A method and apparatus for predicting influenza

Abstract

The application relates to an influenza prediction method and device based on multi-source data fusion and deep learning, which realizes seamless connection of prediction and decision-making by acquiring multi-source heterogeneous data, constructing a spatio-temporal multi-topology dynamic graph neural network, training an influenza prediction core model, outputting a prediction result and generating a decision suggestion, has a complete technical closed loop from multi-modal perception, spatio-temporal modeling, intelligent prediction to decision support, improves prediction accuracy and robustness through cooperation of a dynamic multi-topology graph network, decision-driven learning and a rule engine, and ensures that the prediction result can be automatically and explainably converted into a decision suggestion with actual operation value, and systematically solves the core pain point of disconnection between prediction and decision-making.

Description

Technical Field

[0001] This application relates to the field of public health technology, and more specifically, to an influenza prediction method and device based on multi-source data fusion and deep learning. Background Technology

[0002] As a recurring respiratory infectious disease worldwide, accurate prediction of influenza is crucial for the scientific allocation of public health resources and the effective implementation of epidemic prevention and control strategies.

[0003] Current influenza prediction techniques mainly rely on traditional statistical models, but these techniques generally tend to prioritize accuracy over application. Traditional statistical models, such as the autoregressive integral moving average model and its seasonal variants, are limited by linear assumptions and stationarity requirements, making it difficult to capture the highly nonlinear relationship between meteorological factors and influenza transmission.

[0004] To address the aforementioned issues, existing technologies urgently need improvement. Summary of the Invention

[0005] This application provides an influenza prediction method and apparatus based on multi-source data fusion and deep learning, which improves prediction accuracy and robustness.

[0006] To achieve the above objectives, embodiments of this application provide an influenza prediction method based on multi-source data fusion and deep learning, which includes: Acquire multi-source heterogeneous data of multiple geographic units within the target area; the multi-source heterogeneous data includes influenza surveillance time-series data, multi-dimensional meteorological time-series data, static geographic attribute data, cross-regional population flow data, and Internet public behavior time-series data; Based on the static geographic attribute data, the multi-dimensional meteorological time series data, and the cross-regional population flow data, a geographic adjacency matrix, a meteorological similarity matrix, and a functional association matrix are determined. Based on the learnable dynamic fusion results of the geographic adjacency matrix, meteorological similarity matrix, and functional association matrix, a spatiotemporal multi-topology dynamic graph neural network is constructed, and spatial dependency modeling is performed using the spatiotemporal multi-topology dynamic graph neural network to obtain spatial enhancement features. Based on the historical data in the multi-source heterogeneous data and the spatial enhancement features, a trained influenza prediction core model is obtained. The influenza prediction core model includes a multi-scale feature extraction network and a meta-learning adapter. The trained influenza prediction core model processes real-time data from the multi-source heterogeneous data and outputs prediction results including point prediction and interval prediction of influenza incidence rates for future multiple periods. Based on a preset rule engine, dynamic risk assessment levels and corresponding prevention and control decision suggestions are generated according to the prediction results and real-time external context data.

[0007] This application also provides an influenza prediction device based on multi-source data fusion and deep learning, which includes: The acquisition module is used to acquire multi-source heterogeneous data of multiple geographic units within the target area; the multi-source heterogeneous data includes influenza surveillance time-series data, multi-dimensional meteorological time-series data, static geographic attribute data, cross-regional population flow data, and Internet public behavior time-series data; The association module is used to determine the geographic adjacency matrix, meteorological similarity matrix, and functional association matrix based on the static geographic attribute data, the multi-dimensional meteorological time series data, and the cross-regional population flow data. The enhancement module is used to construct a spatiotemporal multi-topology dynamic graph neural network based on the learnable dynamic fusion results of the geographic adjacency matrix, meteorological similarity matrix and functional association matrix, and to use the spatiotemporal multi-topology dynamic graph neural network to model spatial dependencies and obtain spatial enhancement features. The training module is used to obtain a trained influenza prediction core model based on historical data in the multi-source heterogeneous data and the spatial enhancement features. The influenza prediction core model includes a multi-scale feature extraction network and a meta-learning adapter. The prediction module is used to process real-time data in the multi-source heterogeneous data through the trained influenza prediction core model, and output prediction results including point prediction and interval prediction of influenza incidence rate in the future multiple periods. The output module is used to generate dynamic risk assessment levels and corresponding prevention and control decision suggestions based on the prediction results and real-time external context data, using a preset rule engine.

[0008] This application provides an influenza prediction method and apparatus based on multi-source data fusion and deep learning. By acquiring multi-source heterogeneous data, constructing a spatiotemporal multi-topology dynamic graph neural network, training a core influenza prediction model, outputting prediction results, and generating decision suggestions, it achieves seamless integration of prediction and decision-making. It has a complete technical closed loop from multimodal perception, spatiotemporal modeling, intelligent prediction to decision support. Through the synergy of dynamic multi-topology graph networks, decision-driven learning, and rule engines, it improves prediction accuracy and robustness, and ensures that prediction results can be automatically and interpretably transformed into decision suggestions with practical operational value, systematically solving the core pain point of the disconnect between prediction and decision-making.

[0009] Other features and advantages of this application will be described in detail in the following detailed description section. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] To gain a more complete understanding of this application and its beneficial effects, the following description will be provided in conjunction with the accompanying drawings, wherein the same reference numerals in the following description denote the same parts.

[0012] Figure 1 This is a schematic flowchart of an influenza prediction method provided in an embodiment of this application; Figure 2 This is a schematic diagram of an influenza prediction device provided in an embodiment of this application. Detailed Implementation

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the protection scope of this application.

[0014] In the field of influenza prediction technology, existing methods mainly focus on improving the statistical accuracy of prediction models. However, there are significant obstacles to the connection between prediction results and public health decision-making. Specifically, the influenza incidence rate values ​​output by prediction models lack an effective integration mechanism with contextual factors such as real-time medical resource status, population vulnerability indicators, and current prevention and control measures. This results in prediction results not being directly transformed into actionable decision-making recommendations. This disconnect between prediction and decision-making makes it difficult for public health agencies to adjust resource allocation strategies and early warning response levels in a timely manner based on prediction information, thus affecting the timeliness and accuracy of epidemic prevention and control. In particular, prediction models only provide statistically significant incidence rate values ​​without establishing a dynamic correlation with medical resource occupancy rates, high-risk population distribution, and the intensity of prevention and control measures. This prevents decision-makers from quantitatively assessing the actual risk level of the prediction results in specific contexts.

[0015] To address the aforementioned problems, this application provides an influenza prediction method based on multi-source data fusion and deep learning, such as... Figure 1 As shown, the method includes the following steps: S101: Obtain multi-source heterogeneous data of multiple geographic units within the target area.

[0016] This multi-source heterogeneous data can include influenza surveillance time-series data, multi-dimensional meteorological time-series data, static geographic attribute data, cross-regional population flow data, and internet public behavior time-series data. For example, influenza surveillance time-series data can consist of influenza-like illness (ILI) data regularly reported by disease control and prevention centers at all levels; multi-dimensional meteorological time-series data can be obtained from weather stations or satellites, including indicators such as temperature, humidity, air pressure, and wind speed; static geographic attribute data can include population density, area, and number of medical institutions in geographic units; cross-regional population flow data can be obtained from operator signaling data or transportation department data; and internet public behavior time-series data can come from keyword search volume on social media platforms or search engines. This data can be collected in batches from different data sources through manual entry, file import, or application programming interfaces (APIs).

[0017] S102: Based on the static geographic attribute data, the multi-dimensional meteorological time series data, and the cross-regional population flow data, determine the geographic adjacency matrix, the meteorological similarity matrix, and the functional association matrix.

[0018] Specifically, a geographic adjacency matrix can be constructed by calculating the Euclidean distance between geographic units or determining whether they share boundaries. For example, a distance threshold can be set; if the distance between the centroids of two geographic units is less than this threshold, they are considered geographically adjacent. A meteorological similarity matrix can be generated by comparing the statistical differences in meteorological indicators such as average temperature and average humidity between different geographic units over a period of time. For example, the similarity can be measured by calculating the Pearson correlation coefficient or the dynamic time warping (DTW) distance between the meteorological time series of two geographic units. A functional association matrix can be constructed by analyzing cross-regional population flow data; for example, if the population flow between two geographic units exceeds a preset threshold, they are considered to have a functional association.

[0019] S103: Based on the learnable dynamic fusion results of the geographic adjacency matrix, meteorological similarity matrix and functional correlation matrix, a spatiotemporal multi-topology dynamic graph neural network is constructed, and spatial dependency modeling is performed using the spatiotemporal multi-topology dynamic graph neural network to obtain spatial enhancement features.

[0020] Specifically, the geographic adjacency matrix, meteorological similarity matrix, and functional association matrix can be viewed as different graph topologies. These topologies can be fused using a simple weighted summation method, for example, assigning a fixed weight to each matrix and then summing them to form a composite graph. This composite graph can be used as input to a graph neural network, aggregating node information through graph convolution operations to capture spatial dependencies. Thus, the feature representation of each geographic unit will incorporate information from its neighboring geographic units and meteorologically similar or functionally associated geographic units, forming spatially enhanced features.

[0021] S104: Based on the historical data in this multi-source heterogeneous data and the spatial enhancement features, obtain the trained influenza prediction core model.

[0022] The core influenza prediction model comprises a multi-scale feature extraction network and a meta-learning adapter. Specifically, the multi-scale feature extraction network can consist of multiple parallel convolutional layers, each with a different receptive field to capture features at different time scales. For example, one convolutional layer might focus on short-term fluctuations, while another might focus on medium-term trends. The meta-learning adapter can be a small feedforward neural network that receives a small amount of new task data as input and outputs a bias or scaling factor to adjust the core model parameters. The core model can be trained using traditional supervised learning methods, such as using historical influenza incidence data as labels, to optimize model parameters by minimizing prediction error.

[0023] S105: The trained influenza prediction core model processes the real-time data in the multi-source heterogeneous data and outputs prediction results including point predictions and interval predictions of influenza incidence rates for future multiple periods.

[0024] Specifically, real-time data can include the latest influenza surveillance data, meteorological data, population movement data, and so on. This real-time data is input into the trained influenza prediction core model, and after processing by a multi-scale feature extraction network and a meta-learning adapter, it generates influenza incidence rate predictions for multiple future time periods (e.g., the next week, two weeks, or four weeks). Simultaneously, the model can also provide a confidence interval for each predicted value to quantify the uncertainty of the prediction through methods such as Monte Carlo sampling or Bayesian inference.

[0025] S106: Based on a preset rule engine, dynamic risk assessment levels and corresponding prevention and control decision recommendations are generated according to the prediction results and real-time external context data.

[0026] Specifically, the rule engine can pre-store a series of IF-THEN rules defined by public health experts. For example, a rule could be "If the predicted influenza incidence rate exceeds X% and the medical resource occupancy rate exceeds Y% in the coming week, the risk level is high." Real-time external contextual data can include current hospital bed occupancy rates, vaccine inventory, and healthcare worker availability. The rule engine matches the prediction results (including point predictions and interval predictions) with this real-time contextual data, triggering the corresponding rules to output a dynamic risk assessment level (e.g., low, medium, high, very high) and a series of clear prevention and control decision recommendations, such as "It is recommended to activate the regional joint prevention and control mechanism" or "It is recommended to increase vaccine reserves."

[0027] For ease of understanding, the following explains some key terms in this embodiment: A geographic unit refers to an analytical unit within a target area that has clearly defined geographic boundaries and administrative divisions, such as a city, district, or street. These units are the basic spatial nodes for influenza transmission modeling and prediction.

[0028] Multi-source heterogeneous data refers to various types of data collected from different sources and in different formats. It encompasses influenza surveillance time-series data, multi-dimensional meteorological time-series data, static geographic attribute data, cross-regional population flow data, and internet public behavior time-series data. These data collectively provide comprehensive information input for understanding the mechanisms of influenza transmission.

[0029] A geographic adjacency matrix is ​​a matrix that represents the spatial proximity relationships between geographic units. The elements in the matrix typically indicate whether two geographic units are directly adjacent or within a certain distance.

[0030] A meteorological similarity matrix is ​​a matrix that quantifies the degree of similarity in meteorological conditions between different geographical units. This matrix can be constructed based on the statistical characteristics of multiple meteorological indicators (such as temperature, humidity, and air pressure) over a specific time period.

[0031] A functional association matrix is ​​a matrix that reflects functional connections between geographical units that are not geographically adjacent. For example, two geographical units may have personnel exchanges due to commuting, tourism, or economic activities, even if they are not geographically adjacent.

[0032] Spatiotemporal multi-topology dynamic graph neural networks are deep learning models capable of simultaneously processing time-series data and graph-structured data, and dynamically adjusting the graph topology. This network can capture the temporal and spatial dependencies of influenza transmission and adapt to dynamic changes in transmission paths.

[0033] Spatial augmentation features refer to feature representations containing spatial context information extracted after modeling the spatial dependencies of the original data using a spatiotemporal multi-topological dynamic graph neural network. These features can effectively reflect the transmission patterns of influenza across different geographical units.

[0034] The influenza prediction core model is the core predictive component of this method, responsible for learning influenza transmission patterns from multi-source heterogeneous data and making predictions. This model includes a multi-scale feature extraction network and a meta-learning adapter to address the multi-scale characteristics of influenza transmission and variations in data distribution.

[0035] Multiscale feature extraction networks are a component of the core influenza prediction model, designed to capture features from input data at different time granularities, such as short-term fluctuations, medium-term trends, and long-term cycles.

[0036] The meta-learning adapter is another component of the core influenza prediction model, giving the model the ability to quickly adapt to new tasks or environments. This adapter can adjust the model based on small amounts of new data to cope with scenarios with limited data or data distribution shifts.

[0037] A rules engine is a software component that performs logical reasoning and decision-making based on preset rules. It receives prediction results and real-time external contextual data as input, and generates risk assessment levels and decision recommendations according to the defined IF-THEN rules.

[0038] Real-time external contextual data refers to external information related to the development and control of the epidemic that is acquired in real time during the process of influenza prediction and decision support, such as the status of medical resources, population vulnerability indicators, and current prevention and control measures.

[0039] The following example will provide a more detailed explanation of the above technical solution: Considering the risk of influenza outbreaks at location A during the winter, the complex meteorological conditions and frequent population movement in the area pose challenges to influenza forecasting and decision support.

[0040] First, the system acquires multi-source heterogeneous data from multiple geographic units within Location A. This includes influenza-like illness (ILI) surveillance data from the local CDC, meteorological time-series data such as temperature, humidity, and air pressure provided by weather stations, static geographic attribute data such as population density and distribution of medical institutions, and cross-regional population flow data obtained through mobile phone signaling data analysis. In addition, the system also acquires search volume data for keywords related to influenza symptoms from internet search engines as public behavior time-series data. All this data is collected and preprocessed to ensure data quality and format consistency.

[0041] Next, based on this data, the system determined the geographic adjacency matrix, meteorological similarity matrix, and functional association matrix among the geographic units within location A. For example, the geographic adjacency matrix is ​​determined by calculating the distance between the center points of each geographic unit; units with a distance less than a certain threshold are considered adjacent. The meteorological similarity matrix is ​​constructed by comparing the average temperature and humidity change trends of each unit over a past period; units with similar meteorological patterns are assigned a higher similarity. The functional association matrix, based on cross-regional population flow data, identifies units that, even if not geographically adjacent, pose a risk of transmission due to close population interaction.

[0042] Subsequently, based on the learnable dynamic fusion results of these matrices, the system constructs and utilizes a spatiotemporal multi-topology dynamic graph neural network to model spatial dependencies. This network can dynamically adjust the connection strength between geographic units. For example, in a sustained low-temperature and low-humidity environment, even geographically distant units will have increased connection weights in the meteorological similarity matrix if their meteorological conditions are similar, thus reflecting correlation in the fused graph structure. Through the aggregation operation of the graph neural network, the features of each geographic unit not only include its own historical data but also incorporate information from its geographically neighboring, meteorologically similar, and functionally related units, thereby generating spatially enhanced features. These features can capture the spread patterns of influenza across different spatial dimensions, for example, identifying the risk of regional synchronous outbreaks driven by specific meteorological conditions.

[0043] Building upon this foundation, the system acquires a trained core influenza prediction model based on historical data from multi-source heterogeneous datasets and these spatially enhanced features. This core model comprises a multi-scale feature extraction network and a meta-learning adapter. The multi-scale feature extraction network captures short-term fluctuations in influenza transmission (e.g., daily case count changes), medium-term trends (e.g., weekly increases or decreases in incidence rates), and long-term seasonal cycles from historical data. The meta-learning adapter enables the model to quickly adapt to the unique influenza patterns of different geographical units within location A, allowing for rapid adjustments even in units with limited historical data, based on experience learned from data-rich units.

[0044] When new real-time data (such as the latest ILI monitoring data, weather forecasts, and population movement updates) is input, the trained influenza prediction core model processes this data immediately. The model outputs point predictions of influenza incidence rates for future multiple periods (e.g., the next 7 days, 14 days, and 30 days), while also providing corresponding interval predictions to quantify the uncertainty of the prediction. For example, the model predicts an ILI rate of 3% for a geographic unit at location A over the next 7 days, with a confidence interval of 2.5% to 3.5%.

[0045] Finally, based on a pre-set rule engine, the system generates dynamic risk assessment levels and corresponding prevention and control decision recommendations according to these prediction results and real-time external contextual data. For example, if the model predicts that the upper limit of the confidence interval for the ILI rate in the next 7 days reaches 4% (higher than the pre-set threshold), and real-time external contextual data shows that the hospital bed occupancy rate in that geographical unit has exceeded 80%, the rule engine will determine the risk level of the area as "high" and automatically generate a decision recommendation to "activate the emergency plan, increase the reserve of medical supplies, and issue a health warning to the public." This process is dynamic; as new predictions and contextual data are continuously updated, the risk assessment and decision recommendations will also be adjusted in real time.

[0046] Through the above process, this method can comprehensively perceive the influenza transmission trend at location A, not only providing predictions, but more importantly, transforming the prediction results into actionable decision-making suggestions, effectively solving the disconnect between "prediction and decision-making" in existing influenza prediction methods.

[0047] Based on the influenza prediction example at location A above, this embodiment includes at least the following beneficial effects: Firstly, regarding data acquisition and feature construction, this method constructs an information input by integrating multi-source heterogeneous data, including influenza surveillance time-series data, multi-dimensional meteorological time-series data, static geographic attribute data, cross-regional population flow data, and internet public behavior time-series data. Compared to existing technologies that often rely on a single data source or simply piece together multiple data sources, this method can comprehensively capture the driving factors of influenza transmission. For example, in the example of location A, internet public behavior data provides early warning signals, while cross-regional population flow data reveals the transmission path—features that traditional monitoring data cannot provide.

[0048] Secondly, regarding spatial dependency modeling, this method constructs a spatiotemporal multi-topological dynamic graph neural network by dynamically fusing geographic adjacency matrices, meteorological similarity matrices, and functional correlation matrices. This network is capable of modeling spatial dependencies and obtaining spatially enhanced features. This addresses the problem of insufficient dynamic modeling of spatial transmission paths in existing influenza prediction methods. Traditional graph neural networks typically only consider geographic proximity, making it difficult to capture the moderating effect of meteorological conditions on influenza transmission paths. For example, in the example of location A, even if two geographic units are not geographically adjacent, if their winter meteorological conditions are similar, this method can identify the risk of synchronous outbreaks between them driven by the environment, thereby generating accurate spatially enhanced features, which is impossible with existing technologies.

[0049] Furthermore, the core influenza prediction model obtained by this method includes a multi-scale feature extraction network and a meta-learning adapter, which can process real-time data from multi-source heterogeneous datasets and output prediction results including point predictions and interval predictions of influenza incidence rates over multiple future periods. This effectively solves the problems of existing models' insufficient ability to predict influenza peaks and high-frequency fluctuations, as well as the potential for smoothing due to model learning strategies. For example, in the example of location A, the multi-scale feature extraction network can simultaneously capture both short-term fluctuations and long-term trends in influenza transmission, while the meta-learning adapter enables the model to quickly adapt to the unique influenza patterns of different geographical units within location A, achieving effective predictions even in areas with limited data, thus avoiding the problems of poor prediction results or smoothing in traditional models when facing events.

[0050] Finally, this method, based on a pre-defined rule engine, generates dynamic risk assessment levels and corresponding prevention and control decision recommendations based on the prediction results and real-time external contextual data. This directly solves the problem of the disconnect between prediction and decision-making in existing influenza prediction methods. The numerical information output by existing prediction models often requires manual interpretation to form decisions, which is inefficient and lacks real-time performance. In the example of location A, the rule engine can combine the predicted ILI rate and its confidence interval with contextual data such as real-time medical resource occupancy rates to automatically generate clear and actionable decision recommendations, such as "activate the emergency plan" or "increase medical supply reserves." This automatic generation and context-relevant decision support capability improves the efficiency and scientific nature of public health decision-making, ensuring that prediction results can be automatically and interpretably transformed into decision recommendations with practical operational value.

[0051] In some embodiments, the step of constructing a spatiotemporal multi-topology dynamic graph neural network includes: identifying a set of geographical units with similar meteorological characteristics based on the multi-dimensional meteorological time-series data to form meteorological clusters; constructing a meteorological hypergraph based on the meteorological clusters, wherein each meteorological cluster serves as a hyperedge connecting all nodes within the cluster; aggregating environmental risk information of nodes within the hyperedges through a hypergraph convolutional network, learning cluster-based environmental-driven propagation characteristics, and constructing the spatiotemporal multi-topology dynamic graph neural network.

[0052] Specifically, identifying sets of geographical units with similar meteorological characteristics to form meteorological clusters aims to discover geographical units with similar meteorological patterns from multi-dimensional meteorological time-series data and group them together to form "meteorological clusters." The purpose is to identify areas that may face similar influenza risks due to shared meteorological conditions, laying the foundation for subsequent cluster-based risk modeling. This identification process can employ clustering algorithms, such as K-means, DBSCAN, or hierarchical clustering, to extract features and measure similarity from the multi-dimensional meteorological time-series data (such as temperature, humidity, wind speed, precipitation, etc.) of each geographical unit, and then group geographical units with similar meteorological feature vectors into one category. Alternatively, identification can be performed by setting meteorological threshold rules. For example, when multiple geographical units have their average temperature, relative humidity, and other key meteorological indicators falling within a preset similarity range over a continuous period, these geographical units are classified into a meteorological cluster.

[0053] A meteorological hypergraph is constructed based on the meteorological clusters, where each meteorological cluster serves as a hyperedge connecting all nodes within the cluster. This step aims to build a graph structure capable of representing many-to-many relationships. Unlike traditional graphs where edges connect two nodes, hyperedges in a hypergraph can connect any number of nodes. Here, each meteorological cluster is abstracted as a hyperedge, connecting all geographic units (nodes) within the cluster that share similar meteorological characteristics. This structure can explicitly model the combined effects of meteorological conditions on multiple geographic units, overcoming the limitation of traditional graph models that can only represent pairwise relationships. At the data structure level, a unique hyperedge identifier can be created for each meteorological cluster, and a list containing the identifiers of all geographic unit nodes belonging to that hyperedge can be maintained. The construction of the hypergraph involves defining these hyperedges and the set of nodes they contain. Alternatively, the hypergraph can be represented using an extended form of an adjacency matrix. For example, a bipartite graph can be constructed, where one type of node represents geographic units, and the other type represents meteorological clusters (hyperedges). When a geographic unit belongs to a meteorological cluster, a connection is established between the two.

[0054] By aggregating environmental risk information of nodes within hyperedges using a hypergraph convolutional network (HGCN), and learning clustered environment-driven propagation characteristics, the spatiotemporal multi-topology dynamic graph neural network is constructed. This step utilizes a HGCN for information propagation and feature learning on the hypergraph structure. By applying a HGCN to a meteorological hypergraph, the model can aggregate environmental risk information of all geographic units within the same hyperedge (i.e., the same meteorological cluster), thereby learning clustered influenza propagation characteristics driven by similar meteorological conditions and affecting multiple geographic units. Finally, this clustered propagation characteristic is integrated into the overall spatiotemporal multi-topology dynamic graph neural network, enhancing its ability to model complex spatial dependencies. A message-passing mechanism-based hypergraph convolutional layer can be used, where each node receives aggregated feature information from its superedge, while each superedge also aggregates feature information from its connected nodes. This bidirectional information passing helps capture complex relationships within the hyperedge. Alternatively, the hypergraph can be converted into a traditional bipartite graph, and then standard graph convolutional networks (GCN) or graph attention networks (GAT) can be applied for feature aggregation. In this transformation, connections are established between geographic unit nodes and hyperedge nodes. GCN / GAT can perform convolution operations on this bipartite graph to aggregate information of nodes within the hyperedge.

[0055] The following is a concrete example to illustrate this. Assume there are five geographical units within a target area: city A, city B, city C, city D, and city E. The system first acquires multi-dimensional meteorological time-series data for these cities, such as the average daily temperature, relative humidity, and wind speed over the past week. By analyzing this meteorological data, the system identifies that cities A, B, and C experienced sustained low temperatures and high humidity over a certain period, while cities D and E experienced relatively higher temperatures. At this point, the system identifies cities A, B, and C as a meteorological cluster. Based on this meteorological cluster, the system constructs a meteorological hypergraph, where cities A, B, and C are nodes connected by a hyperedge. This hyperedge represents the common meteorological condition of "sustained low temperatures and high humidity." When the influenza prediction core model runs, the hypergraph convolutional network aggregates the current environmental risk information of cities A, B, and C, such as their respective influenza incidence rates, population densities, and medical resource status. Through hypergraph convolution operations, the network can learn that under the common meteorological conditions of "sustained low temperature and high humidity," the influenza transmission patterns in cities A, B, and C may show a synchronous upward trend. This cluster-like, environment-driven transmission characteristic is then incorporated into the overall spatiotemporal multi-topology dynamic graph neural network to more accurately predict future influenza incidence rates in these cities.

[0056] By introducing a meteorological hypergraph structure, this approach overcomes the limitation of traditional graph neural networks, which can only model one-to-one propagation relationships between geographic units. This hypergraph structure can explicitly learn and characterize the clustered impacts of regional environmental events (such as widespread cold waves or sustained high temperatures) on multiple geographic units. This enables the model to provide earlier warnings and identify areas that may experience synchronous outbreaks due to shared meteorological conditions. This provides crucial technical support for public health departments to implement regional joint prevention and control strategies, effectively solving the problem that traditional methods struggle to characterize the synchronous, clustered health risk impacts of large-scale similar meteorological conditions on multiple geographic units.

[0057] In some embodiments, this application further includes: real-time monitoring of environmental risk characteristic values ​​of nodes within each superedge of the meteorological supermap; when a significant increase in the risk characteristics of nodes exceeding a threshold proportion within a meteorological cluster is identified, generating a regional environmental risk early warning signal; and fusing the early warning signal with individual city prediction results to conduct a risk reassessment and generate decision recommendations including cluster outbreak risks.

[0058] The real-time monitoring of environmental risk characteristic values ​​of nodes within each hyperedge of the meteorological hypermap refers to the continuous acquisition and observation of the numerical changes in the environmental risk characteristics associated with each geographic unit (node) in the meteorological hypermap (a hypermap composed of meteorological clusters connecting geographic unit nodes). Its purpose is to provide foundational data for subsequent identification of cluster risks. For example, the predicted incidence rate, incidence rate trend, or risk index related to meteorological anomalies for each geographic unit can be extracted periodically from the core influenza prediction model as environmental risk characteristic values. Alternatively, the real-time environmental risk score for each geographic unit can be calculated by analyzing the abnormal fluctuations in multi-dimensional meteorological time-series data (such as temperature, humidity, air pressure, and air pollution) of each geographic unit in real time, combined with the correlation between the influenza incidence rate in historical data and these meteorological factors.

[0059] When a significant and synchronous increase in risk characteristics is detected in more than a threshold proportion of nodes within a meteorological cluster, a regional environmental risk warning signal is generated. This is the core mechanism for identifying cluster risks. It uses a set criterion (threshold proportion, significant synchronous increase) to detect whether multiple geographical units within a meteorological cluster (i.e., a hyperedge) simultaneously experience a sharp increase in risk characteristics. Once the conditions are met, a clear warning is issued. Its function is to aggregate dispersed individual risk signals into a regional overall risk signal. For example, a fixed threshold can be set; if more than 70% of the geographical units within a meteorological cluster have environmental risk characteristic values ​​(such as predicted incidence rate growth rate) exceeding two standard deviations of the historical average growth rate simultaneously in the past 24 hours, a warning is triggered. Alternatively, a dynamic threshold can be used, based on historical data to statistically determine the baseline probability of synchronous increases in risk characteristics for each node within the meteorological cluster. When the probability of a synchronous increase event detected in real time is significantly higher than the baseline probability (e.g., P-value less than 0.01), a warning signal is generated.

[0060] By integrating the aforementioned early warning signals with individual city forecasts, a risk reassessment is conducted to generate decision-making recommendations, including those addressing cluster outbreak risks. This step combines regional early warnings with individual forecasts to create more comprehensive and instructive decision-making recommendations. It integrates macro-level regional risk information with micro-level city forecast information, correcting and supplementing risks, and transforming them into specific action recommendations. For example, regional environmental risk early warning signals can be used as a high-priority input to the rule engine. When an early warning signal is triggered, the rule engine will increase the risk assessment level of all geographical units within that meteorological cluster and automatically generate joint prevention and control recommendations for the region, such as "It is recommended that all cities within this meteorological cluster activate regional joint prevention and control mechanisms, strengthen medical resource reserves, and conduct joint health education." Alternatively, a risk correction module can be added after the output layer of the influenza prediction core model. When a regional environmental risk warning signal is generated, this module will adjust the predicted incidence rate of each geographical unit in the region upward based on the intensity and scope of the warning. Combined with the adjusted prediction results, the module will generate more targeted decision-making suggestions through the rule engine, such as "It is recommended that all cities in the region strengthen the monitoring of respiratory diseases and consider activating emergency plans."

[0061] The following is a specific example to illustrate this. As a concrete implementation method, a meteorological cluster can be defined, for example, including cities A, B, and C, which are frequently affected by the same cold wave in winter. The system continuously monitors the environmental risk characteristic values ​​of these three cities. These characteristic values ​​can specifically be the predicted influenza incidence rate for the next week output by the influenza prediction core model. When it is detected that within three consecutive days, the predicted incidence rate in at least two of cities A, B, and C simultaneously increases by more than 30% week-on-week, and their growth rates are both higher than two standard deviations above the historical average for the same period, the system generates a regional environmental risk warning signal for that meteorological cluster. This warning signal is then input into the rule engine for comprehensive judgment along with the individual influenza prediction results for cities A, B, and C. Based on the triggering of the warning signal, the rule engine can automatically upgrade the risk assessment level of these three cities from "medium" to "high" and generate specific decision-making recommendations, such as "It is recommended that cities A, B, and C immediately activate the regional joint prevention and control emergency plan, strengthen the coordination of cross-city medical resource allocation, and jointly carry out winter influenza prevention and control health education activities."

[0062] Through the aforementioned technical solution, this application can proactively detect and identify the risk of regional influenza outbreaks driven by shared environmental factors by utilizing the topological structure of meteorological supermaps. This mechanism overcomes the limitations of traditional forecasting based on individual cities, achieving an upgrade in prevention and control perspective from "city-level response" to "regional early warning." It enables public health departments to obtain collaborative risk information across administrative boundaries earlier and more comprehensively, thus providing crucial technical support for formulating more macro-level and forward-looking regional joint prevention and control strategies, significantly improving the overall efficiency and response capability of influenza prevention and control.

[0063] In some embodiments, the trained influenza prediction core model processes real-time data from the multi-source heterogeneous data and outputs prediction results including point predictions and interval predictions of influenza incidence rates for future multiple periods. Specific steps include: for the current prediction task, constructing an anti-smoothing support set from historical data based on the fluctuating influenza values ​​and meteorological anomalies of the samples; encoding the statistical features of the support set samples to generate a task embedding vector; dynamically generating a set of feature modulation parameters based on the task embedding vector using a meta-learning adapter; and adaptively adjusting the backbone network features of the influenza prediction core model using the modulation parameters through feature linear modulation technology.

[0064] In this context, for the current prediction task, an anti-smoothing support set is constructed from historical data based on the fluctuation and meteorological anomalies of the samples. The aim is to select a subset of samples from a large amount of historical data that share similar or representative characteristics with the current prediction task in terms of flu fluctuations and meteorological conditions. This helps the model quickly focus on the most relevant historical experience when facing new scenarios, rather than being distracted by a large amount of irrelevant historical data. One possible implementation is to calculate the week-on-week change rate of flu incidence rates in historical samples as an indicator of flu fluctuation, and to calculate the standard deviation of meteorological data (e.g., temperature, humidity, air pressure) from historical averages as an indicator of meteorological anomalies. Then, a threshold is set or a clustering algorithm is used to select samples with high fluctuation and large meteorological anomalies to form the support set. Another implementation can use methods based on information entropy or KL divergence to evaluate the differences between historical samples and the current prediction task in the distribution of flu incidence rates and meteorological characteristics, selecting samples with significant differences but still representative—the "anti-smoothing" samples—to capture potential anomalies or changing trends.

[0065] The statistical features of the support set samples are encoded to generate a task embedding vector, which is a compact, low-dimensional representation of the characteristics of the current prediction task. By encoding the statistical features of the support set samples, this vector can capture unique contextual information of the current task, such as the intensity of an influenza outbreak or the extreme nature of weather conditions, thereby guiding the meta-learning adapter to adjust the model. One possible implementation is to perform statistical analysis on the influenza incidence rate, meteorological characteristics (e.g., average temperature, humidity, precipitation), and population flow data of the support set samples, extracting statistics such as mean, variance, kurtosis, and skewness. These statistics are then concatenated into a vector, which is then encoded through a small neural network (e.g., a fully connected layer) to generate the task embedding vector. Alternatively, an autoencoder or variational autoencoder (VAE) can be used to learn the original features of the support set samples, compressing them into a vector in a latent space. This vector is the task embedding vector, effectively capturing the key information of the support set.

[0066] A meta-learning adapter dynamically generates a set of feature modulation parameters based on the task embedding vector. This meta-learning adapter is a key component enabling the model to quickly adapt to new tasks. It receives the task embedding vector as input and outputs a set of feature modulation parameters. These parameters are used to adjust the feature representation within the backbone network of the influenza prediction core model to better adapt it to the specific distribution or pattern of the current prediction task. As one possible implementation, the meta-learning adapter can be a small feedforward neural network whose input is the task embedding vector and output is multiple scalars or vectors. These outputs are directly used as feature modulation parameters, such as scaling factors and shifts, for subsequent linear feature modulation. Alternatively, a more complex meta-learning architecture, such as a variant of MAML (Model-Agnostic Meta-Learning) or Reptile, can be employed. In this implementation, the meta-learning adapter learns how to generate or adjust the initialization parameters of the backbone network, or generate gradient directions for updating the backbone network parameters, thereby achieving finer modulation.

[0067] The feature linear modulation technique is used to adaptively adjust the backbone network features of the influenza prediction core model using the modulation parameters. This technique is a lightweight yet effective model adaptation mechanism. It utilizes modulation parameters generated by a meta-learning adapter to linearly transform (e.g., scale and shift) the features extracted from different layers or stages of the backbone network, thereby dynamically adapting its feature representation to the characteristics of the current task without altering the main structure of the backbone network. One possible implementation is to introduce a batch normalization layer after each convolutional or fully connected layer of the backbone network, replacing the learnable parameters (gamma and beta) in the batch normalization layer with modulation parameters to achieve feature scaling and shifting. Alternatively, element-wise multiplication (scaling) and addition (shifting) operations can be directly applied to specific feature maps of the backbone network, where the multiplication factor and addition bias are the modulation parameters generated by the meta-learning adapter. This approach allows for more flexible control over the modulation position and intensity.

[0068] The following example illustrates this. Suppose we need to predict the impact of a newly discovered influenza outbreak in a city, but historical data for that city is limited. First, the system filters historical influenza data nationwide, selecting regions and time periods similar to the current outbreak intensity and meteorological conditions (e.g., sudden temperature drop, abnormal humidity) in that city, constructing an anti-smoothing support set. For example, if the current influenza incidence rate in the city is increasing by 20% week-on-week, and the average temperature is 8 degrees Celsius lower than the historical average for the same period, the support set will include data from other cities that have historically experienced similar high growth rates and meteorological anomalies. Next, the system statistically analyzes the influenza incidence rate, meteorological indicators (e.g., daily average temperature, relative humidity, wind speed), and population movement data in this support set, calculating their mean, standard deviation, etc., and inputs these statistics into a small neural network to generate a 128-dimensional task embedding vector. Subsequently, this task embedding vector is fed into a meta-learning adapter, which is an MLP containing a two-layer fully connected network. This adapter outputs the scaling factor (gamma) and offset (beta) for the feature maps of each convolutional layer in the backbone network of the influenza prediction core model. For example, for the feature map output by the third convolutional layer in the backbone network, the adapter generates a specific set of gamma and beta values. Finally, during the forward propagation of the backbone network, these dynamically generated gamma and beta values ​​are used to perform element-wise scaling and offset operations on the feature map output by the convolutional layer through feature linear modulation techniques. This allows the features extracted by the backbone network to better capture the unique patterns of cities currently experiencing new outbreaks, enabling relatively accurate predictions even when data for those cities is scarce.

[0069] Through the above technical solution, this application effectively addresses the problem of insufficient generalization ability and difficulty in rapid adaptation leading to prediction failure in existing influenza prediction models when facing scenarios with scarce or changing data, such as newly emerging epidemic areas, new strains, or data reporting delays. By constructing an anti-smoothing support set, the model can accurately locate the most relevant experience for the current task from massive historical data; the task embedding vector efficiently encodes the unique contextual information of the current task; the meta-learning adapter dynamically generates feature modulation parameters based on this and adaptively adjusts the backbone network features of the core influenza prediction model through feature linear modulation technology. This mechanism endows the model with powerful small-sample rapid adaptation and generalization capabilities, enabling it to quickly adjust itself using a very small amount of target scenario data, significantly improving its practicality and robustness in the face of sudden and mutated situations, and ensuring that the prediction results maintain high accuracy and reliability in a dynamically changing environment.

[0070] In some embodiments, the method for obtaining the trained influenza prediction core model includes: in a first training phase, training the backbone network of the influenza prediction core model using mean squared error and peak-aware weighted loss functions; in a second training phase, freezing the backbone network parameters and introducing a differentiable decision simulation layer; determining the estimated resource demand generated by the decision simulation layer based on the prediction results and comparing it with the actual resource state to obtain the decision loss; and combining the decision loss with the prediction loss in a weighted manner, and jointly optimizing the remaining parameters of the influenza prediction core model and the parameters of the decision simulation layer through backpropagation.

[0071] The first training phase aims to lay the foundation for the predictive capabilities of the backbone network of the influenza prediction core model. Mean squared error (MSE) loss is a commonly used regression loss function that measures the model's prediction accuracy by calculating the average of the squared differences between predicted and true values. Its role is to enable the model to learn general trends and patterns in the data. Besides MSE, other regression loss functions such as mean absolute error (MAE) or Huber loss can also be used. Peak-aware weighted loss is used to give greater attention and weight to peak areas or high-risk periods of influenza incidence during training, ensuring the model's accuracy in predicting critical moments of an outbreak. For example, this can be achieved by assigning greater loss weight to samples above a certain incidence threshold, or by using an asymmetric loss function to impose a greater penalty on predictions that underestimate the peak. The backbone network of the influenza prediction core model is the foundation of the entire prediction model, responsible for extracting key features from multi-source heterogeneous data and making preliminary influenza incidence predictions. This backbone network can be a deep neural network, such as a convolutional neural network-based architecture, to capture local patterns in time-series data; or a recurrent neural network or Transformer-based architecture to process sequential data and capture long-term dependencies.

[0072] The second training phase further optimizes the model's decision-making utility based on the first phase. Freezing the backbone network parameters means maintaining the weights and biases of the backbone network trained in the first phase unchanged during the second phase of training, without updating them. This aims to preserve the basic predictive capabilities already learned by the backbone network, avoiding damage to its accurate prediction of influenza incidence during decision optimization, thus focusing the optimization on how to make the prediction results better serve the decision. The differentiable decision simulation layer is a module that can transform the prediction results into specific decision indicators and perform gradient calculations. The role of this simulation layer is to mathematize the complex decision-making process, enabling its integration into the end-to-end training framework of the neural network. For example, this layer could be a simplified model based on linear programming or integer programming, whose objective function and constraints can be approximated as differentiable; or a neural network composed of multilayer perceptrons that simulates the decision-making process by learning the mapping relationship between prediction results and actual decisions in historical data.

[0073] The step of determining the estimated resource demand generated by the decision simulation layer based on the prediction results and comparing it with the actual resource status to obtain the decision loss aims to quantify the impact of the prediction results on actual decision-making. Estimated resource demand refers to the amount of medical resources needed in the future, such as the number of beds, medical staff, and vaccine reserves, calculated by the differentiable decision simulation layer based on the influenza incidence rate prediction results output by the core influenza prediction model. Actual resource status refers to the current or future availability of actual medical resources in the real-world application scenario. Decision loss is an indicator that measures the gap between the estimated resource demand and the actual resource status, guiding the model to learn how to make predictions that better align with actual resource constraints and optimization objectives. For example, decision loss can be defined as the penalty when the estimated resource demand exceeds the actual resource status, or the penalty when the estimated resource demand is significantly lower than the actual resource status, or a weighted combination of both. Specific loss functions can be L1 loss, L2 loss, or an asymmetric loss function customized according to actual business needs.

[0074] The key to achieving decision-driven learning lies in the weighted combination of the decision loss and prediction loss, followed by backpropagation to jointly optimize the remaining parameters of the core influenza prediction model and the parameters of the decision simulation layer. The prediction loss refers to the error between the predicted and actual influenza incidence rates, calculated continuously during the first or second training phase, such as mean squared error. Weighting the decision loss and prediction loss means that the overall optimization objective considers both the accuracy of prediction and the effectiveness of decision-making; adjusting the weights balances their importance. For example, a hyperparameter can be set. , Joint optimization via backpropagation refers to using optimization algorithms such as gradient descent to calculate the gradients of the remaining parameters of the core influenza prediction model (e.g., the parameters of the meta-learning adapter) and the parameters of the differentiable decision simulation layer based on the weighted combined total loss, and then updating these parameters in the opposite direction of the gradients. This joint optimization enables the model not only to pursue statistical predictive accuracy, but more importantly, it to learn how to generate predictive results that are most beneficial to subsequent decisions, thereby improving the practical value of the predictions.

[0075] Specifically, the training method is implemented as follows: In the first training phase, the backbone network of the influenza prediction core model can adopt a temporal prediction network based on the Transformer architecture. Its input consists of preprocessed multi-source heterogeneous data and spatially augmented features. This network captures complex spatiotemporal dependencies through multiple self-attention layers and feedforward network layers. During training, in addition to calculating the mean squared error between the predicted incidence rate and the actual incidence rate, a peak-aware weighted loss function is introduced. For example, when the actual incidence rate exceeds a certain preset threshold (e.g., 100 cases per 100,000 people), the corresponding sample loss weight increases by 5 times to encourage the model to more accurately predict the peak of the epidemic. In the second training phase, the parameters of the aforementioned Transformer backbone network are set to be non-trainable. At this time, a differentiable decision simulation layer is introduced, which can be a module composed of two fully connected neural networks. The input of this module is the future multi-period influenza incidence rate point prediction and interval prediction output by the influenza prediction core model, and its output is the daily bed demand curve for the next week. For example, the simulation layer could contain a function that considers factors such as predicted incidence rate, population size, and historical visit rates to estimate the potential number of daily visits. Then, a differentiable bed turnover model (e.g., a simplified model based on queuing theory with learnable parameters) is used to calculate bed occupancy and demand. The decision loss can be defined as the penalty for estimated bed demand exceeding the actual number of available beds, and the penalty for estimated bed demand being significantly lower than the actual number of available beds. For example, an asymmetric L1 loss function could be used, with a higher penalty coefficient for bed shortages than for bed redundancy. Subsequently, the mean squared error prediction loss from the first stage is weighted and combined with the decision loss from the second stage, for example, with a weight of 0.6 for the prediction loss and 0.4 for the decision loss. Using the Adam optimizer and backpropagation algorithm, the parameters of the meta-learning adapter in the influenza prediction core model and the weights and biases of the fully connected network within the differentiable decision simulation layer are jointly updated. In this way, the model is guided to generate predictions that are not only statistically accurate but also effective in guiding the allocation of medical resources (such as hospital beds).

[0076] Through the above technical solution, this application effectively solves the problem of inconsistency between the training objectives and actual decision-making effectiveness of influenza prediction models. By introducing decision loss and performing end-to-end joint optimization, the model's training objectives and final decision-making effects are highly aligned. This enables the core influenza prediction model to learn to make predictions that are "most beneficial to subsequent resource allocation and prevention and control decisions," thereby fundamentally improving the practical value and application effectiveness of the prediction results. This decision-driven learning paradigm ensures that the prediction results can be used more directly and reliably to guide actual public health resource allocation and prevention and control decisions, significantly compensating for the shortcomings of the "prediction-decision" disconnect in traditional prediction methods.

[0077] In some embodiments, this application further includes: estimating the potential number of new outpatient visits based on the predicted incidence rate, population base, and proportion of high-risk groups; converting the number of outpatient visits into the estimated number of new inpatients based on the historical hospitalization rate; and simulating the bed occupancy and turnover process using the differentiable decision simulation layer to output the future daily bed demand curve.

[0078] The predicted incidence rate is a key indicator output by the core influenza prediction model, reflecting the intensity of influenza prevalence in the target area during a specific future period. This incidence rate can be a percentage, representing the proportion of the population expected to be infected. The population base refers to the total population within the target geographical unit, serving as the basis for estimating the actual number of patients. This data can be derived from official statistical yearbooks or census data. The high-risk population proportion refers to the proportion of the total population who, due to factors such as age and underlying diseases, are more likely to develop severe illness or require medical attention after contracting influenza. This proportion can be determined based on epidemiological survey data or expert experience. The purpose of estimating the potential increase in patient visits is to transform the abstract incidence rate into a concrete medical demand. One approach is to multiply the predicted incidence rate by the population base, then multiply by the high-risk population proportion and a certain patient visit willingness coefficient. Another approach is to train a regression model using historical data, inputting the incidence rate, population base, and high-risk population proportion to directly output the potential increase in patient visits.

[0079] The historical hospitalization rate refers to the proportion of past influenza patients who required hospitalization, reflecting the severity of the influenza disease and the need for medical intervention. This rate can be calculated based on historical medical records, disease surveillance reports, or public health statistics. Converting outpatient visits into projected new hospitalizations aims to further refine the demand for medical resources, particularly hospital beds. One approach is to directly multiply the estimated potential new outpatient visits by the historical hospitalization rate; another approach is to consider the differentiated hospitalization rates across different age groups or high-risk populations, performing stratified calculations to obtain a more accurate estimate.

[0080] The differentiable decision simulation layer is a computational module that models the dynamic changes in medical resources (such as hospital beds) as a series of differentiable mathematical operations. Its core is ensuring that all operations in the simulation process are differentiable, enabling gradient backpropagation during neural network training. This layer can be a set of mathematical expressions based on a state-space model, where bed states (occupied, vacant, cleaning) evolve over time, and these evolution rules are designed as differentiable functions. Simulating bed occupancy and turnover refers to modeling the entire lifecycle of a patient from admission, hospitalization, discharge, bed cleaning, and re-availability. This includes considering factors such as the average length of hospital stay, the time required for bed cleaning, and the demand for different bed types for patients with different conditions. The simulation process can employ the idea of ​​discrete event simulation, but represents the occurrence of events and state transitions using continuously differentiable functions (e.g., using a sigmoid function or a softplus function to approximate a step function), thus achieving overall differentiability.

[0081] The bed demand curve is a time series data point representing the number of beds needed each day over a future period (e.g., the next 7 or 14 days). This curve visually illustrates the dynamic trend of pressure on healthcare resources. The purpose of outputting the daily bed demand curve is to provide healthcare institutions with specific data for resource allocation. This curve can also be used as the final output of the decision-making simulation layer, directly inputting into the calculation of decision loss to assess the impact of the forecast on actual resource management.

[0082] The following example illustrates this. Suppose that the influenza prediction core model predicts that the influenza incidence rate in a city next week will be 0.5%. The city has a population of 1 million, and the proportion of high-risk groups (such as the elderly over 65 years old and patients with chronic diseases) is 15%. First, by multiplying the 0.5% incidence rate by the population of 1 million, then by the proportion of high-risk groups of 15%, and considering an empirical consultation willingness coefficient (e.g., 0.8), the potential number of new consultations can be estimated at 600. Subsequently, based on the city's average hospitalization rate in past influenza seasons (e.g., 10%), these 600 potential new consultations are converted into an estimated number of new hospitalizations, namely 60. These 60 estimated new hospitalizations will be input into a differentiable decision simulation layer. This simulation layer can be constructed as a discrete time-step model based on queuing theory, where each time step (e.g., one day) simulates the patient's admission, length of hospital stay (e.g., an average of 7 days, following a Poisson distribution), discharge, and bed cleaning (e.g., 12 hours). To ensure differentiability, patient admission and discharge events can be approximated by smooth activation functions (such as the Softplus function), and bed status transitions can be described by differentiable matrix operations. For example, bed occupancy status can be represented as a vector, updated by multiplying it by the admission and discharge rate matrices. Ultimately, this differentiable decision simulation layer outputs a sequence of daily bed demand for the next 7 or 14 days; for example, 50 beds are needed on day one, 65 on day two, 70 on day three, and so on, forming a daily bed demand curve. This curve can be directly used to calculate the decision loss and guide the optimization of the core influenza prediction model.

[0083] Through the above technical solution, this application provides a specific and differentiable mathematical modeling method for medical resource demand, effectively solving the problem that decision simulation processes are usually not differentiable, thus making it impossible to perform gradient joint optimization with neural networks. By estimating the potential number of new outpatient visits based on the predicted incidence rate, population base, and proportion of high-risk groups, and converting this into an estimated number of new hospitalizations based on historical hospitalization rates, quantitative input is provided for subsequent resource simulation. More importantly, a differentiable decision simulation layer is used to simulate the bed occupancy and turnover process, and outputs the bed demand curve for each day in the future, making the entire chain from prediction to resource demand assessment differentiable. This enables the decision-driven learning paradigm proposed in the above implementation to be truly realized, that is, the decision loss can be accurately calculated and backpropagated, thereby driving the core influenza prediction model to learn and generate prediction results that are not only statistically accurate but also most efficient in guiding the allocation of medical resources (such as beds), significantly improving the practical value and application utility of the prediction results.

[0084] In some embodiments, after acquiring multi-source heterogeneous data of multiple geographic units within the target area, this application further includes: acquiring time-series data of search engine query volume related to influenza symptoms through an application programming interface; performing noise reduction, smoothing, and standardization preprocessing on the search engine query volume time-series data; and determining dynamic characteristics reflecting public health concerns based on the long-term trend term and short-term residual fluctuation term of the preprocessed search engine query volume time-series data.

[0085] One method involves acquiring time-series data on search engine queries related to influenza symptoms through application programming interfaces (APIs). An API is a predefined set of functions or protocols that allows communication and data exchange between different software applications. Here, APIs can automatically retrieve the frequency or index of user searches for information related to influenza symptoms (such as keywords like "fever," "cough," "flu," and "cold medicine") within a specific time period from large search engine platforms (e.g., Baidu Index, Google Trends). This data reflects public attention to the influenza epidemic and the potential occurrence of symptoms, offering strong real-time data. Another method involves using web crawling technology to periodically crawl publicly available data related to influenza symptoms from specific search engines or health forums and store it in a structured format. The time-series data on search engine queries undergoes denoising, smoothing, and standardization preprocessing. Denoising aims to eliminate random errors or outliers in the data. For example, outliers can be identified and removed using statistical methods (such as Z-score outlier detection and IQR methods), or signal-noise separation can be achieved using wavelet transform-based methods. Smoothing aims to reduce short-term fluctuations in data and reveal underlying trends. For example, moving averages (such as simple moving averages or exponential moving averages) can be used to smooth the data, or a Savitzky-Golay filter can be used for multinomial fitting smoothing. Standardization aims to eliminate dimensional differences between different data sources or time series, making them comparable. For example, Z-score standardization (converting data to a distribution with a mean of 0 and a standard deviation of 1) or Min-Max standardization (scaling data to the [0,1] interval) can be used. Based on the long-term trend and short-term residual fluctuation terms of the preprocessed search engine query volume time series data, dynamic characteristics reflecting public health concerns are determined. The long-term trend reflects the sustained, slowly changing public concern about influenza or related health issues; for example, the general increase in public concern during seasonal influenza peaks. The short-term residual fluctuation term reflects rapid and drastic changes in public concern caused by sudden events or specific outbreaks, outside of the long-term trend; for example, a sudden increase in influenza cases in a region leading to a significant increase in related search volume in a short period. This can be achieved through time series decomposition techniques, such as using methods like Seasonal-Trend Decomposition (STL decomposition), Wavelet Decomposition, or Empirical Mode Decomposition (EMD) to decompose time series data into trend, seasonality, and residual components. The trend component can be used as a long-term trend term, and the residual component can be used as a short-term residual fluctuation term.

[0086] The following is a concrete example to illustrate this. After acquiring multi-source heterogeneous data from multiple geographic units within the target area, the system can utilize Baidu Index's open API to periodically (e.g., daily or weekly) crawl search indices for keywords such as "flu symptoms," "cold and fever," and "cough medicine" within the target area. The acquired raw search index time-series data may contain abnormal peaks or troughs. In such cases, a method based on median absolute deviation (MAD) can be used for outlier detection and removal to denoise the data. Subsequently, to smooth the data, a 7-day exponential moving average filter can be applied to reduce daily fluctuations and highlight weekly trends. Next, the smoothed data is standardized using Min-Max, unifying its numerical range to between 0 and 1. Finally, the STL decomposition algorithm is used to decompose the standardized time-series data into trend components, seasonal components, and residual components. The trend component is extracted as a feature reflecting the public's long-term health concerns, while the residual component is considered as a dynamic feature reflecting the public's short-term symptom-focused searches. These dynamic features are then integrated into the multi-source heterogeneous data as input to the core influenza prediction model.

[0087] Through the aforementioned technical solution, this method can transform previously noisy and unusable internet search data into high-quality early warning signals via a standardized data processing workflow. The decomposed long-term trend term effectively reflects long-term changes in public health awareness, while the short-term residual fluctuation term sensitively captures short-term symptom-focused search behavior, providing the core influenza prediction model with more sensitive early warning information than traditional clinical reports. This significantly improves the real-time performance and early warning capabilities of influenza prediction, enabling the acquisition of crucial information in the early stages of an outbreak, thus buying valuable response time for public health departments and effectively mitigating the decision-making lag caused by delays in traditional monitoring data reporting.

[0088] In some embodiments, the real-time data in the multi-source heterogeneous data processed by the trained influenza prediction core model in this application includes: extracting short-term high-frequency fluctuation features from meteorological and influenza data using the dilated causal convolutional network of the multi-scale feature extraction network; extracting medium-term trend features using the temporal convolutional network of the multi-scale feature extraction network; capturing long-term dependency and periodic features using the encoder of the multi-scale feature extraction network; and adaptively fusing the short-term high-frequency fluctuation features, medium-term trend features, and long-term dependency and periodic features through the gating attention mechanism of the multi-scale feature extraction network.

[0089] The multi-scale feature extraction network is a neural network architecture specifically designed to identify and separate patterns at different temporal granularities from input data. This network typically contains multiple parallel or cascaded sub-modules, each responsible for capturing information at a specific time scale. For example, one sub-module might focus on rapidly changing daytime data, while another focuses on slowly evolving seasonal patterns. The dilated causal convolutional network is a special type of convolutional neural network that expands the receptive field by inserting holes (dilation) into the convolutional kernel, thereby capturing broader temporal context information without increasing the number of parameters. Simultaneously, causal convolution ensures that when predicting the current moment, the model relies only on past and current inputs, avoiding the leakage of future information. This network is primarily used to extract rapidly changing, short-term, high-frequency fluctuation features from data, such as the rapid increase or decrease in influenza cases caused by sudden weather events or short-term behavioral changes. The temporal convolutional network is a neural network suitable for processing sequential data. It typically utilizes one-dimensional convolutional layers to capture local patterns in the sequence and learns longer-term dependencies by stacking multiple convolutions or combining residual connections. This network is used here to extract relatively stable, sustained medium-term trend features from the data, such as weekly or monthly increases or decreases in influenza incidence. The encoder, a common component in deep learning models, maps input sequences (such as time-series data) to a fixed-dimensional vector representation or a series of context-dependent embedding vectors. This encoder captures long-term dependencies and periodic features in the data, such as annual periodicity of influenza outbreaks, cross-year epidemic patterns, or long-term trends related to climate change. The gated attention mechanism is a mechanism that allows the model to dynamically select and weight different input information or features. It typically controls the flow of information by introducing learnable gating units (e.g., using a sigmoid activation function), enabling the model to adaptively focus on the most important features and fuse information from different scales.

[0090] As a specific implementation method, the above-mentioned technical means can be implemented with reference to the following example. In a multi-scale feature extraction network, an architecture containing three parallel branches can be constructed. The first branch can adopt a network composed of multiple layers of dilated causal convolutional layers stacked together, where the dilation rate increases layer by layer, for example, starting from 1, 2, 4, to effectively capture drastic intraday temperature changes or high-frequency weekly fluctuations in influenza cases in meteorological data. The second branch can adopt a temporal convolutional network composed of multiple residual temporal convolutional blocks, each block containing multiple causal convolutional layers, to capture the medium-term upward or downward trend of influenza incidence. The third branch can adopt an encoder based on the Transformer architecture, whose self-attention mechanism can capture long-term dependencies spanning months or even years, such as the seasonal evolution of influenza virus strains or the long-term effects of climate patterns. At the output of these branches, a gated attention layer can be set, which contains multiple linear transformations and a sigmoid activation function to generate a dynamic weight vector for the features extracted by each branch. These weight vectors are then multiplied element-wise with the corresponding feature vectors, and the weighted features are concatenated or summed to form the final fused features, which are then used by subsequent layers of the influenza prediction core model.

[0091] Through the aforementioned technical solution, this application effectively addresses the problem that influenza transmission is influenced by a combination of factors across short, medium, and long-term time scales, and that a single time-series model struggles to fully capture the complex temporal dynamics. By using dilated causal convolutional networks, temporal convolutional networks, and encoders in parallel, the model can extract short-term high-frequency fluctuation features, medium-term trend features, and long-term dependence and periodicity features from meteorological and influenza data, respectively. Furthermore, the introduction of a gated attention mechanism allows the model to adaptively fuse these features at different scales based on actual conditions, thereby comprehensively and meticulously characterizing the complex temporal patterns of influenza transmission. This significantly improves the model's sensitivity to short-term outbreaks and its accuracy in predicting long-term epidemic trends, enabling the prediction results to more accurately reflect the true dynamics of the influenza epidemic and providing a more reliable basis for public health decision-making.

[0092] In some embodiments, the steps of generating dynamic risk assessment levels and corresponding prevention and control decision recommendations in this application specifically include: pre-setting a risk assessment rule knowledge base composed of IF-THEN conditional rules; inputting the real-time predicted incidence rate, its upper confidence interval limit, and real-time medical resource data into the rule engine; performing rule matching, prioritizing and resolving conflicts of the outputs that trigger the rule engine, and generating structured prevention and control decision recommendations.

[0093] The pre-built risk assessment rule knowledge base, composed of IF-THEN conditional rules, refers to the system that, before deployment, encodes and stores the decision-making logic for influenza prevention and control in the form of "if (IF) certain conditions are met, then (THEN) take a certain action or reach a certain conclusion," based on expert knowledge and historical experience in the field of public health. The purpose of this knowledge base is to formalize and standardize expert experience, providing a basis for automated decision-making. For example, this knowledge base can be stored in a relational database, with each record containing a rule ID, a conditional expression (such as "incidence rate is higher than X and medical resources are strained"), and a corresponding decision suggestion template (such as "initiate a Level Y emergency response"). Alternatively, this knowledge base can also be configured using a specific rule language (such as XML or JSON-based rule definitions), allowing for flexible definition and modification of rules.

[0094] Inputting the real-time predicted incidence rate, its upper confidence interval, and real-time medical resource data into the rule engine means that after obtaining the prediction results output by the core influenza prediction model, this key information, along with real-time data reflecting the current capacity of the healthcare system, is used as input for the rule engine's decision-making reasoning. This input data forms the basis for the rule engine to assess current risks and formulate decision recommendations.

[0095] The execution of rule matching, prioritizing and resolving conflicts in the output of the rule engine, and generating structured prevention and control decision recommendations means that after receiving input data, the rule engine traverses all rules in the risk assessment rule knowledge base and evaluates whether the IF condition of each rule is met based on the input data. Once the condition of a rule is met, the rule is triggered. When multiple rules are triggered simultaneously, the rule engine prioritizes these triggered rules according to a preset priority strategy (e.g., rules with higher urgency are prioritized, rules with broader coverage are prioritized). Simultaneously, to avoid conflicting decision recommendations between different rules, the rule engine also performs conflict resolution. For example, if both "mild warning" and "moderate warning" rules are triggered simultaneously, the system will select "moderate warning" as the final decision. Finally, based on the results of priority ranking and conflict resolution, the rule engine generates clear, specific, and actionable structured prevention and control decision recommendations, outputting them in the form of text reports, instruction lists, or visual charts.

[0096] The following is a concrete example. Suppose that in a certain geographical unit, the influenza prediction core model predicts an influenza incidence rate of 150 per 100,000 population for the coming week, with a 95% confidence interval upper limit of 180 per 100,000 population. Simultaneously, real-time medical resource data shows that the bed occupancy rate in this area has reached 85%, and there is a 12% shortage of medical staff. At this point, the rules engine receives this data. The following rules may exist in the pre-built risk assessment rule knowledge base:

[0097] Rule A: IF (real-time predicted incidence rate > 100 / 100,000 AND upper confidence interval > 150 / 100,000 AND bed occupancy rate > 80%) THEN (Recommendation: Activate Level II emergency response, increase bed reserves, issue public health warning; priority: high).

[0098] Rule B: IF (real-time predicted incidence rate > 80 / 100,000 AND healthcare staff shortage > 10%) THEN (Recommendation: Emergency deployment of healthcare staff, initiation of telemedicine support; Priority: Medium).

[0099] Rule C: IF (real-time predicted incidence rate < 50 / 100,000) THEN (Recommendation: Maintain routine monitoring and strengthen health education; Priority: Low).

[0100] Based on the input data above, the conditions for both Rule A and Rule B are met, therefore these two rules are triggered. Since Rule A has a higher priority than Rule B, the rule engine will prioritize the suggestion from Rule A. Simultaneously, because the suggestions from Rule A and Rule B do not directly conflict, the rule engine will integrate them to generate the final structured prevention and control decision-making recommendation, such as: "The current influenza incidence rate is high and is expected to continue to rise, putting significant pressure on medical resources. It is recommended to immediately activate a Level II emergency response, increase bed capacity, issue a public health warning, urgently deploy medical personnel, and initiate telemedicine support."

[0101] Through the aforementioned technical solution, this application can automatically and systematically transform the complex predicted values ​​output by the core influenza prediction model, such as incidence rate predictions and upper limits of confidence intervals, into clear and actionable prevention and control decision-making recommendations. This significantly improves the efficiency and real-time nature of the transformation from prediction results to actual decisions, reduces reliance on expert interpretation, and thus avoids decision inconsistencies caused by differences in individual experience. By encapsulating expert knowledge, this solution makes the decision-making process transparent and interpretable, and ensures that prediction results can be directly and reliably used to guide the allocation of public health resources and the formulation of prevention and control strategies, fundamentally solving the problem of the disconnect between influenza prediction and decision-making.

[0102] In some embodiments, the method of the preceding embodiment further includes: recording the predicted data, actual epidemic data and measures taken for each warning; periodically evaluating the triggering accuracy and decision-making effect of each rule to form feedback data; and using a reinforcement learning algorithm to automatically adjust the judgment threshold parameters in the rule engine using the feedback data.

[0103] The process of recording the predicted data, actual epidemic data, and measures taken for each warning refers to the system's structured storage of its predicted output (e.g., predictions of future multi-cycle influenza incidence rate point and interval predictions, dynamic risk assessment levels, and prevention and control decision recommendations), actual epidemic data (e.g., subsequent actual influenza incidence rates and medical resource occupancy), and actual intervention measures taken based on the system's recommendations (e.g., resource allocation instructions and adjustments to prevention and control policies) each time a warning signal or decision recommendation is generated. This step aims to establish a comprehensive historical operational archive, providing a real and traceable data foundation for subsequent performance evaluation and optimization. Regularly evaluating the trigger accuracy and post-decision effect of each rule to generate feedback data involves periodically analyzing the recorded historical data to quantitatively evaluate the performance of each rule in the rule engine. Trigger accuracy measures the precision with which a rule identifies risks or predicts events; for example, whether an actual epidemic has occurred or reached the warning level when a rule triggers a warning. Post-decision effect assesses the impact and effectiveness of measures taken based on the rules on the actual development of the epidemic; for example, whether resource allocation effectively alleviated medical pressure or whether prevention and control measures successfully curbed the spread of the epidemic. These assessment results are compiled into structured feedback data, serving as the basis for subsequent rule adjustments. This assessment process can be implemented through the development of a dedicated data analysis module, which periodically extracts data from historical records and calculates a performance score for each rule based on predefined assessment indicators (such as true positive rate, false positive rate, resource utilization rate, and epidemic control effectiveness). Furthermore, an expert review interface can be built, allowing public health experts to manually evaluate and score the system-generated decision recommendations and their subsequent effects, transforming expert judgment into quantifiable feedback signals.

[0104] Employing reinforcement learning algorithms to automatically adjust the decision threshold parameters in the rule engine using feedback data involves introducing an adaptive learning mechanism, treating the optimization of rule engine parameters as a reinforcement learning problem. The reinforcement learning algorithm learns how to dynamically adjust the decision thresholds within the rule engine (e.g., the incidence rate threshold triggering risk warnings, the upper limit of the confidence interval for resource allocation) through interaction with the environment (i.e., historical operational data and evaluation feedback) to maximize preset long-term rewards (e.g., maximizing prediction accuracy, minimizing resource waste, and optimizing epidemic control effectiveness). This step enables the rule engine to self-optimize and evolve based on actual operational results. This reinforcement learning algorithm can be implemented using methods such as Q-learning, SARSA, Deep Q-Networks (DQN), or Policy Gradient. For example, the set of decision thresholds in the rule engine can be defined as the action space of a reinforcement learning agent, with periodically evaluated trigger accuracy and decision aftereffects serving as reward signals. The agent learns the optimal decision thresholds for different scenarios by continuously trying different threshold combinations and updating its policy based on the rewards obtained.

[0105] The following example illustrates this. Suppose a rule engine contains the following rule: "If the predicted incidence rate of a geographic unit exceeds threshold A and the upper limit of the confidence interval exceeds threshold B, it is recommended to activate a Level II emergency response and allocate X units of medical resources." During system operation, each time this rule is triggered, the system records the predicted incidence rate, the upper limit of the confidence interval, the actual incidence rate, the actual amount of resources allocated, and the subsequent epidemic control effect. For example, the system records that this rule was triggered 10 times in the past month, with 8 of those instances reaching the Level II response standard (high accuracy), but 3 instances where the allocated X units of resources failed to fully meet the needs (the post-decision effectiveness needs improvement). The system periodically (e.g., weekly) runs an evaluation module to calculate indicators such as the true positive rate, false positive rate, and resource utilization efficiency of this rule, and uses these indicators as feedback data. Subsequently, a reinforcement learning agent (e.g., a Q-learning-based algorithm) uses this feedback data to attempt to adjust thresholds A and B. If it is found that slightly lowering threshold A can detect outbreaks earlier without significantly increasing false alarms, while slightly raising threshold B can more accurately determine resource demand, then the reinforcement learning agent will learn this adjustment strategy and automatically update the values ​​of threshold A and threshold B in the rule engine.

[0106] The above technical solution establishes a complete learning loop of "decision-feedback-optimization," endowing the decision-making system with the ability to continuously evolve. This enables risk assessment standards to adaptively adjust to dynamic factors such as virus mutations, changes in medical conditions, and alterations in social behavior in the real world, thereby maintaining the scientific validity and timeliness of decision recommendations in the long term. It effectively solves the technical problem that static rules defined by experts are difficult to adapt to dynamic realities and are prone to gradually becoming ineffective.

[0107] In some embodiments, Figure 1 The method also includes: extracting source region knowledge with similar features to the target new region from the influenza prediction core model trained in the data-rich region; activating the meta-learning adapter based on a small amount of sample data from the target new region; and rapidly fine-tuning the model based on the transferred prior knowledge to generate an influenza prediction core model suitable for the new region.

[0108] For example, suppose a public health department plans to deploy an influenza prediction system at a newly established monitoring site (e.g., a newly developed area). Because this new development is relatively new, it only has influenza monitoring data from the most recent month. To address its "cold start" problem, the system first uses an existing library of influenza prediction core models trained on multiple cities with rich historical data (such as provincial capitals and large industrial cities). By comparing the new development's static geographical attributes (geographical location, population density, climate type, etc.) with the corresponding features of these cities, the system identifies the source region most similar to the new development. For example, the system might identify "City A," which has similar climate conditions and population structure to the new development, as the source region. Subsequently, the system extracts the pre-trained weight parameters of the multi-scale feature extraction network from the influenza prediction core model corresponding to "City A" as prior knowledge. Next, the limited amount of influenza monitoring time-series data from the new development over the past month, along with multi-dimensional meteorological time-series data, is input into the meta-learning adapter of the influenza prediction core model. This meta-learning adapter, for example, could be a module containing a two-layer fully connected network. Based on a small amount of data from the past month, it learns and outputs a set of feature modulation parameters specific to the data distribution of the newly developed area. Finally, these modulation parameters are used to quickly fine-tune the core influenza prediction model, based on the weight parameters transferred from "City A". Specifically, most layers of the multi-scale feature extraction network can be frozen, and optimization can be performed only on the meta-learning adapter and the prediction layer at the top of the model for a few training epochs, using a very small learning rate (e.g., 0.0001). In this way, the system can generate a core influenza prediction model specifically for the newly developed area with basic predictive capabilities in a short time using limited data, thereby enabling rapid deployment of influenza prediction capabilities.

[0109] Through the above technical solution, this application can systematically solve the "cold start" problem faced by newly established monitoring sites or areas lacking historical data. This solution combines the advantages of transfer learning and meta-learning, enabling the core influenza prediction model to quickly acquire prior knowledge from data-rich source areas and efficiently adapt using limited sample data from the target new area. This significantly shortens the time cycle from data accumulation to model usability in new areas, substantially improving the deployment efficiency and coverage of the influenza prediction system. Therefore, even in scenarios with scarce or changing data, such as newly emerging epidemic areas, new strains, or delayed data reporting, the model can quickly adjust and maintain good generalization ability, thus ensuring the practicality and robustness of the prediction results and providing timely and reliable evidence for public health decision-making.

[0110] In some embodiments, the multi-source heterogeneous data further includes spatiotemporal raster data of surface air pollution concentration retrieved by satellite remote sensing; Figure 1 The method shown further includes, after acquiring multi-source heterogeneous data of multiple geographic units within the target area: processing the spatiotemporal raster data using a lightweight convolutional neural network; extracting the spatial distribution and temporal variation features of regional air pollution exposure; and fusing the extracted spatial distribution and temporal variation features as environmental driving factors with real-time data in the multi-source heterogeneous data.

[0111] The following is a concrete example. As a specific implementation method, when acquiring multi-source heterogeneous data for multiple geographic units within a target area, data from the European Space Agency's (ESA) Sentinel-5P satellite can be incorporated. This satellite provides surface concentration inversion data for various atmospheric pollutants (such as PM2.5, NO2, SO2, O3, etc.) globally, typically published in high-resolution spatiotemporal raster format. For example, daily PM2.5 concentration data in 1km × 1km raster format for the target area can be acquired. To process this data, a lightweight convolutional neural network can be constructed. This network can contain multiple depthwise separable convolutional layers and global average pooling layers to reduce the number of parameters and computational complexity. For example, the network can first capture local spatial and temporal patterns of PM2.5 concentration through several 3D convolutional layers (or a combination of 2D convolutions and GRU / LSTM layers), and then compress the pollution raster data for each geographic unit into a low-dimensional feature vector through a global average pooling layer. These feature vectors can represent the average pollution exposure level, peak pollution intensity, or pollution duration of a geographic unit within a specific time period. Subsequently, these extracted spatial distribution and temporal variation features of regional air pollution exposure, such as the daily average PM2.5 exposure index and PM2.5 concentration change rate for each geographic unit, can be concatenated with other real-time data such as real-time influenza surveillance data, temperature, humidity, and population migration data to form a unified input tensor for subsequent processing by the core influenza prediction model.

[0112] By introducing and processing spatiotemporal raster data of surface air pollution concentration retrieved from satellite remote sensing, this method overcomes the limitations of traditional ground monitoring station data in terms of limited spatial coverage. This enables the model to acquire large-scale, continuous, and refined environmental exposure information, particularly regarding air pollution, a crucial risk factor for respiratory infectious diseases. By efficiently extracting the spatial distribution and temporal variation characteristics of air pollution exposure using a lightweight convolutional neural network and fusing it with real-time data from multi-source heterogeneous datasets as an environmental driving factor, this method can more comprehensively assess the combined risks to environment and health. This not only provides richer and more accurate input information for influenza prediction but also significantly improves the robustness and accuracy of the prediction model under the influence of complex environmental factors, thus providing a solid data foundation for public health departments to formulate more precise prevention and control strategies.

[0113] In some embodiments, the multi-source heterogeneous data further includes respiratory pathogen co-circulation data from multi-source monitoring. Based on this, Figure 1The method also includes integrating the synergistic prediction needs of influenza, COVID-19, and respiratory syncytial virus (RSV). This feature refers to the unified planning and processing of prediction tasks for these three major respiratory pathogens—influenza, SARS-CoV-2, and RSV—rather than treating them as independent prediction problems. This means that the epidemiological interactions, common risk factors, and shared demand for medical resources among these three pathogens will be considered during the model design and data processing phases. For example, a multi-task learning model can be built, or an input layer capable of simultaneously receiving and processing data related to multiple pathogens can be designed to meet the need for predicting their synergistic epidemic situation.

[0114] Furthermore, the solution proposed in this application also includes dynamically adjusting the model weights for different pathogens using a hierarchical ensemble framework. This feature refers to employing a structured ensemble learning method, which can flexibly adjust the contribution of different pathogen prediction models to the overall prediction results based on real-time data or preset strategies. For example, when a pathogen (such as COVID-19) exhibits stronger transmissibility or higher pathogenicity during a specific period, the weight of its corresponding prediction model can be dynamically increased to give its prediction results a more dominant position in the final comprehensive evaluation. This dynamic adjustment can be achieved through adaptive algorithms based on historical performance, rule systems based on expert knowledge, or feedback mechanisms based on real-time epidemic data.

[0115] Finally, the proposed solution also includes a comprehensive risk assessment and resource pressure prediction for multi-pathogen co-emergence. This means that the final prediction is no longer merely a prediction of the incidence rate of a single pathogen, but rather an assessment of the overall risk posed by the co-emergence of multiple pathogens, predicting the potential comprehensive pressure on medical resources (such as beds, ICU beds, and healthcare personnel). The comprehensive risk assessment can be a multi-dimensional indicator, including, for example, overall incidence rate, severe illness rate, and mortality rate, weighted according to the characteristics of different pathogens. The resource pressure prediction can quantitatively provide key indicators such as bed occupancy rate and healthcare personnel workload that the healthcare system may face under different scenarios, thus providing public health departments with more operational decision-making support.

[0116] As a specific implementation method, when acquiring multi-source heterogeneous data from multiple geographical units within a target area, in addition to conventional influenza surveillance data and meteorological data, respiratory pathogen co-circulation data can also be obtained through the following channels: obtaining the positive rate of influenza virus nucleic acid testing from the national influenza surveillance network; obtaining the daily number of newly confirmed cases and hospitalized cases of COVID-19 from local CDC centers; and obtaining the number of positive samples for respiratory syncytial virus (RSV) testing from pediatric hospital laboratory systems. After time alignment and standardization, this data, along with the original multi-source heterogeneous data, is input into the core influenza prediction model. During the prediction phase, the system integrates the prediction requirements for influenza, COVID-19, and RSV. For example, a multi-head output neural network structure can be designed, with each output head responsible for predicting the incidence rate of one pathogen. To dynamically adjust model weights, a hierarchical integration framework can be used. The upper layer of this framework is a decision module that assigns different weights to the prediction models for influenza, COVID-19, and RSV based on the current season, historical epidemic trends, and real-time monitored dominant pathogen strains. For example, the weight of respiratory syncytial virus (RSV) may be increased during winter; while the weight of COVID-19 in the model will be increased accordingly during the peak of the COVID-19 pandemic. Ultimately, the system integrates these weighted predictions to output a comprehensive risk assessment and resource pressure prediction report on multi-pathogen co-epidemic events, including total morbidity, severe illness rate estimates, and the demand for ICU beds, general hospital beds, and healthcare personnel.

[0117] Through the aforementioned technical solutions, this application expands the prediction scope from a single influenza pathogen to a multi-pathogen complex epidemic system, including influenza, COVID-19, and respiratory syncytial virus. This expansion enables the model to more comprehensively and realistically reflect the overall burden of respiratory diseases during winter and spring, overcoming the limitations of single-pathogen prediction in assessing the comprehensive impact on medical resources. By integrating collaborative prediction needs and dynamically adjusting model weights using a hierarchical integration framework, the system can flexibly allocate prediction resources and influence based on the real-time epidemic situation and importance of different pathogens, thereby generating a more accurate and comprehensive comprehensive risk assessment. The final output of the comprehensive risk assessment and resource pressure prediction for multi-pathogen complex epidemics provides a more practically valuable basis for the overall resource planning, emergency response, and precise prevention and control of the medical system, significantly improving the scientific nature and effectiveness of public health decision-making.

[0118] To facilitate better implementation of the methods in the embodiments of this application, the embodiments of this application also provide an influenza prediction device based on multi-source data fusion and deep learning, such as... Figure 2 As shown, the device includes: The acquisition module 201 is used to acquire multi-source heterogeneous data of multiple geographic units within the target area; the multi-source heterogeneous data includes influenza surveillance time-series data, multi-dimensional meteorological time-series data, static geographic attribute data, cross-regional population flow data, and Internet public behavior time-series data. The association module 202 is used to determine the geographic adjacency matrix, the meteorological similarity matrix, and the functional association matrix based on the static geographic attribute data, the multi-dimensional meteorological time series data, and the cross-regional population flow data. The enhancement module 203 is used to construct a spatiotemporal multi-topology dynamic graph neural network based on the learnable dynamic fusion results of the geographic adjacency matrix, meteorological similarity matrix and functional association matrix, and to use the spatiotemporal multi-topology dynamic graph neural network to model spatial dependencies and obtain spatial enhancement features. Training module 204 is used to obtain a trained influenza prediction core model based on historical data in the multi-source heterogeneous data and the spatial enhancement features. The influenza prediction core model includes a multi-scale feature extraction network and a meta-learning adapter. Prediction module 205 is used to process real-time data in the multi-source heterogeneous data through the trained influenza prediction core model, and output prediction results including point prediction and interval prediction of influenza incidence rate in the future multiple cycles. The output module 206 is used to generate a dynamic risk assessment level and corresponding prevention and control decision suggestions based on the prediction results and real-time external context data, using a preset rule engine.

[0119] All of the above technical solutions can be combined in any way to form optional embodiments of this application, and will not be described in detail here.

[0120] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0121] The embodiments, implementation methods, and related technical features of this application can be combined and substituted for each other without conflict.

[0122] The above are merely preferred embodiments of this application and are not intended to limit this application in any way. Although the descriptions of each embodiment in this application have different focuses, and parts not described in detail in a certain embodiment can be referred to the relevant descriptions of other embodiments, any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of this application without departing from the content of the technical solution of this application shall still fall within the scope of the technical solution of this application.

Claims

1. An influenza prediction method based on multi-source data fusion and deep learning, characterized in that, include: Acquire multi-source heterogeneous data of multiple geographic units within the target area; the multi-source heterogeneous data includes influenza surveillance time-series data, multi-dimensional meteorological time-series data, static geographic attribute data, cross-regional population flow data, and Internet public behavior time-series data; Based on the static geographic attribute data, the multi-dimensional meteorological time series data, and the cross-regional population flow data, a geographic adjacency matrix, a meteorological similarity matrix, and a functional association matrix are determined. Based on the learnable dynamic fusion results of the geographic adjacency matrix, meteorological similarity matrix, and functional association matrix, a spatiotemporal multi-topology dynamic graph neural network is constructed, and spatial dependency modeling is performed using the spatiotemporal multi-topology dynamic graph neural network to obtain spatial enhancement features. Based on the historical data in the multi-source heterogeneous data and the spatial enhancement features, a trained influenza prediction core model is obtained. The influenza prediction core model includes a multi-scale feature extraction network and a meta-learning adapter. The trained influenza prediction core model processes real-time data from the multi-source heterogeneous data and outputs prediction results including point prediction and interval prediction of influenza incidence rates for future multiple periods. Based on a preset rule engine, dynamic risk assessment levels and corresponding prevention and control decision suggestions are generated according to the prediction results and real-time external context data.

2. The influenza prediction method according to claim 1, characterized in that, The construction of the spatiotemporal multi-topology dynamic graph neural network includes: Based on the multi-dimensional meteorological time-series data, a set of geographical units with similar meteorological characteristics is identified to form a meteorological cluster; A meteorological hypergraph is constructed based on the meteorological clusters, where each meteorological cluster serves as a hyperedge connecting all nodes within the cluster. By aggregating environmental risk information of nodes within hyperedges through a hypergraph convolutional network, and learning clustered environmental-driven propagation characteristics, the spatiotemporal multi-topology dynamic graph neural network is constructed.

3. The influenza prediction method according to claim 2, characterized in that, Also includes: Real-time monitoring of environmental risk characteristic values ​​of nodes within each hyperedge of the meteorological hypermap; When the risk characteristics of nodes within a meteorological cluster that exceed a threshold proportion simultaneously and significantly increase, a regional environmental risk warning signal is generated. By integrating the regional environmental risk early warning signals with individual city prediction results, a risk reassessment is conducted to generate decision-making recommendations, including those related to cluster outbreak risks.

4. The influenza prediction method according to claim 1, characterized in that, The trained influenza prediction core model processes real-time data from the multi-source heterogeneous data and outputs prediction results including point predictions and interval predictions of influenza incidence rates for future multiple periods, including: For the current prediction task, an anti-smoothing support set is constructed from historical data based on the fluctuation of influenza values ​​and the degree of meteorological anomalies in the samples; Encode the statistical features of the support set samples to generate task embedding vectors; A set of feature modulation parameters is dynamically generated based on the task embedding vector using a meta-learning adapter; The backbone network features of the influenza prediction core model are adaptively adjusted using the modulation parameters through characteristic linear modulation technology.

5. The influenza prediction method according to claim 1, characterized in that, The process of obtaining the trained influenza prediction core model includes: In the first training phase, the backbone network of the influenza prediction core model was trained using mean squared error and peak-sensing weighted loss function; In the second training phase, the backbone network parameters are frozen, and a differentiable decision simulation layer is introduced. Determine the estimated resource demand generated by the decision simulation layer based on the prediction results, and compare it with the actual resource status to obtain the decision loss; The decision loss and prediction loss are weighted and combined, and the remaining parameters of the influenza prediction core model and the decision simulation layer parameters are jointly optimized through backpropagation.

6. The influenza prediction method according to claim 5, characterized in that, Also includes: Estimate the potential number of new outpatient visits based on the predicted incidence rate, population size, and proportion of high-risk groups; Based on historical hospitalization rates, the potential number of new outpatient visits is converted into an estimated number of new hospitalizations. The differentiable decision simulation layer is used to simulate the occupancy and turnover process of beds, and outputs the bed demand curve for each day in the future.

7. The influenza prediction method according to claim 1, characterized in that, After acquiring multi-source heterogeneous data of multiple geographic units within the target area, the method further includes: Obtain time-series data on search engine query volume related to flu symptoms through application programming interfaces; The time-series data of search engine query volume is preprocessed by denoising, smoothing, and standardization. Based on the long-term trend and short-term residual fluctuation terms of the preprocessed search engine query volume time series data, dynamic characteristics reflecting public health concerns are determined.

8. The influenza prediction method according to claim 1, characterized in that, After acquiring multi-source heterogeneous data of multiple geographic units within the target area, the method further includes: Calculate the daily temperature difference, temperature change rate, and perceived temperature-derived characteristics of the multi-dimensional meteorological time-series data; All raw data, derived features, and decomposed trend and fluctuation terms are time-aligned and standardized to generate a unified input tensor for the core influenza prediction model.

9. The influenza prediction method according to any one of claims 1 to 8, characterized in that, The process of processing real-time data from the multi-source heterogeneous data using the trained influenza prediction core model includes: The dilated causal convolutional network of the multi-scale feature extraction network is used to extract short-term high-frequency fluctuation features in meteorological and influenza data. The temporal convolutional network of the multi-scale feature extraction network is used to extract mid-term trend features. The encoder of the multi-scale feature extraction network is used to capture long-term dependencies and periodic features; The gating attention mechanism of the multi-scale feature extraction network adaptively fuses the short-term high-frequency fluctuation features, medium-term trend features, and long-term dependence and periodic features.

10. An influenza prediction device based on multi-source data fusion and deep learning, characterized in that, include: The acquisition module is used to acquire multi-source heterogeneous data of multiple geographic units within the target area; the multi-source heterogeneous data includes influenza surveillance time-series data, multi-dimensional meteorological time-series data, static geographic attribute data, cross-regional population flow data, and Internet public behavior time-series data; The association module is used to determine the geographic adjacency matrix, meteorological similarity matrix, and functional association matrix based on the static geographic attribute data, the multi-dimensional meteorological time series data, and the cross-regional population flow data. The enhancement module is used to construct a spatiotemporal multi-topology dynamic graph neural network based on the learnable dynamic fusion results of the geographic adjacency matrix, meteorological similarity matrix and functional association matrix, and to use the spatiotemporal multi-topology dynamic graph neural network to model spatial dependencies and obtain spatial enhancement features. The training module is used to obtain a trained influenza prediction core model based on historical data in the multi-source heterogeneous data and the spatial enhancement features. The influenza prediction core model includes a multi-scale feature extraction network and a meta-learning adapter. The prediction module is used to process real-time data in the multi-source heterogeneous data through the trained influenza prediction core model, and output prediction results including point prediction and interval prediction of influenza incidence rate in the future multiple periods. The output module is used to generate dynamic risk assessment levels and corresponding prevention and control decision suggestions based on the prediction results and real-time external context data, using a preset rule engine.

Images