchmarks of each management strategy in the management strategy combination are extracted from the health management resource library. The feature offset is compared and analyzed with the typical intervention effect benchmarks to determine the actual intervention effect of each management strategy. Based on the matching consistency results and the actual intervention effects, the effectiveness of each management strategy in the management strategy combination is scored to obtain the strategy effectiveness evaluation results.

8. A health management system for multiple subtypes of Alzheimer's disease based on five databases, used to implement the method as described in any one of claims 1-7, characterized in that, include: The data acquisition unit is used to acquire multi-dimensional health characteristic data of the target individual. The feature mapping unit is used to extract features from the multi-dimensional health feature data to obtain a set of feature vectors, and to map each feature vector in the set of feature vectors to a unified representation space to obtain a normalized feature representation. The subtype identification unit is used to construct a cross-modal feature representation matrix based on the temporal correlation and spatial distribution constraints between features in the normalized feature representation, and to perform pattern matching between the cross-modal feature representation matrix and a preset disease progression pattern library to identify the Alzheimer's disease subtype category corresponding to the target individual. The solution generation unit is used to extract a combination of management strategies that are compatible with the Alzheimer's disease subtype from the health management resource library based on the Alzheimer's disease subtype category, and generate a personalized health management solution. The feedback analysis unit is used to collect real-time feedback data of the target individual during the execution of the personalized health management plan, and to perform correlation analysis between the real-time feedback data and the cross-modal feature representation matrix to obtain the strategy effectiveness evaluation result. The strategy effectiveness evaluation result is used to iteratively optimize the personalized health management plan.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images