AI-assisted schistosomiasis intelligent blood prevention information monitoring and early warning management method and system

Abstract

This invention discloses an AI-assisted intelligent schistosomiasis control information monitoring, early warning, and management method and system, relating to the field of deep learning technology. The method includes: acquiring multi-source monitoring data through an intelligent data acquisition terminal network; constructing a multimodal spatiotemporal knowledge graph and a multi-timescale feature library; constructing a deep temporal network with enhanced spatiotemporal attention based on the multimodal spatiotemporal knowledge graph to capture the spatiotemporal evolution pattern of media propagation and generate predicted infection numbers for each grid unit; extracting historical weather features and infection number sequences from the multi-timescale feature library, calculating weighted baseline values ​​for weather-similar days and moving percentile anomaly reference values, and fusing them into a dynamic early warning baseline. This invention significantly improves prediction accuracy and interpretability in data-sparse areas by capturing key historical fragments through an attention mechanism in the temporal dimension and aggregating neighborhood propagation information through a graph structure in the spatial dimension.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and in particular to an AI-assisted intelligent schistosomiasis prevention and control information monitoring, early warning and management method and system. Background Technology

[0002] Schistosomiasis, a zoonotic parasitic disease that seriously endangers human health, is closely related to the spatiotemporal distribution and breeding environment of Oncomelania snails, as well as human activities. In recent years, deep learning technology has been widely used in the field of infectious disease surveillance and early warning. Existing methods mainly involve constructing long short-term memory networks or spatiotemporal graph convolutional networks to perform time-series modeling of historical case data, or using dynamic models to simulate the transformation process between susceptible individuals, infected individuals, and those who have migrated out, thereby achieving macro-level predictions of epidemic trends. In addition, some studies have attempted to incorporate multi-source data such as meteorological and population movement data into the prediction framework to improve the model's generalization ability.

[0003] However, existing technologies still have the following shortcomings: First, the transmission of schistosomiasis has a unique and complex biological vector chain. As the only intermediate host, the density, infection rate and spatial distribution of Oncomelania snails are the core factors that determine the transmission risk. However, existing prediction models mostly use population case data as a single input source and fail to effectively couple Oncomelania snail ecological monitoring data with population transmission dynamics models. Second, the parameter settings of traditional dynamic models rely on idealized assumptions and are difficult to adapt to complex and ever-changing natural environments. While pure data-driven deep learning methods can fit complex patterns, they lack explicit constraints on the physical laws of disease transmission, resulting in insufficient interpretability of prediction results and limited generalization ability in small sample or sparse data scenarios. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an AI-assisted intelligent schistosomiasis control information monitoring and early warning management method to solve the problem of the difficulty in accurately coupling and predicting the dynamics of Oncomelania snail breeding and the risk of human transmission in schistosomiasis control.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides an AI-assisted intelligent schistosomiasis prevention and control information monitoring and early warning management method, which includes, Multi-source monitoring data is acquired through an intelligent data acquisition terminal network to construct a multimodal spatiotemporal knowledge graph and a multi-timescale feature library; A deep temporal network with enhanced spatiotemporal attention is constructed based on a multimodal spatiotemporal knowledge graph to capture the spatiotemporal evolution pattern of media dissemination and generate the predicted number of infections for each grid unit. Based on a multi-timescale feature library, historical weather features and infection number sequences are extracted for the same period. Weighted baseline values ​​for weather-similar days and moving percentile anomaly reference values ​​are calculated and fused into a dynamic early warning baseline. Grid cells whose predicted infection numbers exceed the dynamic early warning baseline are marked as early warning triggering units and the extent of the exceedance is recorded. A multi-radius spatial scanning window is constructed with each early warning triggering unit as the center. The log-likelihood ratio of the predicted number of infected people within the window is calculated, and clustered windows are screened through Monte Carlo simulation. After merging adjacent windows, contiguous early warning areas and risk levels are generated, and a dynamic schistosomiasis transmission risk map is output.

[0007] As a preferred embodiment of the AI-assisted intelligent schistosomiasis control information monitoring and early warning management method described in this invention, the multi-source monitoring data includes snail image data, surface environment spectral data, microenvironment meteorological data, weather forecast data, water level and flow data, population flow heat map, outpatient and follow-up text data, and treatment action record data.

[0008] As a preferred embodiment of the AI-assisted intelligent schistosomiasis control information monitoring and early warning management method described in this invention, the construction of a multimodal spatiotemporal knowledge graph and a multi-timescale feature library specifically includes: The multi-source monitoring data are aligned according to the geographic grid code to form a basic feature vector set; Fine-grained recognition is performed on snail image data, and the density level of Oncomelania snails, the infection probability of Oncomelania snails and the occurrence of similar snails are output and written into the basic feature vector set; The basic feature vector set is organized into a multi-timescale feature library containing daily, weekly and monthly scales according to the timestamp, and a relationship graph containing spatial adjacency and human flow connection strength is constructed. By associating the node attributes of the relationship graph with the corresponding grid features in the multi-timescale feature library, a multimodal spatiotemporal knowledge graph is formed.

[0009] As a preferred embodiment of the AI-assisted intelligent schistosomiasis prevention and control information monitoring and early warning management method described in this invention, the construction of a deep temporal network with spatiotemporal dual attention enhancement specifically includes: The daily-scale features in the multi-timescale feature library are mapped to a temporal hidden state sequence to obtain the temporal memory representation of each time step. Temporal attention selection is performed on the temporal memory representation, and a temporal context vector is output. The temporal context vector is then converted into temporal gating coefficients for spatial attention, which are used to limit the effective time period of neighborhood propagation, resulting in a temporally gated neighborhood candidate set. Spatial attention propagation of neighborhood candidate sets is calculated along the spatial adjacency relations and human flow connection strength of the multimodal spatiotemporal knowledge graph, and spatial context vector is output. The spatial context vector is associated and matched with the temporal memory representation to generate a propagation cue vector. The propagation cue vector is then used to correct the temporal attention selection to strengthen the historical fragments related to diffusion, resulting in a coupled spatiotemporal dual attention representation. By concatenating temporal memory representations, temporal context vectors, and spatial context vectors, and completing the mapping output layer construction, a deep temporal network with enhanced spatiotemporal attention is obtained.

[0010] As a preferred embodiment of the AI-assisted intelligent schistosomiasis prevention and control information monitoring and early warning management method described in this invention, wherein: the generation of the predicted number of infected persons for each grid unit specifically comprises: By invoking a deep temporal network enhanced with spatiotemporal dual attention, temporal attention is converged on the temporal memory representation of each grid cell to obtain a list of key historical fragments; Perform spatial attention propagation on the temporally gated neighborhood candidate set and output the neighborhood contribution vector; The neighborhood contribution vectors are aggregated into a neighborhood propagation contribution list according to the neighborhood grid cells, and together with the list of key historical fragments, they form a spatiotemporal evolution evidence pair. The evidence of spatiotemporal evolution is mapped to the predicted number of infections in future time windows.

[0011] As a preferred embodiment of the AI-assisted intelligent schistosomiasis control information monitoring and early warning management method described in this invention, the generation of the dynamic early warning baseline specifically involves: Historical weather features and corresponding infection numbers for the same period across a preset historical span are extracted from a multi-timescale feature database to form a historical sample set for the same period. Calculate the similarity between the current weather characteristics and the historical sample set for the same period, filter the set of weather-similar days, and aggregate the number of infections corresponding to the weather-similar days according to the similarity coefficient to obtain the weighted benchmark value of weather-similar days; Extract the number of infected persons within a preset sliding window and calculate the moving percentile anomaly reference value to obtain the time-series anomaly reference quantity; Compare the weighted baseline value of similar weather days with the time-series anomaly reference value, and select the larger one as the dynamic early warning baseline.

[0012] As a preferred embodiment of the AI-assisted intelligent schistosomiasis prevention and control information monitoring and early warning management method described in this invention, wherein: the log-likelihood ratio of the predicted number of infected persons within the calculation window is specifically: Locate the center point of the grid for the early warning triggering unit on the electronic map, and generate a set of circular scanning windows with preset radii; The sum of the predicted number of infections for each grid cell covered by the scanning window is calculated, and the reference value of the number of infections corresponding to the population base of the window is calculated simultaneously to form a window statistical pair; The log-likelihood ratio statistic is calculated based on the sum of the predicted number of infections in the window statistics and the reference value of the number of infections in the population base, and then summarized into a scan window log-likelihood ratio sequence.

[0013] As a preferred embodiment of the AI-assisted intelligent schistosomiasis control information monitoring and early warning management method described in this invention, the step of screening the aggregation window through Monte Carlo simulation specifically involves: Keeping the set of grid cells covered by each scanning window unchanged, the predicted number of infected people in the grid cells within the scanning window is randomly rearranged in multiple rounds to form a random log-likelihood ratio distribution; The sequence of log-likelihood ratios of the scanning window is compared with the random log-likelihood ratio distribution window by window for significance, and non-significant windows are filtered out to generate a set of significant clustered windows.

[0014] As a preferred embodiment of the AI-assisted intelligent schistosomiasis prevention and control information monitoring and early warning management method described in this invention, the step of outputting a dynamic schistosomiasis transmission risk map specifically includes: Perform overlap and edge adjacency determination on the set of significantly clustered windows to obtain a set of contiguous candidate regions; Each contiguous set of candidate areas is converted into a regional boundary, and the number of covered grids, the total number of predicted infections, and the average exceedance are counted to form a regional risk assessment record. Based on regional risk assessment records, risk levels are classified into concern level, warning level, and high-risk level. Regional boundaries and risk levels are overlaid on an electronic map to output a dynamic schistosomiasis transmission risk map.

[0015] Secondly, this invention provides an AI-assisted intelligent schistosomiasis prevention and control information monitoring and early warning management system, comprising, The data acquisition module acquires multi-source monitoring data through an intelligent acquisition terminal network, and constructs a multimodal spatiotemporal knowledge graph and a multi-timescale feature library. The infection prediction module constructs a deep temporal network with enhanced spatiotemporal attention based on a multimodal spatiotemporal knowledge graph, captures the spatiotemporal evolution pattern of media transmission, and generates the predicted number of infections for each grid cell. The threshold generation module extracts historical weather features and infection number sequences based on a multi-timescale feature library, calculates the weighted baseline value of weather-similar days and the moving percentile anomaly reference value and merges them into a dynamic early warning baseline, and marks grid cells whose predicted number of infections exceeds the dynamic early warning baseline as early warning triggering units and records the extent of the exceedance. The risk output module constructs a multi-radius spatial scanning window centered on each early warning triggering unit, calculates the log-likelihood ratio of the predicted number of infected people within the window, filters clustered windows through Monte Carlo simulation, merges adjacent windows to generate contiguous early warning areas and risk levels, and outputs a dynamic schistosomiasis transmission risk map.

[0016] The beneficial effects of this invention are as follows: By constructing a multimodal spatiotemporal knowledge graph that includes snail infestation status, environmental background, and human activities, and using the fine-grained identification results of Oncomelania snails as core features, the problem of the separation between pathogen host monitoring data and human infection data is solved; the spatiotemporal dual attention-enhanced deep temporal network constructed based on the knowledge graph captures key historical fragments in the time dimension through an attention mechanism, and aggregates neighborhood propagation information in the spatial dimension through a graph structure, so that the prediction results not only conform to the data distribution pattern but also follow the epidemiological transmission mechanism, significantly improving the prediction accuracy and interpretability in data-sparse areas; by integrating weather similarity matching and moving percentile method to generate dynamic early warning baselines, the problem that fixed thresholds cannot adapt to seasonal fluctuations is overcome; by using multi-radius spatial scanning and Monte Carlo simulation to identify significant clustering windows, the prediction results are transformed into contiguous early warning areas and graded risk maps, providing spatially explicit decision-making basis for precise snail eradication and human intervention, realizing the leap from single-point identification to regional early warning. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of an AI-assisted intelligent schistosomiasis prevention and control information monitoring, early warning, and management method.

[0019] Figure 2 This is a schematic diagram of data acquisition and preprocessing.

[0020] Figure 3 This is a schematic diagram for propagation pattern recognition and prediction.

[0021] Figure 4 This is a diagram illustrating risk warning and output. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides an AI-assisted intelligent schistosomiasis prevention and control information monitoring and early warning management method, including the following steps: S1. Acquire multi-source monitoring data through an intelligent data acquisition terminal network, and construct a multimodal spatiotemporal knowledge graph and a multi-timescale feature library.

[0026] S1.1. Acquire multi-source monitoring data through an intelligent data acquisition terminal network. This network includes hyperspectral imaging equipment at fixed monitoring points, macro-photography smartphones equipped by grassroots snail inspectors, and drone inspection platforms equipped with multispectral sensors. It collects snail image data, surface environmental spectral data, and micro-environmental meteorological data in real time within each monitoring grid unit. Simultaneously, it accesses gridded weather forecast data from meteorological departments, water level and flow data from water conservancy departments, population flow heat maps from telecommunications operators, outpatient and follow-up text data from medical institution information systems, and treatment action records from schistosomiasis control departments via data interfaces. All acquired multi-source monitoring data are aligned according to a unified geographic grid code, ensuring that each data point corresponds to a specific monitoring grid unit, forming the original dataset.

[0027] S1.2. Vegetation index and soil moisture content are extracted from the surface environmental spectral data in the original dataset as environmental background features; temperature, humidity and precipitation are extracted from microenvironment meteorological data and weather forecast data as meteorological features; water area change rate and flow velocity are extracted from water level and flow data as hydrological features; inflow and outflow intensity are extracted from population flow heat maps as population activity features; symptom description keywords and visit frequency are extracted from outpatient and follow-up text data using natural language processing methods as health monitoring features; and the area and number of molluscicide sprayings are extracted from treatment action record data as intervention measure features. All features are arranged according to grid cells and collection time to construct a basic feature vector set.

[0028] S1.3. Scale and normalize the snail images in the original dataset to a uniform pixel size to obtain standardized snail images. Input the standardized snail images into an improved lightweight deep neural network deployed on edge computing nodes. The improved lightweight deep neural network is based on GhostNet, and channel attention and spatial attention mechanisms are set sequentially after the bottleneck layer of GhostNet. The bottleneck layer of GhostNet extracts features from the standardized snail images and outputs intermediate feature maps.

[0029] The channel attention mechanism performs global average pooling on the intermediate feature map to obtain the channel descriptor. After nonlinear transformation by two fully connected layers, it is multiplied with the intermediate feature map to generate the channel-weighted feature map. The spatial attention mechanism performs global average pooling and global max pooling on the channel-weighted feature map in the channel dimension. The pooling results are concatenated and then passed through a convolutional layer to generate spatial attention weights. The spatial attention weights are multiplied with the channel-weighted feature map to obtain the enhanced feature map.

[0030] The enhanced feature map is input into the classification layer of GhostNet. After global average pooling and fully connected layers, the probability values ​​of snail density level, schistosomiasis infection status, and easily confused similar snail species are calculated by the Softmax function. Based on the probability values, the snail density level, schistosomiasis infection probability, and occurrence of easily confused similar snail species in each grid cell are determined. The above results are written as snail situation features into the basic feature vector set of the corresponding grid cell and timestamp, forming a basic feature vector set containing complete feature information.

[0031] S1.4. The basic feature vector set containing complete feature information is categorized and organized by timestamp. For each monitoring grid unit, the daily feature vectors are arranged in date order to form a daily-scale feature vector sequence. The daily-scale feature vectors within each week are aggregated and statistically analyzed to form a weekly-scale feature vector sequence. The weekly-scale feature vectors within each month are aggregated and statistically analyzed to form a monthly-scale feature vector sequence. The daily-scale, weekly-scale, and monthly-scale feature vector sequences of all grid units are stored uniformly according to grid coding and time dimension to construct a multi-time-scale feature library.

[0032] S1.5. Construct a spatial adjacency graph based on the geographical location of the monitoring grid units. Treat each monitoring grid unit as a node. If two grid units have a common geographical grid boundary, establish an edge between the corresponding nodes to represent the geographical adjacency relationship. Assign a population flow connection strength value to the established edge based on the population flow intensity between each grid unit extracted from the population flow heat map, forming a relationship graph that includes spatial adjacency relationship and population flow connection strength.

[0033] Each node in the relational graph is associated with a corresponding grid cell in a multi-timescale feature library. The feature vector sequence of each grid cell at the daily, weekly, and monthly scales is stored as the node's attribute information, so that each node simultaneously possesses information about its neighboring nodes and its own feature evolution information at multiple time scales, ultimately forming a dynamically updatable multimodal spatiotemporal knowledge graph.

[0034] S2. Construct a deep temporal network with enhanced spatiotemporal attention based on a multimodal spatiotemporal knowledge graph to capture the spatiotemporal evolution pattern of media dissemination and generate the predicted number of infections for each grid cell.

[0035] S2.1. Extract the daily-scale feature vector sequence of each monitoring grid unit from the multi-timescale feature library, and input the daily-scale feature vector sequence into the bidirectional long short-term memory network encoder. The bidirectional long short-term memory network encoder consists of a forward long short-term memory network and a backward long short-term memory network. Extract the hidden state of each time step in forward and reverse time order, respectively. Concatenate the forward hidden state and the backward hidden state to obtain the temporal memory representation of the corresponding time step.

[0036] S2.2. Perform temporal attention selection on the temporal memory representation sequence of each grid cell, and compute at each time step The expression for the current predicted attention score is: ; in, Represents grid cells At time step The original attention score for the current prediction; This represents the learnable row vector used in the temporal attention mechanism to map hidden states to scalar scores; This represents the learnable weight matrix that performs a linear transformation on the input hidden state in the temporal attention mechanism. This represents the learnable bias in the time attention mechanism; It is a grid cell At time step Temporal memory representation; The hyperbolic tangent activation function is used. This is a transpose operation.

[0037] The attention weights are obtained by normalizing the attention scores at all time steps using the Softmax function, expressed as follows: : ; in, Represents grid cells At time step The normalized attention weights are used to measure the importance of the corresponding time step to the final prediction. Indicates the total length of the time series; To find the sum index, iterate from 1 to... All time steps.

[0038] The temporal memory representation is weighted and summed according to attention weights to obtain a temporal context vector. Simultaneously, the temporal context vector is input into a linear layer and converted into temporal gating coefficients using a sigmoid function, expressed as follows: ; in, Represents grid cells The time gating coefficient is used to filter historical time steps; The Sigmoid activation function compresses the output value to between 0 and 1. Map the temporal context vector to a scalar learnable row vector; It is a grid cell The time context vector; It is a learnable bias scalar for linear layers.

[0039] Historical time steps are selected based on the time gating coefficient. If the attention weight of a certain time step is greater than the time gating coefficient, the temporal memory representation of the current time step is retained in the neighborhood candidate set for subsequent spatial propagation.

[0040] S2.3. Along the spatial adjacency relationships and pedestrian connection strength recorded in the multimodal spatiotemporal knowledge graph, perform spatial attention propagation on the neighborhood candidate set of each grid cell. First, obtain all neighboring grid cells of the grid cell (i.e., grids that have geographically adjacent edges to the grid cell). For each neighboring grid cell, take the temporal memory representation of all selected time steps, and calculate the spatial attention score of the neighboring grid cell to the grid cell within the time step. The expression is: ; in, Represents grid cells Neighboring grid cells At time step Spatial attention score; It is a learnable row vector used in spatial attention mechanisms to map concatenated vectors to scalar fractions; It is a learnable weight matrix that performs a linear transformation on the input hidden state in the spatial attention mechanism; The current grid cell At time step Temporal memory representation; Neighboring grid cell At time step Temporal memory representation; This represents a vector concatenation operation; It is a connecting grid cell in a multimodal spatiotemporal knowledge graph. and grid cells The feature vector of the connection strength of the human flow attached to the edge.

[0041] The spatial attention scores of all neighbors at all time steps are normalized using the Softmax function to obtain the spatial attention weights. The temporal memory representations of the neighbors are then weighted and summed to obtain the spatial context vector.

[0042] S2.4. Associate and match the spatial context vector with the temporal memory representation of the grid cell at the latest time step, concatenate the two and input them into a linear layer and pass through Tanh activation to generate a propagation cue vector.

[0043] The propagation cue vector is used to correct the temporal attention selection. The propagation cue vector is used as the query vector to recalculate the corrected attention score for each time step, and then the corrected attention weight is obtained through Softmax, thus obtaining the corrected temporal context vector. At this time, the temporal memory representation, the corrected temporal context vector and the spatial context vector together constitute a coupled spatiotemporal dual attention representation.

[0044] S2.5. Concatenate the temporal memory representation, the corrected temporal context vector, and the spatial context vector to obtain a joint feature vector; input the joint feature vector into the fully connected layer, and output the predicted number of infected individuals for each grid cell within a future time window through a linear transformation, as expressed in: ; in, Represents grid cells The projected number of infections within a future time window is a scalar value. Represents grid cells The joint eigenvectors; This represents the learnable weight vector in a fully connected layer, along with its dimension and joint feature vector. same; This represents a learnable bias scalar in a fully connected layer, used to adjust the output value.

[0045] S3. Based on the multi-timescale feature library, extract historical weather features and infection number sequences for the same period, calculate the weighted baseline value of weather-similar days and the moving percentile anomaly reference value and merge them into a dynamic early warning baseline. Mark the grid cells where the predicted number of infections exceeds the dynamic early warning baseline as early warning triggering units and record the extent of the exceedance.

[0046] S3.1. Extract historical data from the multi-timescale feature library for each monitoring grid unit within the same calendar week as the current date over the past three years. The past three years are chosen as the time span because the three-year cycle can cover different meteorological year types, such as wet years, dry years, and extreme weather years, providing a sufficient sample basis for weather similarity matching, while avoiding environmental bias caused by outdated data. The historical data includes daily weather feature vectors and the corresponding actual number of infections. The weather feature vectors consist of values ​​of three dimensions: temperature, humidity, and precipitation. Arrange all extracted historical data in order of year and date to form the historical sample set for each grid unit.

[0047] For the monitoring grid units that require early warning judgment, the predicted number of infections for the corresponding grid unit within a future time window is obtained. At the same time, the weather feature vector of the corresponding grid unit for the current date is extracted from the multi-timescale feature library. The weather feature vector for the current date is also composed of values ​​of three dimensions: temperature, humidity, and precipitation. The similarity between the weather feature vector of the corresponding grid unit for the current date and the weather feature vector of each historical date in the historical sample set is calculated. The similarity is obtained by calculating the cosine distance between the two vectors. The larger the value of the cosine distance, the more similar the weather conditions of the two dates are.

[0048] S3.2. Based on the similarity calculation results, sort all historical dates in the historical same period sample set from high to low according to their similarity with the current date, and select the top N historical dates with the highest similarity as the weather similarity day set, for example, N is 10; obtain the actual number of infections corresponding to each historical date in the weather similarity day set, and use the similarity of each historical date as a weight coefficient to calculate the weighted average of the actual number of infections of the selected N historical dates to obtain the weighted benchmark value of the weather similarity day. The benchmark value reflects the average infection level of the grid cell in history under similar weather conditions.

[0049] The actual number of infections in the past four weeks was extracted from the multi-timescale feature library for the corresponding grid cells. The past four weeks were chosen as the sliding window because the incubation period of schistosomiasis is generally 4 to 6 weeks, and this time window can effectively capture the recent potential transmission trend. A total of 28 historical dates in the past four weeks were used, and the actual number of infections on these 28 historical dates were arranged in chronological order. The 90th percentile of the 28 infection number sequences was calculated using the moving percentile method. The 90th percentile was chosen as the outlier reference value because this value represents a high cutoff point in statistics, which can identify situations where the number of infections exceeds the normal fluctuation range, while avoiding false triggers caused by occasional increases on a single day. In the specific calculation, the 28 infection numbers were first sorted from smallest to largest, and the value corresponding to the 90th position after sorting was taken as the 90th percentile. This value is the time series outlier reference value, used to measure whether there has been an abnormal increase in the number of infections recently.

[0050] S3.3. Compare the weighted baseline value of similar weather days with the time-series anomaly reference value, and select the larger value of the two as the current dynamic warning baseline for the corresponding grid cell; the dynamic warning baseline is a dynamically adjusted threshold that can adapt to both seasonal fluctuations caused by weather changes and abnormal trends shown in recent data.

[0051] The predicted number of infections in a grid cell within a future time window is compared with the dynamic early warning baseline to determine whether the predicted number of infections exceeds a preset multiple of the dynamic early warning baseline. The preset multiple is 1.5 times, which is obtained based on backtesting analysis of past epidemic data. A multiple of 1.5 times can achieve a balance between sensitivity and specificity, providing timely early warning when the number of infections increases substantially, without frequent false alarms due to small seasonal fluctuations. If the predicted number of infections is greater than 1.5 times the dynamic early warning baseline, the corresponding grid cell is marked as an early warning triggering unit. At the same time, the difference between the predicted number of infections and the dynamic early warning baseline is calculated as the excess range, which is recorded for subsequent spatial clustering analysis. If the predicted number of infections does not exceed 1.5 times the dynamic early warning baseline, the corresponding grid cell is not marked, and the next grid cell is processed.

[0052] S4. Construct a multi-radius spatial scanning window centered on each early warning triggering unit, calculate the log-likelihood ratio of the predicted number of infected persons within the window, and filter clustered windows through Monte Carlo simulation. After merging adjacent windows, generate contiguous early warning areas and risk levels, and output a dynamic schistosomiasis transmission risk map.

[0053] S4.1. Locate the center point of the grid corresponding to each early warning triggering unit marked on the electronic map. Using each grid center point as the center, generate circular scanning windows with three radii of 1 km, 3 km, and 5 km respectively. The selection of 1 km, 3 km, and 5 km as radii is based on the spread ability of Oncomelania snails and on-site snail detection practices. 1 km corresponds to the common range of natural spread of Oncomelania snails, 3 km covers the typical distance of Oncomelania snails spread by water flow, and 5 km covers the potential range of Oncomelania snails that may be spread by human activities. The three radii together constitute a multi-scale detection system to identify clusters of different sizes. Each radii corresponds to a set of scanning windows, and the scanning windows generated by all radii together constitute a multi-radius spatial scanning window set. For each scanning window, record the grid code of all monitoring grid units covered to form a window coverage grid set.

[0054] For each scanning window, the predicted number of infections for all grid cells within the corresponding window's coverage grid set is obtained from the predicted number of infections for each grid cell. The predicted number of infections for all grid cells is summed to obtain the sum of the predicted number of infections for the corresponding scanning window. Simultaneously, based on the population base of each grid cell within the corresponding window's coverage grid set, a population base infection reference value is calculated for each grid cell. The population base infection reference value is equal to the product of the grid cell's population base and the background value of the overall population infection rate. The background value of the overall population infection rate is calculated based on the average infection rate of schistosomiasis in the same period over the past few years in this region. The population base infection reference values ​​for all grid cells within the window are summed to obtain the total population base infection reference value for the corresponding window. The sum of the predicted number of infections for each scanning window is combined with the total population base infection reference value to form a window statistical pair.

[0055] S4.2. Based on the sum of the predicted number of infections in each scan window's window statistics pair and the sum of the reference value of the number of infections in the population base, calculate the log-likelihood ratio statistic for the scan window. The expression is: ; in, Indicates the first Log-likelihood ratio statistic for each scan window; Indicates the first The sum of the predicted number of infections for each scanning window; Indicates the first The sum of the reference values ​​for the number of infected persons based on the population of each scanned window; This represents the total predicted number of infections across all monitored grid units.

[0056] The log-likelihood ratio statistics of all scan windows are summed to form the scan window log-likelihood ratio sequence.

[0057] Keeping the window coverage grid set of each scanning window unchanged, the predicted number of infections in each grid cell within the window coverage grid set is randomly rearranged, that is, the predicted number of infections in all grid cells is randomly swapped within the set. After rearrangement, the log-likelihood ratio statistic for each window is recalculated. The random rearrangement and calculation process is repeated one thousand times to obtain the set of log-likelihood ratio statistics for each scanning window under one thousand random simulations, which serves as the random log-likelihood ratio distribution.

[0058] To further explain, the choice of one thousand simulations is based on the accuracy requirements of Monte Carlo simulations. One thousand simulations can bring the generated random distribution to a stable state, while keeping the computational cost within an acceptable range. One thousand simulations is a commonly used number of simulations in epidemiological spatial scan statistics.

[0059] The log-likelihood ratio statistic of each scanning window is compared with the corresponding random log-likelihood ratio distribution. If the log-likelihood ratio statistic is greater than the value corresponding to the 95th percentile of the random log-likelihood ratio distribution, the scanning window is considered to have statistical clustering and is retained in the clustered window set. Otherwise, the scanning window is filtered out. The 95th percentile is chosen as the significance threshold because this value corresponds to a significance level of 0.05 in statistics, which can control the false positive clustering caused by random fluctuations to within 5%, and is the most commonly used confidence level in spatial scanning statistics.

[0060] S4.3. Determine the overlap and edge adjacency of all scanning windows in the aggregation window set. If the distance between the centers of two scanning windows is less than the sum of the radii of the two windows, or if there is an overlapping area at the boundaries of the two windows, then the two windows are determined to be adjacent windows. Merge all windows with adjacent relationships to form several contiguous candidate region sets. Each contiguous candidate region set consists of a set of interconnected or overlapping scanning windows.

[0061] Each contiguous candidate area set is converted into a contiguous warning area. During the conversion, the union of all scan windows covering the grid cells in the contiguous candidate area set is taken as the spatial range of the contiguous warning area, and the geographical boundary of the corresponding area is generated. The number of grid cells covered in each contiguous warning area, the sum of the predicted number of infected people in all grid cells, and the average of the exceedance of all grid cells are counted and combined to form the regional risk assessment record of the contiguous warning area.

[0062] Risk levels are determined based on the regional risk assessment records of each contiguous warning area. If the total number of predicted infections in the regional risk assessment records is lower than the 50th percentile of the total number of predicted infections in all contiguous warning areas, the corresponding contiguous warning area is classified as a concern level. If the total number of predicted infections is between the 50th percentile and the 90th percentile, it is classified as a warning level. If the total number of predicted infections is higher than the 90th percentile, it is classified as a high-risk level.

[0063] To further explain, the 50th percentile was chosen as the dividing line between the concern level and the warning level because this value represents the median risk level of all contiguous warning areas, which can filter out areas with a risk level higher than half of the areas; the 90th percentile was chosen as the dividing line between the warning level and the high-risk level because this value represents areas with a risk level in the top 10%, and these areas require immediate and intensive intervention measures.

[0064] The geographical boundaries of each contiguous warning area and its corresponding risk level are overlaid on the electronic map. At the same time, key monitoring points within the area are marked on the map. These key monitoring points include historical snail breeding waters and densely populated areas, ultimately generating a dynamic schistosomiasis transmission risk map.

[0065] This embodiment also provides an AI-assisted intelligent schistosomiasis prevention and control information monitoring and early warning management system, including: The data acquisition module acquires multi-source monitoring data through an intelligent acquisition terminal network, and constructs a multimodal spatiotemporal knowledge graph and a multi-timescale feature library. The infection prediction module constructs a deep temporal network with enhanced spatiotemporal attention based on a multimodal spatiotemporal knowledge graph, captures the spatiotemporal evolution pattern of media transmission, and generates the predicted number of infections for each grid cell. The threshold generation module extracts historical weather features and infection number sequences based on a multi-timescale feature library, calculates the weighted baseline value of weather-similar days and the moving percentile anomaly reference value and merges them into a dynamic early warning baseline, and marks grid cells whose predicted number of infections exceeds the dynamic early warning baseline as early warning triggering units and records the extent of the exceedance. The risk output module constructs a multi-radius spatial scanning window centered on each early warning triggering unit, calculates the log-likelihood ratio of the predicted number of infected people within the window, filters clustered windows through Monte Carlo simulation, merges adjacent windows to generate contiguous early warning areas and risk levels, and outputs a dynamic schistosomiasis transmission risk map.

[0066] In summary, this invention addresses the disconnect between pathogen host monitoring data and human infection data by constructing a multimodal spatiotemporal knowledge graph encompassing snail infestation status, environmental background, and human activity, using fine-grained snail identification results as core features. Based on the knowledge graph, a spatiotemporal dual-attention enhanced deep temporal network captures key historical fragments through an attention mechanism in the temporal dimension and aggregates neighborhood propagation information through a graph structure in the spatial dimension. This ensures that the prediction results conform to both data distribution patterns and epidemiological transmission mechanisms, significantly improving prediction accuracy and interpretability in data-sparse areas. Furthermore, by fusing weather similarity matching and moving percentile methods to generate dynamic early warning baselines, the invention overcomes the problem of fixed thresholds failing to adapt to seasonal fluctuations. Finally, by employing multi-radius spatial scanning and Monte Carlo simulation to identify significant clustering windows, the prediction results are transformed into contiguous early warning areas and graded risk maps, providing a spatially explicit decision-making basis for precise snail eradication and human intervention, achieving a leap from single-point identification to regional early warning.

[0067] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An AI-assisted schistosomiasis intelligent blood prevention information monitoring and early warning management method, characterized in that: include, Multi-source monitoring data is acquired through an intelligent data acquisition terminal network to construct a multimodal spatiotemporal knowledge graph and a multi-timescale feature library; A deep temporal network with enhanced spatiotemporal attention is constructed based on a multimodal spatiotemporal knowledge graph to capture the spatiotemporal evolution pattern of media dissemination and generate the predicted number of infections for each grid unit. Based on a multi-timescale feature library, historical weather features and infection number sequences are extracted for the same period. Weighted baseline values ​​for weather-similar days and moving percentile anomaly reference values ​​are calculated and fused into a dynamic early warning baseline. Grid cells whose predicted infection numbers exceed the dynamic early warning baseline are marked as early warning triggering units and the extent of the exceedance is recorded. A multi-radius spatial scanning window is constructed with each early warning triggering unit as the center. The log-likelihood ratio of the predicted number of infected people within the window is calculated, and clustered windows are screened through Monte Carlo simulation. After merging adjacent windows, contiguous early warning areas and risk levels are generated, and a dynamic schistosomiasis transmission risk map is output.

2. The AI-assisted schistosomiasis intelligent blood prevention information monitoring and early warning management method according to claim 1, characterized in that: The multi-source monitoring data includes snail image data, surface environment spectral data, microenvironment meteorological data, weather forecast data, water level and flow data, population flow heat map, outpatient and follow-up text data, and treatment action record data.

3. The AI-assisted schistosomiasis intelligent blood prevention information monitoring and early warning management method according to claim 2, characterized in that: The construction of the multimodal spatiotemporal knowledge graph and multi-timescale feature library specifically includes: The multi-source monitoring data are aligned according to the geographic grid code to form a basic feature vector set; Fine-grained recognition is performed on snail image data, and the density level of Oncomelania snails, the infection probability of Oncomelania snails and the occurrence of similar snails are output and written into the basic feature vector set; The basic feature vector set is organized into a multi-timescale feature library containing daily, weekly and monthly scales according to the timestamp, and a relationship graph containing spatial adjacency and human flow connection strength is constructed. By associating the node attributes of the relationship graph with the corresponding grid features in the multi-timescale feature library, a multimodal spatiotemporal knowledge graph is formed.

4. The AI-assisted schistosomiasis intelligent blood prevention information monitoring and early warning management method according to claim 3, characterized in that: The construction of the deep temporal network with spatiotemporal dual attention enhancement is specifically as follows: The daily-scale features in the multi-timescale feature library are mapped to a temporal hidden state sequence to obtain the temporal memory representation of each time step. Temporal attention selection is performed on the temporal memory representation, and a temporal context vector is output. The temporal context vector is then converted into temporal gating coefficients for spatial attention, which are used to limit the effective time period of neighborhood propagation, resulting in a temporally gated neighborhood candidate set. Spatial attention propagation of neighborhood candidate sets is calculated along the spatial adjacency relations and human flow connection strength of the multimodal spatiotemporal knowledge graph, and spatial context vector is output. The spatial context vector is associated and matched with the temporal memory representation to generate a propagation cue vector. The propagation cue vector is then used to correct the temporal attention selection to strengthen the historical fragments related to diffusion, resulting in a coupled spatiotemporal dual attention representation. By concatenating temporal memory representations, temporal context vectors, and spatial context vectors, and completing the mapping output layer construction, a deep temporal network with enhanced spatiotemporal attention is obtained.

5. The AI-assisted schistosomiasis intelligent blood prevention information monitoring and early warning management method according to claim 4, characterized in that: The process of generating the predicted number of infections for each grid cell is as follows: By invoking a deep temporal network enhanced with spatiotemporal dual attention, temporal attention is converged on the temporal memory representation of each grid cell to obtain a list of key historical fragments; Perform spatial attention propagation on the temporally gated neighborhood candidate set and output the neighborhood contribution vector; The neighborhood contribution vectors are aggregated into a neighborhood propagation contribution list according to the neighborhood grid cells, and together with the list of key historical fragments, they form a spatiotemporal evolution evidence pair. The evidence of spatiotemporal evolution is mapped to the predicted number of infections in future time windows.

6. The AI-assisted intelligent schistosomiasis control information monitoring and early warning management method as described in claim 3, characterized in that: The generation of the dynamic early warning baseline is specifically as follows: Historical weather features and corresponding infection numbers for the same period across a preset historical span are extracted from a multi-timescale feature database to form a historical sample set for the same period. Calculate the similarity between the current weather characteristics and the historical sample set for the same period, filter the set of weather-similar days, and aggregate the number of infections corresponding to the weather-similar days according to the similarity coefficient to obtain the weighted benchmark value of weather-similar days; Extract the number of infected persons within a preset sliding window and calculate the moving percentile anomaly reference value to obtain the time-series anomaly reference quantity; Compare the weighted baseline value of similar weather days with the time-series anomaly reference value, and select the larger one as the dynamic early warning baseline.

7. The AI-assisted intelligent schistosomiasis control information monitoring and early warning management method as described in claim 5, characterized in that: The log-likelihood ratio of the predicted number of infections within the calculation window is specifically as follows: Locate the center point of the grid for the early warning triggering unit on the electronic map, and generate a set of circular scanning windows with preset radii; The sum of the predicted number of infections for each grid cell covered by the scanning window is calculated, and the reference value of the number of infections corresponding to the population base of the window is calculated simultaneously to form a window statistical pair; The log-likelihood ratio statistic is calculated based on the sum of the predicted number of infections in the window statistics and the reference value of the number of infections in the population base, and then summarized into a scan window log-likelihood ratio sequence.

8. The AI-assisted intelligent schistosomiasis control information monitoring and early warning management method as described in claim 7, characterized in that: The filtering of the clustering window through Monte Carlo simulation is specifically as follows: Keeping the set of grid cells covered by each scanning window unchanged, the predicted number of infected people in the grid cells within the scanning window is randomly rearranged in multiple rounds to form a random log-likelihood ratio distribution; The sequence of log-likelihood ratios of the scanning window is compared with the random log-likelihood ratio distribution window by window for significance, and non-significant windows are filtered out to generate a set of significant clustered windows.

9. The AI-assisted intelligent schistosomiasis control information monitoring and early warning management method as described in claim 8, characterized in that: The output of the dynamic schistosomiasis transmission risk map is specifically as follows: Perform overlap and edge adjacency determination on the set of significantly clustered windows to obtain a set of contiguous candidate regions; Each contiguous set of candidate areas is converted into a regional boundary, and the number of covered grids, the total number of predicted infections, and the average exceedance are counted to form a regional risk assessment record. Based on regional risk assessment records, risk levels are classified into concern level, warning level, and high-risk level. Regional boundaries and risk levels are overlaid on an electronic map to output a dynamic schistosomiasis transmission risk map.

10. An AI-assisted intelligent schistosomiasis control information monitoring and early warning management system, based on the AI-assisted intelligent schistosomiasis control information monitoring and early warning management method according to any one of claims 1 to 9, characterized in that: include, The data acquisition module acquires multi-source monitoring data through an intelligent acquisition terminal network, and constructs a multimodal spatiotemporal knowledge graph and a multi-timescale feature library. The infection prediction module constructs a deep temporal network with enhanced spatiotemporal attention based on a multimodal spatiotemporal knowledge graph, captures the spatiotemporal evolution pattern of media transmission, and generates the predicted number of infections for each grid cell. The threshold generation module extracts historical weather features and infection number sequences based on a multi-timescale feature library, calculates the weighted baseline value of weather-similar days and the moving percentile anomaly reference value and merges them into a dynamic early warning baseline, and marks grid cells whose predicted number of infections exceeds the dynamic early warning baseline as early warning triggering units and records the extent of the exceedance. The risk output module constructs a multi-radius spatial scanning window centered on each early warning triggering unit, calculates the log-likelihood ratio of the predicted number of infected people within the window, filters clustered windows through Monte Carlo simulation, merges adjacent windows to generate contiguous early warning areas and risk levels, and outputs a dynamic schistosomiasis transmission risk map.

Images