Method and device for dynamic early warning of environmental health risk based on K-ANAM

Abstract

The application discloses a kind of based on K-ANAM's environmental health risk dynamic early warning method and device, belong to environmental health risk assessment field.The method includes: collecting environmental factors and health outcome data to construct original characteristic vector;Apply positive part truncation operator to extract excess stress part to form non-stationary stress characteristic pool;Based on correlation coefficient, filter effective interaction feature pair set;Decoupling addition network architecture is constructed and includes main effect subnet and synergistic effect subnet;Decoupling addition network architecture is trained, and K-ANAM network is obtained;Real-time data are input into the network after training to obtain total health risk value and single-factor and multi-factor risk contribution;Synergistic contribution ratio is calculated, and the comparison result of total health risk value and preset safety threshold, synergistic contribution ratio and preset synergistic critical value triggers graded early warning.The application overcomes the problems that traditional model has insufficient extrapolation ability under non-stationary climate and cannot explicitly decouple multi-factor synergistic effect.

Description

Technical Field

[0001] This invention belongs to the field of environmental health risk assessment, and in particular relates to a dynamic early warning method and device for environmental health risks based on K-ANAM. Background Technology

[0002] As global climate change intensifies, cities are facing increasingly complex and unstable extreme weather events. However, existing urban emergency response systems often rely on independent threshold triggering mechanisms, such as issuing heatstroke warnings based on temperature or pollution warnings based on AQI. This fragmented approach to early warning, lacking decoupling mechanisms, ignores the nonlinear synergistic amplification effect on human health caused by the superposition of multiple factors. Consequently, in the event of complex disasters, the allocation of medical emergency resources is severely lagging, leading to systemic risks.

[0003] Currently, assessment and early warning technologies for environmental health risks are mainly divided into two categories: traditional epidemiological statistical models (such as GAM / DLNM) and data-driven deep learning models (such as LSTM / CNN / RF). However, neither can meet the needs for accurate quantification under complex extreme weather conditions.

[0004] Traditional epidemiological statistical models are based on historical observation data and use splines for smooth fitting and linear extrapolation. These models have the following shortcomings: they rely on the assumption of stationarity and additive structural constraints. When faced with extreme events that break historical extremes, the model is constrained by linear boundary constraints and tends to regress the risk to the mean, systematically underestimating catastrophic consequences and causing early warning failures. Furthermore, the traditional additive framework struggles to adaptively capture the multiplicative interactions between multiple factors, leading to underreporting of compound risks.

[0005] Data-driven deep learning models are based on end-to-end large-scale data training to construct black-box mapping relationships. These models have the following drawbacks: strong fitting ability, but opaque internal decision-making processes and a lack of interpretability. Decision-makers cannot determine whether high risk is caused by high temperatures alone or by a combination of pollutants, thus failing to guide precise intervention measures; sparse sample dependence—extreme complex disasters are small-sample sparse events, and conventional deep learning models are prone to overfitting on normal data, while exhibiting poor generalization ability in extreme value regions lacking training samples. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a dynamic early warning method for environmental health risks based on K-ANAM, so as to solve the problems that the existing technology is insufficient in extrapolating the health risks of extreme compound disasters under non-stationary climate conditions and cannot explicitly decouple the synergistic effects of multiple factors.

[0007] The objective of this invention is achieved through the following technical solution:

[0008] The first aspect of this invention discloses a dynamic early warning method for environmental health risks based on K-ANAM, comprising:

[0009] Historical time-series data of meteorological dynamic factors and chemical pollutant factors in the target area were collected to form raw feature vectors, and time-series data of health outcome labels were collected simultaneously.

[0010] For each factor in the original feature vector, a preset quantile threshold is used as the normal baseline. The positive part truncation operator is applied to extract the excess pressure part, and the corresponding historical non-stationary pressure features are obtained. A historical non-stationary environmental pressure feature pool is then constructed.

[0011] Calculate the correlation coefficient between any two different factors in the historical non-stationary environmental pressure feature pool, and determine the factor combination with the correlation coefficient greater than the preset collaborative threshold as the historical effective interaction feature pair, forming a set of historical effective interaction feature pairs.

[0012] A decoupled additive network architecture based on the Kolmogorov-Arnold representation theorem is constructed. This decoupled additive network architecture includes main effect subnetworks that correspond one-to-one with the factors in the historical non-stationary environmental pressure feature pool and synergistic effect subnetworks that correspond one-to-one with the factor combinations in the historical effective interaction feature pair set. Its total health risk output is composed of the linear superposition of the outputs of all main effect subnetworks and all synergistic effect subnetworks.

[0013] Based on historical non-stationary stress characteristics, using health outcome label time series data as supervision signals, a decoupled additive network architecture is trained using a loss function with sparse regularization terms to obtain the trained K-ANAM network.

[0014] Real-time collected data of similar factors are processed by the positive part truncation operator and then input into the trained K-ANAM network to output the total health risk value and decompose it into single-factor additive risk contribution and multi-factor synergistic risk contribution.

[0015] The collaborative contribution ratio is calculated based on the total health risk value and the collaborative risk contribution of multiple factors. Based on the comparison results of the total health risk value with the preset safety threshold and the collaborative contribution ratio with the preset collaborative threshold, a graded early warning signal is triggered and output.

[0016] Furthermore, the meteorological dynamic factors include daily maximum temperature, daily minimum temperature, daily average temperature, daily temperature range, relative humidity, atmospheric pressure, and average wind speed;

[0017] The chemical pollutants include ozone, nitrogen dioxide, Particulate matter Particulate matter, sulfur dioxide, and carbon monoxide;

[0018] The health outcome label time-series data consists of daily counts of cardiovascular disease morbidity or mortality.

[0019] Furthermore, the calculation formula for extracting the excess pressure portion using the positive truncation operator is as follows:

[0020]

[0021] In the formula, The first eigenvector in the original feature vector The values ​​of each factor; As a factor The corresponding preset quantile threshold; After the positive part is truncated and extracted, the first The non-stationary pressure characteristics corresponding to each factor.

[0022] Furthermore, the decoupled additive network architecture performs data flow control at the input end, distributing each non-stationary pressure feature in the historical non-stationary environmental pressure feature pool into mutually isolated single-factor feature streams according to the factor dimension, and distributing the factor combinations in the historical effective interaction feature pair set into corresponding interaction feature pairs. The single-factor feature streams are input to the corresponding main effect subnets, and the interaction feature pairs are input to the corresponding synergistic effect subnets.

[0023] Furthermore, the synergistic effect subnet is a double-layer nested topology, which includes a first-level inner layer mapping and a second-level outer layer activation. The first-level inner layer mapping maps a pair of non-stationary pressure features into the hidden layer space and performs additive aggregation through the inner layer spline basis function. The second-level outer layer activation performs nonlinear curvature activation on the aggregation result through the outer layer spline basis function, so as to characterize the multiplicative synergistic multiplication effect without destroying the additive structure of the network.

[0024] Furthermore, the formula for calculating the total health risk of the decoupled additive network architecture is as follows:

[0025]

[0026] In the formula, To decouple the overall health risks of additive network architectures; Basic risk bias; The output of the main effect subnet; The output of the synergistic effect subnet; For a set of effective interaction feature pairs;

[0027] The formula for calculating the output of the main effect subnet is as follows:

[0028]

[0029] In the formula, The output of the main effect subnet; Use the baseline linear weights; This is the residual activation function, which is used to ensure the stable propagation of the basic gradient; For setting B-spline basis functions of order B; These are the control point coefficients that are dynamically updated during backpropagation.

[0030] The formula for calculating the output of the synergistic effect subnet is as follows:

[0031]

[0032] In the formula, The output of the synergistic effect subnet; These are the spline basis functions for the outermost layer with higher degrees of freedom; and These are respectively for non-stationary pressure characteristics and Independently parameterized inner spline basis functions.

[0033] Furthermore, the loss function with sparse regularization term is:

[0034]

[0035] In the formula, Here, N is the loss function; N is the number of samples. The total health risk of the decoupled additive network architecture corresponding to the t-th sample; Let t be the health outcome label corresponding to the t-th sample; is the regularization intensity coefficient; c is the set of all spline control point coefficients in the decoupled additive network architecture.

[0036] Furthermore, the formula for calculating the collaborative contribution ratio is as follows:

[0037]

[0038] In the formula, The contribution ratio of collaboration; The output of the synergistic effect subnet; For the first The total health risk value corresponding to each sample.

[0039] Furthermore, based on the comparison results of the total health risk value with the preset safety threshold and the synergistic contribution ratio with the preset synergistic threshold, a graded early warning signal is triggered and output, including:

[0040] When the total health risk value is less than the preset safety threshold, a green warning signal is triggered and output; this state indicates that the environmental pressure is within the range that the human body can tolerate, the high-risk spline region is not activated in the K-ANAM network, and normal data collection is maintained.

[0041] When the total health risk value is greater than or equal to the preset safety threshold and the synergistic contribution ratio is less than the preset synergistic threshold, a yellow warning signal is triggered and output, and all main effect subnet outputs are retrieved to lock the single factor with the largest contribution; this state indicates that the total risk exceeds the standard but the synergistic term is not significantly activated, indicating that the current disaster is a linear accumulation caused by a single dominant factor;

[0042] When the total health risk value is greater than or equal to a preset safety threshold and the synergistic contribution ratio is greater than or equal to a preset synergistic critical value, a red warning signal is triggered and output, and all synergistic effect subnet outputs are retrieved to lock the factor combination with the largest contribution; this state indicates that the total risk exceeds the standard and the synergistic item is significantly activated, indicating that the current disaster is a mechanism mutation caused by the nonlinear synergistic effect of multiple factors;

[0043] The preset safety threshold is set based on the baseline risk line determined by historical epidemiological data, and the preset collaborative threshold is set according to the calculation result of the collaborative contribution ratio of the trained K-ANAM network to historical environmental data.

[0044] A second aspect of the present invention discloses a dynamic early warning device for environmental health risks based on K-ANAM, comprising:

[0045] The data acquisition module is used to collect historical time-series data of meteorological dynamic factors and chemical pollutant factors in the target area, form raw feature vectors, and simultaneously collect time-series data of health outcome labels;

[0046] The non-stationary pressure reconstruction module is used to extract the excess pressure part of each factor in the original feature vector by using a preset quantile threshold as the normal baseline and applying the positive part truncation operator to obtain the corresponding historical non-stationary pressure features and construct a historical non-stationary environmental pressure feature pool.

[0047] The interaction feature filtering module is used to calculate the correlation coefficient between any two different factors in the historical non-stationary environmental pressure feature pool, and to determine the factor combination with a correlation coefficient greater than a preset collaboration threshold as a historical effective interaction feature pair, forming a set of historical effective interaction feature pairs.

[0048] The network architecture construction module is used to construct a decoupled additive network architecture based on the Kolmogorov-Arnold representation theorem. This decoupled additive network architecture includes main effect subnetworks that correspond one-to-one with the factors in the historical non-stationary environmental pressure feature pool and synergistic effect subnetworks that correspond one-to-one with the factor combinations in the historical effective interaction feature pair set. Its total health risk output is composed of the linear superposition of the outputs of all main effect subnetworks and all synergistic effect subnetworks.

[0049] The network training module is used to train the decoupled additive network architecture based on historical non-stationary stress characteristics, using health outcome label time series data as supervision signals, and employing a loss function with sparse regularization terms to obtain the trained K-ANAM network.

[0050] The real-time risk prediction module is used to process real-time collected data of the same factors through the positive part truncation operator and input it into the trained K-ANAM network, outputting the total health risk value and decomposing it into single-factor additive risk contribution and multi-factor synergistic risk contribution.

[0051] The early warning decision module is used to calculate the collaborative contribution ratio based on the total health risk value and the multi-factor collaborative risk contribution. Based on the comparison results of the total health risk value with the preset safety threshold and the collaborative contribution ratio with the preset collaborative threshold, it triggers and outputs a graded early warning signal.

[0052] The beneficial effects of this invention are:

[0053] (1) This invention employs a positive truncation operator based on a preset quantile threshold to perform background-free pressure reconstruction of the original environmental features. This operator is an extreme value feature extraction operation, which does not change the absolute magnitude of the original extreme data. Instead, it absolutely zeros the normal background that does not exceed the safety threshold, explicitly filters background values ​​that do not pose a significant health hazard, and retains only abnormal increments with pathogenic risks. Unlike the conventional data preprocessing method in the prior art that treats composite weather events that exceed historical extreme values ​​as outliers and smooths them out, this reconstruction step avoids the loss of disaster mutation signals. After positive truncation, the feature pool is in normal climate Under certain conditions, it exhibits high sparsity, only showing high numerical impulse characteristics when compound disasters occur, guiding the network to concentrate computational resources on non-zero feature paths. At the same time, filtering background noise improves the signal-to-noise ratio of extreme events, enabling the subsequent K-ANAM network to fully utilize the high-order derivative information retained by the learnable B-spline basis functions for adaptive curvature fitting, effectively capturing the nonlinear variation trend of environmental factors in the extreme tail interval. Even when facing extreme compound disasters that have not occurred due to a lack of historical prior samples, it can still maintain a high confidence trend extrapolation capability, overcoming the problem of underestimation caused by linear boundary constraints in traditional models.

[0054] (2) The decoupled additive network architecture based on the Kolmogorov-Arnold representation theorem constructed in this invention introduces the decoupling idea of ​​the neural additive model at a macroscopic level. It implements data flow control at the network input end, absolutely isolates the single-factor feature flow from the interaction feature flow, and forcibly cuts off the cross-connection of unrelated factors in the hidden layer, so that each main effect subnetwork and synergistic effect subnetwork remains mathematically independent. Unlike the fully connected or convolutionally wrapped deep structure used in the traditional KAN model and its variants, this white-box decoupling architecture makes the total health risk value mathematically accurate. The solution is divided into single-factor additive risk contribution and multi-factor synergistic risk contribution, from which a synergistic contribution ratio index is derived. For the selected interactive feature pairs, the synergistic effect subnet adopts a double-layer nested spline structure. The heterogeneous environmental pressure features are mapped to implicit biological loads and additively aggregated through the inner spline basis function. Then, the aggregation result is activated by high-order curvature through the outer spline basis function. In this way, the multiplicative synergistic multiplier effect between multiple factors is approximated and quantified without destroying the overall additive structure, providing a quantitative theoretical basis for distinguishing between linear cumulative disasters and mechanism mutation disasters.

[0055] (3) This invention calculates the correlation coefficient between historical non-stationary pressure features through a sliding window mechanism, pre-screens effective interactive feature pairs that exceed the preset collaborative threshold, removes noise combinations with weak correlation, and reduces the feature dimension and parameter quantity of subsequent network inputs; during the model training stage, a Poisson negative log-likelihood loss function with L1 sparse regularization term is used, and the L1 regularization penalty term forces the optimizer to automatically compress the control point coefficients of subnets that do not significantly contribute to the health outcome to zero; the screening and training mechanisms work together to fit the high sparse feature environment formed after positive part truncation processing, which improves the computational efficiency while enhancing the model's anti-overfitting ability and generalization robustness under extreme compound disaster small sample data conditions;

[0056] (4) Based on the decoupled architecture and the synergistic cooperation of nested interactive modules, this invention further extracts the synergistic contribution ratio index, and executes the three-level dynamic early warning triggering logic according to the comparison results of the total health risk value and the preset safety threshold, the synergistic contribution ratio and the preset synergistic critical value. This mechanism can effectively identify whether the current risk is caused by the linear accumulation of a single dominant factor or by the nonlinear synergistic mutation of multiple factors, and locks the single factor or interactive feature pair combination with the largest contribution when triggering yellow or red early warning, providing accurate attribution diagnosis and decision support for urban emergency response. Attached Figure Description

[0057] Figure 1 This is a flowchart of the dynamic early warning method for environmental health risks in this invention;

[0058] Figure 2 This is a flowchart of the data processing of the model in this invention;

[0059] Figure 3 This is a schematic diagram of the graded early warning system in this invention;

[0060] Figure 4 This is a block diagram of the environmental health risk dynamic early warning device in this invention. Detailed Implementation

[0061] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] See Figures 1-4 This embodiment provides a dynamic early warning method for environmental health risks based on K-ANAM:

[0063] The first aspect of this invention discloses a dynamic early warning method for environmental health risks based on K-ANAM, such as... Figure 1 As shown, the dynamic early warning method for environmental health risks includes steps S100 to S700.

[0064] Step S100. Collect historical time-series data of meteorological dynamic factors and chemical pollutant factors in the target area to form an original feature vector, and simultaneously collect time-series data of health outcome labels.

[0065] The original feature vector in this embodiment consists of historical time-series data of meteorological dynamic factors and chemical pollutant factors.

[0066] In some embodiments of this example, the meteorological dynamic factors include a basic temperature control factor, a composite fluctuation factor, and an environmental background factor, wherein the basic temperature control factor includes the daily maximum temperature. Daily minimum temperature and daily average temperature The composite fluctuation factor includes diurnal temperature range. The environmental background factors include relative humidity. Atmospheric pressure and average wind speed .

[0067] In some embodiments of this example, the chemical pollutant factors include oxidizing pollutants, particulate matter, and toxic gases, wherein the oxidizing pollutants include ozone. and nitrogen dioxide The particulate matter includes fine particulate matter. and inhalable particulate matter The toxic gas includes sulfur dioxide. and carbon monoxide .

[0068] The health outcome label time-series data are daily counts of cardiovascular disease (ICD-10 codes I00-I99) morbidity or mortality.

[0069] Step S200. For each factor in the original feature vector, using a preset quantile threshold as the normal baseline, the positive part truncation operator is applied to extract the excess pressure part, thereby obtaining the corresponding historical non-stationary pressure features, and constructing a historical non-stationary environmental pressure feature pool.

[0070] In this embodiment, the preset quantile threshold for each factor in the original feature vector is calculated based on historical time-series data of the target region. The historical non-stationary environmental pressure feature pool consists of the historical non-stationary pressure features of all factors.

[0071] This embodiment uses a positive truncation operator to reconstruct the original feature vector using background-removed pressure, thereby eliminating the interference of normal climate fluctuations on extreme value prediction. Specifically, for each factor, only the excess pressure portion exceeding a preset quantile threshold is extracted as a non-stationary pressure feature, while the portion below the threshold is zeroed out. This processing method does not change the absolute magnitude of the original extreme data, explicitly filters out background fluctuations that do not significantly harm health outcomes, and retains only abnormal increments with pathogenic risks, thus significantly improving the signal-to-noise ratio of the subsequent network model for extreme event inputs.

[0072] This embodiment performs non-stationary pressure reconstruction on the original feature vectors. This process, along with the learnable B-spline basis functions in the subsequent K-ANAM network, forms a close technical synergy, jointly improving the model's extrapolation performance under extreme event conditions. Specifically, after processing with the positive part truncation operator, the constructed non-stationary environmental pressure feature pool exhibits a large number of zero values ​​under normal climate conditions (corresponding to normal climate), and only shows high numerical impulse features when compound disasters occur. This highly sparse feature input to the network allows the network to perform effective calculations only on the subnet paths corresponding to non-zero features during training and inference, improving the efficiency of computational resource utilization. At the same time, because the reconstruction process filters out background noise components, it improves the signal-to-noise ratio in the extreme event interval, enabling the subsequent K-ANAM network to fully utilize the higher-order derivative (curvature) information retained by the learnable B-splines. The non-stationary pressure features extracted by reconstruction, combined with the curvature fitting of the B-splines, enable the model to adaptively fit the nonlinear trend of environmental factors in the extreme tail interval, overcoming the limitations of smooth fitting and the underestimation of extrapolation in traditional models when facing historically unseen extreme weather.

[0073] In some embodiments of this example, the preset quantile threshold is the 90th percentile or the 95th percentile.

[0074] In some embodiments of this example, the calculation formula for extracting the excess pressure portion using the positive truncation operator is as follows:

[0075]

[0076] In the formula, The first eigenvector in the original feature vector The values ​​of each factor; As a factor The corresponding preset quantile threshold; After the positive part is truncated and extracted, the first The non-stationary pressure characteristics corresponding to each factor.

[0077] Step S300. Calculate the correlation coefficient between any two different factors in the historical non-stationary environmental pressure feature pool, and determine the factor combination with the correlation coefficient greater than the preset collaborative threshold as the historical effective interaction feature pair, forming a set of historical effective interaction feature pairs.

[0078] In this embodiment, the correlation coefficient is the Pearson correlation coefficient. The set of historical effective interaction feature pairs consists of all historical effective interaction feature pairs.

[0079] In some embodiments of this example, the correlation coefficient is the Pearson correlation coefficient, which is calculated using a sliding window mechanism.

[0080] This embodiment calculates the Pearson correlation coefficient between time-series data of non-stationary pressure features of any two different factors in the non-stationary environmental pressure feature pool using a sliding window mechanism, and sets a preset collaboration threshold. When the correlation coefficient between two factors is greater than the preset collaboration threshold, the factor combination is identified as a valid interaction feature pair and included in the set of valid interaction feature pairs for subsequent calculation in the collaboration effect subnet; when the correlation coefficient is less than or equal to the preset collaboration threshold, the factor combination is discarded as invalid noise. This screening mechanism pre-emptively eliminates noisy combinations with weak correlation, effectively reducing the feature dimension and number of parameters of the network input, and reducing the complexity of subsequent network models in processing high-dimensional environmental features.

[0081] Step S400. Construct a decoupled additive network architecture based on the Kolmogorov-Arnold representation theorem. This decoupled additive network architecture includes main effect subnetworks that correspond one-to-one with the factors in the historical non-stationary environmental pressure feature pool and synergistic effect subnetworks that correspond one-to-one with the factor combinations in the historical effective interaction feature pair set. Its total health risk output is composed of the linear superposition of the outputs of all main effect subnetworks and all synergistic effect subnetworks.

[0082] Specifically, a decoupled additive network architecture based on the Kolmogorov-Arnold representation theorem is constructed. The decoupled additive network architecture includes a main effect subnetwork and a synergistic effect subnetwork. The total health risk output of the decoupled additive network architecture is composed of the linear superposition of the outputs of all main effect subnetworks and all synergistic effect subnetworks. The main effect subnetworks correspond one-to-one with the factors in the historical non-stationary environmental pressure feature pool, and the synergistic effect subnetworks correspond one-to-one with the factor combinations in the historical effective interaction feature pair set. The main effect subnetworks are used to fit the independent nonlinear health risk curve of a single non-stationary pressure feature, and the synergistic effect subnetworks are used to fit the multiplicative synergistic multiplication effect of paired non-stationary pressure features.

[0083] In some implementations of this embodiment, such as Figure 2 As shown, the decoupled additive network architecture performs data flow control at the input end, distributing each non-stationary pressure feature in the historical non-stationary environmental pressure feature pool into mutually isolated single-factor feature streams according to the factor dimension, and distributing the factor combinations in the historical effective interaction feature pair set into corresponding interaction feature pairs. The single-factor feature streams are input to the corresponding main effect subnets, and the interaction feature pairs are input to the corresponding synergistic effect subnets.

[0084] In some embodiments of this example, the synergistic effect subnet is a two-layer nested topology, which includes a first-level inner layer mapping and a second-level outer layer activation. The first-level inner layer mapping maps a pair of input non-stationary pressure features to the hidden layer space and performs additive aggregation through inner layer spline basis functions. The second-level outer layer activation performs nonlinear curvature activation on the aggregation result through outer layer spline basis functions to characterize the multiplicative synergistic multiplication effect without destroying the additive structure of the network.

[0085] The decoupled additive network architecture constructed in this embodiment introduces the decoupling concept of the neural additive model (NAM), a variant of the generalized additive model (GAM), into its macroscopic architecture, without adopting the fully connected paradigm of KAN. This decoupled additive network architecture forcibly severs the cross-connections of unrelated factors in the hidden layers, ensuring that each main effect subnetwork and synergistic effect subnetwork remains mathematically independent. This white-box architecture, where specific features enter specific subnetworks, is not present in traditional KAN and its variants. Furthermore, this embodiment designs a double-layer nested interactive spline multiplication module, realizing the quantification and physical separation of multi-factor nonlinear synergistic multiplication effects. This overcomes the technical bottleneck of traditional generalized additive models (GAM) and neural additive models (NAM), which are limited by the independent superposition of single variables and struggle to characterize complex interactions.

[0086] In some embodiments of this example, the formula for calculating the total health risk of the decoupled additive network architecture is as follows:

[0087]

[0088] In the formula, To decouple the overall health risks of additive network architectures; Basic risk bias; The output of the main effect subnet; The output of the synergistic effect subnet; For a set of effective interactive feature pairs.

[0089] In some embodiments of this example, the calculation formula for the output of the main effect subnet is as follows:

[0090]

[0091] In the formula, The output of the main effect subnet; Use the baseline linear weights; This is the residual activation function, which is used to ensure the stable propagation of the basic gradient; For setting B-spline basis functions of order B; These are the control point coefficients that are dynamically updated during backpropagation.

[0092] In this embodiment, when a single environmental factor data flows directly through the edge of the network, the main effect subnetwork adaptively adjusts the curvature of the control points to fit an independent nonlinear pathogenic accumulation curve for that factor, overcoming the problem of insufficient fitting ability of multilayer perceptron (MLP) for extreme tail data in traditional neural additive models.

[0093] In some embodiments of this example, the calculation formula for the output of the synergistic effect subnet is as follows:

[0094]

[0095] In the formula, The output of the synergistic effect subnet; These are the spline basis functions for the outermost layer with higher degrees of freedom; and These are respectively for non-stationary pressure characteristics and Independently parameterized inner spline basis functions are used to map heterogeneous features to implicit biological loads of uniform dimensions.

[0096] In this embodiment, the synergistic effect subnet is a two-layer nested interactive spline multiplication module, which solves the underlying deficiency of traditional additive frameworks in handling multivariate nonlinear interactions. Specifically, the first level of the synergistic effect subnet (inner implicit biological load mapping) processes the input heterogeneous environmental pressure features through inner spline basis functions. The data is nonlinearly mapped to a uniform-dimensional implicit biological load in the hidden layer dimension, and then unweighted pure additive aggregation is performed at the intermediate hidden layer nodes. The second level of the synergistic effect subnet (outer layer nonlinear doubling activation) immediately flows the aggregated scalar data into the outer layer spline basis functions with higher degrees of freedom. Higher-order curvature activation is performed. This topology utilizes nested two-level purely additive spline structures to approximate complex multiplicative interactions, enabling the computation and physical separation of multi-factor synergistic multiplication effects without destroying the additive white-box properties of the overall model.

[0097] Step S500. Based on historical non-stationary pressure characteristics, using health outcome label time series data as supervision signals, a decoupled additive network architecture is trained using a loss function with sparse regularization terms to obtain the trained K-ANAM network.

[0098] Specifically, each historical non-stationary environmental pressure feature in the historical non-stationary environmental pressure feature pool is input into the corresponding main effect subnet, and the two historical non-stationary pressure features corresponding to each pair of factor combinations in the historical effective interaction feature pair set are input into the corresponding synergistic effect subnet. The health outcome label time series data is used as the supervision signal, and the decoupled additive network architecture is trained using a loss function with sparse regularization term to obtain the trained K-ANAM network.

[0099] In some embodiments of this example, the loss function with sparse regularization is:

[0100]

[0101] In the formula, Here, N is the loss function; N is the number of samples. The total health risk of the decoupled additive network architecture corresponding to the t-th sample; Let be the health outcome label corresponding to the t-th sample, which follows a discrete count distribution and is guided by a Poisson distribution loss. is the regularization intensity coefficient; c is the set of all spline control point coefficients in the decoupled additive network architecture.

[0102] In this embodiment, a Poisson negative log-likelihood loss function with L1 sparse regularization is constructed during the model training phase. The regularization penalty term forces the optimizer to automatically adjust the control point coefficients of subnets that do not significantly contribute to healthy outcomes. Compressed to zero. This training mechanism fits the highly sparse feature environment required by the model, improving computational efficiency while enhancing the model's anti-overfitting ability and generalization robustness when facing small sample data from extreme and complex disasters.

[0103] Step S600. The real-time collected data of the same factor is processed by the positive part truncation operator and then input into the trained K-ANAM network. The total health risk value is output and decomposed into single-factor additive risk contribution and multi-factor synergistic risk contribution.

[0104] In some embodiments of this example, real-time collected data of the same type of factors are processed by a positive truncation operator and then input into a trained K-ANAM network. The total health risk value is output and decomposed into single-factor additive risk contribution and multi-factor synergistic risk contribution. This includes: combining real-time collected meteorological dynamics factor and chemical pollutant factor data to form a real-time feature vector; performing a positive truncation operator on each factor in the real-time feature vector to obtain a real-time non-stationary environmental pressure feature pool; inputting each non-stationary pressure feature in the real-time non-stationary environmental pressure feature pool into the corresponding main effect subnet; and inputting the two real-time non-stationary pressure features corresponding to each pair of factor combinations in the historical effective interaction feature pair set into the corresponding synergistic effect subnet. The trained K-ANAM network performs forward calculation to output the total health risk value. The decoupled additive network architecture decomposes the total health risk value into single-factor additive risk contribution output by all main effect subnets and multi-factor synergistic risk contribution output by all synergistic effect subnets.

[0105] Step S700. Calculate the collaborative contribution ratio based on the total health risk value and the multi-factor collaborative risk contribution. Based on the comparison results of the total health risk value with the preset safety threshold and the collaborative contribution ratio with the preset collaborative threshold, trigger and output a graded early warning signal.

[0106] In this embodiment, the synergistic contribution ratio represents the proportion of the current risk contributed by the nonlinear synergistic effect of multiple factors (rather than the linear summation of a single factor).

[0107] In some embodiments of this example, the formula for calculating the collaborative contribution ratio is:

[0108]

[0109] In the formula, The contribution ratio of collaboration; The output of the synergistic effect subnet; For the first The total health risk value corresponding to each sample.

[0110] In some implementations of this embodiment, such as Figure 3 As shown, based on the comparison results of the total health risk value with the preset safety threshold and the collaborative contribution ratio with the preset collaborative threshold, a graded early warning signal is triggered and output, including steps S710 to S730.

[0111] Step S710. When the total health risk value is less than the preset safety threshold, a green warning signal is triggered and output; this state indicates that the environmental pressure is within the range that the human body can tolerate, the high-risk spline region is not activated in the K-ANAM network, and normalized data collection is maintained.

[0112] Step S720. When the total health risk value is greater than or equal to the preset safety threshold and the synergistic contribution ratio is less than the preset synergistic critical value, a yellow warning signal is triggered and output, and all main effect subnet outputs are retrieved to lock the single factor with the largest contribution; this state indicates that the total risk exceeds the standard but the synergistic term is not significantly activated, indicating that the current disaster is a linear accumulation caused by a single dominant factor.

[0113] Step S730. When the total health risk value is greater than or equal to the preset safety threshold and the synergistic contribution ratio is greater than or equal to the preset synergistic critical value, a red warning signal is triggered and output, and all synergistic effect subnet outputs are retrieved to lock the factor combination with the largest contribution; this state indicates that the total risk exceeds the standard and the synergistic term is significantly activated, indicating that the current disaster is a mechanism mutation caused by the nonlinear synergistic effect of multiple factors.

[0114] The preset safety threshold is set based on the baseline risk line determined by historical epidemiological data, and the preset collaborative threshold is set according to the calculation result of the collaborative contribution ratio of the trained K-ANAM network to historical environmental data.

[0115] The second aspect of this embodiment discloses a dynamic early warning device for environmental health risks based on K-ANAM, such as... Figure 4 As shown, the dynamic early warning device for environmental health risks includes a data acquisition module, a non-stationary pressure reconstruction module, an interactive feature screening module, a network architecture construction module, a network training module, a real-time risk prediction module, and an early warning decision module.

[0116] The data acquisition module is used to collect historical time-series data of meteorological dynamic factors and chemical pollutant factors in the target area to form an original feature vector, and simultaneously collect time-series data of health outcome labels.

[0117] The non-stationary pressure reconstruction module is used to extract the excess pressure part of each factor in the original feature vector by using a preset quantile threshold as the normal baseline and applying a positive part truncation operator to obtain the corresponding historical non-stationary pressure features and construct a historical non-stationary environmental pressure feature pool.

[0118] The interaction feature filtering module is used to calculate the correlation coefficient between any two different factors in the historical non-stationary environmental pressure feature pool, and to determine the factor combination with the correlation coefficient greater than the preset collaboration threshold as the historical effective interaction feature pair, thus forming a set of historical effective interaction feature pairs.

[0119] The network architecture construction module is used to construct a decoupled additive network architecture based on the Kolmogorov-Arnold representation theorem. This decoupled additive network architecture includes main effect subnetworks that correspond one-to-one with the factors in the historical non-stationary environmental pressure feature pool and synergistic effect subnetworks that correspond one-to-one with the factor combinations in the historical effective interaction feature pair set. Its total health risk output is composed of the linear superposition of the outputs of all main effect subnetworks and all synergistic effect subnetworks.

[0120] The network training module is used to train a decoupled additive network architecture based on historical non-stationary stress characteristics, using time-series data of health outcome labels as supervision signals, and employing a loss function with a sparse regularization term to obtain the trained K-ANAM network.

[0121] The real-time risk prediction module is used to process real-time collected data of the same factors through a positive part truncation operator and input it into the trained K-ANAM network, outputting the total health risk value and decomposing it into single-factor additive risk contribution and multi-factor synergistic risk contribution.

[0122] The early warning decision module is used to calculate the collaborative contribution ratio based on the total health risk value and the multi-factor collaborative risk contribution, and to trigger and output a graded early warning signal based on the comparison results of the total health risk value with the preset safety threshold and the collaborative contribution ratio with the preset collaborative critical value.

[0123] It should be noted that those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system or device described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0124] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A dynamic early warning method for environmental health risks based on K-ANAM, characterized in that, include: Historical time-series data of meteorological dynamic factors and chemical pollutant factors in the target area were collected to form raw feature vectors, and time-series data of health outcome labels were collected simultaneously. For each factor in the original feature vector, a preset quantile threshold is used as the normal baseline. The positive part truncation operator is applied to extract the excess pressure part, and the corresponding historical non-stationary pressure features are obtained. A historical non-stationary environmental pressure feature pool is then constructed. Calculate the correlation coefficient between any two different factors in the historical non-stationary environmental pressure feature pool, and determine the factor combination with the correlation coefficient greater than the preset collaborative threshold as the historical effective interaction feature pair, forming a set of historical effective interaction feature pairs. A decoupled additive network architecture based on the Kolmogorov-Arnold representation theorem is constructed. This decoupled additive network architecture includes main effect subnetworks that correspond one-to-one with the factors in the historical non-stationary environmental pressure feature pool and synergistic effect subnetworks that correspond one-to-one with the factor combinations in the historical effective interaction feature pair set. Its total health risk output is composed of the linear superposition of the outputs of all main effect subnetworks and all synergistic effect subnetworks. Based on historical non-stationary stress characteristics, using health outcome label time series data as supervision signals, a decoupled additive network architecture is trained using a loss function with sparse regularization terms to obtain the trained K-ANAM network. Real-time collected data of similar factors are processed by the positive part truncation operator and then input into the trained K-ANAM network to output the total health risk value and decompose it into single-factor additive risk contribution and multi-factor synergistic risk contribution. The collaborative contribution ratio is calculated based on the total health risk value and the collaborative risk contribution of multiple factors. Based on the comparison results between the total health risk value and the preset safety threshold, and the collaborative contribution ratio and the preset collaborative threshold, a graded early warning signal is triggered and output. The formula for calculating the total health risk of the decoupled additive network architecture is as follows: In the formula, To decouple the overall health risks of additive network architectures; Basic risk bias; The output of the main effect subnet; The output of the synergistic effect subnet; For a set of effective interaction feature pairs; The formula for calculating the output of the main effect subnet is as follows: In the formula, The output of the main effect subnet; Use the baseline linear weights; The residual activation function is used to ensure the stable propagation of the basic gradient; For setting B-spline basis functions of order B; These are the control point coefficients that are dynamically updated during backpropagation; The formula for calculating the output of the synergistic effect subnet is as follows: In the formula, The output of the synergistic effect subnet; These are the spline basis functions for the outermost layer with higher degrees of freedom; and These are respectively for non-stationary pressure characteristics and Independently parameterized inner spline basis functions.

2. The method for dynamic early warning of environmental health risks based on K-ANAM according to claim 1, characterized in that, The meteorological dynamic factors include daily maximum temperature, daily minimum temperature, daily average temperature, daily temperature range, relative humidity, atmospheric pressure, and average wind speed; The chemical pollutants include ozone, nitrogen dioxide, Particulate matter Particulate matter, sulfur dioxide, and carbon monoxide; The health outcome label time-series data consists of daily counts of cardiovascular disease morbidity or mortality.

3. The method for dynamic early warning of environmental health risks based on K-ANAM according to claim 1, characterized in that, The calculation formula for extracting the excess pressure portion using the positive truncation operator is as follows: In the formula, The first eigenvector in the original feature vector The values ​​of each factor; As a factor The corresponding preset quantile threshold; After the positive part is truncated and extracted, the first The non-stationary pressure characteristics corresponding to each factor.

4. The method for dynamic early warning of environmental health risks based on K-ANAM according to claim 1, characterized in that, The decoupled additive network architecture performs data flow control at the input end, distributing each non-stationary pressure feature in the historical non-stationary environmental pressure feature pool into mutually isolated single-factor feature streams according to the factor dimension, and distributing the factor combinations in the historical effective interaction feature pair set into corresponding interaction feature pairs. The single-factor feature streams are input to the corresponding main effect subnets, and the interaction feature pairs are input to the corresponding synergistic effect subnets.

5. The method for dynamic early warning of environmental health risks based on K-ANAM according to claim 1, characterized in that, The synergistic effect subnet is a double-layer nested topology, which includes a first-level inner layer mapping and a second-level outer layer activation. The first-level inner layer mapping maps a pair of non-stationary pressure features into the hidden layer space and performs additive aggregation through inner layer spline basis functions. The second-level outer layer activation performs nonlinear curvature activation on the aggregation result through outer layer spline basis functions to characterize the multiplicative synergistic multiplication effect without destroying the additive structure of the network.

6. The method for dynamic early warning of environmental health risks based on K-ANAM according to claim 1, characterized in that, The loss function with sparse regularization is: In the formula, Here, N is the loss function; N is the number of samples. The total health risk of the decoupled additive network architecture corresponding to the t-th sample; Let t be the health outcome label corresponding to the t-th sample; is the regularization intensity coefficient; c is the set of all spline control point coefficients in the decoupled additive network architecture.

7. The method for dynamic early warning of environmental health risks based on K-ANAM according to claim 1, characterized in that, The formula for calculating the collaborative contribution ratio is as follows: In the formula, The contribution ratio of collaboration; The output of the synergistic effect subnet; For the first The total health risk value corresponding to each sample.

8. The method for dynamic early warning of environmental health risks based on K-ANAM according to claim 1, characterized in that, Based on the comparison results of the total health risk value with the preset safety threshold and the synergistic contribution ratio with the preset synergistic threshold, a graded early warning signal is triggered and output, including: When the total health risk value is less than the preset safety threshold, a green warning signal is triggered and output; this state indicates that the environmental pressure is within the range that the human body can tolerate, the high-risk spline region is not activated in the K-ANAM network, and normal data collection is maintained. When the total health risk value is greater than or equal to the preset safety threshold and the synergistic contribution ratio is less than the preset synergistic threshold, a yellow warning signal is triggered and output, and all main effect subnet outputs are retrieved to lock the single factor with the largest contribution; this state indicates that the total risk exceeds the standard but the synergistic term is not significantly activated, indicating that the current disaster is a linear accumulation caused by a single dominant factor; When the total health risk value is greater than or equal to a preset safety threshold and the synergistic contribution ratio is greater than or equal to a preset synergistic critical value, a red warning signal is triggered and output, and all synergistic effect subnet outputs are retrieved to lock the factor combination with the largest contribution; this state indicates that the total risk exceeds the standard and the synergistic item is significantly activated, indicating that the current disaster is a mechanism mutation caused by the nonlinear synergistic effect of multiple factors; The preset safety threshold is set based on the baseline risk line determined by historical epidemiological data, and the preset collaborative threshold is set according to the calculation result of the collaborative contribution ratio of the trained K-ANAM network to historical environmental data.

9. A dynamic early warning device for environmental health risks based on K-ANAM, characterized in that, include: The data acquisition module is used to collect historical time-series data of meteorological dynamic factors and chemical pollutant factors in the target area, form raw feature vectors, and simultaneously collect time-series data of health outcome labels; The non-stationary pressure reconstruction module is used to extract the excess pressure part of each factor in the original feature vector by using a preset quantile threshold as the normal baseline and applying the positive part truncation operator to obtain the corresponding historical non-stationary pressure features and construct a historical non-stationary environmental pressure feature pool. The interaction feature filtering module is used to calculate the correlation coefficient between any two different factors in the historical non-stationary environmental pressure feature pool, and to determine the factor combination with a correlation coefficient greater than a preset collaboration threshold as a historical effective interaction feature pair, forming a set of historical effective interaction feature pairs. The network architecture construction module is used to construct a decoupled additive network architecture based on the Kolmogorov-Arnold representation theorem. This decoupled additive network architecture includes main effect subnetworks that correspond one-to-one with the factors in the historical non-stationary environmental pressure feature pool and synergistic effect subnetworks that correspond one-to-one with the factor combinations in the historical effective interaction feature pair set. Its total health risk output is composed of the linear superposition of the outputs of all main effect subnetworks and all synergistic effect subnetworks. The network training module is used to train the decoupled additive network architecture based on historical non-stationary stress characteristics, using health outcome label time series data as supervision signals, and employing a loss function with sparse regularization terms to obtain the trained K-ANAM network. The real-time risk prediction module is used to process real-time collected data of the same factors through the positive part truncation operator and input it into the trained K-ANAM network, outputting the total health risk value and decomposing it into single-factor additive risk contribution and multi-factor synergistic risk contribution. The early warning decision module is used to calculate the collaborative contribution ratio based on the total health risk value and the multi-factor collaborative risk contribution, and to trigger and output a graded early warning signal based on the comparison results of the total health risk value with the preset safety threshold and the collaborative contribution ratio with the preset collaborative threshold. The formula for calculating the total health risk of the decoupled additive network architecture is as follows: In the formula, To decouple the overall health risks of additive network architectures; Basic risk bias; The output of the main effect subnet; The output of the synergistic effect subnet; For a set of effective interaction feature pairs; The formula for calculating the output of the main effect subnet is as follows: In the formula, The output of the main effect subnet; Use the baseline linear weights; The residual activation function is used to ensure the stable propagation of the basic gradient; For setting B-spline basis functions of order B; These are the control point coefficients that are dynamically updated during backpropagation; The formula for calculating the output of the synergistic effect subnet is as follows: In the formula, The output of the synergistic effect subnet; These are the spline basis functions for the outermost layer with higher degrees of freedom; and These are respectively for non-stationary pressure characteristics and Independently parameterized inner spline basis functions.

Images