Influenza prediction methods, systems, devices, and media based on epidemiological dynamics

Abstract

The application provides an influenza prediction method, system, device and medium based on epidemic dynamics, and relates to the technical field of epidemic transmission prediction, which realizes collaborative modeling of data-driven and mechanism-driven by fusing multiple source monitoring data and epidemic transmission mechanism; in the feature engineering stage, the influenza monitoring data is structured and features are extracted, forming influenza transmission features with physical meaning; in the prediction stage, a physical information neural network model based on a deep neural network framework is used, which includes an input layer, a multi-layer feedforward network, a time embedding representation layer and an output layer, and is trained by fusing an observation data fitting and a composite loss function constrained by the epidemic dynamics mechanism, so that the model can fit the historical monitoring data and strictly follow the internal dynamics law of disease transmission, thereby significantly improving the accuracy, interpretability, extrapolation stability and robustness to small samples and noise interference of the prediction result.

Description

Technical Field

[0001] This application relates to the field of epidemic transmission prediction technology, and in particular to an influenza prediction method, system, device and medium based on epidemic dynamics. Background Technology

[0002] Seasonal influenza is a respiratory infectious disease with high morbidity and transmission rates worldwide. Its rapid mutation rate, strong transmissibility, and tendency to induce moderate to severe clinical symptoms make it a key target for influenza prevention and control. Influenza epidemics exhibit significant seasonality and regional variations, influenced by a combination of factors including climate conditions, population mobility, contact patterns, and population immunity levels, leading to considerable uncertainty in public health decision-making. Existing influenza prediction methods mainly include mechanism-driven models and data-driven models. While the former has epidemiological significance, it makes idealized assumptions and has poor adaptability. The latter has strong fitting ability but lacks mechanistic constraints, has weak interpretability, and insufficient robustness and extrapolation ability under small sample or non-stationary conditions. Summary of the Invention

[0003] This application provides an influenza prediction method, system, device, and medium based on epidemic dynamics to solve one or more technical problems existing in the prior art, and at least provides a beneficial option or creates conditions that can improve prediction stability and extrapolation reliability while ensuring model interpretability.

[0004] On the one hand, this application provides an influenza prediction method based on epidemic dynamics, including the following steps:

[0005] Acquire influenza surveillance data, which includes influenza time-series data, meteorological data, holiday data, and population movement data; The influenza surveillance data were normalized, vectorized, and time-coded to obtain influenza transmission characteristics; The influenza transmission characteristics are input into a physical information neural network model, and the model outputs an influenza transmission trend prediction result. The physical information neural network model is based on a deep neural network architecture and includes, in sequence, an input layer, a multi-layer feedforward network, a temporal embedding representation layer, and an output layer. The physical information neural network model is trained using a composite loss function that combines observational data fitting with constraints from epidemic dynamics mechanisms, in order to simultaneously meet the fitting requirements of historical influenza surveillance data and the constraints of preset disease transmission dynamics laws.

[0006] Furthermore, the process of normalizing, vectorizing, and temporally encoding the influenza surveillance data to obtain influenza transmission characteristics includes the following steps: The influenza time series data is subjected to time series structuring processing. By unifying the weekly series, correcting the reporting delay, and smoothing short-term fluctuations, a continuous and stable influenza epidemic time series is generated. The meteorological data, holiday data, and population flow data are time-aligned and resampled to obtain influenza aligned data. Based on the influenza aligned data, exogenous features related to influenza transmission are extracted, including seasonal harmonic features, lag variation features of temperature and humidity, holiday interference features, population mobility intensity features, previous epidemic intensity and historical peak information. All extracted features are normalized, encoded and vectorized to obtain an influenza feature vector set. The influenza feature vector set is screened and its dimensionality reduced. Redundant and noisy features are removed by using correlation analysis, principal component analysis, or feature importance assessment methods, and the influenza transmission features are retained.

[0007] Furthermore, the specific structure of the physical information neural network model is as follows: The input layer is used to receive the influenza transmission characteristics as input features, and input the input features into the multilayer feedforward network in the form of a fixed-length feature vector. The multi-layer feedforward network consists of multiple hidden layers, each containing a linear transformation and a non-linear activation function, used to extract high-dimensional non-linear propagation patterns from the input features. The time embedding representation layer adopts a Fourier time embedding structure to map the time variables in the input features into harmonic features in the form of sine and cosine, and parametrically models the influenza transmission rate to capture the periodic transmission characteristics of influenza. The output layer is used to predict key parameters related to epidemic dynamics, including transmission rate, incubation period parameters, recovery rate parameters, and number of infections. The output layer does not use an activation function or uses a positive activation function to ensure that the predicted parameters satisfy biological non-negativity.

[0008] Furthermore, the construction of the composite loss function specifically includes: A data fitting error term is constructed to measure the difference between the output of the physical information neural network model and historical influenza surveillance data; A physical residual term is constructed, which is calculated by substituting the derivatives of each state variable with respect to time from the output of the physical information neural network model into the system of differential equations of epidemic dynamics. This residual term is used to measure the degree of deviation of the model output from the dynamic laws of disease transmission. The system of differential equations of epidemic dynamics is used to describe the dynamic transformation relationship between susceptible, latent, infected and recovered populations. Initial condition residuals and boundary condition residuals are constructed to constrain the state variables of the model at the initial time and at the prediction boundary to satisfy the epidemiological prior, respectively. The data fitting error term, physical residual term, initial condition residual term, and boundary condition residual term are weighted and combined to form the composite loss function. The composite loss function is then minimized by an optimization algorithm, so that the model can simultaneously improve the fitting accuracy of the observed data and the degree of adherence to the epidemic transmission mechanism during the training process.

[0009] Furthermore, the physical information neural network model is trained using the composite loss function, specifically including: The influenza transmission characteristics obtained by processing historical influenza surveillance data through the feature engineering module are used as training samples and input into the physical information neural network model to obtain the state variables output by the network. Using automatic differentiation techniques, the derivatives of each state variable with respect to time are calculated; Substituting the derivatives and corresponding state variables into the set of epidemiological dynamics differential equations used to describe the dynamic transformation relationship between susceptible, latent, infected and recovered populations, the physical residual term is calculated. Based on the observed values ​​in the historical influenza surveillance data and the corresponding state variables output by the network, the data fitting error term is calculated; Based on the preset initial flow state and boundary conditions, calculate the initial condition residual term and the boundary condition residual term respectively; The data fitting error term, physical residual term, initial condition residual term, and boundary condition residual term are weighted and summed to form a composite loss value. Based on the composite loss value, the parameters of the physical information neural network model are updated using a gradient descent-type optimization algorithm, and the above operations are iteratively executed until the convergence condition is met, thus completing the training of the physical information neural network model.

[0010] Furthermore, the method also includes: Based on the influenza transmission trend prediction results, and using preset influenza epidemic intensity thresholds and epidemic duration thresholds, multi-level early warning signals are generated. The multi-level early warning signals include low-risk alerts, medium-risk monitoring, high-risk warnings, and extremely high-risk emergency responses, which are used to provide tiered intervention recommendations.

[0011] Furthermore, the method also includes: Based on the predicted influenza transmission trend and the multi-level early warning signals, the data is dynamically visualized in the form of time series graphs, heat maps, or geographic information system maps. The visualization content includes a predicted curve of the number of infections within a preset time period, a trend of changes in the transmission rate, a distribution of regional risk levels, and the spatiotemporal evolution of warning levels, which is used to support an intuitive understanding of the influenza transmission situation and risk distribution.

[0012] On the other hand, this application provides an influenza prediction system based on epidemic dynamics, including a data acquisition module, a feature engineering module, and a transmission prediction module; The data acquisition module is configured to acquire influenza surveillance data, which includes influenza time-series data, meteorological data, holiday data, and population flow data. The feature engineering module is configured to: normalize, vectorize, and time-encode the influenza surveillance data to obtain influenza transmission characteristics; The transmission prediction module is configured to: input the influenza transmission characteristics into a physical information neural network model, and output the influenza transmission trend prediction result through the physical information neural network model; The physical information neural network model is based on a deep neural network architecture and includes, in sequence, an input layer, a multi-layer feedforward network, a temporal embedding representation layer, and an output layer. The physical information neural network model is trained using a composite loss function that combines observational data fitting with constraints from epidemic dynamics mechanisms, in order to simultaneously meet the fitting requirements of historical influenza surveillance data and the constraints of preset disease transmission dynamics laws.

[0013] On the other hand, this application provides an electronic device including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the aforementioned influenza prediction method based on epidemic dynamics.

[0014] On the other hand, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned influenza prediction method based on epidemic dynamics.

[0015] The beneficial effects of this application are as follows: This application provides an influenza prediction method based on epidemic dynamics. This method achieves collaborative modeling driven by data and mechanism by fusing multi-source monitoring data with epidemic transmission mechanisms. In the feature engineering stage, influenza time series, meteorological, holiday, and population flow data are structured and feature extracted to form influenza transmission characteristics with physical meaning. In the prediction stage, a physical information neural network model based on a deep neural network architecture is used, which includes an input layer, a multi-layer feedforward network, a temporal embedding representation layer, and an output layer. It is trained by a composite loss function that fuses observational data fitting with constraints from epidemic dynamic mechanisms. This allows the model to fully fit historical monitoring data while strictly following the inherent dynamic laws of disease transmission, thereby significantly improving the accuracy, interpretability, extrapolation stability, and robustness to small samples and noise interference of the prediction results. This application also provides corresponding systems, devices, and media. The beneficial effects of the systems, devices, and media are similar to those of the method and will not be repeated here.

[0016] Other features and advantages of this application will be set forth in the following description. Some features and advantages will be obvious to those skilled in the art from the description, or may be understood and realized by implementing the technical solutions of this application. Attached Figure Description

[0017] The accompanying drawings are provided to further understand the technical solutions of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the technical solutions of the present invention, and do not constitute a limitation on the technical solutions of the present invention.

[0018] Figure 1 This is a flowchart of the influenza prediction method based on epidemic dynamics provided in this application; Figure 2 This is a structural diagram of the influenza prediction system based on epidemic dynamics provided in this application. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0020] The present application will be further described below with reference to the accompanying drawings and specific embodiments. The described embodiments should not be considered as limitations on the present application, and all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of the present application.

[0021] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0022] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0023] In the current global public health prevention and control system, seasonal influenza, especially the H3N2 subtype, remains a key focus of disease surveillance and early warning due to its high variability, strong transmissibility, and tendency to cause severe illness. Accurately predicting its transmission trends is crucial for optimizing vaccine deployment, allocating medical resources, and developing non-pharmaceutical interventions. Therefore, the academic community and public health institutions have long relied on two main prediction paradigms: mechanism-driven models and data-driven models.

[0024] Mechanism-driven models, represented by ordinary differential equation compartment models such as SIR, SEIR, and SEIRS, use mathematical equations to characterize the dynamic transformation relationships between susceptible individuals, latent individuals, infected individuals, and recovered individuals. These models have a solid epidemiological theoretical foundation, their parameters possess clear biological significance, and they can reflect the basic laws of disease transmission, providing long-term trend projection capabilities under ideal conditions. However, these models are typically based on strong assumptions such as uniform population mixing, constant contact rates, and static parameters, making it difficult to effectively integrate multi-source external driving factors such as meteorology, holidays, and population movement. Furthermore, they are highly sensitive to the quality of input data; their predictive performance significantly deteriorates when faced with deficiencies in real-world monitoring data, such as reporting delays, underreporting, and noise interference, and they lack the ability to characterize local spatiotemporal heterogeneity.

[0025] Data-driven models primarily rely on statistical learning or deep learning techniques, including time series models such as ARIMA and neural network architectures such as RNN, LSTM, GRU, and Transformer. These methods excel at automatically extracting complex nonlinear patterns from massive amounts of historical data, demonstrating strong fitting capabilities in large-sample, stationary environments. However, they are essentially black-box models, lacking explicit constraints on the intrinsic mechanisms of epidemic transmission, resulting in poor model interpretability. More importantly, in real-world scenarios where influenza surveillance data typically presents small samples, high noise, non-stationarity, or even abrupt changes, purely data-driven models are prone to overfitting the training data, exhibiting weak extrapolation capabilities. The predicted results may severely violate basic epidemiological common sense, such as the emergence of negative infection numbers or unbounded growth in transmission rates, making it difficult to meet the requirements of stability, credibility, and scientific rigor for public health decision-making.

[0026] In recent years, although some studies have attempted to combine the two approaches—for example, by using the output of the mechanistic model as a feature input to the machine learning model, or by adding a simple regularization term to the loss function—these solutions often only achieve shallow fusion. They fail to embed the dynamic laws represented by the differential equations as hard constraints into the entire model training process, and cannot fundamentally guarantee that the neural network output satisfies the basic conservation laws and evolutionary logic of epidemic transmission at any point in time and space. Therefore, there is an urgent need for a new predictive framework that can fully utilize multi-source heterogeneous monitoring data while strictly adhering to the principles of epidemic dynamics, in order to overcome the comprehensive bottlenecks of existing technologies in terms of accuracy, robustness, interpretability, and extrapolation ability.

[0027] To address the aforementioned issues, this application provides an influenza prediction method, system, device, and medium based on epidemic dynamics. By deeply integrating the differential equations describing the disease transmission process as physical constraints into the neural network training mechanism, it achieves an organic fusion of data-driven and mechanism-driven approaches. The system first performs time alignment, feature extraction, and dimensionality reduction on multi-source influenza surveillance data to form structured influenza transmission characteristics. These characteristics are then input into a Physical Information Neural Network (PINN) model based on a deep neural network architecture. This model includes an input layer, a multi-layer feedforward network, a Fourier time embedding representation layer, and an output layer, capable of automatically learning the nonlinear coupling relationship between exogenous factors and periodic transmission patterns. Crucially, the model is trained using a composite loss function that integrates the fitting error of observed data with the residuals of epidemic dynamics. Automatic differentiation techniques are used to calculate the time derivatives of state variables and substitute them into the differential equations to construct physical constraint terms. Simultaneously, initial and boundary condition residuals are introduced to ensure that the prediction results not only highly fit historical surveillance data but also strictly follow the basic dynamic laws of disease transmission at any given time. This significantly improves the accuracy, stability, interpretability, and extrapolation ability in small sample and non-stationary scenarios.

[0028] First, the influenza prediction method based on epidemic dynamics provided in the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0029] Reference Figure 1 The implementation process of the influenza prediction method based on epidemic dynamics provided in this application includes, but is not limited to, the following steps.

[0030] Step S110: Obtain influenza surveillance data.

[0031] The influenza surveillance data includes influenza time-series data, meteorological data, holiday data, and population movement data.

[0032] In step S110, multi-source heterogeneous data closely related to influenza transmission are systematically collected, providing a comprehensive, realistic, and driving foundation of raw information for subsequent modeling. The influenza surveillance data acquired in this step not only includes time-series data reflecting the actual epidemic situation but also integrates key external environmental and social behavioral factors that may influence the dynamics of virus transmission, specifically covering meteorological data, holiday data, and population movement data. These data collectively constitute a multi-dimensional input space describing the complex driving mechanisms of influenza transmission, enabling the predictive model to capture the spatiotemporal evolution of epidemic transmission from a more realistic perspective, thus laying a data foundation for improving prediction accuracy and robustness.

[0033] Optionally, obtaining influenza surveillance data specifically includes: obtaining H3N2 influenza surveillance data from national influenza centers of the Global Influenza Surveillance and Response System and reference laboratories of relevant partner countries, or directly downloading information such as the number of cases, test results, and influenza positivity rates uploaded by various regions from the World Health Organization regional database; subsequently, the obtained raw data undergoes systematic cleaning, integration, and standardization processing, covering key operations such as outlier removal, time alignment, missing value completion, and reporting delay correction, thereby generating a continuous, stable, and structurally consistent time-series dataset, providing a high-quality and reliable data foundation for subsequent model training.

[0034] Step S120: Normalize, vectorize and time-encode the influenza surveillance data to obtain influenza transmission characteristics.

[0035] In step S120, the raw multi-source influenza surveillance data is structured, standardized, and semantically enhanced to generate feature representations that are suitable for the input requirements of deep learning models and rich in epidemiological significance. This step eliminates biases caused by different dimensions and scales through normalization, transforms discrete or categorical variables into computable forms in continuous numerical space through vectorization, and explicitly characterizes the periodicity, trends, and event correlations in the time dimension through time encoding. This series of processes not only ensures the consistency and stability of the input data in mathematical form, but more importantly, it transforms the raw observation data into high-order features that can effectively characterize the coupling effect between the intrinsic mechanism and exogenous driving force of influenza transmission, thereby providing high-quality feature input for the model to accurately learn the transmission patterns.

[0036] Step S130: Input the influenza transmission characteristics into the physical information neural network model, and output the influenza transmission trend prediction results through the physical information neural network model.

[0037] In step S130, a neural network architecture designed with physical mechanism constraints is used to perform nonlinear mapping and dynamic modeling on the extracted influenza transmission characteristics, ultimately outputting a transmission trend prediction result with epidemiological interpretability. The physical information neural network model used in this step not only possesses the powerful fitting ability of deep neural networks, but more importantly, its training process incorporates epidemiological dynamic equations as physical constraints, ensuring that the model follows the basic biological laws of disease transmission while learning data patterns. By inputting the preprocessed features into this model, the system can output key state variables such as the number of infections and the transmission rate in future periods, thereby achieving a scientific, stable, and extrapolation-capable prediction of influenza transmission trends, providing a reliable basis for public health decision-making.

[0038] In some embodiments of this application, the physical information neural network model is based on a deep neural network architecture and includes an input layer, a multi-layer feedforward network, a temporal embedding representation layer, and an output layer. The input layer is used to receive preprocessed influenza transmission features, the multi-layer feedforward network extracts high-order abstract features through nonlinear transformation, the temporal embedding representation layer explicitly models the dynamic evolution law in the time dimension, and the output layer generates the prediction result of the future influenza transmission trend.

[0039] Furthermore, the model is trained using a composite loss function that combines observational data fitting with constraints from epidemiological dynamics mechanisms. This allows the model to not only accurately fit the actual transmission patterns reflected by historical influenza surveillance data during the optimization process, but also strictly follow the preset laws of disease transmission dynamics, thus balancing the flexibility of data-driven approaches with the interpretability and generalization ability of mechanistic models.

[0040] In some embodiments of this application, step S120 involves normalizing, vectorizing, and time-coding the influenza surveillance data to obtain influenza transmission characteristics, including the following steps.

[0041] Step S210: Perform time series structuring processing on the influenza time series data. By unifying the weekly series, correcting the reporting delay, and smoothing short-term fluctuations, a continuous and stable influenza epidemic time series is generated. Time alignment and resampling are performed on meteorological data, holiday data, and population flow data to obtain influenza aligned data.

[0042] In step S210, the raw influenza surveillance data undergoes time-dimensional standardization and structuring to construct a high-quality, modelable time-series input foundation. This step first addresses issues in influenza time-series data such as inconsistent weekly reporting cycles, reporting delays, and short-term drastic fluctuations caused by random factors. It achieves this by adopting a unified weekly sequence format, correcting distortions caused by delayed reporting, and applying smoothing algorithms to suppress abnormal disturbances, thereby generating a continuous, stable time series that accurately reflects the influenza epidemic trend. Simultaneously, to ensure comparability and consistency of multi-source heterogeneous data across the time axis, this step also performs time alignment and resampling operations on meteorological data, holiday data, and population movement data, ensuring precise matching with the influenza time-series data at the same time granularity and time nodes. This ultimately forms structured, time-synchronized influenza aligned data, providing reliable data support for subsequent feature engineering.

[0043] Step S220: Based on influenza aligned data, extract exogenous features related to influenza transmission, including seasonal harmonic features, lag variation features of temperature and humidity, holiday interference features, population mobility intensity features, previous epidemic intensity and historical peak information, and normalize, encode and vectorize all extracted features to obtain an influenza feature vector set.

[0044] In step S220, exogenous driving features closely related to the influenza transmission mechanism are systematically mined from the aligned multi-source data and transformed into numerical representations that the model can process. This step extracts semantic features across multiple dimensions based on the aligned influenza data, including seasonal harmonic features characterizing annual cyclical patterns, lagged temperature and humidity variations reflecting the impact of climate and environment, holiday disturbance features representing social activity disturbances, population flow intensity features reflecting population migration dynamics, and information on previous epidemic intensity and historical peak values. Subsequently, all extracted features are normalized to eliminate dimensional differences, and categorical or event-type variables are converted into numerical forms through encoding. These are further vectorized into fixed-dimensional feature representations, ultimately integrating to form a structurally complete, semantically rich influenza feature vector set that meets the input requirements of neural networks, providing a high-information-density input foundation for the model to learn transmission patterns.

[0045] Step S230: The influenza feature vector set is screened and its dimensionality reduced. Redundant and noisy features are removed by using correlation analysis, principal component analysis or feature importance assessment methods, and the influenza transmission features are retained.

[0046] In step S230, the quality and efficiency of the feature set are optimized to improve the model's generalization ability and computational stability. This step identifies and removes redundant or noisy components from the high-dimensional influenza feature vector set generated in step S220 to avoid irrelevant features interfering with model training or causing overfitting. Specifically, correlation analysis is used to measure the strength of the linear or nonlinear association between each feature and the target variable; principal component analysis is used to compress the feature space and retain the main directions of variation; or feature importance assessment methods, such as weight scoring based on tree models, are used to quantify the actual contribution of each feature to the prediction task. Based on this, the most explanatory and predictive key feature subset is selected, ultimately retaining a concise, robust, and highly correlated influenza transmission feature set, laying an efficient feature foundation for accurate modeling and reliable prediction of the subsequent physical information neural network.

[0047] In some embodiments of this application, feature engineering processing may also employ a multimodal data credibility-aware fusion method, specifically including: before the vectorized multi-source features are input into the neural network, a gated attention layer is added. This layer dynamically allocates fusion weights based on real-time quality indicators of each data source, such as the number of days of delay in medical reports, the confidence level of meteorological sensors, and the noise level of web searches. When the missing data of a certain type exceeds a threshold, a generative completion submodule based on the SEIR differential equation is activated to use physical constraints to deduce the number of influenza-like cases in the missing period, ensuring the integrity and reliability of the input features.

[0048] In some embodiments of this application, the specific structure of the physical information neural network model is as follows.

[0049] First, the input layer receives influenza transmission characteristics as input features and inputs these features into a multi-layer feedforward network in the form of a fixed-length feature vector.

[0050] Specifically, the input layer receives the influenza transmission features obtained from the previous processing and organizes them into a fixed-length feature vector form that the model can process, serving as the initial input to the entire physical information neural network. This layer ensures that all exogenous variables and historical state information related to influenza transmission enter the network in a unified and structured numerical format, providing a stable and complete data foundation for feature abstraction and dynamic modeling in subsequent layers. It is the key interface connecting the data preprocessing stage and the deep modeling stage.

[0051] Secondly, the multilayer feedforward network consists of multiple hidden layers, each containing linear transformations and nonlinear activation functions, used to extract high-dimensional nonlinear propagation patterns from input features.

[0052] Specifically, the role of a multilayer feedforward network is to extract high-dimensional nonlinear propagation patterns layer by layer from the features passed from the input layer. It consists of multiple stacked hidden layers, each containing a combination of linear transformations and nonlinear activation functions. Through this hierarchical nonlinear mapping mechanism, the network can automatically learn the complex interactions and potential patterns in the influenza transmission process, such as the combined effects of climate factors and population behavior on infection rates, and changes in epidemic intensity at different time scales, thereby constructing an internal representation that is highly expressive of the real transmission dynamics.

[0053] Then, the temporal embedding representation layer adopts a Fourier temporal embedding structure to map the time variables in the input features into harmonic features in the form of sine and cosine, and parametrically models the influenza transmission rate to capture the periodic transmission characteristics of influenza.

[0054] Specifically, the temporal embedding layer explicitly models the structural impact of time variables on influenza transmission. This layer employs a Fourier temporal embedding structure, transforming the original timestamps or time indices into a series of harmonic features in sine and cosine forms. This periodic encoding method effectively captures the inherent seasonal and interannual cyclical characteristics of influenza transmission, allowing the model to fully consider temporal phase and periodic patterns when parameterizing the transmission rate. This improves the accuracy of predicting long-term trends and recurring epidemic peaks, and enhances the model's generalization ability in cross-year scenarios.

[0055] Furthermore, the output layer is used to predict key parameters related to epidemic dynamics, including transmission rate, incubation period parameters, recovery rate parameters, and number of infections. The output layer does not use an activation function or uses a positive activation function to ensure that the predicted parameters satisfy biological non-negativity.

[0056] Specifically, the output layer generates predictions of key parameters directly related to the dynamics of the epidemic, including the basic transmission rate, incubation period parameters, recovery rate parameters, and the number of infections in future periods. To ensure the biological and epidemiological validity of these parameters, the output layer is designed to use no activation functions or only positive activation functions that guarantee non-negative output values, such as Softplus or exponential functions. This constraint mechanism effectively avoids meaningless results such as negative transmission or recovery rates in the model output, thus ensuring that the predictions conform to both data-driven principles and the basic scientific principles of disease transmission.

[0057] In some embodiments of this application, the construction of the composite loss function of the physical information neural network model specifically includes the following steps.

[0058] Step S310: Construct a data fitting error term to measure the difference between the output of the physical information neural network model and historical influenza surveillance data.

[0059] In step S310, a data fitting error term is constructed to quantify the inconsistency between the predicted output of the physical information neural network model and the historical influenza surveillance data actually observed. This error term is usually measured using methods such as mean squared error or mean absolute error, directly reflecting the model's accuracy in reproducing known epidemiological phenomena. It is the core supervisory signal driving the model to learn real-world data patterns, ensuring that the model output closely matches the actual trajectory of the epidemic.

[0060] Step S320: Construct the physical residual term. This is calculated by substituting the derivatives of each state variable output by the physical information neural network model with respect to time into the epidemiological dynamics differential equations. This residual term measures the degree of deviation of the model output from the dynamics of disease transmission. The epidemiological dynamics differential equations describe the dynamic transformation relationships between susceptible, latent, infected, and recovered populations.

[0061] In step S320, a physical residual term is constructed. By substituting the derivatives of each state variable output by the model with respect to time into a pre-defined set of differential equations for epidemic dynamics, the deviation between the residual term and the theoretical dynamics on the right side of the equations is calculated. This measures the extent to which the model's predictions deviate from the fundamental physical laws of disease transmission. This set of differential equations characterizes the dynamic transformation relationships between susceptible populations, latent populations, infected populations, and recovered populations over time, reflecting the intrinsic mechanism of infectious disease transmission. Introducing this residual term allows the model to not only rely on data-driven approaches but also embed scientific priors during training, enhancing its physical consistency and extrapolation reliability.

[0062] Step S330: Construct initial condition residual terms and boundary condition residual terms, which are used to constrain the state variables of the model at the initial time and at the prediction boundary to satisfy the epidemiological prior, respectively.

[0063] In step S330, initial condition residuals and boundary condition residuals are constructed to force the model's state variables to remain consistent with the known initial epidemiological situation at the initial time and to satisfy reasonable epidemiological constraints at the boundaries of the prediction time interval. The initial condition residuals ensure that the model evolves from a realistic starting point, such as the accurate setting of the initial number of infections or latent cases; the boundary condition residuals prevent unreasonable state jumps or violations of long-term steady-state assumptions at the end of the prediction window, thereby improving the continuity, rationality, and credibility of the solution throughout the entire time domain.

[0064] Step S340: The data fitting error term, physical residual term, initial condition residual term, and boundary condition residual term are weighted and combined to form a composite loss function. The composite loss function is then minimized by an optimization algorithm, so that the model can simultaneously improve the fitting accuracy of the observed data and the degree of adherence to the epidemic transmission mechanism during the training process.

[0065] In step S340, the aforementioned errors and residuals—including data fitting error terms, physical residual terms, initial condition residual terms, and boundary condition residual terms—are linearly combined according to preset weights to form a unified composite loss function. This composite loss function comprehensively reflects the model's overall performance in terms of both data fitting ability and adherence to mechanisms. By minimizing it using optimization algorithms such as gradient descent, the model can be guided to simultaneously consider high-precision fitting of historical observation data and strict adherence to the dynamic laws of epidemic transmission during training, ultimately achieving influenza transmission trend prediction that combines data adaptability and physical interpretability.

[0066] In some embodiments of this application, the physical information neural network model is trained using a composite loss function, specifically including the following steps.

[0067] Step S410: The influenza transmission characteristics obtained by processing historical influenza surveillance data through the feature engineering module are used as training samples and input into the physical information neural network model to obtain the state variables output by the network.

[0068] In step S410, the influenza transmission characteristics formed from historical influenza surveillance data processed by the feature engineering module are used as training inputs and fed into the physical information neural network model to obtain the model's internal state estimates of the epidemic transmission process. These state variables typically include the number or proportion of susceptible, latent, infected, and recovered individuals. They are a structured expression of the current epidemic situation by the model, providing the basic output for subsequent calculations of various loss terms, and are a key link connecting data preprocessing and model training.

[0069] Step S420: Calculate the derivatives of each state variable with respect to time using automatic differentiation techniques.

[0070] In step S420, automatic differentiation is used to accurately calculate the time derivatives of each state variable output by the physical information neural network. Since epidemic dynamics are essentially described by differential equations, the time rate of change of state variables is the core criterion for judging whether the model conforms to physical laws. Automatic differentiation can efficiently obtain these derivatives without introducing numerical errors, providing a mathematical basis for subsequent substitution into the dynamic equations and the construction of physical constraints, ensuring that the dynamic evolution process of the model is differentiable and verifiable.

[0071] Step S430: Substitute the derivatives and corresponding state variables into the set of epidemiological dynamics differential equations used to describe the dynamic transformation relationship between susceptible, latent, infected and recovered populations, and calculate the physical residual term.

[0072] In step S430, the time derivatives of the state variables obtained in step S420 and their corresponding state variables are substituted into the epidemiological dynamics differential equations describing the dynamic transformation relationship between susceptible, latent, infected, and recovered populations. By comparing the difference between the model derivatives and the theoretical expression on the right side of the equations, the physical residual term is calculated. This residual term quantifies the degree to which the model's predicted trajectory deviates from the actual disease transmission mechanism, ensuring that the training process not only relies on data fitting but also forces the model to follow known epidemiological principles, thereby improving its scientific rigor and generalization ability.

[0073] Step S440: Calculate the data fitting error term based on the observed values ​​in the historical influenza surveillance data and the corresponding state variables output by the network.

[0074] In step S440, the actual number of infections or other measurable state quantities observed in historical influenza surveillance data are compared with the corresponding state variables output by the physical information neural network model. The deviation between the two is calculated to form a data fitting error term. This error term directly reflects the model's ability to reproduce real epidemic data and is a core signal in supervised learning. It ensures that the model continuously adjusts its parameters during training to approximate the actual observation results and maintain a faithful mapping to the real world.

[0075] Step S450: Calculate the initial condition residual term and the boundary condition residual term according to the preset initial flow state and boundary conditions.

[0076] In step S450, based on the initial epidemic state and the boundary conditions of the prediction time window set by prior epidemiological knowledge, the initial condition residual term and the boundary condition residual term are calculated respectively. The initial condition residual term is used to constrain the model's output at the initial time to be consistent with the known initial epidemic situation, such as the initial number of infected persons; the boundary condition residual term is used to regulate the model's behavior at the endpoints of the prediction interval, preventing inconsistencies that do not conform to the long-term trend of the epidemic or the steady-state assumption, thereby enhancing the physical rationality and continuity of the solution throughout the entire time domain.

[0077] Step S460: The data fitting error term, physical residual term, initial condition residual term, and boundary condition residual term are weighted and summed to form a composite loss value.

[0078] In step S460, the aforementioned errors and residuals—including data fitting error, physical residual, initial condition residual, and boundary condition residual—are weighted and summed according to preset weights to form a unified composite loss value. This loss value comprehensively reflects the model's overall performance in fitting observed data, adhering to dynamic laws, and satisfying initial and boundary constraints. It is the sole objective function guiding parameter optimization, enabling the model to achieve an organic integration of data-driven and mechanism-driven approaches during training.

[0079] Step S470: Based on the composite loss value, update the parameters of the physical information neural network model using a gradient descent-type optimization algorithm, and iteratively execute the above operations until the convergence condition is met, thus completing the training of the physical information neural network model.

[0080] In step S470, based on the composite loss value generated in step S460, a gradient descent-type optimization algorithm is used to backpropagate the error signal, updating all trainable parameters of the physical information neural network model layer by layer. This process is repeated from feature input to loss calculation and parameter update until the loss value converges to a stable level or reaches the preset number of iterations. This step is the core mechanism of model learning. Through continuous iterative optimization, the network can ultimately achieve high-precision fitting of historical data and strictly follow the inherent dynamics of epidemic transmission, thus fulfilling the training objective of possessing scientific explanatory and predictive capabilities.

[0081] In some embodiments of this application, a physical information neural network model based on a deep neural network and incorporating epidemiological physical knowledge is constructed. Its structure specifically includes four parts: an input layer receives multi-source data after feature engineering, including historical monitoring time series of H3N2 influenza, meteorological variables such as temperature and humidity, and exogenous variables such as seasonal characteristics. This data, after normalization and vectorization, is input into the neural network as a fixed-length feature vector; a multi-layer feedforward network serves as the main body of the model, employing multiple fully connected hidden layers, each containing linear transformations and nonlinear activation functions such as Tanh, to extract high-dimensional nonlinear propagation patterns layer by layer from the input features, fully capturing the complex temporal dynamics and exogenous characteristics during the spread of H3N2 influenza. The interaction of driving factors ensures the model's strong fitting ability to the epidemic trend. The time embedding and seasonal representation layer introduces a Fourier time embedding structure, mapping the time variable to harmonic features in sine and cosine forms, and parameterizing the transmission rate β(t) to enable the network to automatically learn and characterize the significant periodic and seasonal transmission characteristics of H3N2 influenza. The output layer is used to predict key parameters directly related to epidemic dynamics, including the transmission rate β(t), the incubation period parameter σ, the recovery rate parameter γ, and the number of infected people I(t). This layer does not use activation functions or uses positive activation functions such as softplus to ensure that all output parameters remain non-negative in a biological sense, thereby ensuring the scientific rationality and practical interpretability of the prediction results.

[0082] Furthermore, the core physical constraints of the model are derived from the SEIR propagation dynamics differential equations, specifically expressed as shown in the following formula (1): (1); Where S(t) represents the number of susceptible individuals, i.e., those who are not yet infected but may be infected; E(t) represents the number of latent individuals, i.e., those who are infected but still in the incubation period and do not yet have the ability to transmit the virus; I(t) represents the number of infected individuals, i.e., individuals who are infectious and may spread the virus; and R(t) represents the number of recovered individuals, i.e., those who have recovered from the disease and are no longer infectious. β(t) is the time-dependent transmission rate, reflecting the probability of a new infection resulting from contact between an infected person and a susceptible person per unit time; σ is the latent conversion rate, representing the proportion of individuals who become infectious after the incubation period ends; and γ is the recovery rate, representing the proportion of infected individuals who recover and enter the recovery group each day.

[0083] Furthermore, the derivative of the neural network output state variable with respect to time is calculated using automatic differentiation techniques, and substituted into the aforementioned differential equation to construct the physical residual term. During training, the total loss function... It is composed of data fitting error, differential equation residual, initial condition residual, and boundary condition residual, and is specifically expressed as shown in the following formula (2): (2); in, To monitor the fitting error of the data and measure the deviation between the model output and the actual observed values; The residuals of the SEIR differential equation reflect whether the model output satisfies the laws of epidemic dynamics. The initial condition residuals are used to ensure that the model's state at the start time is consistent with the actual epidemic situation. The boundary condition residuals ensure that the model satisfies reasonable epidemiological assumptions at the end of the prediction time interval; , , , These are the weighting coefficients for each loss term, used to balance the influence of different constraints during the optimization process.

[0084] By employing backpropagation algorithms and iterative optimization, the model accurately fits historical monitoring data while strictly adhering to biological transmission patterns in the process of minimizing the total loss function. This significantly improves the stability, extrapolation ability, and scientific interpretability of the prediction results, making it suitable for public health departments to formulate influenza outbreak early warning and control strategies.

[0085] In some embodiments of this application, the physical information neural network model may also adopt a community heterogeneity adaptive calibration mechanism, specifically including: when constructing the composite loss function, introducing a vulnerability weight factor calculated from community population structure and medical resource data. This factor dynamically adjusts the local constraint strength of the transmission rate β(t) and recovery rate γ in the SEIR equation according to the proportion of elderly population, the prevalence of chronic diseases and the number of beds per thousand people in each community, so that the model outputs differentiated transmission parameters that conform to local realities in different communities (such as aging communities or densely populated school areas), thereby improving the regional prediction accuracy and intervention adaptability.

[0086] In some embodiments of this application, after the model training is completed, the trained physical information neural network model is used to generate preliminary predictions of the spread trend of H3N2 influenza, including future case trends, epidemic intensity, potential peak time, and risk of increase. The effectiveness and reliability of the model are further systematically verified through historical backtesting, cross-validation, and error analysis, thus providing a scientific basis for the optimization and improvement of the final model. The system further determines whether the predicted trend is abnormal: if the number of infections or the positivity rate rises rapidly, exceeds historical levels, or enters a high-risk range, an early warning is triggered; otherwise, routine monitoring is initiated, and the input and prediction results are continuously updated.

[0087] In some embodiments of this application, the influenza transmission prediction process can also be coupled with an online update mechanism for the dynamic parameters of viral variants. Specifically, this includes: the system accesses the H3N2 virus antigen characteristics and gene sequencing data stream released by the National Influenza Center in real time; when the antigen drift index is detected to exceed a preset threshold, a local fine-tuning process for the transmission rate β(t) and the incubation period conversion rate σ in the physical information neural network is automatically triggered. This process only updates the network substructure parameters related to viral infectivity, without retraining the entire model, thereby maintaining the timeliness and accuracy of the prediction results in the context of viral evolution.

[0088] In some embodiments of this application, the method further includes: generating multi-level early warning signals based on influenza transmission trend prediction results and preset influenza epidemic intensity and duration thresholds. The multi-level early warning signals include low-risk alerts, medium-risk monitoring, high-risk warnings, and extremely high-risk emergency responses, used to provide tiered intervention recommendations.

[0089] Specifically, the influenza transmission trend prediction results output by the physical information neural network model are transformed into public health early warning information with practical significance. By setting preset thresholds for influenza epidemic intensity and duration, the system can quantitatively assess the predicted infection scale, transmission speed, and epidemic duration, and determine the risk level of the current or future epidemic accordingly. This transformation process realizes the mapping from abstract predicted values ​​to specific risk classifications, making the prediction results interpretable and actionable.

[0090] Moreover, the multi-level early warning signals provide clearly structured and categorized intervention recommendations. A low-risk warning indicates that influenza activity is at a normal level and no special measures are needed; a medium-risk warning indicates a potential increase in the epidemic, recommending strengthened monitoring and resource preparedness; a high-risk warning indicates that the epidemic has reached a significant epidemic level, requiring the implementation of targeted prevention and control strategies; and an extremely high-risk emergency response corresponds to a large-scale outbreak scenario, requiring immediate mobilization of emergency resources and the implementation of strict intervention measures. This tiered mechanism not only improves the precision of the early warning system but also effectively supports the scientific allocation and timely response of public health resources.

[0091] In some embodiments of this application, the influenza early warning process can also adopt an intervention-based reverse simulation optimization mechanism, specifically including: when the public health department implements interventions such as flow restriction, school closures, or booster vaccination, the system encodes the intervention type, intensity, and start and end time as external control variables and feeds them back to the physical information neural network. By solving the inverse dynamics problem, the system simulates the infection curve under the "no intervention" scenario, calculates the number of infections avoided by the actual intervention, and stores the effect evaluation result in the strategy knowledge base for dynamically adjusting the epidemic intensity threshold and response level of the next round of early warning, thereby achieving closed-loop optimization of the prevention and control strategy.

[0092] In some embodiments of this application, the method further includes: dynamically visualizing the influenza transmission trend prediction results and multi-level early warning signals in the form of time series graphs, heat maps, or geographic information system maps. The visualized content includes the predicted number of infections within a preset future time period, the trend of transmission rate changes, the distribution of regional risk levels, and the spatiotemporal evolution of early warning levels, to support an intuitive understanding of the influenza transmission situation and risk distribution.

[0093] Specifically, the system transforms influenza transmission trend predictions and multi-level early warning signals into intuitive and easily understandable visualizations to enhance public health managers and relevant decision-makers' awareness of the epidemic's development. By employing dynamic visualization techniques such as time series graphs, heat maps, or geographic information system maps, the system can present complex numerical prediction results graphically, allowing users to quickly grasp key information without delving into technical details, thereby improving decision-making efficiency and response speed.

[0094] Time series plots are used to display the predicted number of infections and the trend of transmission rate changes within a preset future time period, clearly reflecting the development path and intensity fluctuations of the epidemic over time, helping to identify potential epidemic peaks and acceleration points of transmission. Heat maps use color intensity to intuitively express the degree of risk concentration in different regions or time periods, highlighting epidemic hotspots. Geographic Information System (GIS) maps further combine warning levels with regional spatial location, presenting the distribution of regional risk levels and the spatiotemporal evolution of warning levels, enabling decision-makers to accurately locate high-risk areas and track the direction of risk spread. Overall, this visualization mechanism effectively integrates forecasting, early warning, and spatial information, providing comprehensive, dynamic, and actionable intelligence support for influenza prevention and control.

[0095] In some embodiments of this application, a complete H3N2 influenza epidemic trend prediction model is first constructed. This model can effectively infer changes in future epidemic intensity based on multi-source historical data and accurately predict potential outbreak risks. Secondly, an influenza outbreak early warning system is established based on this prediction model. When an abnormal upward trend of the epidemic is detected or an upcoming high-risk area is identified, an early warning is issued to the public health department in a timely manner, providing a scientific basis for emergency resource allocation and intervention strategy formulation. Finally, a practical data visualization tool is developed to present the prediction results intuitively through various channels such as websites, mobile applications, or public accounts, enabling decision-makers and the public to clearly understand the development trend of influenza and take corresponding countermeasures in advance, thereby realizing the integrated function of prediction, early warning, and information service.

[0096] Secondly, refer to Figure 2 This application provides an influenza prediction system based on epidemic dynamics, including a data acquisition module, a feature engineering module, and a transmission prediction module.

[0097] The data acquisition module is configured to acquire influenza surveillance data, which includes influenza time-series data, meteorological data, holiday data, and population movement data.

[0098] The feature engineering module is configured to normalize, vectorize, and time-encode influenza surveillance data to obtain influenza transmission characteristics.

[0099] The transmission prediction module is configured to input influenza transmission characteristics into a physical information neural network model and output influenza transmission trend prediction results through the physical information neural network model.

[0100] The physical information neural network model is based on a deep neural network architecture, which includes an input layer, a multi-layer feedforward network, a temporal embedding representation layer, and an output layer.

[0101] The physical information neural network model is trained using a composite loss function that combines observational data fitting with constraints from epidemic dynamics mechanisms, in order to simultaneously meet the fitting requirements of historical influenza surveillance data and the constraints of pre-set disease transmission dynamics laws.

[0102] In some embodiments of this application, the system further includes an influenza prediction and early warning module, configured to generate multi-level early warning signals based on influenza transmission trend prediction results and preset influenza epidemic intensity and duration thresholds. These multi-level early warning signals include low-risk alerts, medium-risk alerts, high-risk alerts, and extremely high-risk emergency responses, used to provide tiered intervention recommendations.

[0103] In some embodiments of this application, the system further includes an influenza early warning visualization module, configured to dynamically visualize the influenza transmission trend prediction results and multi-level early warning signals in the form of time series graphs, heat maps, or geographic information system maps. The visualized content includes a predicted curve of the number of infections within a preset future time period, the trend of transmission rate changes, the distribution of regional risk levels, and the spatiotemporal evolution of early warning levels, used to support an intuitive understanding of the influenza transmission situation and risk distribution.

[0104] Furthermore, embodiments of this application provide an electronic device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the aforementioned influenza prediction method based on epidemic dynamics.

[0105] Furthermore, embodiments of this application provide a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the aforementioned influenza prediction method.

[0106] In summary, the influenza prediction method, system, device, and medium based on epidemic dynamics provided in this application have the following technical effects.

[0107] This approach integrates multi-source heterogeneous monitoring data with epidemiological dynamics to construct a predictive framework that synergistically optimizes data-driven and physical constraints, significantly improving the accuracy, robustness, and interpretability of influenza transmission trend prediction. By introducing a physical information neural network model, it can not only automatically learn complex nonlinear transmission patterns from historical data but also embed differential equation constraints for dynamic transitions in susceptible, latent, infected, and recovered populations during training. This ensures that the prediction results strictly adhere to the basic scientific laws of disease transmission, effectively avoiding the distortion problem of purely data-driven models in extrapolation scenarios.

[0108] Meanwhile, the system uses a feature engineering module to structure and align multi-dimensional factors such as influenza time-series data, meteorology, holidays, and population movement, fully exploring the driving role of exogenous variables in the evolution of the epidemic and enhancing the model's adaptability to complex real-world environments. Based on this, a dual threshold mechanism for epidemic intensity and duration is used to generate four levels of early warning signals: low-risk alert, medium-risk attention, high-risk warning, and extremely high-risk emergency response. This achieves a closed-loop connection from prediction to decision support, providing public health departments with tiered and precise intervention guidelines.

[0109] Furthermore, by dynamically displaying the predicted infection rate curve, transmission rate trend, regional risk level distribution, and spatiotemporal evolution of warning levels through various visualization methods such as time series graphs, heat maps, and geographic information system maps, this application enables prevention and control personnel to intuitively and efficiently grasp the development trend of the epidemic and the spatial risk pattern, significantly improving the efficiency of emergency response and the scientific nature of resource allocation. Overall, this application enhances the consistency of the model's mechanisms, the practicality of its early warning, and the intuitiveness of its interaction while ensuring the accuracy of predictions, providing reliable technical support for the early warning and scientific prevention and control of respiratory infectious diseases such as influenza.

[0110] It should be noted that in all specific embodiments of this application, all data processing activities related to user identity or personal characteristics, such as user information, user behavior data, historical data, and location information, will be conducted in accordance with the principles of legality, legitimacy, and necessity. All data collection, use, storage, and processing will be subject to compliance with applicable national and regional laws, regulations, and industry standards, and informed consent from users will be obtained in a clear and explicit manner before processing. For the processing of sensitive personal information, separate consent from users will be obtained through prominent means such as pop-up prompts and independent confirmation pages. If any processing conflicts with laws and regulations, the laws and regulations will prevail, and necessary data processing will only be carried out within the scope permitted by laws and regulations, ensuring that all data-based applications, analyses, and technical implementations are conducted within the scope permitted by laws and regulations.

[0111] In some alternative embodiments, the functions / operations mentioned in the block diagrams may not occur in the order shown in the operation diagrams. For example, depending on the functions / operations involved, two consecutively shown blocks may actually be executed substantially simultaneously, or the blocks may sometimes be executed in reverse order. Furthermore, the embodiments presented and described in the flowcharts of this application are provided by way of example to provide a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logic flows presented herein. Alternative embodiments are contemplated in which the order of various operations is changed and sub-operations described as part of a larger operation are executed independently.

[0112] Furthermore, although this application is described in the context of functional modules, it should be understood that, unless otherwise stated, one or more of the functions and / or features may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in a separate physical device or software module. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary for understanding this application. Rather, given the properties, functions, and internal relationships of the various functional modules in the apparatus disclosed herein, the actual implementation of the module will be understood within the scope of ordinary skill of an engineer. Therefore, those skilled in the art can implement the application set forth in the claims using ordinary skill. It is also understood that the specific concepts disclosed are merely illustrative and are not intended to limit the scope of this application, which is determined by the full scope of the appended claims and their equivalents.

[0113] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several programs to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0114] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequential list of executable programs for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, a program execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can retrieve and execute a program from or in conjunction with such a program execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can mean any means that can contain, store, communicate, propagate, or transmit a program for use by or in conjunction with a program execution system, apparatus, or device.

[0115] More specific examples (a non-exhaustive list) of computer-readable media include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Additionally, computer-readable media can even be paper or other suitable media on which programs can be printed, for example, by optically scanning the paper or other media, then editing, interpreting, or, if necessary, processing it in a suitable manner to obtain the program electronically, and then storing it in computer memory.

[0116] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable program execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0117] In the foregoing description of this specification, the reference to terms such as "one embodiment / implementation," "another embodiment / implementation," or "certain embodiments / implementations," etc., indicates that a specific feature, structure, material, or characteristic described in connection with an embodiment or example is included in an embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0118] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

[0119] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of the present invention.

Claims

1. An influenza prediction method based on epidemic dynamics, characterized in that, Includes the following steps: Acquire influenza surveillance data, which includes influenza time-series data, meteorological data, holiday data, and population movement data; The influenza surveillance data were normalized, vectorized, and time-coded to obtain influenza transmission characteristics; The influenza transmission characteristics are input into a physical information neural network model, and the model outputs an influenza transmission trend prediction result. The physical information neural network model is based on a deep neural network architecture and includes, in sequence, an input layer, a multi-layer feedforward network, a temporal embedding representation layer, and an output layer. The physical information neural network model is trained using a composite loss function that combines observational data fitting with constraints from epidemic dynamics mechanisms, in order to simultaneously meet the fitting requirements of historical influenza surveillance data and the constraints of preset disease transmission dynamics laws.

2. The influenza prediction method based on epidemic dynamics according to claim 1, characterized in that, The process of normalizing, vectorizing, and time-coding the influenza surveillance data to obtain influenza transmission characteristics includes the following steps: The influenza time series data is subjected to time series structuring processing. By unifying the weekly series, correcting the reporting delay, and smoothing short-term fluctuations, a continuous and stable influenza epidemic time series is generated. The meteorological data, holiday data, and population flow data are time-aligned and resampled to obtain influenza aligned data. Based on the influenza aligned data, exogenous features related to influenza transmission are extracted, including seasonal harmonic features, lag variation features of temperature and humidity, holiday interference features, population mobility intensity features, previous epidemic intensity and historical peak information. All extracted features are normalized, encoded and vectorized to obtain an influenza feature vector set. The influenza feature vector set is screened and its dimensionality reduced. Redundant and noisy features are removed by using correlation analysis, principal component analysis, or feature importance assessment methods, and the influenza transmission features are retained.

3. The influenza prediction method based on epidemic dynamics according to claim 1, characterized in that, The specific structure of the physical information neural network model is as follows: The input layer is used to receive the influenza transmission characteristics as input features, and input the input features into the multilayer feedforward network in the form of a fixed-length feature vector. The multi-layer feedforward network consists of multiple hidden layers, each containing a linear transformation and a non-linear activation function, used to extract high-dimensional non-linear propagation patterns from the input features. The time embedding representation layer adopts a Fourier time embedding structure to map the time variables in the input features into harmonic features in the form of sine and cosine, and parametrically models the influenza transmission rate to capture the periodic transmission characteristics of influenza. The output layer is used to predict key parameters related to epidemic dynamics, including transmission rate, incubation period parameters, recovery rate parameters, and number of infections. The output layer does not use an activation function or uses a positive activation function to ensure that the predicted parameters satisfy biological non-negativity.

4. The influenza prediction method based on epidemic dynamics according to claim 1, characterized in that, The construction of the composite loss function specifically includes: A data fitting error term is constructed to measure the difference between the output of the physical information neural network model and historical influenza surveillance data; A physical residual term is constructed, which is calculated by substituting the derivatives of each state variable with respect to time from the output of the physical information neural network model into the system of differential equations of epidemic dynamics. This residual term is used to measure the degree of deviation of the model output from the dynamic laws of disease transmission. The system of differential equations of epidemic dynamics is used to describe the dynamic transformation relationship between susceptible, latent, infected and recovered populations. Initial condition residuals and boundary condition residuals are constructed to constrain the state variables of the model at the initial time and at the prediction boundary to satisfy the epidemiological prior, respectively. The data fitting error term, physical residual term, initial condition residual term, and boundary condition residual term are weighted and combined to form the composite loss function. The composite loss function is then minimized by an optimization algorithm, so that the model can simultaneously improve the fitting accuracy of the observed data and the degree of adherence to the epidemic transmission mechanism during the training process.

5. The influenza prediction method based on epidemic dynamics according to claim 4, characterized in that, The physical information neural network model is trained using the composite loss function, specifically including: The influenza transmission characteristics obtained by processing historical influenza surveillance data through the feature engineering module are used as training samples and input into the physical information neural network model to obtain the state variables output by the network. Using automatic differentiation techniques, the derivatives of each state variable with respect to time are calculated; Substituting the derivatives and corresponding state variables into the set of epidemiological dynamics differential equations used to describe the dynamic transformation relationship between susceptible, latent, infected and recovered populations, the physical residual term is calculated. Based on the observed values ​​in the historical influenza surveillance data and the corresponding state variables output by the network, the data fitting error term is calculated; Based on the preset initial flow state and boundary conditions, calculate the initial condition residual term and the boundary condition residual term respectively; The data fitting error term, physical residual term, initial condition residual term, and boundary condition residual term are weighted and summed to form a composite loss value. Based on the composite loss value, the parameters of the physical information neural network model are updated using a gradient descent-type optimization algorithm, and the above operations are iteratively executed until the convergence condition is met, thus completing the training of the physical information neural network model.

6. The influenza prediction method based on epidemic dynamics according to claim 1, characterized in that, The method further includes: Based on the influenza transmission trend prediction results, and using preset influenza epidemic intensity thresholds and epidemic duration thresholds, multi-level early warning signals are generated. The multi-level early warning signals include low-risk alerts, medium-risk monitoring, high-risk warnings, and extremely high-risk emergency responses, which are used to provide tiered intervention recommendations.

7. The influenza prediction method based on epidemic dynamics according to claim 6, characterized in that, The method further includes: Based on the predicted influenza transmission trend and the multi-level early warning signals, the data is dynamically visualized in the form of time series graphs, heat maps, or geographic information system maps. The visualization content includes a predicted curve of the number of infections within a preset time period, a trend of changes in the transmission rate, a distribution of regional risk levels, and the spatiotemporal evolution of warning levels, which is used to support an intuitive understanding of the influenza transmission situation and risk distribution.

8. An influenza prediction system based on epidemic dynamics, characterized in that, It includes a data acquisition module, a feature engineering module, and a propagation prediction module; The data acquisition module is configured to acquire influenza surveillance data, which includes influenza time-series data, meteorological data, holiday data, and population flow data. The feature engineering module is configured to: normalize, vectorize, and time-encode the influenza surveillance data to obtain influenza transmission characteristics; The transmission prediction module is configured to: input the influenza transmission characteristics into a physical information neural network model, and output the influenza transmission trend prediction result through the physical information neural network model; The physical information neural network model is based on a deep neural network architecture and includes, in sequence, an input layer, a multi-layer feedforward network, a temporal embedding representation layer, and an output layer. The physical information neural network model is trained using a composite loss function that combines observational data fitting with constraints from epidemic dynamics mechanisms, in order to simultaneously meet the fitting requirements of historical influenza surveillance data and the constraints of preset disease transmission dynamics laws.

9. An electronic device, characterized in that, It includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the influenza prediction method based on epidemic dynamics as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the influenza prediction method based on epidemic dynamics as described in any one of claims 1 to 7.

Images