Public health intelligent management system based on artificial intelligence

Abstract

This invention relates to the field of public health information processing technology and discloses an artificial intelligence-based intelligent public health management system. This system acquires historical medical record text when clinical data in a spatial grid exceeds its boundaries, and generates an exposure semantic payload scalar and sparse state identifier through natural language semantic extraction. Based on this identifier, it performs bidirectional routing control, generating a constrained spatial and temporal computational domain for the tracing chain. Within the constrained boundaries of the computational domain, it calculates the maximum cross-correlation coefficient extreme value between the computational environment and the clinical sequence data. Combining the identifier, it calculates the comprehensive confidence score of the evaluation formula by introducing an automatic decay penalty mechanism, and generates physical intervention commands to drive ventilation or disinfection equipment when the score exceeds the limit. This invention reduces useless computational overhead through dynamic pruning of the computational domain and uses arithmetic penalty decay to filter out model false alarms caused by missing text data, achieving precise linkage intervention of the physical environment.

Description

Technical Field

[0001] This invention relates to the field of public health information processing technology, specifically to a public health intelligent management system based on artificial intelligence. Background Technology

[0002] Existing public health early warning and management systems primarily rely on collecting regional clinical syndrome data and comparing it one-way with predefined baseline thresholds to trigger alarms. This system's operational logic fails to consider the multidimensional spatiotemporal correlation between pathogenic environmental sequences and patient disease progression sequences. Directly employing indiscriminate cross-correlation operations within a global spatial grid to analyze these correlations would result in excessive computational resource consumption and reduced data processing efficiency.

[0003] Furthermore, when tracing a patient's historical exposure trajectory, the system needs to parse clinical outpatient medical records. In practical applications, medical records are often unstructured data and frequently suffer from missing spatiotemporal entity records. Existing correlation models lack mechanisms for identifying and penalizing sparse text states, and when processing medical records with missing features, they cannot differentiate processing paths based on data quality. This leads to the algorithm model easily assigning unreasonable weights to invalid records in spatiotemporal cross-correlation calculations, resulting in biased comprehensive risk scores and false alarms. Ultimately, based on these false alarms, the system issues incorrect intervention commands to the underlying equipment management platform, causing ineffective operation of building ventilation or disinfection equipment within the target area. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an artificial intelligence-based intelligent public health management system, which solves the problem that existing intelligent early warning systems for public health lack a hierarchical processing mechanism to distinguish between complete data with clear epidemiological trajectories and sparse data lacking spatiotemporal information when processing unstructured clinical medical record data.

[0005] The first aspect of this invention provides an artificial intelligence-based intelligent management system for public health, the system comprising: The data access module is used to continuously acquire environmental sequence data and clinical syndrome reporting data from various spatial grids within the target monitoring area; The event wake-up module is used to trigger event wake-up and obtain the historical medical record text of the associated patient group when the clinical syndrome reporting data of a certain spatial grid exceeds a preset baseline threshold. The semantic extraction module is used to perform natural language semantic extraction on the historical medical record text, and generate an exposed semantic payload scalar and sparse state identifier for each patient's tracing chain. The routing control module is used to perform bidirectional routing control based on the sparse state identifier, and generate restricted spatial and temporal computation domains for each of the patient tracing chains. The time series calculation module is used to calculate the maximum cross-correlation coefficient extreme value between the environmental sequence data and the clinical sequence data within the constraint boundary between the spatial calculation domain and the time calculation domain. The early warning decision module is used to establish a comprehensive evaluation formula with an automatic decay penalty mechanism based on the exposed semantic payload scalar, sparse state identifiers and the maximum cross-correlation extremum to calculate the regional comprehensive confidence score, and generate a physical intervention command when the regional comprehensive confidence score exceeds the limit.

[0006] Furthermore, the semantic extraction module includes a feature extraction unit and a mapping quantization unit. The feature extraction unit is used to extract spatiotemporal entity features from medical record text using a pre-trained natural language model. The spatiotemporal entity features include at least location, time, behavior category, and dwell time. The mapping quantization unit is used to perform lookup mapping and weighted operations on the extracted spatiotemporal entity features and a preset public health risk exposure dictionary to obtain the original semantic payload calculation value. If the feature extraction unit fails to extract valid spatiotemporal entity features, the mapping quantization unit forcibly overwrites the original semantic payload calculation value of the tracing chain to zero and sets the corresponding sparse state identifier to 1; otherwise, it sets the sparse state identifier to 0. The above structure, through overwriting and marking mechanisms, realizes the underlying marking of missing states in text data.

[0007] Furthermore, the semantic extraction module also includes a conflict-prevention normalization unit. This unit obtains the theoretical maximum and minimum values ​​of the preset behavioral risk and time coefficients in the risk exposure dictionary as global absolute mapping boundaries. It then uses an extreme value normalization algorithm to linearly map the original semantic payload calculation values ​​to a continuous interval based on these global absolute mapping boundaries, generating a standardized exposure semantic payload scalar. This structure uses the theoretical physical boundaries of the dictionary as the mapping benchmark, avoiding the problem of effective low-risk features being mistakenly calculated as zero due to batch data differences.

[0008] Furthermore, the routing control module includes an internal state machine and a bidirectional routing pruning unit. The internal state machine is used to determine the value of the sparse state identifier for a specific patient. When the sparse state identifier is 0, the bidirectional routing pruning unit switches to the main branch, converging the spatial computation domain to the exposure coordinate grid extracted from the patient's historical medical records, and generating a dynamic time search interval as the time computation domain by combining the error buffer constant. When the sparse state identifier is 1, the bidirectional routing pruning unit switches to the degraded bypass branch, expanding the spatial computation domain to the set of adjacent grids anchored to the patient's residence in the health record, and degrading the time computation domain to a preset static maximum latency interval. Through the above bidirectional routing structure, spatial scaling and computational load control of the computation domain for different data integrity levels are achieved.

[0009] Furthermore, the time-series calculation module is specifically used to calculate the Pearson cross-correlation coefficient between the environmental data sequence and the disease data sequence within the spatial calculation domain only within the determined time calculation domain boundary; retrieve and extract the maximum value in the coefficient set as the maximum cross-correlation coefficient maximum value for a specific patient individual; perform a non-negative truncation operation on the extracted maximum value to obtain the final maximum cross-correlation coefficient extreme value, and simultaneously extract the optimal lag step size corresponding to obtaining the maximum value.

[0010] Furthermore, the comprehensive evaluation formula established in the early warning decision module is specifically expressed as follows: An arithmetic multiplier term containing a specific constant penalty coefficient and sparse state identifiers is calculated. This arithmetic multiplier term is used to perform weighted processing on the extreme value of the maximum cross-correlation coefficient. Then, a linear accumulation is performed by combining the weight constant and the exposed semantic load scalar to obtain the comprehensive confidence score for the source tracing chain of a single patient. This formula structure, by introducing a difference multiplier term, achieves automatic attenuation of the calculation results for sparse text source tracing records.

[0011] Furthermore, the early warning decision module is also connected to the underlying IoT platform. The early warning decision module is used to summarize the comprehensive confidence scores of all related patient tracing chains within the same spatial grid, and perform extreme value aggregation or mean aggregation to obtain the regional comprehensive risk score of the grid. When the score is greater than or equal to the preset high-risk control threshold, a structured alarm data packet is generated and sent to the underlying IoT platform corresponding to the target spatial grid to increase the ventilation load of the target building area and start the associated disinfection hardware.

[0012] A second aspect of this invention provides a public health intelligent early warning method based on artificial intelligence, applied to the public health intelligent management system, the method comprising the following steps: Acquire environmental sequence data and clinical syndrome reporting data of each spatial grid within the target monitoring area, and trigger event wake-up to obtain historical medical record text of the associated patient group when the data exceeds the boundary; Natural language semantic extraction is performed on the historical medical record text to generate an exposed semantic payload scalar and sparse state identifiers; Bidirectional routing control is performed based on the sparse state identifiers to generate a restricted spatial computation domain and a time computation domain. Within the constrained boundaries of the computational domain, calculate the maximum cross-correlation coefficient extreme value between environmental sequence data and clinical sequence data; A comprehensive evaluation formula with an automatic decay penalty mechanism is established to calculate the regional comprehensive confidence score and generate physical intervention instructions when the score exceeds the limit.

[0013] Furthermore, the steps for generating the exposed semantic payload scalar and sparse state identifier include: extracting spatiotemporal entity features from the medical record text and performing dictionary lookup mapping to obtain the original semantic payload calculation value; if no valid entity features are extracted, the original calculation value is forcibly overwritten to zero and the sparse state identifier is set to 1, otherwise it is set to 0.

[0014] Furthermore, after obtaining the original semantic payload calculation value, a conflict-prevention normalization step is also included: using the preset extreme values ​​of the public health risk exposure dictionary as the global absolute mapping boundary, the original semantic payload calculation value is linearly mapped.

[0015] Furthermore, the steps for performing bidirectional routing control include: if the sparse state identifier is 0, then switch to the main route branch, converge the spatial computation domain to the exposed coordinate grid, and generate a dynamic time search interval; if the sparse state identifier is 1, then switch to the degraded bypass branch, expand the spatial computation domain to the residential and adjacent grid sets, and apply the static extreme time interval.

[0016] Furthermore, the step of calculating the maximum cross-correlation coefficient extremum includes: calculating the set of cross-correlation coefficients within the defined boundaries, extracting the maximum value of the current patient individual and performing non-negative truncation, and outputting the local optimal extremum with individual attributes.

[0017] Furthermore, the calculation of the regional comprehensive confidence score includes a penalty multiplier operation: when the corresponding data link has sparse features, a preset arithmetic multiplier term is used to perform forced arithmetic decay on the extracted spatiotemporal cross-correlation extreme values.

[0018] Furthermore, when the regional comprehensive risk score exceeds the limit, the physical intervention command is issued to the building equipment management platform to drive the operation of mechanical ventilation or hardware disinfection equipment.

[0019] A third aspect of the present invention provides an electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the above-described intelligent public health early warning method.

[0020] This invention provides an intelligent public health management system based on artificial intelligence. It has the following beneficial effects: 1. This invention effectively reduces the false alarm rate caused by missing underlying data by introducing sparse state identifiers and an automatic decay penalty mechanism into the early warning assessment formula. When the natural language model cannot extract complete spatiotemporal entity features from historical medical record texts, the system triggers a penalty multiplier term through the identifier, performing forced arithmetic decay on the extreme value of the calculated maximum cross-correlation coefficient. This mechanism directly transforms the data quality defects of unstructured text into reduced weights for relevance assessment, avoiding the system triggering erroneous physical intervention commands due to individual traceable records with no reference value, thus ensuring the objectivity of regional risk scoring.

[0021] 2. This invention utilizes a bidirectional routing control mechanism based on sparse state identifiers to achieve dynamic optimization of system computational overhead and computational domain pruning. The system performs branch calculations based on the data completeness of the medical record text. For data chains with accurate spatiotemporal characteristics, its spatial and temporal computational domains are strictly converged within the actual exposed grid and dynamic search interval; for data chains with missing features, a degraded bypass branch containing the residential adjacent grid is used. This structure, which adaptively scales the spatiotemporal boundaries according to the underlying data state, directly eliminates invalid cross-correlation operations for irrelevant grids, improving the efficiency of environmental and clinical sequence data comparison and processing.

[0022] 3. This invention employs a normalization processing logic based on the extreme values ​​of the risk exposure dictionary theory, ensuring the numerical stability of semantic feature quantization mapping. When generating the exposure semantic payload scalar, the system uses the global absolute mapping boundary as a benchmark, replacing the conventional mapping rules that rely on the extreme values ​​of the input batch data. This technical feature avoids the defect that, when the overall risk of local batch data is low, truly effective low-risk exposure features are incorrectly mapped and erased to zero values ​​by the algorithm, ensuring that the scalar parameters input to the time-series calculation module have a consistent and fixed reference coordinate system. Attached Figure Description

[0023] Figure 1 This is an architecture diagram of a spatiotemporal multi-source cascaded early warning system based on a medical record semantic tensor-driven dynamic computational domain, according to an embodiment of the present invention. Figure 2This is a flowchart of a spatiotemporal multi-source cascaded early warning method based on a medical record semantic tensor-driven dynamic computational domain, according to an embodiment of the present invention. Figure 3 This is a schematic diagram illustrating the spatiotemporal reference alignment and event wake-up principle of an embodiment of the present invention; Figure 4 This is a flowchart illustrating the generation and verification of exposed semantic tensors according to an embodiment of the present invention. Figure 5 This is a schematic diagram illustrating the dynamic computing domain bidirectional routing and generation principle of an embodiment of the present invention. Figure 6 This is a schematic diagram illustrating the principle of dynamic hysteresis cross-correlation calculation within the restricted search boundary of an embodiment of the present invention. Figure 7 This is a schematic diagram illustrating the adaptive penalty decision and cascaded instruction output principle of an embodiment of the present invention; Figure 8 This is a diagram illustrating the dynamic pruning effect of the spatial computing domain in this invention. Figure 9 This is a graph showing the dynamic lag cross-correlation solution within the limited time boundary of an embodiment of the present invention. Figure 10 This is a diagram showing the hedging relationship between the comprehensive performance of three types of algorithms in embodiments of the present invention.

[0024] Among them, 10 is the data access module; 20 is the semantic extraction module; 30 is the routing control module; 40 is the time series calculation module; and 50 is the early warning decision module. Detailed Implementation

[0025] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] See attached document Figure 1 , Figure 1 This is an architecture diagram of a spatiotemporal multi-source cascaded early warning system based on a dynamic computational domain driven by medical record semantic tensors, according to an embodiment of the present invention. The present invention provides an artificial intelligence-based public health intelligent management system, which may include: a data access module 10, a semantic extraction module 20, a routing control module 30, a time-series computation module 40, and an early warning decision module 50.

[0027] The data access module 10 communicates with external environmental monitoring nodes and the medical institution's business system. The data access module 10 is used to receive environmental monitoring time series data and clinical medical record text data, perform data cleaning and spatiotemporal benchmark alignment operations, generate standardized grid data, and load it into the system storage stack.

[0028] The semantic extraction module 20 is connected to the data access module 10 via a data bus. The semantic extraction module 20 is equipped with a pre-trained natural language processing model, used to extract location entities, time entities, and behavioral predicates from clinical medical record text, quantize and calculate the exposure semantic payload scalar, and encapsulate it into a structured exposure tensor. The semantic extraction module 20 is used to verify the completeness of the exposure tensor and output the corresponding Boolean sparse state identifier.

[0029] The routing control module 30 receives the exposure tensor and sparse state identifier output by the semantic extraction module 20. Based on the value of the sparse state identifier, the routing control module 30 selects to execute either a dynamic pruning algorithm or a topology constant degradation compensation algorithm to generate a set of spatial computational domains and a time computational domain interval for use by lower-level computing units.

[0030] The time series calculation module 40 is connected to the routing control module 30 and receives the set of spatial calculation domains and the interval of time calculation domains output by the routing control module 30. Within the received time and spatial boundary constraints, the time series calculation module 40 is used to retrieve the environmental variable series and the clinical variable series, perform cross-correlation function calculation, and search for and extract the maximum cross-correlation coefficient and the corresponding lag time difference.

[0031] The early warning decision module 50 receives the semantic payload data output by the semantic extraction module 20 and the maximum cross-correlation coefficient output by the time series calculation module 40. The early warning decision module 50 executes a comprehensive confidence evaluation algorithm that incorporates an adaptive penalty mechanism, and generates a structured alarm command signal when the evaluation score reaches a set condition.

[0032] See attached document Figure 2 , Figure 2 This is a flowchart of a spatiotemporal multi-source cascaded early warning method based on a medical record semantic tensor-driven dynamic computational domain, according to an embodiment of the present invention. The present invention provides an artificial intelligence-based intelligent public health management method, comprising the following steps: S100 establishes a spatiotemporal discretized coordinate matrix for the target region, receives multi-source heterogeneous data streams, performs time step alignment and normalization processing, and when the clinical outcome sequence within a specific spatial grid reaches the awakening condition, it extracts the associated patient medical record dataset within the target time window and sends it to the subsequent processing link. S200 receives the associated patient medical record dataset, performs joint entity and relation extraction to extract time, location, and behavior triples, combines a dictionary table to quantize and calculate the exposed semantic payload, encapsulates it to form an exposure tensor, and assigns Boolean sparse state identifiers based on the extraction of time and space entities in the tensor. S300 reads sparse state identifiers and controls the computation flow to perform bidirectional routing. When the identifier indicates that the data is complete, a dynamic time computation domain is generated based on a specific timestamp in the tensor. When the identifier indicates that the data is sparse and missing, a degraded time computation domain is generated using a preset default topological coordinate set and a physical epidemiological latency constant. S400 receives the output spatial and temporal computational domain parameters, calculates the dynamic lag cross-correlation function between the preceding environmental variable sequence and the clinical outcome sequence within the boundary range of parameter constraints, and extracts the maximum pathogenic cross-correlation value under the current analysis window; S500 aggregates the cumulative semantic payload value and the maximum cross-correlation value, calculates the comprehensive early warning confidence score with a penalty attenuation term, compares it with the multi-level threshold matrix to generate the final instruction control signal, and sends it to the external collaborative processing platform for execution and scheduling.

[0033] See attached document Figure 3 , Figure 3 This is a schematic diagram illustrating the spatiotemporal reference alignment and event wake-up principle according to an embodiment of the present invention. In this embodiment, a spatiotemporal discretized coordinate matrix of the target region is established, multi-source heterogeneous data streams are received, time step alignment and normalization processing are performed, and when the clinical outcome sequence within a specific spatial grid reaches the wake-up condition, the associated patient medical record dataset within the target time window is extracted and sent to the subsequent processing link. Specifically, this can be implemented through the following steps: The data access module 10 divides the target monitoring area into a spatial grid with fixed resolution and establishes a coordinate index matrix based on the geographic coordinate system.

[0034] The data access module 10 acquires the geographical boundary information of the target area and divides it into discrete blocks in the form of a two-dimensional matrix according to the set spatial resolution. Let the target monitoring area be divided into... Individual block units form a spatial grid set. Each grid cell corresponds to a unique spatial index code. As a preferred approach, the data access module 10 receives the physical location information of equipment from external environmental monitoring nodes and the medical institution's business system, and assigns it to the corresponding spatial grid cell based on the equipment's latitude and longitude coordinates. For the mapping method of converting latitude and longitude coordinates into a spatial grid index matrix, those skilled in the art can use conventional grid coding algorithms or latitude and longitude equal-division projection methods, which are well-known technologies in the field and will not be elaborated upon here.

[0035] The data access module 10 sets a uniform time step, downsamples or interpolates the environmental monitoring time series data and the clinical outcome variable time series to align the clock, and performs numerical normalization processing.

[0036] Because the sampling frequency of IoT sensors in the external environment differs from the reporting frequency of the medical institution's business system, directly performing numerical combination calculations can easily lead to misalignment in the time dimension. Therefore, the data access module 10 establishes a unified data analysis timeline, sets the system's unified time step to a fixed period, and defines the current system's evaluation time node as... The data access module 10 downsamples and aggregates the high-frequency environmental monitoring data according to the time step to generate a grid. Pre-environment variable sequence Simultaneously acquire grids at the same time step. Clinical outcome variable sequence .

[0037] To eliminate the impact of dimensional differences between different types of sensors on subsequent calculations—for example, the numerical scales of water turbidity and airborne particulate matter concentration are significantly different—this module maps heterogeneous physical observations to a dimensionless, unified feature space. The data access module 10 uses an extreme value normalization algorithm to perform a linear transformation on the preceding environmental variable sequence, generating a standardized sequence. Considering that continuous constant sensor outputs within a specific period may lead to algorithm execution errors with a denominator of zero, this embodiment introduces a zero-division prevention parameter in the computational logic. The specific normalization calculation formula is as follows: ; In the formula, Representing spatial grid At the time point Normalized environmental parameter calculation values, These are the raw environmental parameter observations recorded by the sensors. and These represent the minimum and maximum statistical values ​​of this category of environmental parameters within a preset historical observation period (e.g., set to the past 30 baseline time steps), respectively. A buffer term of a positive constant that approaches zero, set for the system.

[0038] The data access module 10 sets a dynamic trigger threshold. When the observed value of the clinical outcome variable exceeds the set wake-up condition, it extracts the set of related patient medical record texts within the target time window.

[0039] To avoid overloading computing power due to continuous high-concurrency operations on global data, this solution relies on the fluctuation of specific indicators to build an underlying event-driven mechanism. Data access module 10 computes spatial grid. Mean of clinical outcome observations over a fixed historical observation period with standard deviation Construct a statistical baseline threshold for this grid. Generally speaking, the logic for establishing this baseline threshold can be expressed as follows: In the formula The sensitivity confidence adjustment coefficient configured for the system typically ranges from 1.5 to 3.0. When the data access module 10 detects a specific spatial grid... At the current time point Clinical outcome variable observations satisfy When this wake-up condition is met, the system's underlying layer generates a trigger event signal.

[0040] After the trigger event signal is generated, the data access module 10 locks the trigger source grid. And the corresponding patient groups that caused the clinical values ​​to exceed the limits. Subsequently, the data access module 10 calls the database interface of the external medical institution's business system, and extracts the chief complaints and present medical history electronic medical record text data of the corresponding patient groups in batches according to the preset historical interception time window range. The time span of the above-mentioned historical interception time window is mainly defined based on the maximum pathogenic incubation period constant of the specific syndrome that triggered the alarm in objective epidemiological statistics. Let the total number of related patients intercepted be... The data access module 10 sends the merged text dataset to the semantic extraction module 20 for natural language processing.

[0041] See attached document Figure 4 , Figure 4 This is a flowchart of the exposure semantic tensor generation and verification process according to an embodiment of the present invention. In this embodiment, the system receives a dataset of associated patient medical records, performs joint entity and relation extraction to extract time-location-behavior triples, quantizes and calculates the exposure semantic payload using a dictionary, encapsulates it to form an exposure tensor, and assigns Boolean sparse state identifiers based on the extraction of time and space entities in the tensor. This can be implemented through the following steps.

[0042] The semantic extraction module 20 receives the associated patient medical record text set sent by the data access module 10, calls the built-in natural language processing model, and performs joint entity and relation extraction on the unstructured text.

[0043] In healthcare IT scenarios, outpatient electronic medical records are typically composed of free-format natural language entered by doctors. Computers often struggle to directly perform numerical spatiotemporal calculations based on such unstructured text. In this embodiment, the semantic extraction module 20 takes the received patient's chief complaint and current medical history text as input data and feeds it into a pre-deployed joint entity relation extraction network model. As a preferred approach, this natural language processing model employs a dual affine joint extraction network based on the Transformer architecture.

[0044] Specifically, the network's internal hierarchical structure and data flow comprise three stages. The text sequence is processed by a front-end word segmenter into a word vector matrix, with a maximum sequence length set at the input (e.g., truncated or zero-padded to a 512-dimensional character sequence length) to unify the input dimension of the underlying tensor. This matrix flows into a text encoding layer stacked with multiple layers of self-attention mechanisms, extracting deep semantic features of the context and outputting a hidden layer feature tensor. This feature tensor is divided into two parallel computational branches at the network output: the first branch connects to a conditional random field sequence labeling layer to identify and extract spatial entities, temporal entities, and behavioral predicate entities from the text; the second branch connects to a dual affine scoring classification layer to infer relationships between the identified entities, ensuring that specific times, locations, and actions semantically belong to the same exposure event. The model ultimately outputs the corresponding set of epidemiological triples.

[0045] To ensure the model can accurately identify medical-specific expressions, the network weights in the semantic extraction module 20 are pre-trained offline on a proprietary dataset. The model training steps include acquiring historical epidemiological survey reports as raw corpora, performing anonymization processing, and then manually annotating the temporal, spatial, and action entities in the corpus using a standardized sequence labeling system to generate realistic labels. The joint loss function during network training is a weighted combination of the cross-entropy loss of the entity recognition branch and the negative log-likelihood loss of the relation classification branch. The model iteratively updates the network layer parameters using the backpropagation algorithm, and the parameters are fixed for online inference after the accuracy metrics on the validation set converge.

[0046] The semantic extraction module 20 calculates the corresponding patient's exposure semantic load scalar based on the extracted semantic triples and pre-defined quantization mapping rules.

[0047] Traditional simple text matching often fails to accurately reflect the true differences in disease transmission risk caused by different patient behaviors. Therefore, the semantic extraction module 20 establishes a behavioral risk weight dictionary and a time decay coefficient mapping table. (For the first...) The system extracts behavioral predicate entities from each patient and matches them with corresponding behavioral risk scalars in the dictionary. This scalar is divided into discrete levels based on the physical contact intensity of the behavior, and its value range is typically set to an integer from 1 to 10. For the duration of the behavior calculated from the temporal entity, the system matches the corresponding time coefficient. This time coefficient is generally a log-normalized value of the duration in hours, with a value range limited to 0 to 1. Meanwhile, to balance the objective differences in the quality of medical record data from different levels of medical institutions, the system assigns a data source reliability factor based on the level of the patient's medical institution, with its value also set to a range of 0 to 1.

[0048] Based on the parameters obtained from the above matching and calculation, the semantic extraction module 20 calculates the patient's exposure semantic load scalar according to the following formula: ; In the formula, It represents a continuous numerical scalar that reflects the degree of risk of exposure to the disease.

[0049] The semantic extraction module 20 encapsulates discrete text attributes and payload scalars into a structured exposure tensor and performs data sparsity checks to generate control identifiers.

[0050] The semantic extraction module 20 calls the geocoding component to convert the extracted spatial location entities into specific spatial grid indexes. Regarding the conversion of time entities, considering that relative time expressions such as "two days ago" often appear in medical record texts, the system extracts the system creation time of the current electronic medical record as the baseline anchor point. The extracted relative time offset is combined with the anchor point for reverse calculation, thereby converting it into a standard system absolute timestamp. The system combines the coordinate index, timestamp, and the exposure semantic payload scalar obtained in step S202, and encapsulates them into a standard-dimensional one-dimensional tensor data structure, namely the exposure tensor. .

[0051] In outpatient settings, medical record entry objectively exhibits data sparsity, with some patient records potentially lacking any epidemiological trajectory description. Directly passing tensors containing null values ​​to the downstream numerical computation engine can easily lead to algorithm addressing out-of-bounds errors or system null pointer exceptions. To ensure the completeness and fault tolerance of the overall algorithm logic, the semantic extraction module 20 adds state machine verification logic to the tensor output, defining Boolean sparse state identifiers. .

[0052] Semantic extraction module 20 examines the grid index in the exposed tensor. With timestamp Field completeness. When both spatial and temporal data are not empty and the model's analytical confidence score is higher than the system's set classification decision threshold (e.g., 0.85), the system determines that the tensor data is complete and sets a sparse state identifier. =0. Conversely, if any one dimension of the attribute, whether temporal or spatial, is missing, or if the resolution confidence is lower than the above threshold, the system determines that the data belongs to a sparse outlier and sets a sparse state identifier. =1. Upon determining the tensor to be in a sparse state, the semantic extraction module 20 forcibly overwrites the exposed semantic payload scalars in the tensor with zero, i.e. =0, thus blocking the mathematical transmission of missing data in subsequent warning score calculations. After verification, the semantic extraction module 20 transmits the exposure tensor carrying the identifier to the routing control module 30.

[0053] See attached document Figure 5 , Figure 5 This is a schematic diagram illustrating the bidirectional routing and generation principle of a dynamic computation domain according to an embodiment of the present invention. In this embodiment, the routing control module 30 reads the sparse state identifier and controls the computation flow to perform bidirectional routing. When the identifier indicates that the data is complete, a dynamic time computation domain is generated based on a specific timestamp in the tensor. When the identifier indicates that the data is sparse and missing, a degraded time computation domain is generated using a preset default topological coordinate set and a physical epidemiological latency constant. This can be implemented through the following steps.

[0054] The routing control module 30 receives the exposed tensor and its carried sparse state identifier, and performs a bidirectional routing decision operation based on Boolean state.

[0055] To balance computational resource consumption and early warning recall rate, the system is configured with a corresponding dynamic scheduling mechanism for differences in data source quality. Specifically, the routing control module 30 reads data from the data source... Sparse state identifiers for each patient And trigger the internal branch selection logic based on the value of the identifier. When the system determines... When the value is 0, it indicates that the medical record text contains clear and highly confident spatiotemporal trajectory information, and the routing control module 30 transfers control to the main route branch to perform high-precision dynamic spatiotemporal boundary pruning. When the system determines... When the value is 1, it indicates that the epidemiological trajectory attribute parsing has failed or the confidence level is not up to standard. The routing control module 30 will transfer control to the bypass branch to perform geographic topology and physical constant degradation compensation.

[0056] The main control branch directly locks the defined spatial grid based on complete exposure tensor information and calculates a dynamic time search interval with error buffer based on the absolute timestamp of the event.

[0057] In the main branch, since the tensor data already provides high-precision exposure locations, the routing control module 30 directly sets the spatial computation domain. Set as the specific mesh to be extracted, i.e. The aforementioned pruning operation based on the spatial dimension of a defined entity can restrict the subsequent computational domain to a single grid.

[0058] To determine the temporal search boundary for subsequent cross-correlation calculations, the routing control module 30 extracts the absolute timestamps of events from the tensor. Considering that time series analysis requires lags to be discrete integer steps, the system needs to perform unit conversion. The system sets the current evaluation time point as... And define the physical time step length of the underlying data aggregation as Because patients' memories of when their symptoms occurred are objectively biased, a system time error tolerance buffer scalar is introduced when determining the time domain. This buffer scalar The specific value is determined by the system hardware infrastructure and the characteristics of disease transmission, and is usually set to 1 to 2 times the system analysis time step. The routing control module 30 combines the current evaluation time point... The time calculation range for estimating the pathogenic exposure lag time. The calculation formula is as follows: ; ; In the formula, and These represent the lower and upper bounds of the dynamic lag time search interval, respectively. This indicates a floor function to ensure the output is an integer step size. The lower bound calculation formula incorporates a function logic that takes the maximum value of zero to prevent the calculated exposure time from being later than the current evaluation time due to an excessively large buffer scalar, thus preventing algorithmic dead zones caused by inverted physical causality.

[0059] When faced with sparse tensors, the bypass degrade branch retrieves the patient's default physical coordinates to perform geographic topology expansion and combines them with known epidemiological incubation period constants to generate a static safe time search interval.

[0060] Upon entering the bypass branch, due to the lack of specific activity trajectories in the medical record, the routing control module 30 initiates a degradation compensation mechanism. As a preferred method, the routing control module 30 utilizes the patient's de-identified unique identifier to initiate a query request to the medical institution's resident health record database, retrieving the coordinates of the patient's default registered residence or regular workplace in the system. The system then maps these regular coordinates to the corresponding permanent grid. Considering that some patients seeking medical treatment in other locations or whose medical records are incomplete may not be able to obtain standard coordinates, the system introduces a fallback fault tolerance mechanism. If the database query returns a null value, the spatial grid of the currently receiving medical institution will be directly set as the permanent grid. Subsequently, the routing control module 30 performs spatial topology expansion based on the resident grid, generating a set of spatial computational domains containing the resident grid and its one-hop adjacent grids, i.e., ... For the construction of the spatial topological relation matrix and the algorithm for retrieving and extracting adjacent grids, those skilled in the art can use conventional adjacency matrix traversal methods or Euclidean distance threshold determination methods, which are well-known technologies in this field and will not be elaborated here.

[0061] For degradation compensation in the time calculation domain, since the specific time of the action cannot be extracted from the medical records, the routing control module 30 retrieves the known maximum latency period constant corresponding to the specific syndrome that triggered the current warning event from the system's built-in medical knowledge base. This physical constant reflects the longest objective lag time from environmental exposure to clinical onset. The routing control module 30 also applies this periodic constant based on the time step. After rounding up, a wider static safety time calculation domain is generated. Its assignment logic is as follows: ; ; After execution, the routing control module 30 will generate a set of spatial computing domains. With time calculation domain interval The data is packaged and the above parameters are transmitted to the timing calculation module 40 to guide subsequent data extraction and correlation calculation.

[0062] See attached document Figure 6 , Figure 6 This is a schematic diagram illustrating the principle of dynamic lag cross-correlation calculation within a constrained search boundary according to an embodiment of the present invention. In this embodiment, the time series calculation module 40 receives the output spatial and temporal calculation domain parameters, calculates the dynamic lag cross-correlation function of the preceding environmental variable sequence and the clinical outcome sequence within the boundary range of the parameter constraints, and extracts the maximum pathogenic cross-correlation value under the current analysis window. This can be implemented through the following steps.

[0063] The timing calculation module 40 receives the set of spatial calculation domains and the interval of time calculation domains output by the routing control module 30, and performs targeted extraction and spatial aggregation of the data sequence in the underlying storage stack based on the above parameters.

[0064] Global time-series computations are prone to background noise and increased computational burden when a large number of irrelevant geographic nodes are introduced. In this embodiment, the time-series computation module 40 strictly follows the spatial computation domain set. The data reading range is limited. The time series calculation module 40 retrieves the current evaluation time node from the system. Previously and in a set Normalized sequence of pre-environment variables for all grids and clinical outcome variable sequence .

[0065] For the set of branches on the main path In the case of a single grid, the system directly extracts the data sequence corresponding to that grid as the target sequence for subsequent analysis. This applies to sets under the bypass degradation branch. In complex cases involving resident grids and their topologically adjacent grids, due to slight fluctuations in environmental parameters within different grids, and to align with the physical law of environmental risk attenuation with spatial distance, the time series computation module 40 utilizes a distance attenuation weighting mechanism to perform spatial aggregation on multiple sets of sequences within the set as a preferred approach. The system centers on the resident grid and performs spatial aggregation on the first sequence within the set. Each grid is assigned a spatial weighting factor. The weight value is inversely proportional to the topological hop count from the grid to the center; for example, the weight of the central resident grid is set to 1.0, and the weight of a one-hop adjacent grid is set to 0.5. The aggregated target environment variable sequence. With clinical outcome sequence The calculation logic is as follows: multiply the sequence values ​​of all grids in the set at the same time step by their respective weight factors. Then, a weighted summation is performed, and the result is divided by the sum of all grid weight factors within the set. This operation converges the local multi-source spatial observations into a one-dimensional time series, providing a unified numerical basis for subsequent calculations.

[0066] The time series calculation module 40 introduces a sliding time window mechanism to calculate the cross-correlation function value between the aggregated environmental variable lag sequence and the clinical outcome sequence.

[0067] Considering the occasional data loss that may occur in the actual operation of IoT sensor networks, the time series calculation module 40 first traverses the target environmental variable sequence and the clinical outcome sequence before performing correlation analysis. When missing data points are detected in the sequence, the system uses an algorithm combining forward observation imputation and local linear interpolation to repair the breakpoints, in order to prevent the calculation engine from producing meaningless output.

[0068] After locking and repairing the target data sequence, to quantify the delayed causal relationship between environmental exposure and clinical onset, the time series calculation module 40 sets a sliding time observation window for truncating the sequence. The length of the sliding time window... This is typically associated with the smoothness and statistical significance of correlation analysis. As a preferred approach, this length is configured in conjunction with the typical outbreak cycle of the specific disease, and its value is generally set to 14 to 30 system time steps.

[0069] For a specific lag time step The timing calculation module 40 has a calculation length of 40. The Pearson cross-correlation coefficient between the lagged environmental variable series and the current clinical outcome series during the observation period. Considering that the variance of the data series within the truncated window may approach zero, potentially causing a division-by-zero error in the algorithm, the system introduces a small positive number in the denominator as a zero-division buffer constant. This constant is typically taken as 10. -6 The cross-correlation function The specific calculation formula is as follows: ; In the formula, This indicates the index of the integral time variable within the sliding time window. and Representing the environmental variable sequences respectively Lagging behind in the historical transition Corresponding observation window The mean and standard deviation within; correspondingly, and These represent the clinical outcome sequences. In the current viewing window The mean and standard deviation within the range.

[0070] The time series calculation module 40 is controlled by the boundary of the time calculation domain. It traverses each lag step within the restricted interval to extract extreme value features and outputs the corresponding pathogenic cross-relationship values.

[0071] The time series calculation module 40 will generate the time calculation domain interval This serves as the range of step sizes for calculating the cross-correlation function in step S402 of the loop execution. The system iterates over the lag variables sequentially with discrete integer step sizes. The system calculates and generates a set of correlation coefficients for the given time interval. By pre-limiting the search upper and lower bounds, the system avoids searching for associations at time points that do not conform to the epidemiological incubation period, which helps reduce false positive matches caused by random data coincidence.

[0072] The system iterates through the generated set of coefficients, retrieves and extracts the maximum value from the set as the value for the individual patient at the current evaluation time point. Maximum cross-correlation coefficient extreme value In actual epidemiological transmission and physical causal logic, an increase in harmful environmental factors often leads to a unidirectional rise in outpatient incidence rates. This means that the pathogenic associations that the system focuses on should exhibit positive correlation characteristics. If the calculated correlation coefficient is negative, it usually indicates that the current fluctuations in environmental variables lack pathogenicity indicators for clinical onset. To shield against such negative correlation interference that has no early warning value, the time series calculation module 40 performs a non-negative truncation operation on the extracted extreme values, the mathematical expression of which is: ; In addition, the system synchronously extracts the optimal lag step size corresponding to the acquisition of the maximum coefficient. ,Right now This parameter, in a physical sense, characterizes the optimal time difference between environmental deterioration and the concentrated outbreak of disease within the current analysis period. After execution, the time series calculation module 40 will output the truncated maximum pathogenicity cross-correlation value. and its corresponding optimal lag step size The data is transmitted to the early warning decision module 50 as a numerical benchmark parameter for the final confidence assessment.

[0073] See attached document Figure 7 , Figure 7 This is a schematic diagram of adaptive penalty decision-making and cascaded instruction output according to an embodiment of the present invention. In this embodiment, the early warning decision module 50 receives the semantic payload features extracted from the front end and the cross-correlation extreme values ​​calculated by time series, fuses sparse state identifiers to perform adaptive penalty scoring calculation, inputs the comprehensive score into a multi-level emergency matrix for comparison to determine the risk level, and generates structured early warning instructions to drive the underlying IoT devices to perform physical responses. Specifically, this can be implemented through the following steps.

[0074] The early warning decision module 50 constructs an adaptive penalty comprehensive confidence model based on sparse labels to calculate the comprehensive pathogenicity risk score of the grid area at the current assessment time node.

[0075] In complex epidemiological tracing scenarios, analysis based on a single data dimension is often insufficient to comprehensively characterize the risk status. As a preferred approach, the early warning decision module 50 receives the exposure semantic payload scalars output by the semantic extraction module 20 for each individual patient. With sparse state identifiers Simultaneously receive the maximum possible correlation values ​​between the disease and the patient's individual condition, output by the time-series calculation module 40. .

[0076] Because the two types of feature parameters mentioned above differ significantly in physical dimensions and numerical ranges, direct weighting would lead to a bias in the evaluation model. Therefore, the system introduces an extreme value normalization algorithm based on absolute boundaries to linearly map the received exposed semantic load scalar to a continuous interval [0,1]. (A zero-buffer constant is also introduced in the denominator of this calculation logic to prevent algorithm errors caused by all extreme values ​​being zero). This generates standardized semantic features, which, for consistency and brevity, will still be referred to as […]. .

[0077] For each specific spatial grid that triggers a wake-up event, the early warning decision module 50 first calculates the score of each patient tracing chain associated within that grid. The early warning decision module 50 establishes a comprehensive evaluation formula incorporating an automatic decay penalty mechanism, the specific mathematical expression of which is as follows: ; In the formula, This represents the overall confidence score calculated by the system for a single patient's traceability chain; and Let represent the prior weight coefficients for the text semantic dimension and the context temporal dimension, respectively, which satisfy . The specific values ​​are usually preset to 0.4 and 0.6; This is the penalty coefficient set by the system for the degradation calculation mode, and its value ranges from 0.2 to 0.5.

[0078] To obtain the related patients within the grid Subsequently, the early warning decision module 50 performs an aggregation calculation (e.g., taking the maximum or average value) on all associated scores within a specific spatial grid, and uses this aggregated value as the final comprehensive confidence score for that grid area.

[0079] Combining the underlying mechanism where semantic features are forcibly overwritten to zero in a sparse state, when the data is in a sparse and missing state (i.e. When this happens, the system not only eliminates the influence of semantic scores, but also... This multiplier term discounts and attenuates the time-series correlation coefficient, which helps suppress overresponse caused by limited data quality.

[0080] The early warning decision module 50 performs multi-level emergency matrix comparison and alarm trigger control based on the comprehensive confidence score.

[0081] After acquiring continuous risk values, the system needs to convert them into discrete action command levels. The early warning decision module 50 has a pre-installed multi-level emergency threshold matrix. This threshold matrix contains three discrete risk thresholds: a low-risk threshold... Medium risk threshold With high risk threshold For continuous daily operation, the specific values ​​of the above thresholds are determined using the statistical quantiles of all risk scores within the historical monitoring period, for example, corresponding to the 80th, 90th, and 95th percentiles of historical data, respectively. In cases of initial system deployment where the accumulated historical assessment samples have not yet reached the system's minimum sample size limit (e.g., 500 calculation records), the early warning decision module 50 internally activates static backup thresholds (e.g., set to 0.6, 0.75, and 0.85) as a safety net, automatically switching to dynamic quantile update mode once the sample size meets the requirements.

[0082] The early warning decision module 50 will use the currently calculated comprehensive confidence score. Compare with the emergency threshold matrix for out-of-bounds errors. When the current environmental fluctuations are determined to be within a safe buffer zone, the system will not trigger any alarm commands; when When, it is determined that a Level 1 Yellow Alert has been triggered; when When, it is determined that a Level II orange alert has been triggered; when At that time, a Level 3 red alert is triggered. Through the aforementioned matrix comparison mechanism based on quantiles, the system can enhance its adaptability to differences in basic incidence rates across different geographical regions.

[0083] The early warning decision module 50 encapsulates the structured early warning dataset and sends external scheduling instructions to the downstream IoT platform. After determining the alarm level, the early warning decision module 50 integrates multi-dimensional key parameters from the source tracing process. The system extracts the set of target spatial grid coordinates that triggered the early warning. The transmission time chain segment corresponding to the optimal pathogenicity time difference (i.e., the optimal lag step size) (Converted to actual traceability timestamps), together with the aforementioned determined risk level status identifiers, are packaged into a structured early warning dataset conforming to standard industrial protocol formats.

[0084] The early warning decision module 50 calls an external communication interface to send the aforementioned structured dataset to the downstream urban environmental IoT control platform via a message queue or IoT communication protocol. For different early warning levels and source tracing grids, the control platform executes physical intervention responses based on built-in automated strategies. In this embodiment, when a red syndrome early warning is received for a specific grid ward area, the underlying control system will adjust the opening of the corresponding grid's fresh air system damper to increase the ventilation rate, or activate the ultraviolet air disinfection equipment in the specific area, achieving automated source tracing and coordinated physical control of environmental risks.

[0085] Taking the operating environment of a municipal-level public health intelligent management and control platform as an example, the target monitoring area is preset as a 1km×1km standard coordinate grid, and the physical time step for data alignment at the system's underlying level is set to... On August 15, 2026, the data access module detected a surge in the number of reported acute respiratory syndrome cases to 25 in spatial grid index G_12 (corresponding to a centrally air-conditioned enclosed commercial complex within the jurisdiction), exceeding the preset statistical baseline threshold. The system immediately triggers an event wake-up, identifies the corresponding patient group, and retrieves the electronic medical record text dataset from the past 14 days (based on the known maximum incubation period constant of aerosol-borne pathogens).

[0086] After extracting the dataset and inputting it into subsequent links, two typical source chain records are extracted for tracking and comparison: For patient A, the text slice contains information about high fever for 2 days and a stay of approximately 4 hours in the underground supermarket of the Central Business District 3 days prior. The semantic extraction module calls the pre-trained model to parse out high-confidence entities and complete the inverse clock inference. After dictionary lookup quantization, the exposed semantic payload scalar is output. Because the spatiotemporal entities are complete, the state machine assigns a location identifier. 0. The routing control module sequentially switches to the main branch, strictly converging the spatial computation domain to... Simultaneously, it combines error buffering to generate an accurate dynamic time search interval. .

[0087] Combined with appendix Figure 8 As shown, the dynamic pruning operation of this spatial computational domain can be visualized using MATLAB R2025a: the geoscatter function under MappingToolbox is called to construct the underlying urban mapping environment. The global computational domain or bypass degradation domain without semantic pruning appears as a large area of ​​light blue semi-transparent topological polygons in the map; while the high-precision computational domain output by the main path pruning in this embodiment... This presents as a highly geographically clustered set of deep red scatter points. The graph visually reflects the system's mechanism for stripping away computational power from sensor nodes in irrelevant environments. For patient B, the medical record states fever and cough for one day, with no history of specific epidemiological contact. Due to the failure of epidemiological trajectory parsing, the semantic payload was forcibly overridden internally within the module. Output sparse identifier The control flow then switches to the degraded bypass branch, retrieves the patient's residence coordinates from the patient's file to generate a spatial topology domain containing a one-hop adjacency grid, and degenerates the time computation domain into a broad static extremum range [0, 336h].

[0088] After receiving the boundary parameters, the time-series calculation module performs parallel cross-correlation calculations on the data columns within the restricted interval. For patient A, the engine calculates the Pearson correlation matrix between the environmental aerosol concentration sequence and the clinical onset sequence in the G_12 region only within a limited window period of [48h, 96h]. Finally, the maximum cross-correlation extreme value is retrieved and extracted. Optimal time difference for disease onset .

[0089] As attached Figure 9 As shown, the system performed time-series tracing on the dynamic optimization process for extracting extreme values. The solid black continuous curve in the figure represents the Pearson cross-correlation coefficient under different lag step sizes. Waveform. The graph uses two vertical dashed lines (xline) to strictly define the search boundary [48h, 96h]. The algorithm skips global traversal and directly searches for peaks within the curve segments bounded by the dashed lines. The peaks are located using red pentagram markers and yline markings. Coordinates of the extreme point.

[0090] Upon entering the early warning decision-making stage, the system pre-sets a downgrade penalty coefficient. and evaluation weight matrix Based on the adaptive penalty model, the overall confidence score of patient A's tracing chain is calculated: ; Patient B experienced arithmetic decay in the multiplier term of the evaluation formula due to triggering the data scarcity and missing data check. ; By aggregating the extreme value aggregates of all associated patient link scores within the G_12 network, a comprehensive risk score for the region is obtained. Directly overflowing the preset high-risk control threshold ( The early warning decision module then generates a standardized structured data packet and distributes it via the Internet, driving the underlying IoT platform to increase the building fresh air push load in the G_12 area to the rated peak, completing the business closed loop from semantic computing to physical blocking.

[0091] In the offline algorithm simulation and engineering testing verification phase, the system accessed environmental IoT sensor snapshot data from a city for the entire period of 2024-2025, and mixed in over 120,000 desensitized respiratory clinic medical records for historical trajectory playback testing. The baseline models included: the traditional single-point boundary crossing alarm method (Baseline 1, which only monitors clinical results and does not calculate correlation) and the brute-force global cross-correlation detection method (Baseline 2, which fully traverses the default maximum latency window to find the correlation extremum, lacking a semantic dimensionality reduction mechanism). The core evaluation indicators are output as shown in the table below: Combined with appendix Figure 10 The performance hedging relationship diagram of the three types of algorithms shows that the gray bars on the left vertical axis represent the model prediction accuracy (F1-Score). The height of the bar represented by this invention (0.905) is absolutely dominant, which is due to... The sparse state machine and automated numerical penalty logic effectively filter out background noise induced by occasional weather fluctuations in complex spatial grids, keeping the false alarm rate (FPR) below 2.3%. On the other hand, by relying on spatial domain topology shearing and temporal domain boundary contraction in the bidirectional routing pre-stage, the scale of temporal tensor multiplication and addition is significantly reduced, and the single trajectory finding operation overhead (ACT) drops dramatically compared to the global solution mode, with a reduction of over 97%, completely eliminating the risk of computing power downtime in multidimensional spatiotemporal continuous analysis under high concurrency conditions.

[0092] The performance hedging relationship of multidimensional evaluation data is presented in a comprehensive comparative rendering chart output by MATLAB R2025a: the `bar` function is used to characterize a group of bar charts based on three types of algorithm models on the horizontal axis, while the `yyaxis` function is used to set the axes on both sides. The left Y-axis is set to the height of the F1-Score bar corresponding to the prediction probability scale, and the right Y-axis uses a logarithmic scaling strategy to map the ACT execution time. The data chart explicitly confirms that the design of this invention establishes an absolute performance advantage in balancing the two essential engineering requirements of intercepting meaningless background calculations and maintaining extremely high tracking accuracy.

[0093] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A public health intelligent management system based on artificial intelligence, characterized in that, The system includes: The data access module is used to acquire environmental sequence data and clinical syndrome reporting data from various spatial grids within the target monitoring area; The event wake-up module is used to retrieve the historical medical record text of the associated patient group when the clinical syndrome reporting data of a certain spatial grid exceeds a preset baseline threshold. The semantic extraction module is used to perform natural language semantic extraction on the historical medical record text, and generate an exposed semantic payload scalar and sparse state identifier for each patient tracing chain. The routing control module is used to generate restricted spatial and temporal computation domains for each of the patient tracing chains based on the sparse state identifiers. The time series calculation module is used to calculate the maximum cross-correlation coefficient extreme value between the environmental sequence data and the clinical sequence data within the constraint boundary between the spatial calculation domain and the time calculation domain. The early warning decision module is used to establish a comprehensive evaluation formula with an automatic decay penalty mechanism based on the exposed semantic payload scalar, sparse state identifiers and the maximum cross-correlation extremum to calculate the regional comprehensive confidence score, and generate a physical intervention command when the regional comprehensive confidence score exceeds the limit.

2. The public health intelligent management system based on artificial intelligence according to claim 1, characterized in that, The semantic extraction module is specifically used for: Spatiotemporal entity features were extracted from the historical medical record text using a natural language model. The extracted spatiotemporal entity features are mapped to a preset risk exposure dictionary to obtain the original semantic payload calculation value; If no valid spatiotemporal entity features are extracted, the original semantic payload calculation value of the tracing chain is overwritten to zero, and the corresponding sparse state identifier is set to 1. Conversely, the sparse state identifier is set to 0.

3. The public health intelligent management system based on artificial intelligence according to claim 2, characterized in that, The semantic extraction module is also specifically used for: The preset theoretical maximum and theoretical minimum values ​​in the risk exposure dictionary are obtained as global absolute mapping boundaries; An extreme value normalization algorithm is used to linearly map the original semantic payload calculation value based on the global absolute mapping boundary to generate a standardized exposed semantic payload scalar.

4. The public health intelligent management system based on artificial intelligence according to claim 1, characterized in that, The routing control module is specifically used for: If the sparse state identifier is 0, the spatial computation domain is converged to the exposure coordinate grid extracted from the patient's historical medical records, and a dynamic time search interval is generated as the time computation domain. If the sparse state identifier is 1, then the spatial computation domain is broadened to the set of adjacent grids anchored to the patient's residence in the health record, and the temporal computation domain is degraded to a preset static maximum latency interval.

5. The public health intelligent management system based on artificial intelligence according to claim 1, characterized in that, The time-series calculation module is specifically used for: Within the boundaries of the time calculation domain, calculate the Pearson cross-correlation coefficient between environmental sequence data and clinical sequence data within the spatial calculation domain; Retrieve and extract the maximum value from the set of Pearson cross-correlation coefficients; Perform a non-negative truncation operation on the extracted maximum value to obtain the final extreme value of the maximum cross-correlation coefficient.

6. The public health intelligent management system based on artificial intelligence according to claim 1, characterized in that, The comprehensive evaluation formula established in the early warning decision module has the following structure: Calculate the arithmetic multiplier term that includes the preset penalty coefficient and the sparse state identifier; The arithmetic multiplier term is used to perform weighted processing on the extreme value of the maximum cross-correlation coefficient; The weighted result is linearly summed with the exposed semantic payload scalar to obtain the comprehensive confidence score for the tracing chain of a single patient.

7. The public health intelligent management system based on artificial intelligence according to claim 6, characterized in that, The early warning decision module is also specifically used for: The comprehensive confidence scores of all related patient tracing chains within the same spatial grid are aggregated, and extreme value aggregation or mean aggregation is performed to obtain the regional comprehensive risk score of the grid. When the comprehensive risk score of the area is greater than or equal to the preset high-risk control threshold, the generation of the physical intervention instruction is triggered.

8. The public health intelligent management system based on artificial intelligence according to claim 1, characterized in that, The physical intervention command is a structured alarm data packet; The system also includes an IoT linkage execution module, which is used to parse the alarm data packet and drive the ventilation equipment or disinfection hardware in the target building area.

9. The public health intelligent management system based on artificial intelligence according to claim 1, characterized in that, The spatial computation domain and the temporal computation domain, after being generated by the routing control module, are input to the temporal computation module as exclusive boundary conditions.

10. A public health intelligent management method based on artificial intelligence, used to execute the public health intelligent management system based on artificial intelligence as described in any one of claims 1-9, characterized in that, The method includes the following steps: Acquire environmental sequence data and clinical syndrome reporting data of each spatial grid within the target monitoring area, and retrieve historical medical record texts of related patient groups when data exceeds the boundaries; Natural language semantic extraction is performed on the historical medical record text to generate an exposure semantic payload scalar and sparse state identifier for each patient's tracing chain. Based on the sparse state identifier, a restricted spatial computation domain and a temporal computation domain are generated for each of the patient tracing chains. Within the constrained boundaries of the spatial computational domain and the temporal computational domain, calculate the maximum extreme value of the cross-correlation coefficient between the environmental sequence data and the clinical sequence data; Based on the exposed semantic payload scalar, sparse state identifiers, and the maximum cross-correlation extremum, a comprehensive evaluation formula with an automatic decay penalty mechanism is established to calculate the regional comprehensive confidence score, and a physical intervention command is generated when the regional comprehensive confidence score exceeds the limit.

Images