Deep learning based multi-parameter disease test prediction and early warning system

Abstract

The application discloses a multi-parameter disease test prediction and early warning system based on deep learning, and relates to the technical field of data processing and artificial intelligence early warning. First, a parameter construction module acquires various test parameters and establishes state data packets; an environment trigger conversion module generates a trigger request by performing semantic feature extraction on environmental data; after receiving the request, an association and arbitration module acquires multi-dimensional real-time data having a historical cooperative relationship, inputs the data into a pre-trained relationship arbitration model, generates a relationship evolution evaluation value, and matches a target interaction strategy; a strategy generation module generates strategy modulation node data and node credibility data according to strategy driving node updates; a correction module adaptively corrects an initial scale mapping rule, performs feature transformation on data, and generates structure-adaptive relationship data; finally, a prediction and early warning module outputs a prediction result and an early warning instruction. The application effectively solves the problem of early warning deviation caused by test parameter association drift in a complex environment.

Description

Technical Field

[0001] This invention relates to the fields of data processing and artificial intelligence early warning technology, specifically a multi-parameter disease detection, prediction and early warning system based on deep learning. Background Technology

[0002] Multi-parameter disease testing prediction and early warning systems are an important component of clinical decision support. Through the synergistic analysis of blood glucose, blood lipids, electrolytes, and multiple biochemical indicators, they aim to achieve early identification of the risk of critical illness or chronic diseases. In recent years, with the application of deep learning algorithms in the medical field, multi-parameter joint prediction models based on neural networks have gradually replaced traditional single-indicator threshold alarms, thus improving the sensitivity of disease early warning.

[0003] In existing technologies, disease prediction schemes mostly construct static mapping models, inputting real-time collected test parameters into pre-trained regression or classification networks to obtain the patient's current disease risk assessment value. Under controlled laboratory environments or stable inpatient monitoring conditions, these methods can effectively extract the nonlinear characteristics between multiple indicators and provide relatively accurate prediction conclusions based on past physiological patterns, demonstrating a certain level of intelligence.

[0004] However, when faced with complex real-world diagnostic and treatment scenarios or home health monitoring environments, the physiological parameters of subjects are highly susceptible to interference from external environmental factors (such as exercise stress, sampling position, dosing intervals, or differences in temperature and humidity during sampling). Such fluctuations in the external environment often lead to non-disease-related "pseudo-drifts" in the intrinsic coupling relationships between various test parameters. For example, under stress conditions, the covariance characteristics of certain biochemical indicators may deviate from historical stable benchmarks.

[0005] Existing technologies, when dealing with such scenarios, typically treat environmental parameters merely as ordinary input dimensions or employ fixed data normalization methods. This "one-size-fits-all" approach fails to delve into the arbitrating role of environmental factors on the "logical correlations" between parameters, resulting in the system's inability to distinguish between pathological abnormalities and parameter shifts caused by environmental stress. In complex and dynamic environments, this lack of logical identification can cause the initial scale mapping rules of the early warning system to fail, leading to biased predictions generated by the model. This can easily result in false positives or false negatives, increasing the risk of clinical diagnosis and reducing the system's technical reliability in real-world, variable scenarios. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a multi-parameter disease detection prediction and early warning system based on deep learning.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows:

[0008] This invention discloses a multi-parameter disease detection prediction and early warning system based on deep learning, comprising:

[0009] The parameter construction module is used to acquire various test parameters and establish a status data packet for parameter nodes. The status data packet includes at least parameter values, historical stable interval data, and initial scale mapping rules.

[0010] The environment trigger conversion module is used to extract semantic features from non-uniformly distributed environmental data and generate trigger requests that characterize the sensitivity of parameter coupling.

[0011] The association and arbitration module is used to, upon receiving the trigger request, obtain the parameter value as the first real-time data and obtain the parameter values ​​of other parameter nodes with historical collaborative relationships as the second real-time data; using a deep neural network architecture, input the first real-time data and the second real-time data together into a pre-trained relationship arbitration model to generate a relationship evolution evaluation value, and select a target interaction strategy from a variety of preset data interaction strategies based on the relationship evolution evaluation value;

[0012] The strategy generation module is used to drive relevant parameter nodes to perform information propagation and update according to the target interaction strategy, generate strategy modulation node data, and generate node credibility data based on the degree of deviation of the strategy modulation node data from the historical stable interval data.

[0013] The correction module is used to adaptively correct the initial scale mapping rules of each parameter node according to the node credibility data, and use the updated scale mapping rules to perform feature transformation on the policy modulation node data to generate structural adaptive relation data.

[0014] The prediction and early warning module is used to generate prediction results based on the adaptive relationship data of the structure and output early warning instructions.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0016] 1. This invention overcomes the limitations of traditional early warning systems that only monitor isolated test parameter values. The system can identify the evolutionary evaluation values ​​of relationships between parameters under different external backgrounds, realizing dynamic verification of historical synergistic relationships between test indicators. Through trigger-request-driven data interaction strategy selection, the system ensures that it can accurately determine the attributes of parameter fluctuations in complex and ever-changing monitoring environments, effectively eliminating interference from non-pathological disturbances on data correlation.

[0017] 2. This invention drives relevant parameter nodes to propagate and update information based on a target interaction strategy, and generates node credibility data by combining historical stable interval data, providing a real-time quality measurement benchmark for subsequent preprocessing stages. Through the generation process of structurally adaptive relational data, multidimensional test features achieve semantic alignment and dimension elimination in a high-dimensional projection space, thereby improving the consistency between feature representation and actual physiological state.

[0018] 3. The risk prediction conclusions generated by this invention can more objectively reflect the true pathological trends of the subjects. Through a closed-loop processing flow from environmental perception to structural adaptation, the robustness of the system under heterogeneous data interference is improved, and the accuracy of multi-parameter disease testing is enhanced. Attached Figure Description

[0019] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts. Wherein:

[0020] Figure 1 This is a system module connection diagram of the present invention;

[0021] Figure 2 This is a flowchart of the workflow steps of the present invention;

[0022] Figure 3 This is a flowchart illustrating the steps of the adaptive correction of the initial scale mapping rule of the present invention.

[0023] Figure 4 This is a connection diagram of the dynamic noise cleaning module of the present invention. Detailed Implementation

[0024] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0025] In the disease early warning field addressed in this solution, existing early warning models largely rely on static biochemical indicator mapping logic, making it difficult to balance robustness and accuracy under different external environments. Traditional methods, in complex contexts such as patient stress, changes in body position during data collection, or fluctuations in dosing intervals, are susceptible to environmental noise interference with the inherent coupling characteristics of test parameters, leading to non-pathological "spurious drift" in parameter correlations and consequently biased risk prediction results. Existing systems cannot synchronously perceive the impact of external environmental fluctuations on the logical correlation of multiple parameters. Especially when increased stress levels cause distortions in parameter synergy, fixed-scale detection models may experience systematic failure, failing to meet the needs of precise clinical early warning and tiered control.

[0026] To address the aforementioned issues, the study discovered a nonlinear mapping between the deviation of environmental stress levels and parameter coupling relationships. Logical error compensation was achieved by establishing an environmental trigger-relationship arbitration coupling model. Further investigation revealed that specific core test parameters are highly sensitive to environmental stress, while auxiliary coordination parameters, although exhibiting small fluctuations, demonstrate strong correlation stability. Therefore, a strategy of dynamically switching data interaction tactics based on relationship evolution evaluation values ​​was proposed.

[0027] After introducing the basic concept of the present invention, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0028] Example:

[0029] like Figure 1 As shown, this deep learning-based multi-parameter disease testing prediction and early warning system is designed to support integration with internationally accepted medical data exchange standards and existing hospital information systems (such as Hospital Information System (HIS), Laboratory Information System (LIS), and Electronic Health Records (EHR)). This includes:

[0030] The parameter construction module is used to acquire various test parameters and establish a status data packet for parameter nodes. The status data packet contains at least the initial parameter values, historical stable interval data, and preset initial scale mapping rules. The initial scale mapping rules perform initial standardization on the original parameter values ​​to obtain standardized parameter values.

[0031] The environment triggering conversion module acquires environmental data, extracts semantic features, and generates trigger requests. Environmental semantic feature extraction is achieved through a pre-trained environment coding network, which is a multi-branch deep neural network structure. This network includes a multilayer perceptron branch for processing continuous numerical environmental parameters and an embedding coding branch for processing discrete or textual environmental descriptions. The feature vectors output from each branch are concatenated in a fusion layer and input to a fully connected mapping layer to generate the environmental semantic vector Et. The network has 2–5 layers and hidden unit dimensions of 32–256.

[0032] The association and arbitration module is used to obtain the standardized parameter value of the current parameter node as the first real-time data when a trigger request is received, and to obtain the standardized parameter values ​​of other parameter nodes with historical collaborative relationships as the second real-time data; the first real-time data and the second real-time data are input together into the pre-trained relationship arbitration model to generate a relationship evolution evaluation value, and the target interaction strategy is selected from a variety of preset data interaction strategies based on the relationship evolution evaluation value;

[0033] The policy generation module performs graph neural network message passing operations on the first and second real-time data based on the graph structure and propagation weights defined by the target interaction policy, generating policy modulation node data. It also generates node credibility data based on the deviation of the policy modulation node data from historical stable interval data. The graph neural network is a multi-layer graph attention network structure, including 2–4 graph convolutional layers. Each graph convolutional layer consists of a weighted product of the adjacency matrix At and the node feature matrix, combined with attention weights for feature aggregation. Each layer outputs features with a dimension of 32–128. Residual connections and layer normalization operations are set between layers to enhance training stability. The final layer output serves as the policy modulation node data. The graph neural network is trained offline, and only forward propagation operations are performed online.

[0034] The correction module is used to calculate the mapping parameter compensation amount based on the node credibility data, update the parameters in the initial scale mapping rule, and use the updated scale mapping rule to perform feature transformation on the policy modulation node data to generate structural adaptive relation data.

[0035] The prediction and early warning module is used to generate prediction results based on the structural adaptive relationship data and the node credibility data, and output early warning instructions.

[0036] like Figure 2 The diagram shows the workflow steps of this invention. The working principle of this application is as follows: The system first acquires multi-dimensional raw test data through a parameter construction module. This module consists of a data acquisition interface, registers, and a preprocessor at the hardware level. After the system receives various test parameters such as biochemical, immunological, or hematological parameters, it encapsulates a status data packet for each parameter node. This data packet not only stores the current standardized parameter values ​​in real time but also extracts historical stable interval data (including mean, standard deviation, and extreme value range) for a specific patient or population by reading historical records stored in the on-chip cache. Simultaneously, the data packet loads a preset initial scale mapping rule, which consists of a set of initialized linear transformation operators used to map the raw dimensions to a standardized space that can be processed by the neural network.

[0037] The environment-triggered conversion module acquires non-uniformly distributed environmental context data (such as sampling temperature, humidity, sample storage time, and patient stress state) through an environmental sensor interface. This module utilizes dedicated hardware logic to perform semantic feature extraction, converting discrete environmental signals into high-dimensional semantic vectors. By evaluating the potential impact of these environmental vectors on the coupling relationships of current pathological parameters, the module generates trigger requests characterizing the sensitivity of parameter coupling. This process achieves a shift from "timed triggering" to "feature-driven triggering," ensuring that the system activates the deep arbitration mechanism only when environmental fluctuations may cause data drift.

[0038] In one implementation, environmental semantic feature extraction can be achieved using a convolutional neural network (CNN), a multilayer perceptron network (MLP), or a Transformer encoder structure based on a self-attention mechanism, for feature embedding and semantic representation of multidimensional environmental data. In other implementations, any network structure that can achieve nonlinear feature extraction and semantic space mapping of environmental data is an equivalent implementation of the semantic feature extraction described in this application.

[0039] Upon receiving a request, the association and arbitration module retrieves the first real-time data of the current node and the second real-time data with historical collaborative relationships (such as metabolic pathway associations or complementary physiological functions) from the cache. The system runs a pre-trained relationship arbitration model through a neural network processing unit (NPU), outputting a relationship evolution evaluation value that quantifies the shift in the real-time association of the current parameter pair relative to the historical baseline. In other implementations, the deep feedforward neural network structure can be replaced by a convolutional neural network, a graph attention network, a Transformer encoding network, or a combination thereof.

[0040] The policy generation module constructs a dynamic graph structure within the ASIC based on a locked target interaction policy. The NPU executes message passing operations of the graph neural network, amplifying the target node's information by aggregating features of neighboring nodes (secondary real-time data), generating policy-modulated node data. The hardware logic then compares the modulated values ​​with historical stable intervals, calculating Euclidean or Mahalanobis distance based on their statistical deviation, thereby generating node credibility data. This data intuitively reflects the probability that the current test value is affected by environmental interference or pathological mutations.

[0041] The correction module receives the confidence level data and calculates the compensation amounts for the mapping parameters (such as slope correction and bias correction values) using a hardware-defined compensation function. This step directly applies to the initial scale mapping rules, completing the dynamic update of the parameters. The updated rules are then used to perform nonlinear mapping and dimensional transformation on the policy modulation node data, generating structurally adaptive relational data. This process eliminates dimensional differences and completes the adaptive transformation from the original numerical space to a high-dimensional structured feature space, giving the feature vectors robustness against environmental interference.

[0042] The prediction and early warning module uses structurally adaptive relationship data and node credibility data as joint inputs to drive the time series prediction model to generate disease risk prediction results. The system combines risk probability and credibility weights, and outputs early warning instructions at different levels through the early warning controller, completing a closed loop from raw data input to decision support output.

[0043] By deeply coupling medical logic with deep learning operators, this application effectively suppresses the "illusion" phenomenon of neural networks when processing noisy data. Through real-time rewriting of the scaling rules by the correction module, it solves the problem of systematic drift of test parameters under different stress environments. The system's output warning commands have statistical confidence and pathological interpretability, reducing false alarm and false negative rates, and providing highly reliable hardware-level technical support for early screening of critically ill patients.

[0044] Preferably, the association and arbitration module further includes a logic verification submodule, used for:

[0045] Obtain a pre-defined medical causal knowledge graph and extract pathological causal paths corresponding to the first real-time data and the second real-time data from it;

[0046] The consistency between the relationship evolution assessment value and the pathological causal path is compared, and the logical deviation coefficient is calculated.

[0047] Determine whether the logical deviation coefficient exceeds the preset conflict threshold; if it does, generate a conflict suppression factor and use the conflict suppression factor to perform nonlinear mapping constraints on the output vector of the relationship arbitration model in order to correct the relationship evolution evaluation value.

[0048] The medical causal knowledge graph is a pre-constructed structured pathological knowledge network. Its construction process includes: extracting entities such as diseases, symptoms, biochemical indicators, and pathological events from standard medical knowledge databases, clinical guidelines, and historical electronic medical records; establishing causal edges between entities based on co-occurrence relationships in medical literature and statistical results of clinical pathways; representing causal directions using a directed graph structure and assigning weights to each edge using a causal strength scoring function, where the causal strength score is weighted by a statistical significance index and expert-annotated confidence level. The knowledge graph is stored in a graph database, and a graph traversal algorithm is used to extract the pathological causal path corresponding to the current parameter node for subsequent logical consistency verification.

[0049] To prevent spurious correlations in the model, the logic verification submodule synchronously retrieves the on-chip stored medical causal knowledge graph and extracts the corresponding pathological causal paths (such as the necessary causality between "elevated creatinine and decreased glomerular filtration rate"). By calculating the consistency deviation between the model output and the causal path, a conflict suppression factor is generated. This conflict suppression factor acts on the NPU output in a nonlinear mapping constraint manner to correct the relation evolution evaluation value. The system then retrieves and locks the target interaction strategy from the policy library based on the corrected relation evolution evaluation value.

[0050] Specifically, when performing logical verification, the association and arbitration module injects the rigid logic of medical causal knowledge into the elastic output of the neural network through nonlinear mapping constraints. The logical verification constraints only apply to the output of the relational arbitration model and do not participate in the remodulation of the weights propagated in the graph neural network. In one embodiment, this constraint is introduced by introducing a logical deviation coefficient... The penalty weight operator is implemented. To avoid unnecessary suppression of effective information when the model output is consistent with medical causal knowledge, the logic verification submodule adopts a threshold-triggered causal constraint mechanism, specifically:

[0051] Let the relation arbitration model output relation evolution evaluation vector E, the medical causal consistency deviation coefficient δ, and the preset logical conflict threshold be... (Preferably set between 0.4 and 0.6), then the corrected evaluation vector Calculated according to the following rules:

[0052] ;

[0053] in, Let be the conflict suppression coefficient, and satisfy 0 < <1.

[0054] When the reasoning path of the medical knowledge graph is consistent with the model output (δ≤ The system maintains the original judgment of the model; only when the logical conflict exceeds the threshold is the linear truncation constraint applied to the relation evolution evaluation value, thereby achieving conflict triggering suppression rather than consistency weakening.

[0055] In another embodiment, for severe scenarios involving highly lethal diseases, the nonlinear mapping constraint can also employ a piecewise smooth truncation algorithm. The system sets a safety boundary. When the logic deviation coefficient Upon entering the warning zone, the output vector is flexibly compressed using the hyperbolic tangent function (tanh). This algorithm avoids the gradient vanishing problem caused by direct truncation, while ensuring that the output remains within the boundaries of the medical causal path.

[0056] Through the aforementioned specific nonlinear mapping algorithm, the system achieves the obedience of the "empirical model" to "theoretical knowledge" at the hardware level. Even if the relationship arbitration model deviates due to data noise (such as incorrectly determining that two irrelevant indicators are highly correlated), the logic verification submodule can use this nonlinear constraint to force the relationship evolution evaluation value back to the baseline that conforms to clinical logic.

[0057] This application further proposes that the relationship arbitration model is trained under supervision using the cross-entropy loss function, and L2 regularization is applied during the training process, with the regularization coefficient preferably being... This is to prevent high-dimensional features from overfitting to specific environmental labels.

[0058] The relationship arbitration model is constructed through the following steps. In the initial stage of model construction, the system acquires a historical dataset containing multiple historical test parameters and corresponding environmental labels. This dataset covers samples under different physiological states and environmental stress conditions. Extracting parameter pairs with time-varying characteristics from the historical dataset is crucial for establishing a benchmark. Preferably, the system uses a preset sliding window to capture the dynamic correlation between parameters. The size of the sliding window is preset to 24 to 72 hours based on the physiological metabolic cycle of the test parameters. In one embodiment, the correlation is quantified by calculating the mutual information value of each parameter pair within the window. The calculation formula is as follows:

[0059] ;

[0060] Where X and Y represent two parameter nodes involved in the calculation;

[0061] It is a joint probability distribution;

[0062] and This represents a marginal probability distribution.

[0063] When mutual information value Exceeding the preset relevant threshold When the parameter is an empirical value (e.g., within the range of 0.6 to 0.8), the system determines that the parameter has a historical collaborative relationship.

[0064] During the model training phase, the system uses the logarithmic values ​​of parameters possessing the aforementioned historical collaborative relationships as input feature vectors and performs supervised learning using the corresponding parameter relationship anomaly identifiers as output labels. The training process is completed in a Neural Processing Unit (NPU) or a general-purpose high-performance computing cluster, comprising a 2-layer GCN with 128 units per layer. The training dataset contains ≥10,000 historical test records, with a disease to healthy sample ratio of 1:1. Training conditions include: using the Adam optimizer, with the learning rate preferably set at [value missing]. to The batch size is recommended to be 64 or 128 bytes based on the hardware memory, and the number of training epochs is set to 50 to 200 epochs depending on the convergence of the loss function on the validation set. To prevent overfitting, a packet loss rate mechanism is introduced during training, with the ratio set to 0.2 to 0.5. The final relationship arbitration model can output a relationship evolution evaluation value, which is used to quantify the degree of drift of the current parameter relationship relative to the historical cooperative relationship.

[0065] In a specific implementation scenario, such as considering the synergistic relationship between creatinine and urea nitrogen, if the environmental triggering module extracts the environmental semantics of "high-temperature dehydration," the relationship arbitration model will combine historical data to determine whether the current covariance drift is within the range of physiological compensation. Through this construction method, the system achieves a transition from purely data-driven to pathological logic constraints.

[0066] By introducing collaborative relationship extraction based on mutual information and deep arbitration model training, this application achieves accurate modeling of the anomaly characteristics of complex medical test parameters. The model can keenly capture subtle parameter relationship imbalances and completes deviation quantification based on historical benchmarks before the data enters the prediction module. This construction method not only improves the system's robustness to cross-environment parameter fluctuations but also provides a precise mathematical basis for subsequent scaling corrections by quantifying the degree of drift. This fundamentally reduces prediction bias caused by sensor errors or physiological stress, ensuring the objectivity and logical rigor of the early warning command generation process.

[0067] This application further proposes several pre-defined data interaction strategies, including a strategy based on dynamic weighted graph propagation. Driven by the target interaction strategy, relevant parameter nodes undergo information propagation and updates. This application, by introducing a graph neural network architecture, achieves dynamic modeling and information fusion of complex coupling relationships between multi-parameter nodes. Specifically, this process aims to leverage the topological correlations between parameters to enhance the system's fault tolerance to local data anomalies and the depth of feature extraction.

[0068] In the specific implementation process, the system first receives the relation evolution evaluation value output by the relation arbitration model. This relation evolution evaluation value is mapped to an edge weight adjustment factor, which is used to dynamically construct the parameter relation adjacency matrix at the current time. Adjacency matrix elements Instead of static binary connections, the system nonlinearly adjusts the connection based on the strength of the real-time collaborative relationship between parameter node i and node j. If the relationship evolution evaluation value indicates a drift in a certain parameter relationship, the adjustment factor will automatically reduce the weight of the corresponding edge in the adjacency matrix to suppress the propagation of erroneous information across nodes. Preferably, the adjacency matrix is ​​initialized based on the historical collaborative relationships determined by mutual information, ensuring the sparsity of the graph structure and the consistency with the pathological logic.

[0069] Subsequently, based on the selected target interaction strategy, the system matches a target attention function from a variety of preset graph attention mechanisms. The attention function calculates the current propagation weight between a parameter node and its neighboring nodes by performing an inner product operation between the query vector and the key vector. In one embodiment, the calculation of this weight incorporates a Softmax operator for normalization, and its specific calculation formula is as follows:

[0070] ;

[0071] in, This represents the weight of the importance of node j to node i;

[0072] For node feature vectors, , , These are the original feature vectors of nodes i, j, and k;

[0073] For shared weight transformation matrix;

[0074] Let i be the set of neighbors of node i;

[0075] The eigenvector of node i after linear transformation.

[0076] The system captures potential correlations between parameters from different subspaces through parallel computation of multiple independent attention heads. Specifically, the number of attention heads is typically preset to 4 to 8 to ensure comprehensive feature extraction.

[0077] After obtaining the current propagation weight and adjacency matrix Subsequently, the system drives the graph convolution operator to perform graph convolution feature aggregation operations on the first and second real-time data. The system uses the Laplacian operator to standardize the adjacency matrix and combines the convolution kernel to perform weighted summation and nonlinear activation on the features of each node and its neighborhood.

[0078] By dynamically constructing an adjacency matrix and introducing a graph attention mechanism, this application achieves cross-dimensional deep fusion of parameter node features. The system can dynamically adjust the contribution ratio of each parameter in the diagnostic features based on the real-time evolution of parameter relationships. When a parameter experiences abnormal fluctuations due to environmental stress, graph convolution operations can utilize neighboring node data with highly reliable collaborative relationships to constrain and compensate for it, effectively correcting the performance error of single-point test values. This processing logic based on dynamic graph propagation enhances the global consistency of feature representation, laying a solid foundation for the accurate assessment of subsequent node credibility and improving the stability of risk prediction results in complex clinical contexts.

[0079] This application further proposes specific steps for generating node credibility data based on the degree of deviation of the policy-modulated node data from historical stable interval data.

[0080] In the specific implementation process, the system first obtains the target interaction strategy determined by the association and arbitration module, and extracts the response sensitivity coefficient under the current environment from the strategy. Preferably, the dynamic adjustment logic of this coefficient depends on the stress level output by the environment-triggered conversion module. Specifically, the stress level of the current environment (e.g., set to level 1 to 5) is determined by semantic feature extraction from non-uniformly distributed environmental data. In one embodiment, the response sensitivity coefficient... The stress level, as represented by environmental data, decreases nonlinearly as the stress level increases. The value ranges from 0.3 to 2.0. This means that under highly unstable stress environments (such as high stress levels), the system will reduce... This proactively reduces sensitivity to data deviations, thereby logically tolerating a wider range of fluctuations and avoiding a blind drop in credibility due to environmental factors.

[0081] The system modulates node data according to the strategy. The degree of deviation from historical stable interval data is used to calculate the node credibility data. The core calculation formula is as follows:

[0082] ;

[0083] in, For response sensitivity coefficient;

[0084] This represents node credibility data;

[0085] This represents the policy modulation node data corresponding to the parameter node whose credibility is to be calculated;

[0086] and These represent the mean and standard deviation of the corresponding parameters extracted from historical stable interval data, respectively.

[0087] During the calculation process, the system uses the standard deviation to express the absolute deviation between the current standardized parameter value and its historical mean. The data is normalized to convert it into a standardized deviation. Through mapping using a negative exponential function, the node credibility data... The range of values ​​is limited to the interval (0,1).

[0088] In a specific implementation scenario, such as when environmental noise increases dramatically, the system recognizes an increase in stress level and automatically adjusts... The value decreases non-linearly from a baseline (e.g., 1.5) to a lower level (e.g., 0.6). At this point, even if the policy modulates the node data... The standardized parameter values ​​showed a significant jump, due to The suppression effect ultimately leads to the calculated node credibility data. It can still be maintained within a reasonable range, thus ensuring the stability of subsequent early warning logic.

[0089] By introducing an environment-adaptive response sensitivity coefficient and an exponential decay model, this application achieves refined calibration of the reliability of each parameter node. The system can dynamically adjust the sensitivity threshold to data fluctuations based on the complexity of the current environment. In a stable environment, even minor parameter anomalies can be captured with high sensitivity; however, in complex stress environments, the system uses the response sensitivity coefficient... The adaptive adjustment implements "logical tolerance," effectively filtering out non-pathological spurious drift. This credibility assessment, based on the dual constraints of statistical distribution and environmental perception, provides a scientific quantitative basis for the accurate correction of subsequent scale mapping rules.

[0090] like Figure 3 The diagram shows the steps of adaptively correcting the initial scale mapping rule of the present invention. This application further proposes specific steps for adaptively correcting the initial scale mapping rule of each parameter node based on node credibility data. In a specific implementation, the correction module first receives the real-time generated node credibility data C and compares it with a preset credibility threshold. Perform logical judgment. Preferably, the preset confidence threshold... The range is typically set between 0.2 and 0.6, and the specific value can be adjusted according to the clinical fluctuation characteristics of a particular test parameter. When the system determines... If the current parameter node is found to have insufficient representation credibility, an adaptive correction process is triggered.

[0091] Specifically, when At that time, the node data is modulated based on the strategy of the current parameter node. Compared with the historical stable interval mean The deviation between them is used to calculate the compensation amount for the scale mapping parameters. The compensation amount includes the mapping slope correction value. Deviation from intercept value Common characteristics.

[0092] The compensation amount satisfies the following:

[0093] ;

[0094] in, The preset maximum compensation magnitude is used to limit the maximum change in the scale mapping parameters. The formula for the corrected compensation amount is:

[0095] ;

[0096] ;

[0097] When satisfied At this point, the current data is marked as potential pathological abnormality data. This data does not participate in the scale mapping compensation calculation, but only enters the subsequent prediction model processing flow. The pathological abnormality judgment coefficient is the preferred one. When the pathological abnormality criteria are met for T consecutive time steps, the scale mapping is frozen and updated, and only prediction is performed without compensation.

[0098] The system utilizes the mapping slope correction value Deviation from intercept value The linear transformation parameters in the initial scale mapping rule are updated to generate the updated scale mapping rule. In one embodiment, the initial scale mapping rule is represented as a linear model. Wherein, the slope parameter k is responsible for amplitude scaling, the bias parameter b is responsible for baseline translation, x represents the original parameter value, i.e., the unprocessed physiological or biochemical index value directly collected from the testing equipment; y represents the standardized parameter value, i.e., the uniform dimensional feature value input to the neural network module after transformation by the initial scale mapping rule. The update formula is as follows:

[0099] ;

[0100] ;

[0101] The initial scale mapping rule is a linear scale mapping model, which includes a slope parameter k for scaling the original test parameters and a bias parameter b for baseline translation.

[0102] By generating updated scale mapping rules, the system constructs an adaptive feature reprojection space. For example, when a patient experiences baseline drift in a certain biochemical indicator due to high heat stress, the correction module can compensate positively or negatively. The interference The water level was brought back to a level consistent with physiological logic to avoid false alarms from the prediction model.

[0103] By introducing a reliability-based parameter compensation algorithm, this application achieves online closed-loop adjustment of the original data mapping logic. The system can dynamically optimize the focal length (slope) and bias (intercept) of its perceived "field of view" based on real-time feedback on data quality. This adaptive correction mechanism weakens the interference of non-pathological fluctuations on model feature extraction, ensuring that the structurally adaptive relational data has undergone deep semantic alignment before being input into the prediction module, thereby enhancing the clinical reference value of early warning instructions.

[0104] Preferably, this application also proposes a specific implementation method for feature transformation of policy modulation node data using updated scale mapping rules. This application achieves a structured transformation from physical observations to deep semantic features by constructing a nonlinear projection space. Specifically, this step aims to eliminate dimensional differences among multi-source heterogeneous parameters and extract core pathological features with environmental robustness.

[0105] In the specific implementation process, the system will modulate node data according to the strategy. As input, through the updated slope parameter and bias parameters The system processes dynamically reconstructed scaling rules. In one embodiment, the system constructs a nonlinear projection space that not only includes linear scaling logic but also performs nonlinear mapping on the data by introducing nonlinear activation operators (such as variants of Sigmoid or ReLU). Through this mapping, the system performs dimensional transformation and standardization, projecting the originally low-dimensional discrete parameter points into a high-dimensional space with higher representational power, generating a high-dimensional feature tensor after eliminating dimensional differences, which serves as structurally adaptive relational data.

[0106] The nonlinear projection space is implemented by a feature reconstruction network, which is a feedforward neural network structure containing an input layer, at least two nonlinear transformation layers, and an output embedding layer. Each transformation layer consists of a combination of a linear transformation with batch normalization and an activation function, either ReLU or GELU. The network output dimension is 64–512, used to form a high-dimensional embedding representation with uniform dimensions. The network parameters are modulated by the mapping coefficients in the updated scaling rule, causing the distribution of features of different parameter nodes in the projection space to adaptively adjust with changes in confidence.

[0107] Specifically, the standardization process uses updated rules to normalize the amplitude of each parameter dimension, enabling indicators with different physical meanings (such as blood glucose in mmol / L and hemoglobin in g / L) to be converted into dimensionless feature components with consistent value ranges. During this process, the nonlinear projection space, based on the nonlinear physiological response characteristics of the parameters, assigns higher weightings to values ​​in critical or abnormal regions, while compressing small fluctuations in the stable range.

[0108] Preferably, the generated structural adaptive relation data after eliminating dimensional differences is defined as structural adaptive relation data. This data form is no longer merely a variation of the original numerical values, but includes node state information, neighboring node collaborative constraint information, and confidence information corrected by environmental confidence. In a specific embodiment, for a cardiovascular parameter set, this tensor will integrate blood pressure fluctuation characteristics, heart rate frequency domain characteristics, and blood oxygen saturation trends to form a multi-dimensional feature vector.

[0109] By constructing a nonlinear projection space based on adaptive scaling, this application achieves precise reshaping of the deep features of test data. The system effectively solves the problem of model weight imbalance caused by inconsistencies in dimensions during multi-parameter joint analysis. By projecting the corrected data into a high-dimensional tensor space, the system enhances its ability to capture implicit cross-parameter correlations, ensuring that subsequent time-series prediction models can be computed at a "pure" and "aligned" feature level. Compared to traditional linear normalization methods, the structurally adaptive relational data generated by this scheme better preserves pathological details under extreme conditions, providing high-quality data input for outputting high-precision disease risk prediction results, thereby reducing the risk of false alarms in complex disease prediction.

[0110] like Figure 4 The diagram shown illustrates the connection of the dynamic noise removal module of this invention. This application further proposes that the system also includes a dynamic noise removal module connected between the parameter construction module and the association and arbitration module. In specific implementation, the dynamic noise removal module first constructs a multi-parameter joint distribution model for a set of parameter pairs with historical cooperative relationships. Preferably, this model employs a Gaussian mixture model (GMM) or a multivariate normal distribution model. Its core logic lies in learning the joint probability distribution law of each test indicator under physiological homeostasis through historical big data. In one embodiment, the joint distribution model determines the parameters using the maximum likelihood estimation (MLE) method, capturing the specific probability topology formed between parameters due to physiological compensation or metabolic coupling.

[0111] For the current input set of parameter nodes with historical cooperative relationships, the system calculates the likelihood probability of this set under the joint distribution model. The likelihood probability reflects the probability of the current combination of node values ​​occurring under normal physiological conditions. The system sets a preset probability threshold. The preset probability threshold The value is typically set between 0.01 and 0.05. When the calculated likelihood probability is lower than this preset probability threshold... When this happens, the system determines that there are suspected abnormal parameter values ​​in the parameter node set that are seriously inconsistent with the overall physiological trend. This determination method is superior to traditional single-point amplitude limiting filtering because it can identify hidden noise that is within a reasonable range but has "unreasonable combinational logic".

[0112] Specifically, once a suspected anomaly is identified, the module does not directly discard the data. Instead, it employs a reconstruction and completion strategy based on a joint distribution model. The system uses the conditional probability distribution formula and other relatively normal parameters in the set as observation benchmarks to derive the expected value of the anomalous node. This value is then used to reconstruct and complete the data, obtaining the cleaned and standardized parameter values. Its calculation logic follows:

[0113] ;

[0114] in, Let E be the reconstructed numerical value, and E be the expected value based on the joint distribution model.

[0115] It is a random vector;

[0116] This corresponds to the observation vector. This completion method ensures the continuity of the data stream while approximating the true physiological water level as closely as possible.

[0117] By introducing a dynamic noise cleaning module, this application achieves deep denoising of test data from the perspective of statistical consistency. The system can effectively identify and correct non-pathological mutations caused by transient sensor failures, operational interference, or sample contamination. Through reconstruction and completion based on joint distribution, the system restores the basic cooperative relationship between parameters before the data enters the association and arbitration module, effectively preventing noise interference from being amplified in the subsequent graph neural network.

[0118] This application further proposes that the system also includes a strategy pool self-evolution module. By introducing a closed-loop feedback learning mechanism, this application achieves continuous iteration and performance optimization of the system's decision-making logic. Specifically, this module aims to dynamically adjust the strategy selection preference through feedback from clinical results, ensuring that the system can adapt to environmental changes or the evolution of pathological characteristics during long-term operation.

[0119] In the specific implementation process, the strategy pool self-evolution module continuously records the feedback verification data after the output of the warning command. Preferably, the feedback data comes from the clinical physician's diagnosis, the gold standard results of subsequent laboratory tests, or the patient's post-treatment follow-up data. The system compares the risk probability of the output warning command with the actual clinical results and calculates the prediction accuracy index. This index covers comprehensive evaluation parameters such as precision, recall, and F1-score (comprehensive index). In one embodiment, the system uses a fixed period (such as weekly or every 30 days) as the statistical unit to continuously evaluate the system's performance.

[0120] When the prediction accuracy metric is detected to be below a preset accuracy threshold (e.g., an empirical value in the range of 0.80 to 0.90) for N consecutive periods (preferably, N is set to 3 to 5 periods), the system automatically triggers self-evolution logic. Specifically, the system resamples and adjusts the policy weight allocation ratios in various preset data interaction strategies. The resampling process utilizes the distribution differences between successful and failed cases, employing policy gradient algorithms or Monte Carlo sampling methods from reinforcement learning to assign higher selection probability weights to policies with superior performance, while suppressing the weights of policies that lead to false positives under specific stress scenarios.

[0121] The adjusted weight allocation ratio is then updated in real time to the association and arbitration modules. This is achieved by rewriting the strategy mapping table in the arbitration logic or adjusting the top-level bias of the classifier, thereby optimizing the selection logic of the target interaction strategy. For example, if the system finds that the "dynamic weighted graph propagation strategy" is more robust than other strategies in a "high humidity environment," the self-evolution module will automatically increase the activation weight of this strategy in that scenario, enabling the system to more accurately lock the optimal strategy when encountering similar environmental requests in the future.

[0122] By introducing a self-evolving strategy pool module, this application achieves a leap from static rules to dynamic evolution capabilities in the system. The system possesses an "experience accumulation" effect similar to that of clinical experts, enabling it to continuously correct its internal strategy allocation logic through autonomous learning of past warning errors. This adaptive evolution mechanism effectively solves the problem of performance degradation of deep learning models after actual deployment, ensuring that the warning logic always keeps pace with the latest pathological data distribution and extending the system's technical lifecycle.

[0123] This application further proposes that the step of generating prediction results based on structural adaptive relational data and combined with node credibility data aims to finely adjust the "memory" and "forgetting" mechanisms of neural networks using credibility data, so as to ensure that the model can still focus on high-value feature information in complex interference environments.

[0124] In the specific implementation process, the system first concatenates the structural adaptive relationship data with the corresponding node credibility data using tensors. Preferably, if the structural adaptive relationship data is a d-dimensional vector, it is concatenated with its corresponding node credibility data (usually a weight vector of the same dimension) to generate an enhanced feature vector. This enhanced feature vector It not only carries the corrected pathological features but also simultaneously includes quality labels for each dimension of the features. Subsequently, the system will enhance the feature vector. The input is fed into a pre-trained time series prediction model. The recurrent computation unit of this model can be a Long Short-Term Memory (LSTM) network, a gated recurrent unit (GRU), or a variant thereof, to capture the nonlinear trend of the test parameters evolving over time.

[0125] Within the internal logic processing of the loop computation unit, the system performs crucial confidence modulation processing. Specifically, taking LSTM as an example, let the original forget gate output be... Its computational logic depends on the hidden state of the previous time step. Enhance the feature vector with the current input The reliability modulation process for the forget gate output includes:

[0126] Let's assume the original forgetting gate The output formula is as follows:

[0127] ;

[0128] in, This is the forget gate weight matrix;

[0129] Use the Sigmoid activation function;

[0130] Forget gate bias term;

[0131] Node credibility data corresponding to the feature dimensions By introducing a modulation function, the modulated output of the forget gate is obtained. The formula is as follows:

[0132] ;

[0133] in, The preset modulation coefficient is preferably set between 0.3 and 0.7. The technical logic of this operator lies in: when the confidence level of a certain feature dimension... At lower levels, Enlargement, leading to Reduce. Utilize the modulated forget gate output. Replaces the original forgetting gate output in cell state The updated computation can prompt the model to actively "forget" those severely disturbed feature components, thereby protecting the cell state. Not polluted by noise.

[0134] Output using the modulated forget gate Replaces the original forgetting gate output in cell state Update the calculation, the formula is as follows:

[0135] ;

[0136] in, This is element-wise multiplication (Hadamard product).

[0137] For input gate output;

[0138] Candidate cell state;

[0139] This represents the cell state at the previous moment.

[0140] The system simultaneously outputs a vector for the hidden state. Apply confidence-weighted residual enhancement by introducing a mapping gain coefficient. (The preferred range is 0.1 to 0.5), the high-confidence input features are directly mapped to the output using residual connections, as shown in the following formula:

[0141] ;

[0142] Extract the hidden state vector at the end time step The data is input to a fully connected classification layer, and the disease risk probability is output as the prediction result through the Softmax operator. This operation enhances the weight gain of the high-confidence dimension in the current hidden state, ensuring the sensitivity of the prediction result to high-quality data.

[0143] By introducing a confidence-based modulation operator within the recurrent units of the time-series model, this application achieves dynamic intervention in the neural network computation path. The system no longer indiscriminately processes historical information and current input, but instead possesses a "selective memory" capability based on data quality. When the input data contains outliers induced by environmental stress, the rapid closure of the forget gate and the confidence weighting of the residual end work synergistically to effectively block the propagation of errors to future time steps. Through a deeply coupled prediction mechanism, the model's robustness in extremely low signal-to-noise ratio environments is improved, ensuring the consistency and accuracy of the final risk prediction results in terms of pathological logic.

[0144] This application further proposes that the step of executing the output warning instruction also includes converting the continuous output of the backend deep learning into decision instructions with clinical guidance significance, and simultaneously feeding back the quality status of the original data.

[0145] In the specific implementation process, the system first obtains the risk probability value corresponding to the prediction result, as well as the set of key parameters and the average node credibility on which the prediction result is based. Preferably, the set of key parameters refers to the top three feature nodes (such as creatinine, blood potassium, pH value, etc.) in terms of weight contribution in the graph neural network propagation and attention mechanism.

[0146] Simultaneously, the system calculates the average node credibility of this parameter set at the current time step. The formula is as follows:

[0147] ;

[0148] Where M is the dimension of the key parameter set;

[0149] This provides real-time node reliability data for each parameter.

[0150] The early warning module determines the risk level based on the probability value falling within a preset risk threshold range (e.g., 0.4-0.6 for low risk, 0.6-0.8 for medium risk, and above 0.8 for high risk), and also considers the average node confidence level. The final warning level is determined by comprehensive analysis. In one embodiment, when the average node credibility... Below the preset confidence threshold When the score is 0.70 (e.g., the system determines that although the current prediction model has a high score, the underlying input is severely affected by the stress environment, and there is a potential risk of false alarm or incomplete information.

[0151] Specifically, for this "high-risk, low-reliability" situation, the system implements a defensive early warning strategy. In one embodiment, the system automatically upgrades the warning level for the same risk probability value (e.g., automatically raising it from medium risk to high risk to attract clinical attention) or adds an "uncertainty indicator" to the output command (e.g., displaying "Due to environmental fluctuations, the prediction reliability is low; it is recommended to verify in conjunction with clinical signs" on the warning interface). This design reflects the principle of prudence in medical diagnosis, ensuring that even in the event of data corruption, the system will not draw overly absolute or misleading conclusions.

[0152] By introducing a reliability-weighted early warning grading mechanism, this application achieves a logical leap from "probability calculation" to "decision support." The early warning instructions output by the system not only include quantitative values ​​of disease risk but also objectively reveal the quality background of the data itself. This multi-dimensional grading logic effectively mitigates the one-sidedness of the output results of black-box models under extreme conditions, providing doctors with comprehensive judgment criteria including both "outcome" and "reliability," and reducing the medical risks caused by blindly trusting the model's output.

[0153] The following is a specific implementation of a multi-parameter disease testing prediction and early warning system based on deep learning:

[0154] We selected white blood cell count (WBC), C-reactive protein (CRP), and blood oxygen saturation (SpO2) from five consecutive days of laboratory data of a patient as parameter nodes. We established stable interval statistics from historical healthy samples: WBC: =7, CRP: , =2;SpO2: =97, =1. The initial scaling rule is the standardization function x′=(x−μ) / σ. The current raw data is WBC=12, CRP=15, SpO2=93, then the standardization result is... = (12−7) / 1.5 = 3.33, = (15−3) / 2 = 6, =(93−97) / 1=−4, forming a state data packet. The environment module collects environmental data such as air pressure and oxygen content, and obtains a stress level S=0.8 through semantic encoding, which is mapped to a response sensitivity λ(t)=1 / (1+e^{2S})≈0.31, generating a trigger request. The relationship arbitration model outputs a relationship evolution evaluation value R=0.72 based on the historical collaborative relationship constructed by mutual information. The system selects a graph attention interaction strategy and constructs an adjacency matrix. =[[1,0.6,0.5],[0.6,1,0.7],[0.5,0.7,1]]. This is achieved through graph convolution propagation. = ·x′ ≈[2.8, 5.1, −3.2]. Calculate node reliability data. Below the preset confidence threshold =0.5, triggering the correction module, the compensation function outputs slope correction Δk=0.2 and intercept Δb=−0.5, updates the scale mapping to x″=1.2x′−0.5, and generates a structural adaptive feature vector F. F is concatenated with the confidence tensor and input into the LSTM prediction model. Under the action of the gating correction operator, the influence of low-node confidence data C dimension is suppressed, and the final output disease risk probability P=0.81 is obtained. A preset confidence threshold is used. =0.5, due to average confidence level =0.46< The warning level was raised from Level II to Level III and an uncertainty label was added. At the same time, the prediction result was recorded for the self-evolution update of the strategy pool weights.

[0155] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. A multi-parameter disease detection prediction and early warning system based on deep learning, characterized in that, include: The parameter construction module is used to acquire various test parameters and establish a status data packet for parameter nodes. The status data packet includes at least the original parameter values, historical stable interval data, and a preset initial scale mapping rule. The initial scale mapping rule performs an initial standardization process on the original parameter values ​​to obtain standardized parameter values. The environment trigger conversion module is used to acquire environmental data, extract semantic features, and generate trigger requests; The association and arbitration module is used to obtain the parameter value of the current parameter node as the first real-time data when the trigger request is received, and to obtain the parameter values ​​of other parameter nodes with historical collaborative relationships as the second real-time data. The first real-time data and the second real-time data are input together into a pre-trained relationship arbitration model to generate a relationship evolution evaluation value. Based on the relationship evolution evaluation value, a target interaction strategy is selected from a variety of preset data interaction strategies. The strategy generation module is used to perform message passing operations on the first real-time data and the second real-time data based on the target interaction strategy to generate strategy modulation node data. Based on the degree of deviation of the modulated node data relative to the historical stable interval data according to the strategy, node credibility data is generated; The correction module is used to calculate the mapping parameter compensation amount based on the node credibility data, update the parameters in the initial scale mapping rule, and use the updated scale mapping rule to perform feature transformation on the policy modulation node data to generate structural adaptive relation data. The prediction and early warning module is used to generate prediction results based on the adaptive relationship data of the structure and the node credibility data, and output early warning instructions.

2. The deep learning-based multi-parameter disease detection prediction and early warning system according to claim 1, characterized in that, The relationship arbitration model is constructed through the following steps: Obtain a historical dataset containing multiple historical test parameters and corresponding environmental labels; Extract parameter pairs with time covariance characteristics from the historical dataset, calculate the mutual information value of each parameter pair within a preset sliding window, and establish the historical collaborative relationship based on the mutual information value; Using the logarithmic values ​​of parameters with the historical collaborative relationship as input features and the corresponding parameter relationship change identifiers as output labels, a deep neural network is trained to obtain the relationship arbitration model.

3. The deep learning-based multi-parameter disease detection prediction and early warning system according to claim 1, characterized in that, Based on the target interaction strategy, relevant parameter nodes are driven to propagate and update information, specifically including: Using the relationship evolution evaluation value as the edge weight adjustment factor, the adjacency matrix of the parameter relationship at the current time is dynamically constructed. ; According to the target interaction strategy, a target attention function is selected from a variety of preset graph attention mechanisms, and the target attention function is used to calculate the current propagation weight between the parameter node and its neighboring nodes; Based on the current propagation weight and the adjacency matrix The first real-time data and the second real-time data are subjected to graph convolution operation to generate the policy modulation node data.

4. The deep learning-based multi-parameter disease detection prediction and early warning system according to claim 1, characterized in that, The specific steps for generating node reliability data based on the deviation of the modulated node data from the historical stable interval data include: Obtain the target interaction strategy determined by the association and arbitration module, and extract the response sensitivity coefficient in the current environment from the target interaction strategy. ; The node credibility data is calculated using the following formula: ； in, This represents the node's credibility data; This represents the policy modulation node data corresponding to the parameter node whose credibility is to be calculated; and These represent the mean and standard deviation of the corresponding parameters extracted from the historical stable interval data, respectively. Wherein, the response sensitivity coefficient The stress level decreases nonlinearly as the stress level represented by the environmental data increases; the current stress level of the environment is determined by extracting semantic features from the non-uniformly distributed environmental data.

5. The deep learning-based multi-parameter disease detection prediction and early warning system according to claim 1, characterized in that, The specific steps for adaptively correcting the initial scale mapping rules of each parameter node based on the node credibility data include: Receive the node credibility data and determine whether it is lower than a preset credibility threshold; If the deviation is below the preset confidence threshold, a preset error compensation function is invoked to calculate the mapping slope correction value and intercept deviation value based on the degree of deviation. The linear transformation parameters in the initial scale mapping rule are updated using the mapping slope correction value and intercept deviation value to generate the updated scale mapping rule.

6. The deep learning-based multi-parameter disease detection prediction and early warning system according to claim 1, characterized in that, The updated scale mapping rules are used to perform feature transformation on the policy modulation node data, specifically as follows: The policy modulation node data is input into the nonlinear projection space constructed by the updated scale mapping rule, and dimensional transformation and standardization are performed to generate a high-dimensional feature tensor after eliminating dimensional differences, which serves as the structure adaptive relation data.

7. The deep learning-based multi-parameter disease detection prediction and early warning system according to claim 1, characterized in that, The system further includes a dynamic noise removal module, connected between the parameter construction module and the association and arbitration module, for: For a set of parameter pairs with historical collaborative relationships, a multi-parameter joint distribution model is constructed. For a set of parameter nodes with historical collaborative relationships, calculate the likelihood probability of the set of parameter nodes under the joint distribution model; When the likelihood probability is lower than a preset probability threshold, it is determined that there are suspected abnormal parameter values ​​in the parameter node set, and the suspected abnormal parameter values ​​are reconstructed and completed based on the joint distribution model to obtain cleaned parameter values.

8. The deep learning-based multi-parameter disease detection prediction and early warning system according to claim 1, characterized in that, The system also includes a policy pool self-evolution module, used for: Record the feedback verification data after the warning command is output, and calculate the prediction accuracy index; When the prediction accuracy index is lower than the preset accuracy threshold for N consecutive periods, the strategy weight allocation ratio in the multiple preset data interaction strategies is resampled and adjusted. The adjusted weight allocation ratio is updated to the association and arbitration module to optimize the selection logic of the target interaction strategy.

9. The deep learning-based multi-parameter disease detection prediction and early warning system according to claim 1, characterized in that, The specific steps for generating prediction results based on the structural adaptive relationship data and combined with the node credibility data include: The structural adaptive relationship data and the corresponding node credibility data are concatenated using tensors to generate an enhanced feature vector. The enhanced feature vector is input into the pre-trained time series prediction model; in the loop calculation unit of the time series prediction model, the node confidence data corresponding to the current time step is obtained; Determine whether the component values ​​of each feature dimension in the node credibility data are lower than the preset credibility threshold; if they are lower than the preset credibility threshold, call the preset gating correction operator to reduce the activation intensity of the forget gate in the loop calculation unit for the corresponding feature dimension and increase the mapping gain of the hidden state to the high credibility dimension. The hidden state vector output by the time-series prediction model at the end time step is extracted and input into the fully connected classification layer, and the prediction result representing the probability of disease risk is output.

10. The deep learning-based multi-parameter disease detection prediction and early warning system according to claim 9, characterized in that, The steps for executing the output warning command also include: Obtain the risk probability value corresponding to the prediction result, as well as the set of key parameters and average node confidence on which the prediction result is generated; The final warning level is determined by comprehensively considering the preset risk threshold range into which the risk probability value falls and the average node credibility. When the average node credibility is lower than the preset credibility threshold, the warning level is increased or an uncertainty label is added for the same risk probability value.

Images