Chronic disease trend prediction system based on deep learning

Abstract

The invention discloses a chronic disease trend prediction system based on deep learning, particularly relates to the field of data feature management, is used for solving the problems of feature leakage and model prediction deviation in chronic disease trend prediction, and is characterized in that multi-source real-time data is subjected to dynamic time window division and grouping sorting to generate structured data blocks; trend and fluctuation components are extracted through depth time sequence decomposition, and feature representation is enhanced in combination with a context difference matrix; calculating a time sequence consistency index and a cross-layer dependence intensity index by using context enhancement features, comprehensively evaluating time consistency and association intensity of the features among different storage layers, and generating a comprehensive feature security score; on the basis of scoring, high-risk features are eliminated, high-context value and high-safety features are reserved preferentially, a finally generated model training feature set prevents prediction deviation caused by feature leakage and a self-circulation effect, and a solid data basis is provided for constructing a chronic disease trend prediction model with high precision and high reliability.

Description

Technical Field

[0001] The present invention relates to the field of data feature management, and more specifically, to a chronic disease trend prediction system based on deep learning. Background Art

[0002] In the analysis and prediction of the development trend of chronic diseases, the integration of multi-dimensional data is one of the key links. These data include dynamic features that are updated in real time (such as exercise steps, social media activity) and high-value features with a long period but far-reaching impact (such as environmental pollution index, economic level indicators). Data is managed through a multi-layer storage architecture. Recent data is usually retained in the hot layer to support real-time analysis, while historical features are archived in the cold layer for long-term preservation. However, in model training, developers often want to improve model accuracy by combining historical features and real-time data. In this process, the management of data time constraints and access rules is particularly important. The existing technology lacks strict review of time validity when referencing feature data, resulting in the historical feature true value being prematurely or improperly referenced to the prediction task of the current period, resulting in feature leakage. This not only causes the model to incorrectly obtain future information during training, but may also trigger the self-circulation effect of the model self-circulating and referencing the cold layer true value on the hot layer data, resulting in falsely high offline performance and poor actual application effect, which seriously affects the accuracy and reliability of chronic disease trend prediction. Therefore, how to effectively manage the timing constraints of feature references in a multi-layer storage architecture, prevent historical data from being used in advance, and block improper interaction between the true value of the cold layer and the real-time data of the hot layer is a technical problem that needs to be solved urgently. Summary of the invention

[0003] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a chronic disease trend prediction system based on deep learning, which generates structured data blocks by dynamically dividing and grouping multi-source real-time data into time windows, extracts trend and fluctuation components through deep time series decomposition, and enhances feature representation in combination with the context difference matrix; uses context-enhanced features to calculate the time series consistency index and cross-layer dependency strength index, comprehensively evaluates the time consistency and correlation strength of features between different storage layers, and generates a comprehensive feature safety score; based on the score, high-risk features are eliminated, and features with high context value and high security are retained preferentially. The final generated model training feature set prevents prediction bias caused by feature leakage and self-loop effects, and provides a solid data foundation for building a high-precision and high-reliability chronic disease trend prediction model to solve the problems raised in the above-mentioned background technology.

[0004] To achieve the above object, the present invention provides the following technical solutions:

[0005] The chronic disease trend prediction system based on deep learning includes: dynamic grouping module, time series analysis module, feature evaluation module, security calculation module and screening optimization module;

[0006] Dynamic grouping module: receives real-time data and sorts the data by segments according to timestamps. It also generates initial data blocks according to data source groups, assigns dynamic time window identifiers to each group of data blocks, and outputs structured initial data blocks and dynamic time window identifiers for the time series parsing module to call.

[0007] Time series parsing module: performs time series decomposition on the data blocks generated by dynamic grouping, extracts long-term trend components and periodic fluctuation components, calculates the context difference matrix based on adjacent data blocks in the time window, generates context enhanced feature representation, and outputs the context enhanced feature representation containing trend components, periodic components and context difference matrix for use by the feature evaluation module;

[0008] Feature evaluation module: Based on context-enhanced feature representation, calculate the temporal consistency index, match the temporal dependency relationship between the cold layer and the hot layer features through the dynamic time warping method, build a feature interaction network, analyze the covariation strength of the cold layer and the hot layer features, generate a cross-layer dependency strength index, and output the temporal consistency index and the cross-layer dependency strength index for comprehensive evaluation by the security computing module;

[0009] Security calculation module: It combines the timing consistency index and the cross-layer dependency strength index, generates feature security scores based on multi-dimensional cross-logic, and marks the timing violation types of high-risk features. It also records the risk information in the metadata management table, outputs feature security scores and high-risk feature tags for use by the screening and optimization module;

[0010] Screening and optimization module: Sort all features according to feature safety scores, combine the contextual relevance of features with cross-layer covariation risks, eliminate or limit the use of low-scoring features, prioritize loading of high-scoring features, generate a feature set for final model training, and output the final high-quality model training feature set for the model training phase.

[0011] In a preferred embodiment, the dynamic grouping module includes the following contents:

[0012] Receive real-time data from multiple sources through the data access interface, and globally sort based on timestamps to ensure that all data is arranged in chronological order;

[0013] In order to adapt to the update frequency and dynamic changes of characteristics of different data sources, information entropy is used to evaluate each data source S i The data complexity within the time window t dynamically adjusts the length of the time window according to the calculated information entropy value;

[0014] According to the dynamically adjusted time window W i (t), divide the data from each data source into several initial data blocks B i (t); each data block B i (t) is contained in the window W i (t) All data records D i (t) and ensure that the data are in chronological order;

[0015] Assign a unique dynamic time window identifier to each generated data block;

[0016] The generated data blocks and their corresponding dynamic time window identifiers are stored in the hot layer database, and the mapping relationship between each dynamic time window identifier and the data block to which it belongs, the time window information and the data source are recorded in the metadata management table.

[0017] In a preferred embodiment, the timing analysis module includes the following contents:

[0018] For each data block B i (t), first extract the long-term trend component and periodic fluctuation component within the data block; use the second-order polynomial approximation method combined with wavelet transform:

[0019] Polynomial approximation: Let B i (t)={x1,x2,…,x N} indicates that in the time window W i (t) N observations collected within a period; the long-term trend component T of the data block is obtained by second-order polynomial approximation i (t): Among them, l is the sequential index of the data in the corresponding data block, a, b, c are the coefficients of the second-order polynomial, and the long-term trend component is saved in the form of a discrete sequence;

[0020] Wavelet transform: To capture the periodic or short-term fluctuation characteristics, the data block is subtracted from the extracted long-term trend component to obtain the original residual sequence: R i (t) = B i (t)-T i (t); then perform continuous wavelet transform on the original residual sequence, use wavelet basis functions and select a set of discrete scales {σ1,σ2,…,σ M},calculate: in, Indicated on the scale σ m The wavelet basis function under τ represents the translation amount, and ν is the integral variable; the main band with concentrated energy in the optimal sub-interval is selected as the main component of the periodic fluctuation, and the corresponding component is recorded as P i (t); Finally, each data block is split into long-term trends Ti (t) and periodic fluctuation P i (t)Two parts.

[0021] In a preferred embodiment, the timing analysis module further includes the following contents:

[0022] After completing the temporal decomposition of each group of data blocks, the differences between adjacent data blocks in the same time window are measured to characterize the dynamic association between local contexts. For this purpose, the context difference matrix M is defined as i (α, t1, t2), where t1, t2 represent the time location of adjacent data blocks, and α is a comprehensive operator for the difference between trend and periodic components; it is calculated as follows: M i (α,t1,t2)=∫(Λ(T i (t1),P i (t1))-Λ(T i (t2),P i (t2))) 2 dα; where Λ(T i (t1),P i (t1)) is used to combine the long-term trend and periodic fluctuation of the data block into a function expression integrated in the α dimension, where α represents the sequential traversal of a series of discrete scales or time domain segments, and finally obtains an integral value similar to the cumulative variance;

[0023] After calculating the context difference matrix, the context difference matrix is ​​integrated with the decomposed long-term trend and periodic fluctuation to construct the context enhanced feature representation CEF. i (t).

[0024] In a preferred embodiment, the feature evaluation module includes the following:

[0025] In order to match the time dependence of the cold layer and hot layer characteristics, the dynamic time warping algorithm is used to transform the cold layer feature sequence C c (t) = {c c1 ,c c2 ,…,c cN} and thermosphere characteristic sequence C h (t) = {c h1 ,c h2 ,…,c hM} for alignment; the distance calculation formula of the dynamic time warping algorithm is: Among them, Π represents all alignment paths, δ(c ci ,c hj ) is the cold layer characteristic c ci With the thermosphere characteristics c hj The distance measurement between them is usually Euclidean distance:

[0026] By normalizing the dynamic time warping distance D DTW To obtain the temporal consistency index TCI i (t): Here, ∈ is a very small positive number that prevents division by zero errors.

[0027] In a preferred embodiment, the feature evaluation module includes the following:

[0028] The cold layer feature C c (t) and thermospheric characteristics C h (t) is used as a node of the network, by calculating the covariation relationship ρ between features c,h (t) is used to determine the edge weights between nodes; the calculation formula for the covariance relationship is: Among them, ρ c,h (t) represents the covariance strength between the cold layer characteristics and the hot layer characteristics at time t, c ck and c hk are the values ​​of the cold layer characteristics and the hot layer characteristics at the kth observation point, and are the means of the cold layer features and the hot layer features, respectively, and N1 is the feature length; if the covariation relationship exceeds the preset threshold, an edge is added to the feature interaction network with a weight of ρ c,h (t);

[0029] Calculation of cross-layer dependency strength index: After the feature interaction network is constructed, the cross-layer dependency strength index is calculated using the network topology analysis method; the specific processing logic is as follows:

[0030] Feature importance score: For each node in the feature interaction network, calculate its feature importance score I i , using PageRank algorithm, the formula is as follows: Where d is the damping factor, represents the set of neighbor nodes of node i, deg(j) is the out-degree of node j;

[0031] Cross-layer dependency strength index calculation: Normalize the feature importance scores of all nodes and define CLDSI i (t) is the weighted average of all edges in the cold layer and hot layer feature interaction network: Among them, E represents all the edges in the feature interaction network, I c and I h Score the importance of cold layer features and hot layer features respectively.

[0032] In a preferred embodiment, the secure computing module includes the following:

[0033] The timing consistency index and cross-layer dependency strength index calculated in the feature evaluation module are fused to generate a comprehensive feature safety score. Based on the generated feature safety score, each feature is risk assessed and high-risk features with timing violations are marked. When the feature safety score is lower than the safety score threshold, the corresponding feature is considered to have a timing violation risk and is marked as a high-risk feature. The identified violation features and their violation types are recorded in the metadata management table.

[0034] In a preferred embodiment, the screening optimization module includes the following contents:

[0035] After excluding high-risk features, the following processing logic is carried out for the remaining feature set; define the context relevance metric CR i (t) represents the average correlation between feature i and its adjacent data blocks in time window t; the inverse index of the context difference matrix is ​​regarded as part of the correlation measure: Among them, Ω(t) represents the relationship between the data block B i (t) The index set of all adjacent or same-window data blocks. Φ is a nonlinear transformation operator used to map larger difference values ​​to lower correlations, thereby obtaining the average context correlation CR between features and neighborhoods. i (t).

[0036] In a preferred embodiment, the screening optimization module further includes the following contents:

[0037] Based on the existing feature security score, the context fusion index UFI is constructed by combining the context relevance metric and the cross-layer dependency strength index. i (t), used to distinguish high-scoring features from potential risk features; the specific calculation formula is (example): UFI i (t) = 1-exp(-FSS i (t) CR i (t)·Θ(CLDSI i (t))); where Θ(CLDSI i (t)) is a correction function for the cross-layer dependency strength index.

[0038] In a preferred embodiment, the screening optimization module further includes the following contents:

[0039] For features whose context fusion index is far below the safety threshold, directly remove them from the candidate feature set or impose usage restrictions, and mark them accordingly in the metadata table to avoid accidental loading during subsequent training or inference;

[0040] Features with a context fusion index higher than the security threshold are considered to be safe and compliant and have high context value, added to the feature whitelist, and registered in the hot layer feature management record to support subsequent fast loading and use;

[0041] After the screening is completed, all high-scoring features that have passed security and context checks are integrated to form the final model training feature set; trainable markers are added to these features in the metadata management table, and their corresponding data block records and time window ranges are synchronously updated to the training database to ensure that features with insufficient scores or illegal markings are not called in actual training.

[0042] Technical effects and advantages of the chronic disease trend prediction system based on deep learning of the present invention:

[0043] The present invention accurately manages real-time and historical data from multiple sources by constructing a multi-level, dynamically adaptive data processing process, ensuring that feature references strictly follow the compliance of timing and dependencies in the prediction of chronic disease trends. First, the received data is dynamically divided into time windows and grouped and sorted to generate structured data blocks, and key trends and fluctuation components are extracted through deep time series decomposition, and the feature representation is enhanced in combination with the context difference matrix; then, the context-enhanced features are used to calculate the time consistency index and the cross-layer dependency strength index, comprehensively evaluate the time and correlation consistency of features between different storage layers, and generate a comprehensive feature safety score; based on these scores, high-risk features can be effectively identified and eliminated, and features with high context value and high security can be retained first; the model training feature set finally generated not only has high context value, but also fundamentally prevents prediction bias caused by feature leakage or self-loop effects, providing a solid data foundation for the realization of a high-reliability, high-precision chronic disease trend prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 This is a schematic diagram of the structure of the chronic disease trend prediction system based on deep learning of the present invention. DETAILED DESCRIPTION

[0045] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0046] Embodiment 1: Figure 1 The present invention provides a chronic disease trend prediction system based on deep learning, including: a dynamic grouping module, a time series analysis module, a feature evaluation module, a security calculation module and a screening optimization module;

[0047] Dynamic grouping module: receives real-time data and sorts the data by segments according to timestamps. It also generates initial data blocks according to data source groups, assigns dynamic time window identifiers to each group of data blocks, and outputs structured initial data blocks and dynamic time window identifiers for the time series parsing module to call.

[0048] Time series parsing module: performs time series decomposition on the data blocks generated by dynamic grouping, extracts long-term trend components and periodic fluctuation components, calculates the context difference matrix based on adjacent data blocks in the time window, generates context enhanced feature representation, and outputs the context enhanced feature representation containing trend components, periodic components and context difference matrix for use by the feature evaluation module;

[0049] Feature evaluation module: Based on context-enhanced feature representation, calculate the temporal consistency index, match the temporal dependency relationship between the cold layer and the hot layer features through the dynamic time warping method, build a feature interaction network, analyze the covariation strength of the cold layer and the hot layer features, generate a cross-layer dependency strength index, and output the temporal consistency index and the cross-layer dependency strength index for comprehensive evaluation by the security computing module;

[0050] Security calculation module: It combines the timing consistency index and the cross-layer dependency strength index, generates feature security scores based on multi-dimensional cross-logic, and marks the timing violation types of high-risk features. It also records the risk information in the metadata management table, outputs feature security scores and high-risk feature tags for use by the screening and optimization module;

[0051] Screening and optimization module: Sort all features according to feature safety scores, combine the contextual relevance of features with cross-layer covariation risks, eliminate or limit the use of low-scoring features, prioritize loading of high-scoring features, generate a feature set for final model training, and output the final high-quality model training feature set for the model training phase.

[0052] In the chronic disease trend prediction system, the multi-source and time series of data are the basis for achieving accurate prediction. The system needs to process real-time data from different sources, such as personal exercise steps, social media activity, and high-value features that are updated periodically, such as environmental pollution index and economic level indicators. These data are managed through a multi-layer storage architecture of "hot → warm → cold" to ensure a balance between real-time analysis and long-term trend research. However, the update frequency of different data sources and the dynamic changes in data characteristics may lead to temporal inconsistency and complexity in data processing. To this end, the core task of the dynamic grouping module is to accurately segment and source the received real-time data, generate initial data blocks with unique identifiers, and ensure that subsequent processing can be carried out on the basis of orderly and high-quality data, thereby effectively preventing feature leakage and self-circulation effects, and improving the overall accuracy and reliability of chronic disease trend prediction.

[0053] The dynamic grouping module includes the following:

[0054] S1.1, data reception and preliminary sorting:

[0055] First, the real-time data from multiple sources is received through the data access interface. Each data record includes data content, timestamp, and data source identification. The received data is globally sorted according to the timestamp to ensure that all data is arranged in chronological order and eliminate time confusion caused by differences in the update frequency of data from different sources.

[0056] S1.2, information entropy calculation and dynamic time window division:

[0057] In order to adapt to the update frequency and dynamic changes of characteristics of different data sources, the information entropy H is used i (t) to evaluate each data source S i The complexity of data in time window t. The calculation formula of information entropy is: Among them, p k (t) represents the probability of the kth feature appearing in the time window t, and K is the total number of feature categories. The time window W is dynamically adjusted according to the calculated information entropy value. i Length of (t): Among them, θ high and θ low is the preset entropy threshold, W small , W medium , W large They correspond to time window lengths of different granularities.

[0058] S1.3: Data grouping and initial data block generation:

[0059] According to the dynamically adjusted time window W i(t), divide the data from each data source into several initial data blocks B i (t). Each data block B i (t) is contained in the window W i (t) All data records D i (t), and ensure that the data is arranged in chronological order: B i (t) = {D i1 ,D i2 ,…,D in where D ij ∈W i (t).

[0060] S1.4: Dynamic time window identifier allocation: Assign a unique dynamic time window identifier DTWI to each generated data block. i (t), which is composed as follows: DTWI i (t) = Hash (S i ∥T start (t)∥T end (t)); where T start (t) and T end (t) are time windows W i (t)’s start and end timestamps, and Hash is a deterministic hash function that ensures the uniqueness and fast retrieval of dynamic time window identifiers.

[0061] S1.5: Metadata Recording and Data Block Storage:

[0062] The generated data blocks and their corresponding dynamic time window identifiers are stored in the hot layer database, and the mapping relationship between each dynamic time window identifier and the data block to which it belongs, the time window information and the data source are recorded in the metadata management table to ensure that these data blocks can be accurately referenced and processed later.

[0063] The dynamic grouping module solves the complex management problem of multi-source heterogeneous data in terms of time and source by introducing dynamic time window division based on information entropy and context-enhanced feature representation. This process not only improves the accuracy of data segmentation and sorting, but also ensures the traceability and manageability of data blocks through unique dynamic time window identification. Ultimately, the dynamic grouping module provides high-quality initial data blocks for subsequent feature evaluation and screening, laying the foundation for preventing feature leakage and self-circulation effects, and ensuring the overall accuracy and reliability of the chronic disease trend prediction system.

[0064] In the dynamic grouping module, the system has divided the real-time data from different sources into dynamic time windows based on indicators such as timestamps and information entropy, forming multiple initial data blocks and assigning them unique dynamic time window identifiers. The data within each data block has been arranged in chronological order to ensure temporal consistency and source independence. Based on this, the timing analysis module aims to perform deeper timing decomposition and context difference measurement on these data blocks, laying a solid foundation for the calculation of the timing consistency index and cross-layer dependency strength index of the subsequent feature evaluation module.

[0065] The timing analysis module includes the following:

[0066] S2.1, extract long-term trend components and periodic fluctuation components:

[0067] For each data block B i (t), first extract the long-term trend component and periodic fluctuation component within the data block. Use the second-order polynomial approximation method combined with wavelet transform:

[0068] Polynomial approximation: Let B i (t)={x1,x2,…,x N} indicates that in the time window W i (t) N observations collected in (each observation may contain multi-dimensional information, which is simplified to a single symbol to represent the record set). The long-term trend component T of the data block is obtained by second-order polynomial approximation i (t): Among them, l is the sequential index of the data in the block (corresponding to the time order), a, b, c are the coefficients of the second-order polynomial, and the long-term trend component is saved in the form of a discrete sequence, which describes the overall upward or downward trend changes in the data block.

[0069] Wavelet transform: To capture the periodic or short-term fluctuation characteristics, the data block is subtracted from the extracted long-term trend component to obtain the original residual sequence: R i (t) = B i (t)-T i (t); then perform continuous wavelet transform on the original residual sequence, select an appropriate wavelet basis function (such as Morlet function) and select a set of discrete scale sets {σ1,σ2,…,σ M},calculate: in, Indicated on the scale σ m The wavelet basis function under τ represents the translation amount and ν is the integral variable. The main band with concentrated energy in the optimal subinterval is selected as the main component of periodic fluctuation, and the corresponding component is recorded as P i (t). Finally, each data block can be split into long-term trends T i(t) and periodic fluctuation P i (t)Two parts.

[0070] Decomposing data into trends and fluctuations can improve the granularity of feature representation, allowing the model to independently analyze long-term changes and short-term fluctuations. This decomposition method is more flexible than using raw data directly, and can capture key potential change patterns in chronic disease trend prediction, such as long-term changes (trends) and short-term fluctuations (such as seasonal activity patterns) in patient behavior patterns.

[0071] S2.2, calculate the context difference matrix:

[0072] After completing the temporal decomposition of each group of data blocks, the differences between adjacent data blocks in the same time window are measured to characterize the dynamic association between local contexts. To this end, the context difference matrix M is defined i (α, t1, t2), where t1, t2 represent the time location of adjacent (or similar in the same window) data blocks, and α is a comprehensive operator for the difference between trend and periodic components. The calculation method is as follows: M i (α,t1,t2)=∫(Λ(T i (t1),P i (t1))-Λ(T i (t2),P i (t2))) 2 dα; where Λ(T i (t1),P i (t1)) is used to combine the long-term trend and periodic fluctuation of the data block into a function expression that can be integrated in the α dimension (for example, the trend sequence and the periodic sequence are spliced ​​or interpolated into a reference curve of the same scale). α represents the sequential traversal of a series of discrete scales or time domain segments, and finally obtains an integral value similar to the cumulative variance, which is measured by the difference of higher-order functions.

[0073] In this matrix, the larger the value of the context difference matrix, the more significant the difference between adjacent data blocks at the trend and cycle levels, and the smaller the value, the higher the degree of context proximity.

[0074] The introduction of the contextual difference matrix can accurately describe the local correlation within the time window. By clarifying the change pattern between adjacent data blocks, the model's ability to understand the temporal characteristics and semantic associations between data can be enhanced, which helps to identify potential feature mismatches or temporal inconsistencies in the subsequent process.

[0075] S2.3, generate context-enhanced feature representation:

[0076] After calculating the context difference matrix, the context difference matrix is ​​integrated with the decomposed long-term trend and periodic fluctuation to construct the context enhanced feature representation CEF. i (t). The final context-enhanced feature representation of each data block can be recorded as: CEF i (t)= <T i (t),P i (t),{M i (α,t,t′)} t′∈Ω(t) >; where Ω(t) represents the reference data block index set related to the data block in the current time window (or adjacent time window), and locates the comparable adjacent data blocks according to the previous dynamic time window identifier. This representation not only retains the structure of the data block in its own time series dimension, but also incorporates the difference measurement results between the adjacent data blocks, realizing the context association across time slices.

[0077] By combining its own characteristics with contextual differences, a more comprehensive feature representation can be generated, improving the model's ability to understand associations across time points and data sources.

[0078] Through the implementation of the above-mentioned time series analysis module, the system uses the time series decomposition method combining second-order polynomial approximation and wavelet transform to split each initial data block into two major components: long-term trend and periodic fluctuation, and introduces the context difference matrix in the time window to quantitatively measure the differences between adjacent data blocks, and finally generates context-enhanced feature representation. This representation will be used to calculate the time series consistency index in the subsequent feature evaluation module, and combine the feature information from the cold layer with the observation data of the hot layer to complete the evaluation of the cross-layer dependency strength.

[0079] The time series analysis module is closely connected with the initial data blocks and dynamic time window identifiers generated by the dynamic grouping module. The trend and periodic components are obtained through the joint decomposition of polynomials and wavelets, and the representation of each data block at the time and proximity level is enriched with the context difference matrix. This context-enhanced feature representation helps to subsequently identify possible feature mismatches and potential leakage points between different levels (hot layers and cold layers), and provides a more complete and detailed time series foundation for preventing the self-circulation effect in the chronic disease trend prediction model. The processing logic of this module provides high-quality input for the calculation of the time series consistency index and cross-layer dependency strength index in the next stage (feature evaluation module), ensuring that the entire system can be further improved in terms of prediction accuracy and data security.

[0080] In the feature evaluation module, each group of initial data blocks has been decomposed into long-term trend components and periodic fluctuation components, and context-enhanced feature representations have been generated through the context difference matrix. These context-enhanced feature representations not only retain the temporal characteristics of the data block itself, but also contain the difference measurement results between adjacent data blocks, providing rich basic information for subsequent temporal consistency evaluation and cross-layer dependency strength analysis. The core task of the feature evaluation module is to further calculate the temporal consistency index and cross-layer dependency strength index based on the context-enhanced feature representation to ensure that the temporal dependency relationship between the cold layer and hot layer features is reasonably matched, and to quantify their covariation strength, thereby providing a key basis for feature safety scoring.

[0081] The temporal consistency index measures the degree of matching between the cold layer features and the hot layer features in terms of temporal dependency. Its core is to align the cold layer and hot layer feature sequences through a dynamic time warping method, and quantify the synchronization between the two in the temporal dimension. The higher the value, the more consistent the temporal dependency of the two layers of features. The cross-layer dependency strength index constructs a feature interaction network and analyzes the covariation strength between the cold layer and hot layer features to characterize the interdependence and importance of features at different levels, taking into account the interaction weights between features and the network topology. The higher the value, the closer and more stable the association between the cold layer and hot layer features. In the present invention, the purpose of calculating these two indices is to comprehensively evaluate the potential impact of the historical features of the cold layer on the immediate features of the hot layer by quantifying temporal consistency and cross-layer dependency, ensure that the time constraints and interaction logic are met when referencing features in a multi-layer storage architecture, and provide a scientific and accurate basis for subsequent feature screening and model training, thereby effectively avoiding the risks of feature leakage and self-circulation effects, and ensuring the accuracy of chronic disease trend prediction and data security.

[0082] The feature assessment module includes the following:

[0083] S3.1, first use the context-enhanced feature representation to calculate the temporal consistency index TCI i (t), which measures the cold layer characteristics C c (t) and thermospheric characteristics C h (t) The specific calculation method is as follows:

[0084] In order to match the time dependence of the cold layer and hot layer characteristics, the dynamic time warping algorithm is used to transform the cold layer feature sequence C c (t) = {c c1 ,c c2 ,…,c cN} and thermosphere characteristic sequence C h (t) = {c h1 ,c h2 ,…,c hM} for alignment. The distance calculation formula of the dynamic time warping algorithm is: Where Π represents all possible alignment paths, δ(c ci ,c hj ) is the cold layer characteristic c ci With the thermosphere characteristics c hj The distance measurement between them is usually Euclidean distance: By normalizing the dynamic time warping distance D DTW To obtain the temporal consistency index TCI i (t): Among them, ∈ is a very small positive number to prevent zero division errors. The value range of the temporal consistency index is [0,1]. The closer the value is to 1, the higher the consistency of the cold layer and hot layer features in the temporal dependency relationship.

[0085] Cold layer features and hot layer features are based on the characteristic classification of data in a multi-layer storage architecture, corresponding to the data types of long-term storage and instant updates respectively. In a multi-layer storage architecture, cold layer features mainly refer to data stored in the cold layer. These data are usually long-term historical data, such as long-term trend variables, multi-year health status indicators of patients (such as historical blood sugar levels, chronic disease attack frequency, environmental pollution index, etc.). They are characterized by a long update cycle and are mainly used for trend analysis or historical pattern mining; while hot layer features refer to data stored in the hot layer. These data are usually instant data with high frequency updates, such as real-time exercise steps collected by wearable devices, current weather conditions, real-time interaction data on social platforms, etc. They are characterized by strong timeliness and dynamic changes, and mainly serve real-time prediction and rapid response needs.

[0086] The concepts of cold layer features and hot layer features are introduced based on the functional division of multi-layer storage architecture to solve the problem of interactive processing of historical data and real-time data in chronic disease trend prediction. The historical features stored in the cold layer provide support for long-term patterns, while the real-time features stored in the hot layer reflect the current state. The interaction of these two types of features is the key to trend prediction by deep learning models. However, the different distribution of cold layer features and hot layer features in the time dimension may lead to inconsistent time dependencies or deviations in interaction logic. Therefore, it is necessary to conduct quantitative evaluation through the temporal consistency index and the cross-layer dependency strength index to ensure that the reference of cold layer historical features meets the temporal constraints, and quantify the degree of mutual dependence between cold layer and hot layer features, so as to provide reliable data support for subsequent feature screening and model training. By accurately processing the two types of features, feature leakage or data misuse can be effectively avoided, and the accuracy of chronic disease trend prediction and the robustness of the model can be improved.

[0087] S3.2, after obtaining the temporal consistency index, further analyze the covariation relationship between the cold layer and the hot layer features, construct a feature interaction network and calculate the cross-layer dependency strength index CLDSI i (t).

[0088] The cold layer feature C c (t) and thermospheric characteristics C h (t) is used as a node of the network, by calculating the covariation relationship ρ between features c,h (t) to determine the edge weights between nodes. The calculation formula for the covariance relationship is: Among them, ρ c,h (t) represents the covariance strength between the cold layer characteristics and the hot layer characteristics at time t, c ck and c hk are the values ​​of the cold layer characteristics and the hot layer characteristics at the kth observation point, and are the means of the cold layer features and the hot layer features, respectively, and N1 is the feature length. If the covariation relationship exceeds the preset threshold, an edge is added to the feature interaction network with a weight of ρ c,h (t).

[0089] Calculation of cross-layer dependency strength index: After constructing the feature interaction network, the cross-layer dependency strength index is calculated using the network topology analysis method. The specific processing logic is as follows:

[0090] Feature importance score: For each node in the feature interaction network, calculate its feature importance score I i , using PageRank algorithm, the formula is as follows: Where d is the damping factor (usually set to 0.85), It represents the set of neighbor nodes of node i, and deg(j) is the out-degree of node j.

[0091] Cross-layer dependency strength index calculation: Normalize the feature importance scores of all nodes and define CLDSI i (t) is the weighted average of all edges in the cold layer and hot layer feature interaction network: Among them, E represents all edges in the feature interaction network, I c and I h The importance scores of cold layer features and hot layer features are given respectively. The value range of the cross-layer dependency strength index is [0,1]. The higher the value, the stronger the dependency between the cold layer and hot layer features.

[0092] The feature evaluation module accurately calculates the temporal consistency index and cross-layer dependency strength index of the cold layer and hot layer features through the dynamic time warping method and feature interaction network construction. In this process, not only the accurate matching of the time dependency relationship of the features of different storage layers is guaranteed, but also the covariation strength between the features is quantified, providing key data support for identifying potential timing violations and cross-layer feature leakage. The results can directly provide high-quality input for the generation of feature security scores, ensuring that the subsequent feature screening links can make decisions based on rigorous and comprehensive basis, significantly improving the system's temporal compliance in feature use and the reliability of cross-layer dependency assessment, thereby optimizing the data processing accuracy and application credibility of the chronic disease trend prediction model.

[0093] In the feature evaluation module, the temporal consistency index and cross-layer dependency strength index have been calculated through context-enhanced feature representation. These indices reflect the matching degree of the temporal dependency relationship between the cold layer and the hot layer features and the covariation strength between them. In order to further evaluate the security and compliance of each feature, the security calculation module aims to combine these two indices to generate a comprehensive feature security score and identify and mark those high-risk features that have the risk of temporal violations.

[0094] The secure computing module includes the following:

[0095] The temporal consistency index and cross-layer dependency strength index calculated in the feature evaluation module are combined to generate a comprehensive feature safety score. The specific method is as follows:

[0096] To ensure that the temporal consistency index and the cross-layer dependency strength index are comparable in the fusion calculation, the system normalizes the two indices using the hyperbolic tangent function and marks the processed temporal consistency index and cross-layer dependency strength index as TCI′, respectively. i (t) and CLDSI i (t).

[0097] By calculating the geometric mean of the two indices, the standardized temporal consistency index and the cross-layer dependency strength index are fused to generate the feature safety score FSS. i (t). For example, the following formula is used: This formula comprehensively reflects the security of the feature in terms of temporal consistency and cross-layer dependency strength by calculating the geometric mean of the two exponents.

[0098] Based on the generated feature security score, each feature is assessed for risk and high-risk features with timing violations are marked: when the feature security score is lower than the security score threshold, the corresponding feature is considered to have a timing violation risk, marked as a high-risk feature, and directly removed; the identified violation features and their violation types are recorded in the metadata management table.

[0099] By comprehensively calculating the timing consistency index and the cross-layer dependency strength index, a feature safety score is generated, and high-risk features with timing violation risks are accurately identified and marked. This process not only ensures the timing compliance and dependency safety of feature references, but also enhances the system's feature management capabilities through detailed violation type classification and recording.

[0100] In the security calculation module, the system generates feature security scores based on the temporal consistency index and the cross-layer dependency strength index, and marks features with scores below the security threshold as high-risk features. The goal of the current screening and optimization module is to use these scoring results to further sort and screen all features, combine contextual relevance and cross-layer covariation risk information, and produce a high-quality feature set for model training.

[0101] The screening optimization module includes the following:

[0102] After initially eliminating high-risk features, the following processing logic is carried out for the remaining feature set;

[0103] S5.1, in the process of temporal decomposition and context difference matrix calculation in the temporal parsing module, a context-enhanced feature representation has been constructed for each feature, which contains the difference measurement results between adjacent data blocks. For the convenience of this module, the context relevance metric CR is defined i (t) represents the average correlation between feature i and its adjacent data blocks in time window t. The inverse index of the context difference matrix can be regarded as part of the correlation measure: Among them, Ω(t) represents the relationship between the data block B i (t) The index set of all adjacent or same-window data blocks, Φ is a type of nonlinear transformation operator (e.g., Φ(x) = exp(-γ·x), γ is a positive real constant), which is used to map the larger the difference value to the lower the correlation, thereby obtaining the average context correlation CR between the feature and the neighborhood. i (t).

[0104] In addition, in the feature evaluation module, the cross-layer dependency strength index obtained through feature interaction network analysis is used as the reverse reference value of cross-layer covariation risk: when the cross-layer dependency strength index is higher, it means that the covariation strength between the cold layer and the hot layer features is greater. Once the feature score is low, it may have a higher risk (that is, it will bring more serious consequences when leaked). On the contrary, if the cross-layer dependency strength index is low, even if the feature security score is not high, its cross-layer risk is relatively small.

[0105] S5.2, based on the existing feature security score, the context fusion index UFI is constructed by combining the context relevance metric and the cross-layer dependency strength index i(t) is used to accurately distinguish high-scoring features from potential risk features. The specific calculation formula is (example): UFI i (t) = 1-exp(-FSS i (t) CR i (t)·Θ(CLDSI i (t))); where Θ(CLDSI i (t)) is a correction function for the cross-layer dependency strength index, for example: Θ(CLDSI i (t)) = 1 + κ·(CLDSI i (t)) η ; Among them, κ and η are positive real parameters, which are used to amplify the penalty effect when the cross-layer dependency intensity is very high, so that the final index of high-risk features is greatly reduced.

[0106] S5.3, all features are globally sorted according to the context fusion index, and sites with larger feature indexes (i.e., when the context fusion index value is low or negative anomalies occur) are considered to have serious security risks. There will be two types of processing actions in this link:

[0107] (1) Eliminate or limit the use of low-scoring features:

[0108] For features whose context fusion index is far below the safety threshold, they are directly removed from the candidate feature set or their usage is restricted (for example, they are only used for review or online reasoning is closed), and corresponding marks are made in the metadata table to avoid accidental loading during subsequent training or reasoning.

[0109] (2) Prioritize loading high-scoring features:

[0110] Features with a context fusion index higher than the security threshold are considered to be safe and compliant and have high context value. They are added to the feature whitelist and registered in the hot layer feature management record to support subsequent fast loading and use.

[0111] S5.4, after the above screening is completed, all high-scoring features that have passed the security and context checks are integrated to form the final model training feature set. Trainable flags are added to these features in the metadata management table, and their corresponding data block records and time window ranges are synchronously updated to the training database to ensure that features with insufficient scores or illegal markings are not called in actual training. Once the output of the training feature set is completed, the model training phase is entered, which fundamentally avoids the risk of possible feature leakage.

[0112] Based on the feature safety score, the screening and optimization module introduces reverse analysis of contextual relevance and cross-layer covariation risk, and completes the final sorting and screening of features through a novel context fusion index, achieving accurate elimination or restricted use of potential high-risk features, and ensuring the rapid call and integration of high-scoring features. This comprehensive judgment strategy can effectively reduce the risk of misjudgment caused by feature leakage and self-circulation in the multi-layer data scenario of chronic disease trend prediction, and significantly improve the overall quality and security level of model training data, providing a technical guarantee of great practical value for building an accurate and reliable chronic disease prediction model.

[0113] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0114] The above description is only by way of illustration of certain exemplary embodiments of the present invention. It is undoubted that those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the above drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

[0115] It should be noted that, in this article, if there are relational terms such as first and second, etc., they are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "including a..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0116] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A chronic disease trend prediction system based on deep learning, characterized by: include: Dynamic grouping module, timing analysis module, feature evaluation module, security calculation module and screening optimization module; Dynamic grouping module: receives real-time data and sorts the data by segments according to timestamps. It also generates initial data blocks according to data source groups, assigns dynamic time window identifiers to each group of data blocks, and outputs structured initial data blocks and dynamic time window identifiers for the time series parsing module to call. Time series parsing module: performs time series decomposition on the data blocks generated by dynamic grouping, extracts long-term trend components and periodic fluctuation components, calculates the context difference matrix based on adjacent data blocks in the time window, generates context enhanced feature representation, and outputs the context enhanced feature representation containing trend components, periodic components and context difference matrix for use by the feature evaluation module; Feature evaluation module: Based on context-enhanced feature representation, calculate the temporal consistency index, match the temporal dependency relationship between the cold layer and the hot layer features through the dynamic time warping method, build a feature interaction network, analyze the covariation strength of the cold layer and the hot layer features, generate a cross-layer dependency strength index, and output the temporal consistency index and the cross-layer dependency strength index for comprehensive evaluation by the security computing module; Security calculation module: It combines the timing consistency index and the cross-layer dependency strength index, generates feature security scores based on multi-dimensional cross-logic, and marks the timing violation types of high-risk features. It also records the risk information in the metadata management table, outputs feature security scores and high-risk feature tags for use by the screening and optimization module; Screening and optimization module: Sort all features according to feature safety scores, combine the contextual relevance of features with cross-layer covariation risks, eliminate or limit the use of low-scoring features, prioritize loading of high-scoring features, generate a feature set for final model training, and output the final high-quality model training feature set for the model training phase.

2. The chronic disease trend prediction system based on deep learning according to claim 1, characterized in that: The dynamic grouping module includes the following: Receive real-time data from multiple sources through the data access interface, and globally sort based on timestamps to ensure that all data is arranged in chronological order; In order to adapt to the update frequency and dynamic changes of characteristics of different data sources, information entropy is used to evaluate each data source S i The data complexity within the time window t dynamically adjusts the length of the time window according to the calculated information entropy value; According to the dynamically adjusted time window W i (t), divide the data from each data source into several initial data blocks B i (t); each data block B i (t) is contained in the window W i (t) All data records D i (t) and ensure that the data are in chronological order; Assign a unique dynamic time window identifier to each generated data block; The generated data blocks and their corresponding dynamic time window identifiers are stored in the hot layer database, and the mapping relationship between each dynamic time window identifier and the data block to which it belongs, the time window information and the data source are recorded in the metadata management table.

3. The chronic disease trend prediction system based on deep learning according to claim 2 is characterized in that: The timing analysis module includes the following: For each data block B i (t), first extract the long-term trend component and periodic fluctuation component within the data block; use the second-order polynomial approximation method combined with wavelet transform: Polynomial approximation: Let B i (t)={x1,x2,…,x N } indicates that in the time window W i (t) N observations collected within a period; the long-term trend component T of the data block is obtained by second-order polynomial approximation i (t): Among them, l is the sequential index of the data in the corresponding data block, a, b, c are the coefficients of the second-order polynomial, and the long-term trend component is saved in the form of a discrete sequence; Wavelet transform: To capture the periodic or short-term fluctuation characteristics, the data block is subtracted from the extracted long-term trend component to obtain the original residual sequence: R i (t) = B i (t)-T i (t); then perform continuous wavelet transform on the original residual sequence, use wavelet basis functions and select a set of discrete scales {σ1,σ2,…,σ M },calculate: in, Indicated on the scale σ m The wavelet basis function under τ represents the translation amount, and v is the integral variable; the main band with concentrated energy in the optimal sub-interval is selected as the main component of the periodic fluctuation, and the corresponding component is recorded as P i (t); Finally, each data block is split into long-term trends T i (t) and periodic fluctuation P i (t)Two parts.

4. The chronic disease trend prediction system based on deep learning according to claim 3 is characterized in that: The timing analysis module also includes the following: After completing the temporal decomposition of each group of data blocks, the differences between adjacent data blocks in the same time window are measured to characterize the dynamic association between local contexts. For this purpose, the context difference matrix M is defined as i (α, t1, t2), where t1, t2 represent the time location of adjacent data blocks, and α is a comprehensive operator for the difference between trend and periodic components; it is calculated as follows: M i (α,t1,t2)=∫(Λ(T i (t1),P i (t1))-Λ(T i (t2),P i (t2))) 2 dα; where Λ(T i (t1),P i (t1)) is used to combine the long-term trend and periodic fluctuation of the data block into a function expression integrated in the α dimension, where α represents the sequential traversal of a series of discrete scales or time domain segments, and finally obtains an integral value similar to the cumulative variance; After calculating the context difference matrix, the context difference matrix is ​​integrated with the decomposed long-term trend and periodic fluctuation to construct the context enhanced feature representation CEF. i (t).

5. The chronic disease trend prediction system based on deep learning according to claim 4 is characterized in that: The feature assessment module includes the following: In order to match the time dependence of the cold layer and hot layer characteristics, the dynamic time warping algorithm is used to transform the cold layer feature sequence C c (t) = {c c1 ,c c2 ,…,c cN } and thermosphere characteristic sequence C h (t) = {c h1 ,c h2 ,…,c hM } for alignment; the distance calculation formula of the dynamic time warping algorithm is: Among them, Π represents all alignment paths, δ(c ci ,c hj ) is the cold layer characteristic c ci With the thermosphere characteristics c hj The distance measurement between them is usually Euclidean distance: By normalizing the dynamic time warping distance D DTW To obtain the temporal consistency index TCI i (t): Here, ∈ is a very small positive number that prevents division by zero errors.

6. The chronic disease trend prediction system based on deep learning according to claim 5 is characterized in that: The feature assessment module also includes the following: The cold layer feature C c (t) and thermospheric characteristics C h (t) is used as a node of the network, by calculating the covariation relationship ρ between features c,h (t) is used to determine the edge weights between nodes; the calculation formula for the covariance relationship is: Among them, ρ c,h (t) represents the covariance strength between the cold layer characteristics and the hot layer characteristics at time t, c ck and c hk are the values ​​of the cold layer characteristics and the hot layer characteristics at the kth observation point, and are the means of the cold layer features and the hot layer features, respectively, and N1 is the feature length; if the covariation relationship exceeds the preset threshold, an edge is added to the feature interaction network with a weight of ρ c,h (t); Calculation of cross-layer dependency strength index: After the feature interaction network is constructed, the cross-layer dependency strength index is calculated using the network topology analysis method; the specific processing logic is as follows: Feature importance score: For each node in the feature interaction network, calculate its feature importance score I i , using PageRank algorithm, the formula is as follows: Where d is the damping factor, represents the set of neighbor nodes of node i, deg(j) is the out-degree of node j; Cross-layer dependency strength index calculation: Normalize the feature importance scores of all nodes and define CLDSI i (t) is the weighted average of all edges in the cold layer and hot layer feature interaction network: Among them, E represents all edges in the feature interaction network, I c and I h Score the importance of cold layer features and hot layer features respectively.

7. The chronic disease trend prediction system based on deep learning according to claim 6, characterized in that: The secure computing module includes the following: The temporal consistency index and cross-layer dependency strength index calculated in the feature evaluation module are fused to generate a comprehensive feature safety score; Based on the generated feature safety score, each feature is assessed for risk and high-risk features with timing violations are marked: When the feature safety score is lower than the safety score threshold, the corresponding feature is considered to have a timing violation risk and is marked as a high-risk feature; The identified violation characteristics and their violation types are recorded in the metadata management table.

8. The chronic disease trend prediction system based on deep learning according to claim 7, characterized in that: The screening optimization module includes the following: After excluding high-risk features, the following processing logic is carried out for the remaining feature set; define the context relevance metric CR i (t) represents the average correlation between feature i and its adjacent data blocks in time window t; Consider the inverse of the contextual discrepancy matrix as part of the relevance measure: Among them, Ω(t) represents the relationship between the data block B i (t) The index set of all adjacent or same-window data blocks. Φ is a nonlinear transformation operator used to map larger difference values ​​to lower correlations, thereby obtaining the average context correlation CR between features and neighborhoods. i (t).

9. The chronic disease trend prediction system based on deep learning according to claim 8, characterized in that: The screening optimization module also includes the following: Based on the existing feature security score, the context fusion index UFI is constructed by combining the context relevance metric and the cross-layer dependency strength index. i (t), used to distinguish high-scoring features from potential risk features; The specific calculation formula is (example): UFI i (t) = 1-exp(-FSS i (t) CR i (t)·Θ(CLDSI i (t))); where Θ(CLDSI i (t)) is a correction function for the cross-layer dependency strength index.

10. The chronic disease trend prediction system based on deep learning according to claim 9, characterized in that: The screening optimization module also includes the following: For features whose context fusion index is far below the safety threshold, directly remove them from the candidate feature set or impose usage restrictions, and mark them accordingly in the metadata table to avoid accidental loading during subsequent training or inference; Features with a context fusion index higher than the security threshold are considered to be safe and compliant and have high context value, added to the feature whitelist, and registered in the hot layer feature management record to support subsequent fast loading and use; After the screening is completed, all high-scoring features that have passed the security and context checks are integrated to form the final model training feature set; Add trainable flags to these features in the metadata management table, and update their corresponding data block records and time window ranges to the training database synchronously to ensure that features with insufficient scores or illegal markings are not called in actual training.

Images