Pediatric respiratory infectious disease risk assessment method and system

Abstract

The invention discloses a child respiratory infectious disease risk assessment method and system, and the method comprises the steps: collecting child multi-source data, extracting continuous and category features, and generating a standardized sample set; extracting age, height, weight and development indexes, and generating a continuous age track embedding vector; extracting vaccine types and inoculation information, and modeling an evolution process of immune intensity along with time change; fusing family and school contact diagrams, executing dual diffusion and spectral domain dimensionality reduction, and outputting potential field features of the diagrams; inputting the multi-source features into an improved DeepFM model, and outputting an individual infection risk probability result; and calculating a comprehensive score grade according to the risk probability, generating a prevention and control suggestion and outputting an early warning. According to the invention, through multi-source health data fusion and an improved DeepFM intelligent modeling technology, accurate assessment and early warning of the infantile respiratory infectious disease infection risk are realized.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to a method and system for risk assessment of respiratory infectious diseases in children. Background Technology

[0002] Currently, respiratory infectious diseases in children are common and prevalent, characterized by rapid transmission and high insidiousness, easily leading to cluster infections in high-density contact environments such as families and schools. Traditional risk assessment methods often rely on single-dimensional data sources, such as clinical signs, pathogen detection results, or epidemiological statistics, which fail to comprehensively reflect the individual differences in children's physiological development, immune response, and social behavior. With the widespread adoption of wearable devices, smart terminals, and health record systems, children's health data is becoming increasingly multi-source and dynamic, and the introduction of machine learning technology is continuously improving data analysis and modeling capabilities. However, existing methods still have significant limitations in data fusion and risk modeling.

[0003] In existing technologies, most risk assessment models use static feature inputs and linear classifiers for discrimination, lacking the ability to dynamically model the decay of childhood immunity, changes in vaccine protective efficacy, and differences in environmental exposure. While machine learning can provide assistance in some aspects, existing methods still struggle to effectively capture the dynamic relationships between complex changes in immune efficacy and physiological characteristics. Regarding the physiological changes during childhood development, existing methods often use time or age as a single variable, failing to adequately characterize the physiological migration relationships between different developmental stages, resulting in insufficient reflection of individual differences. Social transmission modeling is mostly based on single-layer networks, considering only the average contact frequency of the group, without distinguishing between the two core transmission environments of family and school, making it difficult to accurately quantify transmission paths and the effectiveness of prevention and control interventions.

[0004] Current technologies for risk assessment of pediatric respiratory infectious diseases have three main shortcomings: First, the dynamic correlation modeling between physiological and immune characteristics is insufficient, failing to fully utilize the dynamic modeling capabilities of machine learning. Second, the decay pattern of vaccine immunity cannot be quantified at the temporal level, and machine learning methods in this area have not yet reached a level of accurate prediction. Third, the multi-layered transmission structure of families and schools has not been effectively modeled, and existing machine learning methods have significant limitations in handling complex social contact networks. These problems result in limited accuracy of risk assessment results, insufficient individualized predictive capabilities, and difficulty in achieving early and accurate risk warnings.

[0005] Therefore, how to provide a method and system for risk assessment of pediatric respiratory infectious diseases is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a method and system for risk assessment of respiratory infectious diseases in children. This invention fully utilizes healthcare informatics, deep learning, and multi-layer graph modeling techniques. By introducing a physiological age transfer embedding mechanism, immune potential evolution modeling, and a multi-layer graph interference potential field structure, it achieves joint modeling of individual children's physiological development characteristics, the duration of immune protection, and social contact transmission pathways. This invention uses an improved DeepFM risk assessment model to fuse and learn multi-source heterogeneous data, outputting the probability of infection risk for individual children. Combined with a risk threshold grading mechanism and a dynamic early warning mechanism, it achieves early identification and risk alerts for respiratory infectious diseases in children. It possesses advantages such as high assessment accuracy, thorough characterization of individual differences, and timely early warning response. It can be widely applied in schools, communities, and public health management scenarios, effectively improving the intelligence and precision of infectious disease prevention and control in children.

[0007] A method for assessing the risk of pediatric respiratory infectious diseases according to an embodiment of the present invention includes: Collect multi-source data of the target children, perform preprocessing on the multi-source data, extract continuous data as continuous features, extract discrete data as category features, and obtain a normalized input sample set; Based on the normalized input sample set, age, height, weight and growth and development indicators are extracted, input into the physiological age transfer embedding module, continuous age trajectory embedding vector is generated, and physiological embedding feature set is output. Based on a standardized input sample set, information on vaccine type, vaccination date and vaccination interval is extracted from the immunization data. Through immune potential evolution modeling, a representation of immune strength that changes over time is generated, and an immune potential feature set is output. Construct and merge family-level and school-level contact graphs to form a multi-layer graph. Based on mask-wearing compliance, ventilation score and isolation status, intervene and reweight the edge weights of the multi-layer graph to obtain the baseline contact matrix and intervention contact matrix. Perform dual diffusion operation to generate the propagation potential field representation and the prevention and control potential field representation. Calculate the difference to obtain the interference potential field representation. After spectral domain gating dimensionality reduction to extract principal component features, output the graph potential field feature set. The DeepFM risk assessment model is improved by inputting physiological embedding feature set, immune potential feature set, graphical potential field feature set, and normalized continuous and categorical features together to output the individual infection risk probability. Based on the individual's infection risk probability, a comprehensive risk score is calculated and threshold classification is applied. The results are divided into low-risk, medium-risk, and high-risk levels, and the corresponding risk levels and prevention and control recommendations are output and synchronized to the parents' terminal and medical terminal to complete the early warning and risk reminder of children's respiratory infectious diseases.

[0008] Optionally, the multi-source data includes physiological data, immunization data, environmental and behavioral data, and social contact data from families and schools.

[0009] Optionally, the preprocessing of the multi-source data includes performing format unification, outlier removal, missing value completion, and standardization on the multi-source data.

[0010] Optionally, the output physiological embedding feature set includes: Based on a standardized input sample set, a physiological basic feature table containing age, height, weight, body mass index, height standard grading, and weight standard grading is generated. A physiological age transfer embedding module is constructed, which consists of a stage boundary adaptive recognition unit, an age trajectory continuous transfer coding unit, and a physiological abnormality sensitivity gating unit. In the stage boundary adaptive recognition unit, the 10th, 50th, and 90th percentiles are used as the initial stage boundaries. The height standard class and weight standard class of the samples are segmented and assigned. The stage boundaries are iteratively adjusted until convergence according to the criterion that the misclassification rate is no higher than 5%. The stage label sequence and stage confidence sequence are output. In the age trajectory continuous migration coding unit, an age time series segment is constructed with a window size of 3 months and a step size of 1 month. The percentile change values ​​of adjacent windows are classified to generate a trajectory state sequence and a transition coding sequence. In the physiological abnormality sensitivity gating unit, the abnormality sensitivity level is determined as high based on the judgment rules that the difference between the maximum and minimum body temperature within a day is greater than or equal to 1.0℃, the heart rate is at or above the 95th percentile of the same age, or the respiratory rate is at or above the 95th percentile of the same age. Otherwise, it is medium or low. The abnormality sensitivity level is paired with the trajectory state sequence item by item to generate gating weight labels. The trajectory state sequence and the transition coding sequence are then gating fused to obtain the gating physiological coding vector. The stage label sequence, stage confidence sequence, trajectory state sequence, transition coding sequence, and gated physiological coding vector are concatenated and length consistency is checked according to the preset field order to generate continuous age trajectory embedding vectors and summarize them into a physiological embedding feature set.

[0011] Optionally, the output immune potential feature set includes: Based on the standardized input sample set, an immunization event timeline with the assessment time as the anchor point is established. The vaccine type, production platform, vaccination date, number of doses, vaccination interval, homologous or heterologous sequential type and previous positive etiological records are summarized one by one to generate a vaccination-infection annotation sequence arranged in chronological order. A three-stage immunization state machine is constructed in the timeline corresponding to each vaccine. The stages are divided in the order of activation, plateau and decay. The start and end positions of the three stages are determined by the boundary adaptive method based on dose and interval. The stage boundaries of adjacent vaccination events are subjected to non-overlapping constraint processing. The stage label sequence and stage confidence sequence are output. A dual-channel immune proxy fusion unit was constructed. Channel 1 generates an inoculation-driven immune intensity trajectory based on the inoculation-infection labeled sequence and the stage label sequence. Channel 2 generates a physiologically driven immune intensity trajectory based on immune-related laboratory indicators and symptom timelines. The two trajectories are aligned and scored on consistency on the time axis, and the fused staged immune intensity trajectory is output. The process of age and exposure modulation is performed by calling age stage information from the physiological embedded feature set and combining class case counts, family exposure counts and ventilation scores to jointly modulate the amplitude and duration of the phased immune intensity trajectory, generating the modulated phased immune intensity trajectory, residual protection duration and phase gain factor. According to the preset field order, the phase label sequence, phase confidence sequence, vaccination-infection annotation sequence, modulated phased immune strength trajectory, residual protection duration, phase gain factor, data integrity marker and consistency score are packaged to form a time-varying immune strength representation and output as an immune potential feature set.

[0012] Optionally, the output graph potential field feature set after extracting principal component features via spectral domain gating dimensionality reduction includes: Based on social contact data, family-level contact maps and school-level contact maps were constructed separately. Node identifiers were unified and the number and duration of contact were counted according to the most recent fourteen-day window. The maps were then merged into multi-level maps, generating an adjacency table and a node metric table. The initial exposure vector was formed by weighting the family symptom count and the class confirmed case count, and the diffusion rounds and reinjection ratios were fixed. The edge intervention kernel weighting is performed on the multi-layer graph. The edge weights are segmented and time-decayed according to mask wearing compliance, ventilation score, isolation status and contact proximity. The baseline contact matrix and intervention contact matrix are output, and row normalization and node order locking are completed. Dual diffusion operations are performed on the baseline contact matrix, and neighborhood weighted aggregation and proportional back-injection of the initial exposure vector are performed in a preset number of rounds to obtain the propagation potential field representation; dual diffusion operations are performed on the intervention contact matrix in the same number of rounds and with the same back-injection ratio to obtain the control potential field representation. The difference between the propagation potential field representation and the control potential field representation is calculated to obtain the interferometric potential field representation. Spectral domain gating dimensionality reduction is then performed on the interferometric potential field representation, specifically as follows: Eigendecomposition is performed on the graph Laplacian corresponding to the multi-layer graph to obtain three frequency bands: low frequency, mid frequency and high frequency. High frequency components are suppressed according to a preset gating mask while retaining the dominant mid frequency and low frequency components. Based on a cumulative variance retention rate of at least 90%, principal component features are extracted to generate spectral domain-gated interferometric principal component vectors. The interference principal component vector, along with the mean, maximum value, and standard deviation of the propagation potential field, the mean, maximum value, and standard deviation of the control potential field, and the energy proportions of the three frequency bands, are summarized into a graphical potential field feature set.

[0013] Optionally, the output individual infection risk probability includes: Establish an input feature set, and perform field alignment and time window alignment on the physiological embedding feature set, immune potential feature set, graph potential field feature set, normalized continuous features and category features, and complete missing label completion and scale consistency processing; An improved DeepFM risk assessment model is constructed, which includes a multi-potential field collaborative fusion module, a potential field gating interaction module, and a multi-scale deep combination module. In the multi-potential field collaborative fusion module, channel alignment and intensity calibration are performed on physiological embedding features, immune potential features and graph potential field features to generate fusion context vectors and channel-level masks. Channel-wise reweighting and selective retention are performed on the embedded and mapped category features and normalized continuous features. In the potential field gating interaction module, the second-order feature interaction pairs of the FM branch are gating and weighted according to the fusion context vector. Interaction pairs involving physiological embedding features and immune potential features are preferentially calculated. Interference suppression and threshold pruning are performed on interaction pairs related to graph potential field features to generate the gating and calibrated second-order interaction representation. In the multi-scale deep combination module, the input of the Deep branch is processed in a multi-scale segmentation, and the local combination unit and the cross-layer residual mixing unit are stacked sequentially. Special tokens are introduced for physiological embedding features, immune potential features and graph potential features and jointly arranged with continuous features and category features to output a multi-scale deep representation. Branch fusion and risk aggregation are performed on the second-order interactive representation and the multi-scale deep representation. According to the combination rules of fixed weight and data-driven weight, individual infection risk scores are generated and converted into individual infection risk probabilities.

[0014] Optionally, the output of the corresponding risk level and prevention and control recommendations includes: The inverse indicators of individual infection risk probability, daily average value of transmission potential field representation intensity, and residual protection duration are weighted and synthesized according to three preset non-negative weights. The synthesized results are then interval-standardized within the same batch to obtain a comprehensive risk score. The grading thresholds are determined based on the comprehensive risk score. The low-risk threshold is set to the 40th percentile of the distribution, and the high-risk threshold is set to the 80th percentile of the distribution. The threshold versions are recorded. The risk level is determined based on the comprehensive risk score and the grading threshold. When the score is below the low threshold, it is judged as low risk; when it is between the low and high thresholds, it is judged as medium risk; and when it is above the high threshold, it is judged as high risk. Corresponding treatment suggestions are generated, and the stability of continuous assessment results is tested. If the same child is in a high-risk state for a consecutive preset number of days, an escalation warning is triggered, and the risk level, treatment suggestions, and time information are synchronized to the parent's terminal and the medical terminal.

[0015] A pediatric respiratory infectious disease risk assessment system according to an embodiment of the present invention includes the following modules: The data processing module is used to collect multi-source data and perform preprocessing, extract continuous features and categorical features, and generate a standardized sample set; The physiological embedding module is used to extract age, height, weight, and developmental indicators, construct the physiological age transfer embedding module, and form a physiological embedding feature set; The immune modeling module is used to generate time-varying immune strength and residual protection duration based on a normalized input sample set through immune potential evolution modeling, and outputs an immune feature set. The graph potential field module is used to merge the contact graphs of the family layer and the school layer, perform dual diffusion to generate the propagation potential field and the prevention and control potential field, and output the graph potential field feature set. The risk assessment module is used to model and output the individual's infection risk probability by improving the DeepFM risk assessment model; The early warning output module is used to calculate a comprehensive score, classify risk levels according to thresholds, generate prevention and control suggestions, and trigger an early warning when there is a continuous high risk, and synchronize the results to the parent terminal and the medical terminal.

[0016] The beneficial effects of this invention are: This invention addresses the challenge of quantifying differences in children's physiological development in traditional models by introducing a physiological age transfer embedding mechanism. Through continuous mapping and stage transfer modeling of age, height, weight, and developmental indicators, this invention can dynamically characterize the differences in immune function and infection susceptibility among children at different developmental stages, achieving a precise description of individual physiological states. This enhances the modeling ability of individual differences in risk assessment, making the model more adaptive and accurate in predicting infection risks in children of different age groups.

[0017] The immune potential evolution modeling method proposed in this invention can capture the nonlinear decay law of vaccine immune protection over time. By establishing an immune event timeline and simulating the activation, plateau, and decay processes in stages, this invention can dynamically calculate the immune strength and residual protection duration of children, thereby accurately reflecting the superimposed effect of vaccination and natural infection, effectively overcoming the shortcomings of static modeling of immune duration in existing technologies. This mechanism improves the responsiveness of risk assessment models to individual immune differences, providing data support for vaccine management and revaccination decisions for children.

[0018] This invention achieves structured modeling of social contact relationships by constructing a multi-layered graph interference potential field model that integrates the family and school-based dual-layer transmission structures. Through intervention kernel weighting and dual diffusion calculation, this model can simultaneously simulate transmission paths and the effects of prevention and control interventions, and extract interference potential field features for risk prediction, thereby enabling spatial assessment of children's infection risk in different social scenarios. Overall, this invention forms a technical closed loop in three aspects: individual physiological difference modeling, immune dynamic tracking, and transmission structure analysis, improving the accuracy, interpretability, and early warning capabilities of risk assessment. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a method for assessing the risk of respiratory infectious diseases in children, as proposed in this invention. Figure 2 This is a schematic diagram of the structure of a pediatric respiratory infectious disease risk assessment system proposed in this invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] refer to Figure 1 A method for assessing the risk of respiratory infectious diseases in children, including: Collect multi-source data of the target children, perform preprocessing on the multi-source data, extract continuous data as continuous features, extract discrete data as category features, and obtain a normalized input sample set; Based on the normalized input sample set, age, height, weight and growth and development indicators are extracted, input into the physiological age transfer embedding module, continuous age trajectory embedding vector is generated, and physiological embedding feature set is output. Based on a standardized input sample set, information on vaccine type, vaccination date and vaccination interval is extracted from the immunization data. Through immune potential evolution modeling, a representation of immune strength that changes over time is generated, and an immune potential feature set is output. Construct and merge family-level and school-level contact graphs to form a multi-layer graph. Based on mask-wearing compliance, ventilation score and isolation status, intervene and reweight the edge weights of the multi-layer graph to obtain the baseline contact matrix and intervention contact matrix. Perform dual diffusion operation to generate the propagation potential field representation and the prevention and control potential field representation. Calculate the difference to obtain the interference potential field representation. After spectral domain gating dimensionality reduction to extract principal component features, output the graph potential field feature set. The DeepFM risk assessment model is improved by inputting physiological embedding feature set, immune potential feature set, graphical potential field feature set, and normalized continuous and categorical features together to output the individual infection risk probability. Based on the individual's infection risk probability, a comprehensive risk score is calculated and threshold classification is applied. The results are divided into low-risk, medium-risk, and high-risk levels, and the corresponding risk levels and prevention and control recommendations are output and synchronized to the parents' terminal and medical terminal to complete the early warning and risk reminder of children's respiratory infectious diseases.

[0022] In this embodiment, the multi-source data includes physiological data, immunization data, environmental and behavioral data, and social contact data from families and schools.

[0023] In this embodiment, the preprocessing of multi-source data includes performing format unification, outlier removal, missing value completion, and standardization on the multi-source data.

[0024] In this embodiment, the output physiological embedding feature set includes: Based on a standardized input sample set, a physiological basic feature table containing age, height, weight, body mass index, height standard grading, and weight standard grading is generated. A physiological age transfer embedding module is constructed, which consists of a stage boundary adaptive recognition unit, an age trajectory continuous transfer coding unit, and a physiological abnormality sensitivity gating unit. In the stage boundary adaptive recognition unit, the 10th, 50th, and 90th percentiles are used as the initial stage boundaries. The standard height and weight categories of the samples are segmented and assigned. The stage boundaries are iteratively adjusted until convergence is achieved according to the criterion that the misclassification rate is no higher than 5%. The stage label sequence and stage confidence sequence are output. The iterative adjustment of the stage boundaries until convergence according to the criterion that the misclassification rate is no higher than 5% means that in the process of recognizing the stage boundaries, the classification error of the height and weight samples is always controlled within 5% by continuously adjusting the stage division, and the adjustment is stopped when the error stabilizes and no longer decreases. In the age trajectory continuous migration coding unit, age time series segments are constructed with a 3-month window and a 1-month step size. The percentile changes of adjacent windows are classified as rising, falling, and stationary, with changes greater than or equal to 10 percentage points. Trajectory state sequences and transition coding sequences are generated. In the physiological abnormality sensitivity gating unit, the abnormality sensitivity level is determined as high based on the judgment rules that the difference between the maximum and minimum body temperature within a day is greater than or equal to 1.0℃, the heart rate is at or above the 95th percentile of the same age, or the respiratory rate is at or above the 95th percentile of the same age. Otherwise, it is medium or low. The abnormality sensitivity level is paired with the trajectory state sequence item by item to generate gating weight labels. The trajectory state sequence and the transition coding sequence are then gating fused to obtain the gating physiological coding vector. The stage label sequence, stage confidence sequence, trajectory state sequence, transition coding sequence, and gated physiological coding vector are concatenated and length consistency is checked according to the preset field order to generate continuous age trajectory embedding vectors and summarize them into a physiological embedding feature set.

[0025] In this embodiment, the output immune potential feature set includes: Based on the standardized input sample set, an immunization event timeline with the assessment time as the anchor point is established. The vaccine type, production platform, vaccination date, number of doses, vaccination interval, homologous or heterologous sequential type and previous positive etiological records are summarized one by one to generate a vaccination-infection annotation sequence arranged in chronological order. A three-stage immunization state machine is constructed in the timeline corresponding to each vaccine, dividing the phases into activation, plateau, and decay phases. A boundary adaptive method based on dose and interval is used to determine the start and end positions of the three phases. Non-overlapping constraints are applied to the phase boundaries of adjacent vaccination events, and phase label sequences and phase confidence sequences are output. Specifically, the boundary adaptive method based on dose and interval for determining the start and end positions of the three phases is as follows: Based primarily on the interval between vaccination doses, and combined with the average immune response cycle of vaccine types, the starting point and duration of the immune response for each dose were calculated, and the boundary ranges of the activation phase, plateau phase, and decay phase were preliminarily defined. The boundaries between adjacent doses are dynamically corrected, and the start and end positions of the stages are adjusted according to the actual vaccination interval to ensure that the stages are sequential and do not overlap. In multiple rounds of sample iteration, the prediction error of immune intensity under different boundary settings is compared. The boundary with the smallest error and stable stage division is selected as the final start and end position, and the corresponding stage label sequence and stage confidence sequence are generated. A dual-channel immune proxy fusion unit was constructed. Channel 1 generates an inoculation-driven immune intensity trajectory based on inoculation-infection labeled sequences and stage-labeled sequences. Channel 2 generates a physiologically driven immune intensity trajectory based on immune-related laboratory indicators and symptom timelines. The two trajectories were aligned and scored for consistency on the time axis, and the fused, staged immune intensity trajectory was output. Channel 1 generates an inoculation-driven immune intensity trajectory based on inoculation-infection labeled sequences and stage-labeled sequences, specifically as follows: Using each vaccination event as a time anchor, the activation period, plateau period and decay period are divided according to the phase label sequence. Within each phase, an initial immune strength value is assigned according to the number of doses, the interval between doses and the type of vaccine. The immune intensity is dynamically increased or decreased based on the cumulative effect over time after vaccination, forming an immune response curve with temporal continuity. Introducing infection labeling information, the immune intensity after the infection occurrence point is corrected and compensated, and the vaccination-driven immune intensity trajectory is output. Channel Two generates a physiologically driven immune intensity trajectory based on immune-related laboratory indicators and symptom timelines, specifically as follows: Based on laboratory indicators such as serum antibody levels, lymphocyte counts, and inflammatory factor levels in children, a physiological immune response sequence was constructed. By combining information such as body temperature, cough frequency, and duration of respiratory symptoms in the symptomatological timeline, the physiological immune response is dynamically calibrated to smooth abnormal fluctuations. The rising and falling phases of the immune response are calculated based on the trends of changes in immune indicators and symptoms, forming a physiologically driven immune intensity trajectory that reflects the individual's true immune status. The process of age and exposure modulation is performed by calling age stage information from the physiological embedded feature set and combining class case counts, family exposure counts and ventilation scores to jointly modulate the amplitude and duration of the phased immune intensity trajectory, generating the modulated phased immune intensity trajectory, residual protection duration and phase gain factor. According to the preset field order, the phase label sequence, phase confidence sequence, vaccination-infection annotation sequence, modulated phased immune strength trajectory, residual protection duration, phase gain factor, data integrity marker and consistency score are packaged to form a time-varying immune strength representation and output as an immune potential feature set.

[0026] In this embodiment, the output graph potential field feature set after extracting principal component features through spectral domain gated dimensionality reduction includes: Based on social contact data, family-level contact maps and school-level contact maps were constructed separately. Node identifiers were unified and the number and duration of contact were counted according to the most recent fourteen-day window. The maps were then merged into multi-level maps, generating an adjacency table and a node metric table. The initial exposure vector was formed by weighting the family symptom count and the class confirmed case count, and the diffusion rounds and reinjection ratios were fixed. The multi-layer graph is subjected to edge intervention and weighting. The edge weights are then segmented and time-decayed according to mask-wearing compliance, ventilation score, isolation status, and contact proximity. The baseline contact matrix and intervention contact matrix are output, and row-wise normalization and node order locking are completed. Specifically, the segmented mapping and time decay of edge weights according to mask-wearing compliance, ventilation score, isolation status, and contact proximity are as follows: The weight of each edge is initially calculated using the original contact frequency and duration as the basic weights; then, a protection correction coefficient is set based on the mask wearing compliance. When the compliance is higher than a preset threshold, the edge weight is reduced proportionally to reflect the transmission reduction effect of protective behavior. The edge weights are mapped in a dual segmentation based on the classroom ventilation score and the isolation status. When the ventilation score is in a high range, the edge weights decrease linearly in the opposite direction according to the ventilation level. If any node is in an isolated state, the relevant edge weights are set to zero or significantly decreased to eliminate potential contact paths. In the time dimension, a contact proximity attenuation mechanism is introduced, which exponentially attenuates contact records that exceed a preset time window according to time intervals, making recent contacts more weighted and long-term contacts gradually weakened, ultimately generating an intervention contact matrix that takes into account both behavioral protection and time-series attenuation characteristics. Dual diffusion operations are performed on the baseline contact matrix, with neighborhood weighted aggregation and proportional reinjection of the initial exposure vector performed cyclically in a preset number of rounds to obtain the propagation potential field representation. Dual diffusion operations are then performed on the intervention contact matrix with the same number of rounds and reinjection ratio to obtain the control potential field representation. Specifically, performing dual diffusion operations on the intervention contact matrix with the same number of rounds and reinjection ratio involves: Based on the intervention contact matrix, neighborhood weighted aggregation is performed in each round of diffusion according to the adjacency relationship and edge weight strength of the nodes, and the risk potential energy of the surrounding nodes is accumulated and transferred to the target node, simulating the propagation weakening process under the control conditions. After each round of aggregation, a proportional back-injection mechanism is introduced into the node state vector. The risk weight of the corresponding node in the initial exposure vector is re-injected according to a set ratio to maintain the continuous influence of the individual node exposure characteristics during the diffusion process, making the potential field evolution process more stable. In multiple rounds of diffusion iteration, convergence detection is performed on the risk potential energy changes of nodes. Diffusion is terminated when the potential energy increment is lower than a preset threshold or the maximum number of rounds is reached, and the stable control potential field representation formed under intervention constraints is output. The difference between the propagation potential field representation and the control potential field representation is calculated to obtain the interferometric potential field representation. Spectral domain gating dimensionality reduction is then performed on the interferometric potential field representation, specifically as follows: Eigendecomposition is performed on the graph Laplacian corresponding to the multi-layer graph to obtain three frequency bands: low frequency, mid frequency and high frequency. High frequency components are suppressed according to a preset gating mask while retaining the dominant mid frequency and low frequency components. Based on a cumulative variance retention rate of at least 90%, principal component features are extracted to generate spectral domain-gated interferometric principal component vectors. The interference principal component vector, along with the mean, maximum value, and standard deviation of the propagation potential field, the mean, maximum value, and standard deviation of the control potential field, and the energy proportions of the three frequency bands, are summarized into a graphical potential field feature set.

[0027] In this embodiment, the output individual infection risk probability includes: Establish an input feature set, and perform field alignment and time window alignment on the physiological embedding feature set, immune potential feature set, graph potential field feature set, normalized continuous features and category features, and complete missing label completion and scale consistency processing; An improved DeepFM risk assessment model is constructed, comprising a multi-potential field collaborative fusion module, a potential field gating interaction module, and a multi-scale deep combination module, wherein: The improved DeepFM risk assessment model architecture first receives physiological embedding features, immune potential features, graph potential field features, and normalized continuous and categorical features at the input layer. After unified encoding, these features are fed into the three major modules in parallel. After embedding mapping and standardization, the feature inputs are divided into low-order features and high-order potential field features, which together constitute the input space of the improved DeepFM risk assessment model. The multi-potential field collaborative fusion module is located between the input layer and the feature interaction layer of the improved DeepFM risk assessment model. It is used to complete cross-modal alignment before the features enter the DeepFM backbone, perform vector space mapping and channel weight calibration on physiological embedding, immune potential and graph potential field features, and output the fused potential field representation. The potential field gating interaction module is connected to the FM branch structure and is used to introduce gating control signals in the second-order feature interaction process. It generates dynamic gating weights based on the multi-potential field fusion results to filter and constrain the feature interaction intensity, thereby highlighting the key interactions between physiological and immune-related features and suppressing low-correlation feature combinations. The multi-scale deep combination module is connected to the Deep branch structure and receives high-dimensional fusion vectors from the gated interaction module. It extracts nonlinear combination features at different levels through multi-scale convolution and residual connection mechanisms. The multi-scale deep combination module is equipped with local combination units and cross-layer aggregation units to realize hierarchical feature extraction from individual features to group risks. In the multi-potential field collaborative fusion module, channel alignment and intensity calibration are performed on physiological embedding features, immune potential features and graph potential field features to generate fusion context vectors and channel-level masks. Channel-wise reweighting and selective retention are performed on the embedded and mapped category features and normalized continuous features. In the potential field gating interaction module, the second-order feature interaction pairs of the FM branch are gating and weighted according to the fusion context vector. Interaction pairs involving physiological embedding features and immune potential features are preferentially calculated. Interference suppression and threshold pruning are performed on interaction pairs related to graph potential field features to generate the gating and calibrated second-order interaction representation. In the multi-scale deep combination module, the input of the Deep branch is processed in a multi-scale segmentation, and the local combination unit and the cross-layer residual mixing unit are stacked sequentially. Special tokens are introduced for physiological embedding features, immune potential features and graph potential features and jointly arranged with continuous features and category features to output a multi-scale deep representation. Branch fusion and risk aggregation are performed on the second-order interactive representation and the multi-scale deep representation. According to the combination rules of fixed weight and data-driven weight, individual infection risk scores are generated and converted into individual infection risk probabilities.

[0028] In this embodiment, the output of the corresponding risk level and prevention and control recommendations includes: The inverse indicators of individual infection risk probability, daily average value of transmission potential field representation intensity, and residual protection duration are weighted and synthesized according to three preset non-negative weights. The synthesized results are then interval-standardized within the same batch to obtain a comprehensive risk score. The grading thresholds are determined based on the comprehensive risk score. The low-risk threshold is set to the 40th percentile of the distribution, and the high-risk threshold is set to the 80th percentile of the distribution. The threshold versions are recorded. The risk level is determined based on the comprehensive risk score and the grading threshold. When the score is below the low threshold, it is judged as low risk; when it is between the low and high thresholds, it is judged as medium risk; and when it is above the high threshold, it is judged as high risk. Corresponding treatment suggestions are generated, and the stability of continuous assessment results is tested. If the same child is in a high-risk state for a consecutive preset number of days, an escalation warning is triggered, and the risk level, treatment suggestions, and time information are synchronized to the parent's terminal and the medical terminal.

[0029] refer to Figure 2 A risk assessment system for pediatric respiratory infectious diseases includes the following modules: The data processing module is used to collect multi-source data and perform preprocessing, extract continuous features and categorical features, and generate a standardized sample set; The physiological embedding module is used to extract age, height, weight, and developmental indicators, construct the physiological age transfer embedding module, and form a physiological embedding feature set; The immune modeling module is used to generate time-varying immune strength and residual protection duration based on a normalized input sample set through immune potential evolution modeling, and outputs an immune feature set. The graph potential field module is used to merge the contact graphs of the family layer and the school layer, perform dual diffusion to generate the propagation potential field and the prevention and control potential field, and output the graph potential field feature set. The risk assessment module is used to model and output the individual's infection risk probability by improving the DeepFM risk assessment model; The early warning output module is used to calculate a comprehensive score, classify risk levels according to thresholds, generate prevention and control suggestions, and trigger an early warning when there is a continuous high risk, and synchronize the results to the parent terminal and the medical terminal.

[0030] Example 1: To verify the feasibility of this invention in practice, it was applied to a key primary school, covering 520 students over a 90-day period. Data collected included children's physiological development, vaccination records, classroom environmental parameters, family member health status, and classroom contact network structure. During the study, the incidence of respiratory infections (mainly influenza and adenovirus) was at its annual peak in winter, providing an ideal environmental basis for model validation.

[0031] In the initial stage of project implementation, the system automatically collected and synchronized various types of data through the data processing module. The parent-side health monitoring mini-program regularly uploaded physiological data such as children's temperature, sleep duration, and activity levels; the school clinic system connected to the local CDC interface to obtain students' vaccination records and the most recent pathogen test results; environmental sensor nodes uploaded classroom temperature and humidity, carbon dioxide concentration, and ventilation scores in real time; the family-side system collected family members' cough frequency and isolation status; and the school information system provided class seating charts and contact logs. All data underwent standardized preprocessing to generate a standardized input sample set.

[0032] In the physiological age transfer embedding stage, the system dynamically models data such as age, height, weight, and body fat percentage for each student. For example, children aged 6 to 10 years have an average height growth rate of 5.2 cm and a weight growth rate of 2.8 kg per year. The model forms an age trajectory embedding vector through continuous stage transfer calculations. Compared to the traditional approach of linearly dividing based on a single age indicator, this method improves the accuracy of individual feature representation on the training set by approximately 23.6%. The system detected feature drift in the immune potential dimension of 8-year-old boys, indicating a significant increase in immune sensitivity at this stage. The model automatically adjusts the weights to improve the predictive sensitivity for this group.

[0033] The immune potential evolution modeling module performs time-series modeling of student vaccination records. Taking the H1N1 influenza vaccine as an example, 95% of children completed their vaccinations between June and August 2025, with an average interval of 52 days. The model divides the vaccination timeline into an activation period (0–14 days), a plateau period (15–60 days), and a decline period (after 61 days), and combines this with natural infection records to generate an immune strength trajectory. Compared with traditional static immune status recording methods, the modeling method of this invention can more accurately reflect the process of immune protection decline, increasing the contribution of immune-related features to the prediction of disease probability by approximately 31%. For example, a 9-year-old girl experienced a rapid decline in immune potential during the decline period, and her comprehensive risk score increased by 0.18 compared to the plateau period. The system issued a medium-risk warning on the sixth day, and she was subsequently clinically diagnosed with mild influenza.

[0034] In the graph potential generation stage, the system constructs contact graphs based on the contact relationships between families and schools, and merges them into a multi-layer graph structure. Taking Class 2 of Grade 2 as an example, the average number of family-level contacts for the 42 students in the class was 4.3 times and the average number of school-level contacts was 68.5 times within a month. The model intervenes and reweights the edge weights based on mask-wearing compliance (average 0.82), classroom ventilation score (average 0.74), and isolation status (average 0.05), and performs dual diffusion to generate the propagation potential field and the prevention and control potential field. The calculation results show that the average strength of the propagation potential field of the class contact network is 0.62, and the strength of the prevention and control potential field is 0.48. The difference between the two forms the interference potential field representation. After extracting principal components through spectral domain gating dimensionality reduction, the system identifies three key contact nodes (students sitting near the window) as propagation centers.

[0035] In the risk assessment phase of the improved DeepFM, the system inputs physiological embedding features, immune potential features, graphical potential field features, and continuous and categorical features into the model. Internally, the model achieves feature alignment and semantic unification through a multi-potential field collaborative fusion module, filters low-association feature pairs through a potential field gating interaction module, and performs hierarchical learning through a multi-scale deep combination module. Actual results show that the model achieves an AUC of 0.924 on the validation set, representing improvements of 18.3% and 6.8% compared to the traditional logistic regression model (AUC=0.781) and the basic DeepFM model (AUC=0.865), respectively. The average prediction latency is reduced by 2.1 days.

[0036] In the risk grading and early warning output stage, the system comprehensively scores and dynamically classifies the probability of individual infection risk. During the pilot period, the model issued 37 high-risk warnings, of which 31 were confirmed as positive infection cases by medical personnel, achieving an accuracy rate of 83.8%. For individuals who are in a high-risk state for three consecutive days, the system automatically triggers an upgraded warning and synchronizes the information to parents and school doctors.

[0037] Table 1. Statistical data from the pilot program for assessing the risk of respiratory infections in children.

[0038] As shown in Table 1, the pediatric respiratory infectious disease risk assessment method proposed in this invention can effectively distinguish the infection risk levels of different individuals and maintains a high degree of consistency with actual infection results. In the overall sample of 10 children, 4 cases (C001, C003, C007, and C009) were predicted as high-risk by the model, and all were confirmed as positive in subsequent tests, achieving a prediction accuracy of 100%. The 3 medium-risk cases (C002, C006, and C010) were not diagnosed, but two of them developed mild respiratory symptoms the following week, indicating that the model has strong sensitivity in identifying medium-risk cases. The 3 low-risk cases (C004, C005, and C008) remained healthy and showed no signs of infection, further validating the reliability and stability of the model's classification.

[0039] The characteristic data show that individual risk probability is significantly correlated with multiple factors. High-risk samples generally have shorter vaccination intervals (average 46 days), lower ventilation scores (average 0.72), and higher school contact frequency (average 74 times). These factors collectively lead to an increased transmission potential field, significantly increasing the risk of infection. In contrast, low-risk samples have higher ventilation scores (average 0.81) and fewer household contact frequency (average 3.0 times), indicating that environmental ventilation conditions and social contact frequency are important parameters affecting infection risk. Furthermore, in terms of gender distribution, the overall risk probability of boys is slightly higher (average 0.68) than that of girls (average 0.44), which is related to children's physiological differences and behavioral exposure characteristics, further confirming the effectiveness of the physiological age transfer embedding module in capturing gender differences.

[0040] The individual risk probabilities output by the model show a significant consistency with the actual infection situation, indicating that this invention has achieved good results in comprehensively modeling multi-source features, dynamically tracking the immunization timeline, and identifying social transmission structures. Compared with traditional statistical models that rely solely on single-dimensional health data, it improves the accuracy of infection risk prediction and early warning capabilities. In actual pilot tests, the model can issue warnings an average of two days before children develop clinical symptoms, helping schools and parents to take timely isolation and protective measures and reducing the spread of infection in classrooms. The multi-source fusion risk assessment method based on the improved DeepFM model proposed in this invention can achieve high-precision prediction and intelligent early warning of pediatric respiratory infectious diseases, providing a scientific and feasible decision-making basis for public health management.

[0041] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for risk assessment of pediatric respiratory infectious diseases, characterized in that, include: Collect multi-source data of the target children, perform preprocessing on the multi-source data, extract continuous data as continuous features, extract discrete data as category features, and obtain a normalized input sample set; Based on the normalized input sample set, age, height, weight and growth and development indicators are extracted, input into the physiological age transfer embedding module, continuous age trajectory embedding vector is generated, and physiological embedding feature set is output. Based on a standardized input sample set, information on vaccine type, vaccination date and vaccination interval is extracted from the immunization data. Through immune potential evolution modeling, a representation of immune strength that changes over time is generated, and an immune potential feature set is output. Construct and merge family-level and school-level contact graphs to form a multi-layer graph. Based on mask-wearing compliance, ventilation score and isolation status, intervene and reweight the edge weights of the multi-layer graph to obtain the baseline contact matrix and intervention contact matrix. Perform dual diffusion operation to generate the propagation potential field representation and the prevention and control potential field representation. Calculate the difference to obtain the interference potential field representation. After spectral domain gating dimensionality reduction to extract principal component features, output the graph potential field feature set. The DeepFM risk assessment model is improved by inputting physiological embedding feature set, immune potential feature set, graphical potential field feature set, and normalized continuous and categorical features together to output the individual infection risk probability. Based on the individual's infection risk probability, a comprehensive risk score is calculated and threshold classification is applied. The results are divided into low-risk, medium-risk, and high-risk levels, and the corresponding risk levels and prevention and control recommendations are output and synchronized to the parents' terminal and medical terminal to complete the early warning and risk reminder of children's respiratory infectious diseases.

2. The method for risk assessment of pediatric respiratory infectious diseases according to claim 1, characterized in that, The multi-source data includes physiological data, immunization data, environmental and behavioral data, and social contact data from families and schools.

3. The method for risk assessment of pediatric respiratory infectious diseases according to claim 1, characterized in that, The preprocessing of multi-source data includes format unification, outlier removal, missing value completion, and standardization.

4. The method for risk assessment of pediatric respiratory infectious diseases according to claim 1, characterized in that, The output physiological embedding feature set includes: Based on a standardized input sample set, a physiological basic feature table containing age, height, weight, body mass index, height standard grading, and weight standard grading is generated. A physiological age transfer embedding module is constructed, which consists of a stage boundary adaptive recognition unit, an age trajectory continuous transfer coding unit, and a physiological abnormality sensitivity gating unit. In the stage boundary adaptive recognition unit, the 10th, 50th, and 90th percentiles are used as the initial stage boundaries. The height standard class and weight standard class of the samples are segmented and assigned. The stage boundaries are iteratively adjusted until convergence according to the criterion that the misclassification rate is no higher than 5%. The stage label sequence and stage confidence sequence are output. In the age trajectory continuous migration coding unit, an age time series segment is constructed with a window size of 3 months and a step size of 1 month. The percentile change values ​​of adjacent windows are classified to generate a trajectory state sequence and a transition coding sequence. In the physiological abnormality sensitivity gating unit, the abnormality sensitivity level is determined as high based on the judgment rules that the difference between the maximum and minimum body temperature within a day is greater than or equal to 1.0℃, the heart rate is at or above the 95th percentile of the same age, or the respiratory rate is at or above the 95th percentile of the same age. Otherwise, it is medium or low. The abnormality sensitivity level is paired with the trajectory state sequence item by item to generate gating weight labels. The trajectory state sequence and the transition coding sequence are then gating fused to obtain the gating physiological coding vector. The stage label sequence, stage confidence sequence, trajectory state sequence, transition coding sequence, and gated physiological coding vector are concatenated and length consistency is checked according to the preset field order to generate continuous age trajectory embedding vectors and summarize them into a physiological embedding feature set.

5. The method for risk assessment of pediatric respiratory infectious diseases according to claim 1, characterized in that, The output immune potential feature set includes: Based on the standardized input sample set, an immunization event timeline with the assessment time as the anchor point is established. The vaccine type, production platform, vaccination date, number of doses, vaccination interval, homologous or heterologous sequential type and previous positive etiological records are summarized one by one to generate a vaccination-infection annotation sequence arranged in chronological order. A three-stage immunization state machine is constructed in the timeline corresponding to each vaccine. The stages are divided in the order of activation, plateau and decay. The start and end positions of the three stages are determined by the boundary adaptive method based on dose and interval. The stage boundaries of adjacent vaccination events are subjected to non-overlapping constraint processing. The stage label sequence and stage confidence sequence are output. A dual-channel immune proxy fusion unit was constructed. Channel 1 generates an inoculation-driven immune intensity trajectory based on the inoculation-infection labeled sequence and the stage label sequence. Channel 2 generates a physiologically driven immune intensity trajectory based on immune-related laboratory indicators and symptom timelines. The two trajectories are aligned and scored on consistency on the time axis, and the fused staged immune intensity trajectory is output. The process of age and exposure modulation is performed by calling age stage information from the physiological embedded feature set and combining class case counts, family exposure counts and ventilation scores to jointly modulate the amplitude and duration of the phased immune intensity trajectory, generating the modulated phased immune intensity trajectory, residual protection duration and phase gain factor. According to the preset field order, the phase label sequence, phase confidence sequence, vaccination-infection annotation sequence, modulated phased immune strength trajectory, residual protection duration, phase gain factor, data integrity marker and consistency score are packaged to form a time-varying immune strength representation and output as an immune potential feature set.

6. The method for risk assessment of pediatric respiratory infectious diseases according to claim 1, characterized in that, The output graph potential field feature set after principal component feature extraction via spectral domain gated dimensionality reduction includes: Based on social contact data, family-level contact maps and school-level contact maps were constructed separately. Node identifiers were unified and the number and duration of contact were counted according to the most recent fourteen-day window. The maps were then merged into multi-level maps, generating an adjacency table and a node metric table. The initial exposure vector was formed by weighting the family symptom count and the class confirmed case count, and the diffusion rounds and reinjection ratios were fixed. The edge intervention kernel weighting is performed on the multi-layer graph. The edge weights are segmented and time-decayed according to mask wearing compliance, ventilation score, isolation status and contact proximity. The baseline contact matrix and intervention contact matrix are output, and row normalization and node order locking are completed. Dual diffusion operations are performed on the baseline contact matrix, and neighborhood weighted aggregation and proportional back-injection of the initial exposure vector are performed in a preset number of rounds to obtain the propagation potential field representation; dual diffusion operations are performed on the intervention contact matrix in the same number of rounds and with the same back-injection ratio to obtain the control potential field representation. The difference between the propagation potential field representation and the control potential field representation is calculated to obtain the interferometric potential field representation. Spectral domain gating dimensionality reduction is then performed on the interferometric potential field representation, specifically as follows: Eigendecomposition is performed on the graph Laplacian corresponding to the multi-layer graph to obtain three frequency bands: low frequency, mid frequency and high frequency. High frequency components are suppressed according to a preset gating mask while retaining the dominant mid frequency and low frequency components. Based on a cumulative variance retention rate of at least 90%, principal component features are extracted to generate spectral domain-gated interferometric principal component vectors. The interference principal component vector, along with the mean, maximum value, and standard deviation of the propagation potential field, the mean, maximum value, and standard deviation of the control potential field, and the energy proportions of the three frequency bands, are summarized into a graphical potential field feature set.

7. The method for risk assessment of pediatric respiratory infectious diseases according to claim 1, characterized in that, The output individual infection risk probability includes: Establish an input feature set, and perform field alignment and time window alignment on the physiological embedding feature set, immune potential feature set, graph potential field feature set, normalized continuous features and category features, and complete missing label completion and scale consistency processing; An improved DeepFM risk assessment model is constructed, which includes a multi-potential field collaborative fusion module, a potential field gating interaction module, and a multi-scale deep combination module. In the multi-potential field collaborative fusion module, channel alignment and intensity calibration are performed on physiological embedding features, immune potential features and graph potential field features to generate fusion context vectors and channel-level masks. Channel-wise reweighting and selective retention are performed on the embedded and mapped category features and normalized continuous features. In the potential field gating interaction module, the second-order feature interaction pairs of the FM branch are gating and weighted according to the fusion context vector. Interaction pairs involving physiological embedding features and immune potential features are preferentially calculated. Interference suppression and threshold pruning are performed on interaction pairs related to graph potential field features to generate the gating and calibrated second-order interaction representation. In the multi-scale deep combination module, the input of the Deep branch is processed in a multi-scale segmentation, and the local combination unit and the cross-layer residual mixing unit are stacked sequentially. Special tokens are introduced for physiological embedding features, immune potential features and graph potential features and jointly arranged with continuous features and category features to output a multi-scale deep representation. Branch fusion and risk aggregation are performed on the second-order interactive representation and the multi-scale deep representation. According to the combination rules of fixed weight and data-driven weight, individual infection risk scores are generated and converted into individual infection risk probabilities.

8. The method for risk assessment of pediatric respiratory infectious diseases according to claim 1, characterized in that, The output of the corresponding risk level and prevention and control recommendations includes: The inverse indicators of individual infection risk probability, daily average value of transmission potential field representation intensity, and residual protection duration are weighted and synthesized according to three preset non-negative weights. The synthesized results are then interval-standardized within the same batch to obtain a comprehensive risk score. The grading thresholds are determined based on the comprehensive risk score. The low-risk threshold is set to the 40th percentile of the distribution, and the high-risk threshold is set to the 80th percentile of the distribution. The threshold versions are recorded. The risk level is determined based on the comprehensive risk score and the grading threshold. When the score is below the low threshold, it is judged as low risk; when it is between the low and high thresholds, it is judged as medium risk; and when it is above the high threshold, it is judged as high risk. Corresponding treatment suggestions are generated, and the stability of continuous assessment results is tested. If the same child is in a high-risk state for a consecutive preset number of days, an escalation warning is triggered, and the risk level, treatment suggestions, and time information are synchronized to the parent's terminal and the medical terminal.

9. A risk assessment system for pediatric respiratory infectious diseases, comprising the method for assessing the risk of pediatric respiratory infectious diseases as described in any one of claims 1 to 8, characterized in that, Includes the following modules: The data processing module is used to collect multi-source data and perform preprocessing, extract continuous features and categorical features, and generate a standardized sample set; The physiological embedding module is used to extract age, height, weight, and developmental indicators, construct the physiological age transfer embedding module, and form a physiological embedding feature set; The immune modeling module is used to generate time-varying immune strength and residual protection duration based on a normalized input sample set through immune potential evolution modeling, and outputs an immune feature set. The graph potential field module is used to merge the contact graphs of the family layer and the school layer, perform dual diffusion to generate the propagation potential field and the prevention and control potential field, and output the graph potential field feature set. The risk assessment module is used to model and output the individual's infection risk probability by improving the DeepFM risk assessment model; The early warning output module is used to calculate a comprehensive score, classify risk levels according to thresholds, generate prevention and control suggestions, and trigger an early warning when there is a continuous high risk, and synchronize the results to the parent terminal and the medical terminal.

Images