Neuroimmune disease risk prediction method based on artificial intelligence

Abstract

The invention relates to the technical field of artificial intelligence neural immune risk prediction, and discloses a neural immune disease risk prediction method based on artificial intelligence. According to the method, a dynamic health map fusing individual physiological time sequence data and external environment factors is constructed, a risk deconstruction network of a preset pathophysiological pathway is utilized to analyze a multi-pathway risk activation signal, and a to-be-verified risk event is screened out according to a predefined threshold value; triggering an independent thread, and calling a matching mode in the historical pathology database to perform event review and credibility scoring; and integrating the confirmed events, calculating an aggregation risk index, and generating an individualized risk report containing the differentiated monitoring tasks. According to the method, dynamic monitoring, multi-channel deconstruction and evidence verification of disease risks are realized, and the accuracy and interpretability of early warning are improved.

Description

An Artificial Intelligence-Based Method for Predicting the Risk of Neuroimmune Diseases Technical Field

[0001] This invention relates to the field of artificial intelligence neuroimmune risk prediction technology, specifically to a method for predicting the risk of neuroimmune diseases based on artificial intelligence. Background Technology

[0002] Current early risk prediction for neuroimmunological diseases primarily relies on static clinical indicator analysis and single biomarker detection. These methods typically deal with cross-sectional data or isolated physiological parameters, failing to effectively integrate the continuously changing physiological temporal information of individuals with complex and ever-changing external environmental factors. Since the occurrence and development of neuroimmunological diseases are the result of multiple physiological pathways being affected by internal and external factors over time, this assessment method based on fragmented information is ill-suited to capturing the dynamic evolution of disease risk.

[0003] Existing technical solutions have shortcomings. At the data level, there is a lack of mechanisms for effectively integrating individual physiological time-series data with external environmental factors. Data from different sources fail to achieve precise alignment and deep correlation analysis across temporal and spatial scales, resulting in information fragmentation. At the analytical level, most models only output a general comprehensive risk score, failing to analyze the specific pathophysiological pathways from which the risk originates, thus limiting the biological interpretability of risk assessment and its guiding value for precision intervention. Furthermore, the risk signal output of these models often lacks a closed-loop verification process with historical clinical pathological evidence, and the reliability and credibility of their warning results need improvement.

[0004] Existing methods face challenges in achieving early, accurate, and interpretable risk warnings, and key issues such as deep fusion of multi-source dynamic data, pathophysiological pathway-level deconstruction of risk signals, and evidence verification of warning events need to be addressed. Summary of the Invention

[0005] The purpose of this invention is to provide an artificial intelligence-based method for predicting the risk of neuroimmune diseases, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a method for predicting the risk of neuroimmunological diseases based on artificial intelligence, the method comprising:

[0007] A dynamic health data model is initialized, and an individual physiological time-series data queue and an external environmental factor set are configured. A collaborative analysis engine is invoked to perform fusion calculations on the individual physiological time-series data queue and the external environmental factor set, generating a fused health atlas with time-series labels. The fused health atlas is input into a trained risk deconstruction network, which outputs a set of risk activation signals corresponding to different pathophysiological pathways. This includes: the risk deconstruction network comprising a multilayer feature perceptron and a pathway output layer; the fused health atlas is input into the multilayer feature perceptron of the risk deconstruction network, and abstract feature transformation is performed layer by layer to obtain a high-dimensional implicit feature representation; the high-dimensional implicit feature representation is input into the pathway output layer, which predefines several pathophysiological pathways related to neuroimmunological diseases, each pathophysiological pathway corresponding to an independent signal output unit; each signal output unit... The high-dimensional implicit feature representation is calculated to output a continuous numerical signal representing the degree of risk activation of the pathway. The continuous numerical signals output by all signal output units constitute the set of risk activation signals. According to predefined risk signal threshold rules, the set of risk activation signals is filtered, and risk activation signals exceeding the corresponding threshold are marked as risk events to be verified. For each risk event to be verified, an independent verification analysis thread is triggered to retrieve a matching reference pattern from the historical pathology database for event review. All verified risk events are integrated to form a structured risk event set. Based on the risk activation signal intensity and pathophysiological pathway attribution of each event in the risk event set, an aggregated risk index is calculated. Based on the aggregated risk index and the risk event set, an individualized risk prediction report is constructed, and a differentiated monitoring and alert task sequence is generated and executed.

[0008] Preferably, initializing a dynamic health data model and configuring an individual physiological time-series data queue and an external environmental factor set includes: creating a blank architecture for the dynamic health data model, the blank architecture including a data input port, an internal buffer framework, and a data output port; allocating a first data channel and a second data channel to the data input port, the first data channel being dedicated to receiving streaming physiological parameters and the second data channel being dedicated to receiving discrete environmental parameters; instantiating the individual physiological time-series data queue within the internal buffer framework and logically binding the input end of the individual physiological time-series data queue to the first data channel; instantiating the external environmental factor set within the internal buffer framework and logically binding the input end of the external environmental factor set to the second data channel; configuring a trigger condition for the data output port, the trigger condition being that the data volume of the individual physiological time-series data queue reaches a preset capacity or the monitoring period reaches a cutoff point; within the preset monitoring period, continuously injecting multi-dimensional physiological parameters of the target individual into the individual physiological time-series data queue, and synchronously updating the spatiotemporal environmental parameters associated with the target individual into the external environmental factor set.

[0009] Preferably, the step of continuously injecting multi-dimensional physiological parameters of the target individual into the individual physiological time-series data queue within a preset monitoring period, and simultaneously updating the spatiotemporal environmental parameters associated with the target individual into the external environmental factor set, includes: acquiring the raw physiological signal stream of the target individual through a connected biosensor device; performing noise reduction and feature extraction operations on the raw physiological signal stream to obtain standardized multi-dimensional physiological parameters, each dimension corresponding to a physiological function indicator; sequentially pushing the standardized multi-dimensional physiological parameters into the individual physiological time-series data queue according to their timestamps; simultaneously, asynchronously acquiring environmental data of the target individual's geographical location and publicly available health event data from the environmental monitoring network and service interface, wherein the environmental data includes climate indicators and pollutant concentrations, and the health event data includes regional epidemic information; formatting the acquired environmental data and health event data to generate spatiotemporal environmental parameters; and replacing the historical parameters stored in the external environmental factor set with the latest spatiotemporal environmental parameters within the current monitoring period.

[0010] Preferably, the step of calling the collaborative analysis engine to perform fusion calculations on the individual physiological time-series data queue and the external environmental factor set to generate a fused health atlas with time-series labels includes: extracting the current state of the individual physiological time-series data queue and the external environmental factor set from the data output port of the dynamic health data model; the collaborative analysis engine includes a time-series alignment module and a feature cross-calculation module; using the time-series alignment module to map the parameters in the external environmental factor set to each time node of the individual physiological time-series data queue according to time slices, forming a time-aligned mixed data frame; sending the time-aligned mixed data frame into the feature cross-calculation module, which calculates the nonlinear correlation weights between physiological parameters, between environmental parameters, and between physiological and environmental parameters; and re-encoding and concatenating all parameters in the time-aligned mixed data frame according to the calculated nonlinear correlation weights to generate a multidimensional fused health atlas with time-series labels, where each node in the fused health atlas represents an enhanced feature.

[0011] Preferably, the step of filtering the set of risk activation signals according to predefined risk signal threshold rules and marking risk activation signals exceeding the corresponding threshold as risk events to be verified includes: acquiring the set of risk activation signals output by the risk deconstruction network, each risk activation signal having an identity identifier of its respective pathophysiological pathway; reading a predefined risk signal threshold rule library, the risk signal threshold rule library setting a static threshold and a dynamic adjustment coefficient for each pathophysiological pathway; determining the specific value of the dynamic adjustment coefficient based on the target individual's basic health record, and calculating the currently applicable dynamic threshold for each pathophysiological pathway in combination with the static threshold; comparing the value of each risk activation signal with its corresponding dynamic threshold; if the value of a certain risk activation signal is greater than its corresponding dynamic threshold, then the risk activation signal and its complete context features are extracted from the fused health map, encapsulated together into a risk event data packet to be verified, and a unique event identifier is assigned to the data packet.

[0012] Preferably, for each risk event to be verified, triggering an independent verification analysis thread to retrieve a matching reference pattern from the historical pathology database for event review includes: determining the historical pathology database sub-database to be queried based on the pathophysiological pathway identity identifier of the risk activation signal in the risk event data packet; in the independent verification analysis thread, converting the contextual features in the risk event data packet into a query vector; using the query vector to perform a similarity search in the corresponding historical pathology database sub-database to find at least one historical reference pattern whose feature pattern matching degree exceeds a preset matching threshold; extracting the final clinical outcome labels of historical risk events similar to the current risk event to be verified from the historical reference patterns; scoring the credibility of the current risk event to be verified based on the statistical distribution of the final clinical outcome labels to obtain a credibility score; if the credibility score exceeds a preset credibility threshold, then the risk event to be verified is determined to have passed the review and is marked as a confirmed risk event.

[0013] Preferably, the integration of all verified risk events to form a structured risk event set includes: collecting all data packets of risk events marked as confirmed risk events; grouping all confirmed risk events according to pathophysiological pathways; sorting them in descending order based on the risk activation signal values ​​of each confirmed risk event within the same pathophysiological pathway group; supplementing each confirmed risk event with its credibility score obtained in the verification analysis thread; and organizing all grouped, sorted, and credible risk events into a hierarchical, indexable risk event set data structure.

[0014] Preferably, the step of calculating the aggregated risk index based on the risk activation signal intensity and pathophysiological pathway attribution of each event in the risk event set includes: reading the risk activation signal value and confidence score of each confirmed risk event from the structured risk event set; obtaining a pre-configured pathophysiological pathway weight table, which specifies the relative importance weight of different pathways in the overall risk assessment; finding the corresponding relative importance weight based on the pathophysiological pathway attribution of each confirmed risk event; inputting the risk activation signal value, confidence score, and relative importance weight of each pathway of each confirmed risk event into an aggregation calculation function, which outputs a comprehensive scalar value; and normalizing the comprehensive scalar value to a preset index range to obtain the final aggregated risk index.

[0015] Preferably, the step of constructing a personalized risk prediction report based on the aggregated risk index and the risk event set, and generating and executing a differentiated monitoring and alert task sequence, includes: creating a blank template for the personalized risk prediction report, the blank template containing an index summary area, an event details area, and an action suggestion framework; filling the aggregated risk index and its corresponding risk level description into the index summary area; formatting and filling the main content of the structured risk event set, including the confirmed risk event's pathway, signal strength ranking, and credibility score, into the event details area; and adjusting the aggregated risk index according to the risk event set. For specific pathophysiological pathways, a list of corresponding monitoring items and health tips are selected from a pre-defined action rule base and filled into the action suggestion framework to generate a complete personalized risk prediction report. The content of the action suggestion framework in the personalized risk prediction report is parsed, and each monitoring item is converted into an executable monitoring task instruction, and each health tip is converted into a notification message to be sent. According to the urgency and logical dependency of the monitoring tasks, all monitoring task instructions are sorted and scheduled to form a monitoring task sequence. According to the pre-defined triggering time, the instructions in the monitoring task sequence are executed sequentially, and the corresponding notification messages are sent synchronously.

[0016] Compared with existing technologies, the advantages of this invention are as follows: By using a time-series alignment module, external environmental factors are mapped to each node of an individual's physiological time-series data in time slices, forming a time-aligned hybrid data frame. Then, a feature cross-calculation module analyzes the nonlinear correlation weights between physiological parameters, between environmental parameters, and between physiological and environmental parameters, and accordingly re-encodes and connects the features. This method can generate a dynamic, multi-dimensional fused health map, where each node represents the interaction characteristics of endogenous physiological state and exogenous environmental exposure at a specific moment. This overcomes the limitations of traditional methods, such as single data dimension and static analysis, and achieves a more comprehensive and realistic dynamic representation of an individual's health status.

[0017] A risk deconstruction network was designed, comprising a multi-layer feature perceptron and a pathway output layer. The pathway output layer predefines several pathophysiological pathways related to the disease, each corresponding to an independent signal output unit. This network can parse continuous numerical signals of risk activation levels corresponding to different pathways from a fused health atlas. For each risk signal exceeding a threshold, an independent validation analysis thread is triggered, performing a similarity search between its contextual features and reference patterns in a historical pathology database. The credibility of the current risk event is then scored based on the statistical distribution of clinical outcomes from historical cases. This mechanism combines data-driven risk detection with clinical evidence-based verification, enabling risk warnings to not only pinpoint specific biological processes but also possess traceable evidence support, thus enhancing the reliability of risk decisions. Attached Figure Description

[0018] Figure 1 is a schematic diagram of the working principle of the artificial intelligence-based neuroimmunological disease risk prediction method of the present invention; Figure 2 is a flowchart of initializing the dynamic health data model; Figure 3 is a flowchart of generating the fused health atlas; Figure 4 is a comparison diagram of pathophysiological pathway weights and average confidence scores of risk events; Figure 5 is a time-series change diagram of risk index and pathway signal values ​​during the dynamic health monitoring stage. Detailed Implementation

[0019] The technical solutions of 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.

[0020] Referring to Figure 1, this invention provides an artificial intelligence-based method for predicting the risk of neuroimmunological diseases. The method includes: initializing a dynamic health data model, which is configured with an individual physiological time-series data queue for receiving continuous individual physiological data and a set of external environmental factors for receiving external environmental parameters. A collaborative analysis engine is invoked to perform fusion calculations on the individual physiological time-series data queue and the set of external environmental factors to generate a fused health atlas with time-series labels. This fused health atlas is input into a pre-trained risk deconstruction network, which outputs a set of risk activation signals corresponding to different pathophysiological pathways of neuroimmunological diseases. Based on a set of predefined risk signal threshold rules, this set of risk activation signals is filtered, and signals with values ​​exceeding their corresponding thresholds are marked as risk events to be verified. For each risk event to be verified, the system triggers an independent verification analysis thread to retrieve a reference pattern matching feature pattern from a historical pathology database for event verification. All verified risk events are integrated to form a structured risk event set. Based on the risk activation signal intensity of each event in this set and its corresponding pathophysiological pathway, an aggregated risk index is calculated. Based on this aggregated risk index and structured risk event set, an individualized risk prediction report is constructed, and a differentiated monitoring and alert task sequence is generated and executed accordingly.

[0021] In one embodiment of the present invention, referring to Figure 2, the process of initializing the dynamic health data model involves creating a blank architecture for the model, which includes a data input port, an internal buffer framework, and a data output port. A first data channel and a second data channel are allocated to the data input port. The first data channel is dedicated to receiving streaming physiological parameters, and the second data channel is dedicated to receiving discrete environmental parameters. An individual physiological time-series data queue is instantiated within the internal buffer framework, and the input of this queue is logically bound to the first data channel. Simultaneously, an external environmental factor set is instantiated within the internal buffer framework, and the input of this set is logically bound to the second data channel. A trigger condition for the data output port is configured, which is set to the data volume of the individual physiological time-series data queue reaching a preset capacity or the preset monitoring period reaching its end point. Within the preset monitoring period, the raw physiological signal stream of the target individual is acquired through a connected biosensor device. Noise reduction and feature extraction operations are performed on the raw physiological signal stream to obtain standardized multi-dimensional physiological parameters, each dimension corresponding to a physiological function indicator. These standardized multi-dimensional physiological parameters are sequentially pushed into the individual physiological time-series data queue according to their timestamps. Simultaneously, environmental data of the target individual's geographical location and publicly available health event data are asynchronously obtained from the environmental monitoring network and service interface. The environmental data includes climate indicators and pollutant concentrations, and the health event data includes regional epidemic information. The obtained environmental data and health event data are formatted to generate spatiotemporal environmental parameters, and the latest spatiotemporal environmental parameters within the current monitoring period are used to replace the historical parameters stored in the external environmental factor set.

[0022] In practical implementation, the initialization of the dynamic health data model and the configuration of the individual physiological time-series data queue and the external environmental factor set are achieved through the following steps: First, create a blank architecture for the dynamic health data model. This blank architecture includes a data input port, an internal buffer framework, and a data output port. Allocate a first data channel and a second data channel to the data input port. The first data channel is dedicated to receiving streaming physiological parameters, and the second data channel is dedicated to receiving discrete environmental parameters. Second, instantiate the individual physiological time-series data queue within the internal buffer framework and logically bind the input end of the individual physiological time-series data queue to the first data channel. Third, instantiate the external environmental factor set within the internal buffer framework and logically bind the input end of the external environmental factor set to the second data channel. Finally, configure the trigger conditions for the data output port. The trigger conditions are set to the individual physiological time-series data queue reaching a preset capacity or the monitoring period reaching its end point.

[0023] In practice, data injection and updates are performed within a preset monitoring period. The raw physiological signal stream of the target individual is acquired through connected biosensor devices. This raw physiological signal stream includes heart rate variability (HRV) signals and electrodermal activity (EDA) signals. Noise reduction and feature extraction are performed on the raw physiological signal stream. The noise reduction operation uses wavelet transform to filter out power frequency interference. The feature extraction operation calculates the time-domain standard deviation and the low-frequency / high-frequency power ratio from the HRV signals. The non-specific EDA response frequency and average amplitude are extracted from the EDA signals, resulting in standardized multi-dimensional physiological parameters, each corresponding to a physiological function indicator. These standardized multi-dimensional physiological parameters are then sequentially pushed into the individual physiological time-series data queue according to their timestamps. The individual physiological time-series data queue uses a circular buffer structure; when the data volume exceeds a preset capacity, the oldest historical data is overwritten.

[0024] In some embodiments, environmental data is acquired and updated synchronously. Environmental data of the target individual's geographical location and publicly available health event data are asynchronously acquired from the environmental monitoring network and service interface. Environmental data includes temperature, humidity, PM2.5 concentration, and pollen index, while health event data includes regional influenza incidence reports. The acquired environmental and health event data are formatted to generate spatiotemporal environmental parameters. These parameters include geographic coordinates, timestamps, and corresponding parameter vectors. The latest spatiotemporal environmental parameters within the current monitoring period are used to replace the historical parameters stored in the external environmental factor set. The external environmental factor set is designed as a key-value pair storage structure, where the key is the parameter type and the value is the latest parameter vector.

[0025] In some embodiments, the triggering logic involves specific settings for a preset capacity and a monitoring cycle. The preset capacity is set to the number of data points capable of storing 168 consecutive hours of data, based on the sampling frequency of the biosensor and monitoring requirements. The monitoring cycle is set to a complete cycle of 24 hours, with a cutoff time of midnight each day. When the number of data points in the individual physiological time-series data queue reaches the preset capacity, the data output port is immediately triggered. Alternatively, when the system time reaches the monitoring cycle cutoff time of midnight each day, the data output port is also triggered regardless of whether the individual physiological time-series data queue is full, ensuring that the data can be processed periodically or on demand by the downstream collaborative analysis engine.

[0026] It is understandable that the standardization of multi-dimensional physiological parameters in feature extraction follows a unified protocol. The initial feature values ​​calculated from the original signal need to undergo standardization transformation, and the transformation formula is as follows:

[0027] in: This represents the original numerical value of a certain dimension feature calculated from the original physiological signal stream. This represents the mean of this characteristic in a reference healthy population. This represents the standard deviation of the characteristic in a reference healthy population. This represents one of the parameter values ​​in the multidimensional physiological parameters after standardization.

[0028] It is understandable that the update mechanism of the external environmental factor set is overriding. In each monitoring cycle, the system pulls the latest environmental data and health event data from multiple data sources, and after formatting, generates a brand new spatiotemporal environmental parameter object. This brand new spatiotemporal environmental parameter object completely replaces all the content previously saved in the external environmental factor set. The external environmental factor set does not retain historical environmental data sequences, but only maintains the latest and most relevant environmental snapshot at the current moment. This design reduces data storage overhead and focuses on the potential impact of recent environmental exposure on physiological state.

[0029] In one embodiment of the present invention, referring to Figure 3, when the collaborative analysis engine is invoked for fusion calculation, the current state of the individual physiological time-series data queue and the set of external environmental factors are first extracted from the data output port of the dynamic health data model. This collaborative analysis engine includes a time-series alignment module and a feature cross-calculation module. The time-series alignment module maps the parameters in the set of external environmental factors to each time node of the individual physiological time-series data queue according to time slices, forming a time-aligned hybrid data frame. This time-aligned hybrid data frame is then fed into the feature cross-calculation module, which calculates the nonlinear correlation weights between physiological parameters, between environmental parameters, and between physiological and environmental parameters. Based on the calculated nonlinear correlation weights, all parameters in the time-aligned hybrid data frame are re-encoded and concatenated to generate a multidimensional, time-labeled fusion health map. Each node in this fusion health map represents an enhanced feature. The fusion health map is then input into a trained risk deconstruction network, which includes multi-layer feature perceptrons and pathway output layers. The integrated health map first inputs to a multilayer feature perceptron for layer-by-layer abstract feature transformation to obtain a high-dimensional implicit feature representation. This high-dimensional implicit feature representation is then input to the pathway output layer. The pathway output layer predefines several pathophysiological pathways related to neuroimmunological diseases. Each pathophysiological pathway corresponds to an independent signal output unit. Each signal output unit calculates the high-dimensional implicit feature representation and outputs a continuous numerical signal representing the degree of risk activation of that pathway. The continuous numerical signals output by all signal output units constitute the set of risk activation signals.

[0030] In practical implementation, the process of calling the collaborative analysis engine to fuse and calculate the individual physiological time-series data queue and the set of external environmental factors to generate a fused health map, and then inputting the fused health map into the risk deconstruction network to output risk activation signals, is achieved through the following steps: The current state of the individual physiological time-series data queue and the set of external environmental factors are extracted from the data output port of the dynamic health data model. The individual physiological time-series data queue contains a multi-dimensional physiological parameter sequence sorted by timestamp, and the set of external environmental factors contains the latest spatiotemporal environmental parameters. The collaborative analysis engine includes a time-series alignment module and a feature cross-calculation module. The time-series alignment module processes the extracted data, reading the timestamps and parameters from the set of external environmental factors and mapping these parameters to each corresponding time node in the individual physiological time-series data queue according to time slices, thus forming a time-aligned hybrid data frame. Each row of the hybrid data frame represents a unified time point, and its columns contain all physiological and environmental parameters at that time point.

[0031] In practice, time-aligned hybrid data frames are fed into a feature cross-computation module for calculation. This module calculates nonlinear correlation weights between physiological parameters, environmental parameters, and between physiological and environmental parameters based on a pre-trained neural network model. The calculation process involves inputting each row of data from the hybrid data frame into a multilayer perceptron. The hidden layers of the multilayer perceptron perform nonlinear transformations and interactions on the input features, and the output layer generates a scalar weight representing the correlation strength for each pair of features. Based on the calculated nonlinear correlation weights, the feature cross-computation module re-encodes all parameters in the time-aligned hybrid data frame, fusing the original parameter values ​​with the corresponding correlation weights to generate a new, high-dimensional feature representation. Each dimension in this new feature representation is an augmented feature. The augmented features at all time points are combined sequentially to ultimately generate a fused health map with temporal labels. Structurally, the fused health map is a three-dimensional tensor, with its dimensions representing the time step, feature dimension, and augmented feature depth, respectively.

[0032] In some embodiments, the operation of the time-series alignment module involves specific time matching logic. Each physiological parameter data point in the individual physiological time-series data queue has a timestamp accurate to the second, while the spatiotemporal environmental parameters in the external environmental factor set may be updated at an hourly granularity. The time-series alignment module uses the physiological parameter timestamp as a benchmark to find the environmental parameter with the smallest time difference for each physiological parameter data point for pairing. If the difference between the timestamp of a physiological parameter point and the update timestamp of the environmental parameter exceeds a preset tolerance window, such as 30 minutes, the time-series alignment module uses linear interpolation to estimate the environmental value corresponding to that physiological time point based on the two nearest environmental parameter points, ensuring that each time node has complete physiological and environmental information, forming a seamless time-aligned hybrid data frame.

[0033] In some embodiments, the calculation of nonlinear association weights in the feature cross-calculation module relies on a pre-trained model trained on time-series data containing a large amount of known health and disease state annotations. The training objective is to minimize the combination of reconstruction loss and prediction loss. When a new time-aligned mixed data frame is input, the forward propagation process of the neural network in the feature cross-calculation module not only calculates the interaction weights between features but also performs element-wise scaling and bias adjustment on the original features, with the scaling factor determined by the interaction weights. The re-encoding and concatenation process can be formally represented by a transformation function that takes the original feature vector and the calculated full interaction weight matrix as input and outputs an enhanced feature vector, thereby constructing each node of the fused health atlas.

[0034] It is understandable that the process of inputting the fused health map into the trained risk deconstruction network is automated, with the fused health map serving as the direct input to the risk deconstruction network. The risk deconstruction network comprises a multi-layer feature perceptron and a pathway output layer. The fused health map is first fed into the multi-layer feature perceptron, which consists of multiple fully connected layers and non-linear activation function layers stacked together. This perceptron performs layer-by-layer abstract feature transformations on the input fused health map, extracting higher-level and more abstract patterns from each layer, ultimately obtaining a high-dimensional implicit feature representation. This high-dimensional implicit feature representation integrates all the potential information from temporal, physiological, and environmental interactions.

[0035] It is understandable that the design of the pathway output layer corresponds to a specific pathophysiological pathway, and the pathway output layer predefines several pathophysiological pathways related to neuroimmunological diseases. Each signal output unit is essentially a fully connected layer combined with a sigmoid activation function, which performs linear weighted summation on the high-dimensional implicit feature representation from the multilayer feature perceptron and outputs a continuous numerical signal between 0 and 1 through nonlinear mapping. This continuous numerical signal represents the degree to which an individual exhibits risk activation signs in this specific pathophysiological pathway within the current analysis period. All signal output units work simultaneously, and the output continuous numerical signals together constitute the set of risk activation signals.

[0036] Optionally, the multilayer perceptron employs dropout regularization during training. During model training, each hidden layer of the multilayer perceptron randomly discards some neurons and their connections with a certain probability to reduce the risk of overfitting. During inference, when the trained risk deconstruction network processes the newly generated fused health map, dropout is disabled, and all neurons participate in forward propagation computation to ensure the integrity and consistency of feature transformation, thereby obtaining a stable and reliable high-dimensional implicit feature representation.

[0037] Optionally, the parameters of each signal output unit in the pathway output layer are trained independently. Although all signal output units share the same high-dimensional implicit feature representation from the multilayer feature perceptron as input, each signal output unit has an independent weight vector and bias term. During model training, the loss function is designed as a weighted sum of the losses of all pathway output units, enabling the risk deconstruction network to simultaneously learn differentiated risk assessments for multiple pathophysiological pathways. Therefore, in application, the risk deconstruction network can output a set of risk activation signals corresponding to different pathophysiological pathways in parallel, rather than a single comprehensive score. The calculation of the nonlinear correlation weights can be expressed as:

[0038] in: Representation of features With features The non-linear correlation weight scalar value between them and These represent the first and second data frames from time-aligned mixed data frames, respectively. The and the first Embedding vectors of each feature This represents the element-wise multiplication operation between vectors. It is the parameter weight matrix obtained from model training. It is a bias vector. It is the Sigmoid activation function, used to map the output to the (0,1) interval to represent the weight strength.

[0039] In one embodiment of the present invention, when filtering according to predefined risk signal threshold rules, a set of risk activation signals output by the risk deconstruction network is obtained, each risk activation signal having an identifier of its respective pathophysiological pathway. A predefined risk signal threshold rule library is read, which sets a static threshold and a dynamic adjustment coefficient for each pathophysiological pathway. Based on the target individual's basic health record, the specific value of the dynamic adjustment coefficient is determined, and the applicable dynamic threshold for each pathophysiological pathway is calculated in conjunction with the static threshold. The value of each risk activation signal is compared with its corresponding dynamic threshold. If the value of a risk activation signal is greater than its corresponding dynamic threshold, the risk activation signal and its complete contextual features are extracted from the fused health graph, encapsulated together into a data packet of a risk event to be verified, and a unique event identifier is assigned to this data packet. For each risk event to be verified, an independent verification analysis thread is triggered, and the historical pathology database sub-database to be queried is determined based on the pathophysiological pathway identifier of the risk activation signal in the data packet of the risk event to be verified. In a separate verification analysis thread, the contextual features in the risk event data package to be verified are converted into query vectors. These query vectors are then used to perform a similarity search in the corresponding historical pathology database sub-database, identifying at least one historical reference pattern whose feature pattern matching degree exceeds a preset matching threshold. The final clinical outcome labels of historical risk events similar to the current risk event to be verified are extracted from the historical reference patterns. Based on the statistical distribution of these final clinical outcome labels, the credibility of the current risk event to be verified is scored. If the credibility score exceeds a preset credibility threshold, the risk event to be verified is deemed to have passed the review and is marked as a confirmed risk event.

[0040] In practice, the process of filtering a set of risk activation signals output by the risk deconstruction network according to predefined risk signal threshold rules and marking risk activation signals exceeding the corresponding threshold as risk events to be verified is achieved through the following steps: First, a set of risk activation signals output by the risk deconstruction network is obtained. Each risk activation signal has an identifier for its respective pathophysiological pathway, stored in a unique code format. Second, a predefined risk signal threshold rule library is read. This library sets a static threshold and a dynamic adjustment coefficient for each pathophysiological pathway. The static threshold is determined based on large-scale population cohort analysis. The dynamic adjustment coefficient is a scaling factor related to individual characteristics. Based on the target individual's basic health record, which includes fields such as age, genetic risk markers, and past medical history, the specific value of the dynamic adjustment coefficient is determined. This value is obtained by looking up a predefined mapping table, which groups the record fields... The system maps the values ​​to specific coefficients and calculates the applicable dynamic threshold for each pathophysiological pathway using static thresholds. It then compares the value of each risk activation signal with its corresponding dynamic threshold using a greater-than-normal operator. If the value of a risk activation signal exceeds its corresponding dynamic threshold, the system determines the signal to be significantly abnormal and performs a stripping operation to extract the signal with the excessive value and its complete contextual features from the fused health map. The contextual features include all enhanced feature nodes in the fused health map corresponding to the time window used to generate the risk activation signal. These are collectively encapsulated into a data packet for a risk event to be verified. This data packet uses a structured data format and includes fields such as event identifier, timestamp, pathway identification, risk activation signal value, and contextual feature array. A unique event identifier is assigned to each data packet, generated by combining the timestamp, pathway code, and random number.

[0041] In some embodiments, the dynamic threshold is calculated using a linear adjustment model. The static threshold, dynamic adjustment coefficient, and risk factor scores from the basic health record are used together in the calculation. The dynamic threshold calculation ensures the accuracy of individualized risk assessment. The risk signal threshold rule base is stored using version management. The static threshold and dynamic adjustment coefficient are iterated as medical evidence is updated. The system periodically pulls the latest version of the rule base from the central knowledge base to ensure the timeliness and scientific nature of the threshold setting. The execution of independent verification and analysis threads is concurrent. Each risk event data packet to be verified is assigned to an independent thread for processing. The thread manager monitors the execution status and resource usage of all threads. After the thread completes the review, it automatically reclaims resources and summarizes the results into the main process.

[0042] In practice, for each risk event to be verified, an independent verification analysis thread is triggered to retrieve a matching reference pattern from the historical pathology database for event review. This is achieved through the following steps: Based on the pathophysiological pathway identifier of the risk activation signal in the risk event data packet, the relevant historical pathology database sub-database is determined. The historical pathology database is divided into different sub-databases according to pathophysiological pathways. In the independent verification analysis thread, the contextual features in the risk event data packet are converted into query vectors. The conversion process uses the same feature extractor as the intermediate layer of the risk deconstruction network. The query vectors are then used to perform similarity searches in the corresponding historical pathology database sub-databases. The similarity search uses cosine similarity measurement. Find at least one historical reference pattern whose feature pattern matching degree exceeds a preset matching threshold (set to 0.7). Extract the final clinical outcome labels of historical risk events similar to the current risk event to be verified from the historical reference patterns. Final clinical outcome labels include "diagnosed multiple sclerosis" and "no disease". Score the credibility of the current risk event to be verified based on the statistical distribution of the final clinical outcome labels. Statistical distribution analysis calculates the proportion of positive outcomes in the matched historical reference patterns. The credibility score is proportional to this proportion. If the credibility score exceeds a preset credibility threshold (set to 0.65), the risk event to be verified is determined to have passed the review and is marked as a confirmed risk event.

[0043] In some embodiments, the extraction of historical reference patterns involves the preprocessing of historical case data. Each reference pattern in the historical pathology database is associated with a final clinical outcome label. The matching operation retrieves the top K historical reference patterns with the highest feature pattern matching degree. The default value of K is 10. The statistical distribution of the final clinical outcome label calculates the number of each type of outcome in these top K reference patterns. The confidence score is calculated based on the proportion of positive outcomes, and a weighted average is performed considering the matching degree weight to ensure that the score reflects the pattern similarity and outcome correlation.

[0044] It is understandable that the calculation of the dynamic threshold can be formally expressed as a formula: dynamic threshold. From static threshold and individualized adjustment factors The decision was made jointly, and the calculation formula is as follows:

[0045] in: Representing the The current applicable dynamic thresholds for each pathophysiological pathway. Representing the The static threshold of a pathophysiological pathway Represents the global adjustment intensity coefficient. The first term is calculated from the target individual's basic health record. Pathway-specific risk offset, risk offset The calculation is based on the weighted sum of relevant risk factors in the archive.

[0046] In one embodiment of the present invention, when integrating all verified risk events, a data package of all risk events marked as confirmed risk events is collected. Based on pathophysiological pathways, all confirmed risk events are grouped, and within the same pathophysiological pathway group, they are sorted in descending order according to the risk activation signal values ​​of each confirmed risk event. A confidence score obtained in the verification analysis thread is added to each confirmed risk event. All grouped, sorted, and supplemented with confidence scores, confirmed risk events are organized into a hierarchical, indexable risk event set data structure. An aggregated risk index is calculated based on the risk activation signal intensity and pathophysiological pathway affiliation of each event in the risk event set. The risk activation signal value and confidence score of each confirmed risk event are read from the structured risk event set, and a pre-configured pathophysiological pathway weight table is obtained, which specifies the relative importance weight of different pathways in the overall risk assessment. Based on the pathophysiological pathway attribution of each confirmed risk event, its corresponding relative importance weight is found. The risk activation signal value, confidence score, and relative importance weight of each confirmed risk event are input into the aggregation calculation function. The aggregation calculation function outputs a comprehensive scalar value. This comprehensive scalar value is normalized to a preset index range to obtain the final aggregated risk index.

[0047] In practice, the process of integrating all verified risk events into a structured risk event set and calculating an aggregated risk index based on this set is achieved through the following steps: First, all data packets of risk events marked as confirmed risk events are collected and stored in a temporary memory buffer. Second, all confirmed risk events are grouped based on pathophysiological pathways, using the pathophysiological pathway identifiers carried in the data packets. Third, the grouping operation generates a mapping structure with pathway identifiers as keys and event lists as values. Finally, within the same pathophysiological pathway group, the risk activation signal values ​​of each confirmed risk event are used for further analysis. The events are sorted in descending order to ensure that the event with the highest signal strength is at the top of the list. Each confirmed risk event is supplemented with a credibility score obtained in the verification analysis thread. The credibility score is written as a new field in the event data packet. All confirmed risk events that have been grouped, sorted and supplemented with credibility scores are organized into a hierarchical, indexable risk event collection data structure. This data structure uses JSON format. The top level is a list of pathway groups. Each pathway group object contains a pathway identifier, a pathway name, and an array of confirmed risk events. Each event object in the array contains an event identifier, a risk activation signal value, a credibility score, and a context feature reference link.

[0048] In some embodiments, the risk event set data structure is designed to support efficient querying. The hierarchical structure allows for quick location of specific groups through pathway identifiers. The event arrays within each pathway group are sorted in descending order of risk activation signal values, ensuring that the most significant events are prioritized when generating reports or conducting subsequent analysis. The credibility score is stored as an independent field of the event object, with the score value retained to two decimal places. Context feature reference links point to the original context feature data stored in the feature repository for retrieval when in-depth analysis is required. This structure separates event metadata from the massive feature data, optimizing the transmission and processing efficiency of the set structure.

[0049] In practice, the operation of calculating the aggregated risk index based on the risk activation signal intensity and pathophysiological pathway affiliation of each event in the risk event set involves retrieving the risk activation signal value and confidence score of each confirmed risk event from the structured risk event set. These values ​​are directly obtained from the corresponding fields of the risk event set data structure. A pre-configured pathophysiological pathway weight table is then retrieved. This table, stored in a configuration file, specifies the relative importance weights of different pathways in the overall risk assessment. Based on the pathophysiological pathway affiliation of each confirmed risk event, its corresponding relative importance weight is found. This search is performed by matching pathway identifiers in the pathway weight table. The risk activation signal value, confidence score, and relative importance weight of each confirmed risk event are input into the aggregation calculation function. The aggregation calculation function first calculates the weighted risk contribution of each event, then sums the contributions of all events. The aggregation calculation function outputs a comprehensive scalar value, which is then normalized to a preset index range [0, 100] to obtain the final aggregated risk index. See Table 1.

[0050] Table 1: Pathophysiological Pathway Weight Table

[0051] Aggregate Risk Index The calculation formula is:

[0052] in: This represents the calculated aggregate risk index, with a value between 0 and 100. The total number of confirmed risk events in the structured set of risk events. Representing the The risk activation signal value of a confirmed risk event. Representing the Credibility score of each confirmed risk event Representative and the The pathophysiological pathway to which each event belongs The corresponding relative importance weights, This represents the maximum value among all possible relative importance weights, used to scale the sum.

[0053] In some embodiments, the pre-configured pathophysiological pathway weight table is predefined by medical experts based on the known importance of each pathway in the pathogenesis of the target disease. The weight values ​​are loaded during system deployment and can be updated through the management interface. The design of the aggregation calculation function considers the strength, reliability, and pathway importance of the risk signal. The calculation of the weighted risk contribution reflects the greater impact of high-intensity, high-reliability events occurring in high-weight pathways on the overall risk index. The normalization process uses a fixed denominator. To ensure that regardless of the number of events, the final index The value range is stable within the preset interval [0, 100], which facilitates unified interpretation.

[0054] Referring to Figure 4, the correlation between the "relative importance weights" of the four core pathophysiological pathways and the "average confidence scores" of the corresponding risk events is presented. Specifically, the relative importance weights of the pathways (blue bars) are determined based on the contribution of the pathways to the pathogenesis of the disease as defined by medical experts. Among them, the "blood-brain barrier permeability alteration" pathway has the highest weight (1.8), while the "abnormal B cell activation" pathway has a relatively low weight (1.2). The average confidence scores of the risk events (red lines) are statistical results verified by referring to historical pathology databases, showing a trend that is not completely synchronized with the pathway weights: the confidence score of the risk events corresponding to the "abnormal B cell activation" pathway (approximately 0.95) is the highest, while the confidence score of the "glial cell activation" pathway (approximately 0.88) is the lowest. This non-linear correlation between weights and confidence reflects that although risk events of high-weighted pathways are more critical in the pathogenesis, their clinical reference model matching degree is relatively low. Therefore, the dual influence of pathway weights and event confidence needs to be considered in the calculation of the aggregated risk index.

[0055] In one embodiment of the present invention, when constructing a personalized risk prediction report based on an aggregated risk index and a set of risk events, a blank template for the personalized risk prediction report is created. This blank template includes an index summary area, an event details area, and an action suggestion framework. The aggregated risk index and its corresponding risk level description are filled into the index summary area. The main content of the structured set of risk events, including the pathways to which confirmed risk events belong, signal strength ranking, and credibility score, is formatted and filled into the event details area. Based on the level of the aggregated risk index and the specific pathophysiological pathways involved in the set of risk events, a corresponding list of monitoring items and health tips are selected from a preset action rule base and filled into the action suggestion framework, thereby generating a complete personalized risk prediction report. The content of the action suggestion framework in the personalized risk prediction report is parsed, transforming each monitoring item into an executable monitoring task instruction and each health tip into a notification message to be sent. Based on the urgency and logical dependencies of the monitoring tasks, all monitoring task instructions are sorted and scheduled to form a monitoring task sequence. According to preset triggering times, the instructions in the monitoring task sequence are executed sequentially, and the corresponding notification messages are sent synchronously.

[0056] In practical implementation, the operation of constructing individualized risk prediction reports based on aggregated risk indices and risk event sets, and generating differentiated monitoring and alert task sequences, is achieved through the following steps: First, a blank template for the individualized risk prediction report is created. This template contains three structured areas: an index summary area, an event details area, and an action suggestion framework. The index summary area is reserved for fields to be filled in with the aggregated risk index and its corresponding risk level description. The event details area is designed in tabular form to display detailed information on confirmed risk events. The action suggestion framework includes freely configurable text paragraphs and a project list area to carry specific suggestions. Second, the aggregated risk index and its corresponding risk level description are filled into the index summary area. The risk level description is selected from "low risk," "medium risk," "high risk," and "extremely high risk" levels according to the predefined numerical range to which the aggregated risk index belongs. This descriptive label, along with the specific value of the aggregated risk index, is then filled into the designated field.

[0057] In practice, the main content of the structured risk event set is formatted and filled into the event details area. Each pathway group and its subordinate confirmed risk events are read sequentially from the risk event set data structure. The pathway name, risk activation signal value sorting position, and confidence score of each event are extracted. According to the pathway group and signal strength descending order, this information is organized into table rows and columns and filled into the corresponding table template in the event details area. Based on the level of the aggregated risk index and the specific pathophysiological pathways involved in the risk event set, the corresponding monitoring item list and health tips are selected from the preset action rule base. The action rule base is a mapping database whose key is a combination of risk level and pathway identifier, and the value is the associated monitoring item list string and health tips text set. The action rule base is queried to obtain the monitoring item list and health tips text that match the current assessment result, and these text contents are filled into the corresponding areas of the action suggestion framework, thereby generating an individualized risk prediction report document containing all three parts.

[0058] In some embodiments, the action rule base is constructed based on clinical guidelines and expert consensus. The monitoring item list includes specific examination names, monitoring frequencies, and equipment requirements, such as "home heart rate variability monitoring twice a week" and "cerebrospinal fluid oligoclonal band detection within one month." The health tips describe behavioral suggestions in plain language, such as "avoid recent infections" and "maintain regular sleep." After the personalized risk prediction report is generated, it is saved in a standardized format and assigned a unique report identifier. The content of the action suggestion framework in the personalized risk prediction report is parsed. The text parsing engine identifies the structured parts in the action suggestion framework, splits the monitoring item list string into independent monitoring item items by delimiters, and splits the health tips text set into independent prompt statements by items.

[0059] In some embodiments, the conversion of monitoring task instructions follows a predefined instruction template. Each monitoring item is matched with an instruction template, which includes fields such as action type, target device or system, and parameter settings. The parser extracts key information from the monitoring item and fills it into the instruction template to generate an executable monitoring task instruction. The conversion of prompt messages directly uses the key health tips as message content, adds the target user's communication identifier and planned sending timestamp, and forms a prompt message object to be sent. All monitoring task instructions are sorted and scheduled according to the urgency and logical dependencies of the monitoring tasks. The urgency is determined by the urgency label preset for each monitoring item in the action rule base, and the logical dependencies describe the sequential execution constraints between tasks. The sorting algorithm calculates a linear execution sequence based on the urgency label and dependency graph to form a monitoring task sequence.

[0060] It is understandable that the scheduling logic of the monitoring task sequence can be formally expressed, and the ordering priority of the monitoring task instructions can be determined. Due to its urgency coefficient With aggregation risk index The decision was made jointly, and the calculation formula is as follows:

[0061] in: This represents the calculated task priority value; the higher the value, the higher the task's position in the monitoring task sequence. This represents the preset urgency coefficient for this monitoring project, obtained from the action rule base. Its value is a constant between 0 and 1. Represents the aggregate risk index. and This represents an adjustment weighting coefficient used to balance the impact of urgency and overall risk level on task priority. The ranking process is based on the calculated weighting coefficient. The values ​​are sorted in descending order for all monitoring task instructions.

[0062] It is understood that the execution of task instructions and the sending of messages are completed automatically by the system. The instructions in the monitoring task sequence are executed sequentially according to the preset triggering time. The triggering time includes immediate execution, timed execution, or waiting for a specific event to trigger. During the execution process, the corresponding equipment control interface, laboratory reservation system interface, or data acquisition service interface are called, and the corresponding prompt messages are sent to the receiving terminal designated by the target individual through message push service, email, or SMS gateway. The system records the execution status of each instruction and message sending log, and archives them with the unique identifier of the individualized risk prediction report.

[0063] Referring to Figure 5, during the dynamic health monitoring phase of neuroimmunological disease risk prediction, the figure shows the time-series changes of the aggregated risk index and the average pathway signal value (×10) from January to June. Specifically, the aggregated risk index (blue curve) shows a phased upward trend with fluctuations; the average pathway signal value (green dashed line, ×10) shows a correlation with the risk index with fluctuations, with local peaks in February, April, and May, and falling back to a baseline level of around 3.5 in March and June. This change reflects the dynamic correlation between the pathway risk activation signal output by the risk deconstruction network and the aggregated risk index: when the average pathway signal value (representing the mean of the degree of multi-pathway risk activation) increases, the risk signal of the corresponding pathophysiological pathway is more likely to exceed the dynamic threshold, the set of risk events after review is richer, and the aggregated risk index calculated by pathway weight and signal strength increases accordingly; while the decline in the pathway signal value and the synchronous decrease in the risk index in June reflect the timeliness of the system's dynamic monitoring and aggregate assessment of risk events.

[0064] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0065] 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 method for predicting the risk of neuroimmune diseases based on artificial intelligence, characterized in that, include: Initialize the dynamic health data model and configure the individual physiological time-series data queue and the set of external environmental factors; The collaborative analysis engine is invoked to perform fusion calculations on the individual physiological time-series data queue and the external environmental factor set to generate a fused health map with time-series labels; The fused health map is input into a trained risk deconstruction network, which outputs a set of risk activation signals corresponding to different pathophysiological pathways. The risk deconstruction network includes a multilayer feature perceptron and a pathway output layer. The fused health map is input into the multilayer feature perceptron of the risk deconstruction network, and abstract feature transformation is performed layer by layer to obtain a high-dimensional implicit feature representation. The high-dimensional implicit feature representation is input into the pathway output layer, which predefines several pathophysiological pathways related to neuroimmunological diseases, each corresponding to an independent signal output unit. Each signal output unit calculates the high-dimensional implicit feature representation and outputs a continuous numerical signal representing the degree of risk activation of that pathway. All signal outputs... The continuous numerical signals output by the unit constitute the set of risk activation signals; according to predefined risk signal threshold rules, the set of risk activation signals is filtered, and risk activation signals exceeding the corresponding threshold are marked as risk events to be verified; for each risk event to be verified, an independent verification analysis thread is triggered to retrieve a matching reference pattern from the historical pathology database for event review; all verified risk events to be verified are integrated to form a structured risk event set; based on the risk activation signal intensity and pathophysiological pathway attribution of each event in the risk event set, an aggregated risk index is calculated; based on the aggregated risk index and the risk event set, an individualized risk prediction report is constructed, and a differentiated monitoring and alert task sequence is generated and executed.

2. The method for predicting the risk of neuroimmune diseases based on artificial intelligence according to claim 1, characterized in that, Initializing a dynamic health data model and configuring an individual physiological time-series data queue and an external environmental factor set includes: creating a blank architecture for the dynamic health data model, the blank architecture comprising a data input port, an internal buffer framework, and a data output port; allocating a first data channel and a second data channel to the data input port, the first data channel being dedicated to receiving streaming physiological parameters and the second data channel being dedicated to receiving discrete environmental parameters; instantiating the individual physiological time-series data queue within the internal buffer framework and logically binding the input end of the individual physiological time-series data queue to the first data channel; instantiating the external environmental factor set within the internal buffer framework and logically binding the input end of the external environmental factor set to the second data channel; configuring triggering conditions for the data output port, the triggering conditions being that the data volume of the individual physiological time-series data queue reaches a preset capacity or the monitoring period reaches a cutoff point; within the preset monitoring period, continuously injecting multi-dimensional physiological parameters of the target individual into the individual physiological time-series data queue, and synchronously updating the spatiotemporal environmental parameters associated with the target individual into the external environmental factor set.

3. The method for predicting the risk of neuroimmune diseases based on artificial intelligence according to claim 2, characterized in that, The process of continuously injecting multi-dimensional physiological parameters of the target individual into the individual physiological time-series data queue within a preset monitoring period, and simultaneously updating the spatiotemporal environmental parameters associated with the target individual into the external environmental factor set, includes: acquiring the raw physiological signal stream of the target individual through a connected biosensor device; performing noise reduction and feature extraction operations on the raw physiological signal stream to obtain standardized multi-dimensional physiological parameters, each dimension corresponding to a physiological function indicator; sequentially pushing the standardized multi-dimensional physiological parameters into the individual physiological time-series data queue according to their timestamps; simultaneously, asynchronously acquiring environmental data of the target individual's geographical location and publicly available health event data from the environmental monitoring network and service interface, wherein the environmental data includes climate indicators and pollutant concentrations, and the health event data includes regional epidemic information; formatting the acquired environmental data and health event data to generate spatiotemporal environmental parameters; and replacing the historical parameters stored in the external environmental factor set with the latest spatiotemporal environmental parameters within the current monitoring period.

4. The method for predicting the risk of neuroimmune diseases based on artificial intelligence according to claim 1, characterized in that, The process of invoking the collaborative analysis engine to perform fusion calculations on the individual physiological time-series data queue and the external environmental factor set to generate a fused health atlas with time-series labels includes: extracting the current state of the individual physiological time-series data queue and the external environmental factor set from the data output port of the dynamic health data model; the collaborative analysis engine includes a time-series alignment module and a feature cross-calculation module; using the time-series alignment module to map the parameters in the external environmental factor set to each time node of the individual physiological time-series data queue according to time slices, forming a time-aligned mixed data frame; sending the time-aligned mixed data frame into the feature cross-calculation module, which calculates the nonlinear correlation weights between physiological parameters, between environmental parameters, and between physiological and environmental parameters; and re-encoding and concatenating all parameters in the time-aligned mixed data frame according to the calculated nonlinear correlation weights to generate a multidimensional fused health atlas with time-series labels, where each node in the fused health atlas represents an enhanced feature.

5. The method for predicting the risk of neuroimmune diseases based on artificial intelligence according to claim 4, characterized in that, The step of filtering the set of risk activation signals according to predefined risk signal threshold rules, and marking risk activation signals exceeding the corresponding threshold as risk events to be verified, includes: acquiring the set of risk activation signals output by the risk deconstruction network, each risk activation signal having an identity identifier of its respective pathophysiological pathway; reading a predefined risk signal threshold rule library, the risk signal threshold rule library setting a static threshold and a dynamic adjustment coefficient for each pathophysiological pathway; determining the specific value of the dynamic adjustment coefficient based on the target individual's basic health record, and calculating the currently applicable dynamic threshold for each pathophysiological pathway in combination with the static threshold; comparing the value of each risk activation signal with its corresponding dynamic threshold; if the value of a certain risk activation signal is greater than its corresponding dynamic threshold, then the risk activation signal and its complete context features are extracted from the fused health map, encapsulated together into a risk event data packet to be verified, and a unique event identifier is assigned to the data packet.

6. The method for predicting the risk of neuroimmune diseases based on artificial intelligence according to claim 5, characterized in that, For each risk event to be verified, an independent verification analysis thread is triggered to retrieve a matching reference pattern from the historical pathology database for event review. This includes: determining the historical pathology database sub-database to be queried based on the pathophysiological pathway identity of the risk activation signal in the risk event data packet; in the independent verification analysis thread, converting the contextual features in the risk event data packet into a query vector; using the query vector to perform a similarity search in the corresponding historical pathology database sub-database to find at least one historical reference pattern whose feature pattern matching degree exceeds a preset matching threshold; extracting the final clinical outcome labels of historical risk events similar to the current risk event to be verified from the historical reference patterns; scoring the credibility of the current risk event to be verified based on the statistical distribution of the final clinical outcome labels to obtain a credibility score; if the credibility score exceeds a preset credibility threshold, the risk event to be verified is determined to have passed the review and is marked as a confirmed risk event.

7. The method for predicting the risk of neuroimmune diseases based on artificial intelligence according to claim 6, characterized in that, The process of integrating all verified risk events to form a structured risk event set includes: collecting all data packages of risk events marked as confirmed risk events; grouping all confirmed risk events according to pathophysiological pathways; sorting them in descending order based on the risk activation signal values ​​of each confirmed risk event within the same pathophysiological pathway group; supplementing each confirmed risk event with its credibility score obtained in the verification analysis thread; and organizing all grouped, sorted, and credibility-score-supplemented confirmed risk events into a hierarchical, indexable risk event set data structure.

8. The method for predicting the risk of neuroimmune diseases based on artificial intelligence according to claim 7, characterized in that, The step of calculating the aggregated risk index based on the risk activation signal intensity and pathophysiological pathway attribution of each event in the risk event set includes: reading the risk activation signal value and confidence score of each confirmed risk event from the structured risk event set; obtaining a pre-configured pathophysiological pathway weight table, which specifies the relative importance weight of different pathways in the overall risk assessment; finding the corresponding relative importance weight based on the pathophysiological pathway attribution of each confirmed risk event; inputting the risk activation signal value, confidence score, and relative importance weight of each pathway of each confirmed risk event into an aggregation calculation function, which outputs a comprehensive scalar value; and normalizing the comprehensive scalar value to a preset index range to obtain the final aggregated risk index.

9. The method for predicting the risk of neuroimmune diseases based on artificial intelligence according to claim 8, characterized in that, The process of constructing a personalized risk prediction report based on the aggregated risk index and the set of risk events, and generating and executing a differentiated monitoring and alerting task sequence, includes: creating a blank template for the personalized risk prediction report, the blank template containing an index summary area, an event details area, and an action suggestion framework; filling the index summary area with the aggregated risk index and its corresponding risk level description; filling the event details area with the main content of the structured risk event set, including the confirmed risk event's pathway, signal strength ranking, and credibility score, after formatting; and adjusting the aggregated risk index according to the specific risk events involved in the risk event set. The pathophysiological pathway involves selecting a list of corresponding monitoring items and health tips from a pre-defined action rule base and filling them into the action suggestion framework to generate a complete personalized risk prediction report. The content of the action suggestion framework in the personalized risk prediction report is parsed, transforming each monitoring item into an executable monitoring task instruction and each health tip into a notification message to be sent. Based on the urgency and logical dependencies of the monitoring tasks, all monitoring task instructions are sorted and scheduled to form a monitoring task sequence. According to a pre-defined triggering time, the instructions in the monitoring task sequence are executed sequentially, and corresponding notification messages are sent synchronously.

Images