Artificial intelligence-driven integrated platform for early screening and intervention of cognitive impairment

Abstract

The present application relates to the technical field of early screening of cognitive impairment, in particular to an artificial intelligence driven early screening and intervention integrated platform for cognitive impairment, comprising: a multi-modal data sensing and collecting unit; a central AI processing and analyzing unit, introducing a differentiable causal discovery mechanism, adopting an end-to-end structure search algorithm based on Gumbel-Sigmoid to jointly optimize the existence probability and causal direction of edges in a graph, embedding an improved Granger causal loss function and a neurophysiological prior constraint; a personalized intervention strategy generating unit; an effect tracking and feedback closed loop unit. Through the time sequence collaborative processing of multi-modal data based on cognitive task key event timestamps, combined with dynamic heterogeneous graph construction and causal analysis mechanism, the present application forms an individualized causal chain graph reflecting the dynamic evolution characteristics of cognitive function decline, which can accurately capture the core path of cognitive impairment progression, and solves the problems of insufficient time sequence collaboration of multi-modal data and lack of dynamic nature of causal analysis.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of early screening of cognitive impairment, in particular to an artificial intelligence driven early screening and intervention integrated platform for cognitive impairment. BACKGROUND

[0002] Early identification and dynamic intervention of cognitive impairment is the key to reduce the risk of disease progression and improve the quality of life of patients. Its pathological mechanism presents the characteristics of multi-dimension and time sequence evolution, and needs to rely on multi-source data fusion and accurate analysis technology to realize whole-process management.

[0003] In the prior art, related patents have explored in the field of cognitive impairment screening. For example, Chinese patent CN202511189298.X discloses an early screening and diagnosis system for cognitive impairment based on multi-modal data, which includes a multi-modal data acquisition module (acquiring image, gene, behavior and clinical narrative information) and a causal purification module (filtering pseudo-correlation through clinical narrative anchoring and reverse intervention deduction), realizing the upgrade from "correlation" to "causality" in diagnosis and treatment. For another example, Chinese patent CN202510258074.3 discloses a cognitive impairment screening method and system based on face analysis, which collects the facial movement features of the user when watching a video containing a funny point, inputs a preset model to output a screening result, and has the characteristics of non-invasive and less dependence on professional personnel.

[0004] Although the above technical solutions have corresponding design advantages, the above technical solutions still have the following technical defects: first, multi-modal data lacks time sequence collaborative processing, and causal analysis does not reflect the dynamic evolution characteristics: CN202511189298.X realizes the causal purification of multi-modal data, but does not collaboratively process the time sequence consistency of multi-source data, and the causal relationship analysis does not combine the dynamic evolution law of cognitive function decline, so it is difficult to accurately capture the core path of pathological progression; CN202510258074.3 only analyzes single dimension features within a fixed time window, without considering the time sequence association and long-term evolution trend of data, and cannot reflect the dynamic changes of cognitive function. Second, the function coverage is not complete, and the closed-loop linkage of screening and targeted intervention is not realized: CN202511189298.X focuses on screening and causal analysis, and does not design a targeted intervention and effect feedback mechanism directly related to the analysis result; CN202510258074.3 only realizes the screening function, without extending to the personalized intervention and dynamic tracking link, and cannot form whole-process management. In view of this, we propose an artificial intelligence driven early screening and intervention integrated platform for cognitive impairment. SUMMARY

[0005] The present application aims to provide an artificial intelligence driven cognitive impairment early screening and intervention integrated platform to solve the problems of lack of time sequence collaborative processing of multi-modal data, incomplete functional coverage and failure to realize closed-loop linkage of screening and targeted intervention in the background art.

[0006] To achieve the above technical problems, the present application aims to provide an artificial intelligence driven cognitive impairment early screening and intervention integrated platform, comprising:

[0007] A multi-modal data perception and acquisition unit, which acquires synchronous eye movement signals, electroencephalogram signals and behavioral task data streams of a cognitive impairment risk group based on eye movement electroencephalogram synchronous recording technology, associates cognitive task key events to generate time stamps, completes time sequence labeling of multi-modal physiological and behavioral data perception and acquisition, and transmits the output time sequence labeled multi-modal data to a central AI processing and analysis unit;

[0008] A central AI processing and analysis unit, which receives the time sequence labeled multi-modal data transmitted by the multi-modal data perception and acquisition unit, completes cross-modal time sequence alignment of eye movement, electroencephalogram and behavioral data streams with cognitive task key event time stamps as anchor points; constructs a dynamic heterogeneous graph structure, maps physiological and behavioral variables as graph nodes, initializes a directed edge table with time delay weights to represent potential interaction between variables, and generates a time sequence graph sequence evolving with task progress; introduces a differentiable causal discovery mechanism, uses an end-to-end structure search algorithm based on Gumbel-Sigmoid to jointly optimize the existence probability and causal direction of edges in the graph, embeds an improved Granger causal loss function and neurophysiological prior constraint, and automatically learns a time sequence causal weight matrix with biological rationality; identifies a driving decline mode based on the time sequence causal weight matrix, generates an individualized cognitive function decline causal chain graph, establishes a personalized baseline based on the first detection data, and calculates the longitudinal deviation degree of each causal path strength in the subsequent detection data; uses an LSTM-attention network to model a multi-round causal path strength sequence, and outputs an individualized cognitive function decline causal chain graph and key driving path to an individualized intervention strategy generation unit, while outputting a standardized clinical report;

[0009] An individualized intervention strategy generation unit, which receives the individualized cognitive function decline causal chain graph and key driving path output by the central AI processing and analysis unit, generates an individualized cognitive intervention strategy for targeted blocking of the decline path, and synchronizes the generated individualized cognitive intervention strategy to an effect tracking and feedback closed-loop unit;

[0010] An effect tracking and feedback closed-loop unit receives the personalized cognitive intervention strategy synchronized by the personalized intervention strategy generation unit, dynamically monitors the implementation process and effect of the personalized cognitive intervention strategy, collects post-intervention multi-modal physiological and behavioral data, and returns the collected post-intervention multi-modal data to the central AI processing and analysis unit; the central AI processing and analysis unit reiterates the cognitive decline causal chain graph and modeling results based on the returned post-intervention multi-modal data, and feeds back the iterated results to the personalized intervention strategy generation unit, which corrects the intervention strategy parameters.

[0011] As a further improvement of the technical solution, the multi-modal data perception and collection unit includes a signal synchronization collection module, a task event timestamp association module, and a data format standardization module, wherein:

[0012] The signal synchronization collection module synchronously collects the eye movement signal, the electroencephalogram signal, and the behavioral task data stream of the cognitive impairment risk group based on the eye movement electroencephalogram synchronous recording technology;

[0013] The task event timestamp association module associates the cognitive task key events to generate corresponding timestamp labels for the collected eye movement signal, electroencephalogram signal, and behavioral task data stream;

[0014] The data format standardization module performs format standardization processing on the eye movement signal, electroencephalogram signal, and behavioral task data stream labeled with timestamps, and transmits the processed time-series labeled multi-modal physiological and behavioral data to the central AI processing and analysis unit.

[0015] As a further improvement of the technical solution, the central AI processing and analysis unit includes a cross-modal time-series alignment module, a dynamic heterogeneous graph construction module, a differentiable causal discovery module, a cognitive decline causal chain modeling module, and a sequence prediction and output module, wherein:

[0016] The cross-modal time-series alignment module receives the time-series labeled multi-modal data transmitted by the multi-modal data perception and collection unit, and completes the cross-modal time-series alignment of the eye movement, electroencephalogram, and behavioral data stream with the cognitive task key event timestamp as the anchor point;

[0017] The dynamic heterogeneous graph construction module constructs a dynamic heterogeneous graph structure based on the aligned multi-modal data, maps physiological and behavioral variables as graph nodes, initializes a directed edge table with time delay weights to represent the potential interaction relationship between variables, and generates a time-series graph sequence evolving with the task process;

[0018] The differentiable causal discovery module introduces a differentiable causal discovery mechanism to process the time-series graph sequence and jointly optimize the existence probability and causal direction of the edges in the graph;

[0019] The cognitive decline causal chain modeling module, based on the output of the differentiable causal discovery module, identifies driving decline patterns, generates an individualized cognitive function decline causal chain map, establishes a personalized baseline, and calculates the longitudinal deviation of the causal path strength.

[0020] The sequence prediction and output module uses an LSTM-attention network to model the intensity sequence of multi-round causal paths, outputs an individualized causal chain map of cognitive function decline and key driving paths to the personalized intervention strategy generation unit, and outputs a standardized clinical report.

[0021] As a further improvement to this technical solution, the cross-modal temporal alignment module includes a cognitive task anchor reference determination submodule and a multimodal data timestamp calibration submodule, wherein:

[0022] The cognitive task anchor benchmark determination submodule extracts the triggering time of stimulus presentation and response instructions in the cognitive task, and uses the triggering time as the benchmark time point for multimodal data alignment;

[0023] The multimodal data timestamp calibration submodule matches the sampling times of eye-tracking signals, electroencephalogram signals, and behavioral task data streams with the reference time points. After calibration, the time deviation of each modality data does not exceed 1 / 2 of the corresponding data sampling period.

[0024] As a further improvement to this technical solution, the dynamic heterogeneous graph construction module, when constructing a dynamic heterogeneous graph, includes the following steps in the process of initializing the directed edge delay weights and generating the time series graph sequence:

[0025] S22.1 Determine the preset time delay window and obtain the time series data sequence corresponding to the graph node;

[0026] S22.2 Calculate the correlation coefficient of time-series data sequences at different nodes under the time delay window;

[0027] S22.3. Normalize the correlation coefficients of all node pairs to obtain the initial values ​​of the time delay weights of the edges in the graph. ;in, , All are node indices in a dynamic heterogeneous graph;

[0028] S22.4, will As the initial weights of the directed edges in the dynamic heterogeneous graph, a sequence of time graphs that evolves with the progress of the task is generated.

[0029] As a further improvement to this technical solution, the differentiable causal discovery module, based on a time-series graph sequence, optimizes the probability of graph edge existence and causal direction by constructing a joint optimization objective function, including the following steps:

[0030] S23.1 Determine the predicted value of the node time series data at the target time, and the set of parent nodes of the node corresponding to the target time, to clarify the basis for causal relationship fitting;

[0031] S23.2 Construct a prediction error term to characterize the fitting accuracy of causal relationships, ensuring the accuracy of causal inference;

[0032] S23.3, Add a sparsity regularization term to constrain the redundancy of causal relationships;

[0033] S23.4 Integrate the prediction error term and the sparsity regularization term to form an improved Granger causal loss function LGranger, and simultaneously optimize the existence probability and causal direction of edges in the graph based on the improved Granger causal loss function LGranger.

[0034] As a further improvement to this technical solution, the cognitive decline causal chain modeling module, after generating a cognitive decline causal chain map, calculates the longitudinal deviation of the causal path strength based on a personalized baseline, including the following steps:

[0035] S24.1 Based on the initial detection data, determine the first... Initial strength value of causal path , for the first Establish personalized baselines based on causal paths;

[0036] S24.2, Obtain the first The first test Current strength value of the causal path Capture cross-cycle change data;

[0037] S24.3, based on and Calculate the first Longitudinal deviation of causal paths .

[0038] As a further improvement to this technical solution, the sequence prediction and output module, based on a multi-round causal path strength sequence, utilizes an LSTM-attention network to calculate temporal step attention weights to achieve sequence modeling. The process includes the following steps:

[0039] S25.1, Obtaining the multi-round causal path strength sequence after inputting it into the LSTM network to generate the first... Step Hidden State and context vectors representing global features of the sequence. This forms the basis for constructing a two-dimensional input of local temporal features and global features;

[0040] S25.2, based on and The association mapping relationship is calculated for the first Attention score for each time step;

[0041] S25.3. Normalize the attention scores for all time steps to obtain the first... Attention weights for each time step This enables the modeling of multi-round causal path strength sequences.

[0042] As a further improvement to this technical solution, the personalized intervention strategy generation unit includes an intervention input parsing module, a target path blocking strategy generation module, and an intervention strategy synchronization module, wherein:

[0043] The intervention input parsing module receives the individualized cognitive function decline causal chain map and key driving paths output by the central AI processing and analysis unit, and analyzes the path correlation and core features of the key driving paths in the causal chain map.

[0044] The targeted path blocking strategy generation module generates a personalized cognitive intervention strategy that targets and blocks the decline path based on the analysis results of the intervention input analysis module.

[0045] The intervention strategy synchronization module synchronizes the personalized cognitive intervention strategy generated by the target path blocking strategy generation module to the effect tracking and feedback closed-loop unit.

[0046] As a further improvement to this technical solution, the effect tracking and feedback closed-loop unit includes an intervention strategy receiving module, an intervention process monitoring module, a feedback data acquisition module, and a data feedback and iteration connection module, wherein:

[0047] The intervention strategy receiving module receives personalized cognitive intervention strategies synchronized by the personalized intervention strategy generation unit.

[0048] The intervention process monitoring module monitors the implementation process of the personalized cognitive intervention strategy and the corresponding physiological and behavioral changes based on the personalized cognitive intervention strategy obtained by the intervention strategy receiving module.

[0049] The feedback data acquisition module collects post-intervention multimodal physiological and behavioral data within the monitoring range of the intervention process monitoring module;

[0050] The data feedback and iteration connection module feeds back the multimodal data collected by the feedback data acquisition module after intervention to the central AI processing and analysis unit, and receives the cognitive decline causal chain map and modeling results after the central AI processing and analysis unit iterates based on the fed-back data, and synchronizes them to the personalized intervention strategy generation unit.

[0051] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0052] 1. This invention performs time-series collaborative processing on multimodal data based on the timestamps of key events in cognitive tasks, and combines dynamic heterogeneous graph construction and causal analysis mechanisms to form an individualized causal chain map that reflects the dynamic evolution characteristics of cognitive function decline. This can accurately capture the core path of cognitive impairment progression and solve the problems of insufficient time-series collaboration of multimodal data and lack of dynamism in causal analysis.

[0053] 2. Based on the aforementioned individualized causal chain map, this invention constructs an integrated platform of "early screening - targeted intervention - effect feedback", which directly links the screening analysis results with the targeted intervention strategy, and optimizes the scheme through feedback data iteration, fully covering the entire process of cognitive impairment management, and solving the problems of incomplete functional coverage and the lack of closed-loop linkage between screening and targeted intervention. Attached Figure Description

[0054] Fig. 1 This is a schematic diagram of the system framework of the present invention;

[0055] Fig. 2 This is a schematic diagram of the macroscopic technical link process of the central AI processing and analysis unit in this invention;

[0056] Fig. 3 This is a schematic diagram illustrating the application scenarios of the present invention;

[0057] The meanings of the labels in the diagram are as follows:

[0058] 1. Multimodal data sensing and acquisition unit; 11. Signal synchronization acquisition module; 12. Task event timestamp association module; 13. Data format standardization module;

[0059] 2. Central AI Processing and Analysis Unit; 21. Cross-modal Temporal Alignment Module; 22. Dynamic Heterogeneous Graph Construction Module; 23. Differentiable Causality Discovery Module; 24. Cognitive Decline Causal Chain Modeling Module; 25. Sequence Prediction and Output Module;

[0060] 3. Personalized intervention strategy generation unit; 31. Intervention input parsing module; 32. Targeted pathway blocking strategy generation module; 33. Intervention strategy synchronization module;

[0061] 4. Effect tracking and feedback closed-loop unit; 41. Intervention strategy receiving module; 42. Intervention process monitoring module; 43. Feedback data collection module; 44. Data feedback and iterative connection module. Detailed Implementation

[0062] 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0063] like Figs. 1-3 As shown, this embodiment provides an integrated platform for early screening and intervention of cognitive impairment driven by artificial intelligence, including:

[0064] Multimodal data perception and acquisition unit 1, based on eye-tracking and EEG synchronous recording technology, acquires synchronous eye-tracking signals, EEG signals and behavioral task data streams of cognitive impairment risk groups, associates key events of cognitive tasks to generate timestamps, completes the perception and acquisition of time-marked multimodal physiological and behavioral data, and transmits the output time-marked multimodal data to the central AI processing and analysis unit 2.

[0065] In this embodiment, the multimodal data sensing and acquisition unit 1 includes a signal synchronization acquisition module 11, a task event timestamp association module 12, and a data format standardization module 13, wherein:

[0066] The signal synchronization acquisition module 11 is based on eye-tracking and electroencephalogram (EEG) synchronization recording technology to simultaneously acquire eye-tracking signals, EEG signals, and behavioral task data streams of people at risk of cognitive impairment.

[0067] Specifically, the signal synchronization acquisition module 11 adopts a general eye-tracking and EEG synchronization acquisition scheme: it selects eye-tracking acquisition devices that support a sampling rate of ≥100Hz, such as desktop non-invasive eye-tracking acquisition devices, and EEG acquisition devices that support a sampling rate of ≥250Hz, such as scalp electrode-based non-invasive EEG acquisition devices, as the core acquisition devices. At the same time, it deploys cognitive task execution programs, such as digital working memory tasks and visual search tasks, on the PC to acquire behavioral task data streams. The eye-tracking acquisition devices, EEG acquisition devices, and PC-side task programs are connected through hardware trigger lines. When the PC-side task program starts, it synchronously sends trigger signals to the eye-tracking and EEG acquisition devices, so that the three types of devices start data acquisition based on the same trigger signal, ensuring that the acquisition start sequence of eye-tracking signals, EEG signals, and behavioral task data streams is consistent. Among them, the acquisition of behavioral task data streams is linked in real time with the operation events of the task program, such as stimulus presentation and user response, and is recorded synchronously with the task progress.

[0068] The task event timestamp association module 12 associates key events of the cognitive task and generates corresponding timestamp markers for the collected eye movement signals, EEG signals and behavioral task data streams;

[0069] Specifically, the key events of the cognitive task include four core nodes: "task initiation," "stimulus target presentation," "user response submission," and "task phase end." When the above key events occur, the PC-based cognitive task program synchronously generates a marker instruction containing the event type and the system time of the event occurrence, and sends it to the acquisition software of the eye-tracking acquisition device, the EEG acquisition device, and its own behavioral data recording module. After receiving the marker instruction, each device embeds the same event timestamp (format: UNIX timestamp + millisecond offset) into the acquisition data at the corresponding time. For example, when the "stimulus target presentation" event occurs, the task program generates the timestamp "1690000000123_stimulus presentation," and this marker is added to the sampled data of eye-tracking, EEG, and behavioral data at that time, realizing the temporal binding of the three types of data with the key events of the cognitive task.

[0070] The data format standardization module 13 performs format standardization processing on the eye movement signals, EEG signals and behavioral task data streams after the timestamp is marked, and transmits the processed time-marked multimodal physiological and behavioral data to the central AI processing and analysis unit 2.

[0071] Specifically, the raw data formats of each device are as follows: eye-tracking acquisition devices output a .csv file containing "timestamp, X coordinate of fixation point, Y coordinate of fixation point, and fixation duration"; EEG acquisition devices output a .edf file containing "timestamp and potential values ​​of each lead (such as F3, F4, C3)"; and behavioral task programs output a .json file containing "timestamp, reaction time, and task accuracy".

[0072] The data format standardization module 13 uses Python scripts to call corresponding data parsing libraries (such as pandas parsing .csv, mne parsing .edf, and json library parsing .json) to convert the three types of data into structured data tables indexed by "unified task event timestamps". The data table fields include "unified timestamp, data type (eye movement / EEG / behavior)", and data content (such as eye movement: fixation X / Y; EEG: F3 lead potential; behavior: reaction time)". Finally, the standardized data is stored as an HDF5 format file and transmitted to the central AI processing and analysis unit 2.

[0073] The Central AI Processing and Analysis Unit 2 receives time-stamped multimodal data transmitted from the Multimodal Data Perception and Acquisition Unit 1. Using the timestamps of key events in the cognitive task as anchors, it completes cross-modal temporal alignment of eye-tracking, EEG, and behavioral data streams. It constructs a dynamic heterogeneous graph structure, mapping physiological and behavioral variables to graph nodes, initializing directed edges with time-delay weights to represent potential interactions between variables, and generating a time-series graph sequence that evolves with the task progress. A differentiable causal discovery mechanism is introduced, employing an end-to-end structure search algorithm based on Gumbel-Sigmoid to jointly optimize the existence probability and causal direction of edges in the graph. An improved Granger causal loss function and neurophysiological prior constraints are embedded to automatically learn and obtain... A biologically plausible temporal causal weight matrix is ​​used to identify driving decline patterns and generate an individualized cognitive decline causal chain map. A personalized baseline is established based on the initial test data, and the longitudinal deviation of the intensity of each causal path in subsequent test data is calculated. An LSTM-attention network is used to model the multi-round causal path intensity sequence, outputting the individualized cognitive decline causal chain map and key driving paths to the personalized intervention strategy generation unit 3, while also outputting a standardized clinical report. The central AI processing and analysis unit 2 includes a cross-modal temporal alignment module 21, a dynamic heterogeneous graph construction module 22, a differentiable causal discovery module 23, a cognitive decline causal chain modeling module 24, and a sequence prediction and output module 25, wherein:

[0074] In this embodiment, the cross-modal temporal alignment module 21 receives temporally tagged multimodal data transmitted by the multimodal data sensing and acquisition unit 1, and completes cross-modal temporal alignment of eye-tracking, EEG, and behavioral data streams using the timestamps of key events in the cognitive task as anchor points. The cross-modal temporal alignment module 21 includes a cognitive task anchor point benchmark determination submodule and a multimodal data timestamp calibration submodule, wherein:

[0075] The cognitive task anchor benchmark determination submodule extracts the triggering time of stimulus presentation and response instructions in the cognitive task, and uses the triggering time as the benchmark time point for multimodal data alignment;

[0076] Specifically, the cognitive task anchor benchmark determination submodule extracts the trigger times of two core key events in the cognitive task: "stimulus presentation" and "response instruction." The specific implementation is as follows: Event logs during the cognitive task execution process are parsed from time-series-marked multimodal data; the system timestamps corresponding to the "stimulus presentation identifier" and "response instruction identifier" in the logs are filtered and recorded as the benchmark time point set. ,(satisfy ),in Indicates the first The triggering time of each key event ( (This represents the total number of critical events, which is dynamically adjusted according to the cognitive task design.) This set serves as a unified benchmark for multimodal data alignment.

[0077] The multimodal data timestamp calibration submodule matches the sampling times of eye-tracking signals, EEG signals, and behavioral task data streams with the reference time points. After calibration, the time deviation of each modality data does not exceed 1 / 2 of the corresponding data sampling period.

[0078] Specifically, the multimodal data timestamp calibration submodule sets the sampling times of eye-tracking signals, electroencephalogram signals, and behavioral task data streams with the reference time points. To achieve precise matching, the specific implementation is as follows:

[0079] Assume the sampling rate of the eye movement signal is (Unit: Hz), corresponding sampling period The sampling rate of the EEG signal is The corresponding sampling period The sampling rate of the behavioral task data stream is The corresponding sampling period ;

[0080] Sampling time of arbitrary modal data Find the set of reference time points Zhongyu closest reference time point Calculate time deviation ;

[0081] like ( The sampling period for this modal data is... Corresponding to , , Then, the sampling time is corrected using the fixed offset calibration method. The corrected formula is:

[0082] ;

[0083] in This is a sign function used to maintain the temporal direction of the sampling time relative to the reference time point;

[0084] like Then the original sampling time of the modal data is directly retained. No corrections are needed.

[0085] After the sampling time calibration is completed, for the corrected sampling time The corresponding modal data values ​​are calculated using linear interpolation: assuming the original modality is earlier than... The most recent sampling time is The corresponding data value is Originally later than The most recent sampling time is The corresponding data value is ,but The corresponding values ​​are:

[0086] ;

[0087] If sampling time Earlier than the base time point set The first benchmark or later than the last benchmark point Then retain the original sampling time. It is marked as a "non-reference aligned sampling point" and its timing attribute needs to be separately marked when the subsequent module processes it.

[0088] After calibration, the deviations of all sampling times in the eye-tracking signals, EEG signals, and behavioral task data streams from their corresponding reference time points all meet the requirements. This ensures the temporal consistency of multimodal data.

[0089] In this embodiment, the dynamic heterogeneous graph construction module 22 constructs a dynamic heterogeneous graph structure based on aligned multimodal data, maps physiological and behavioral variables to graph nodes, initializes directed edges with time delay weights to represent the potential interaction relationships between variables, and generates a time series graph sequence that evolves with the task progress. The process of initializing the directed edge time delay weights and generating the time series graph sequence during the construction of the dynamic heterogeneous graph by the dynamic heterogeneous graph construction module 22 includes the following steps:

[0090] S22.1 Determine the preset time delay window and obtain the time series data sequence corresponding to the graph node;

[0091] Specifically, the implementation of determining the preset time delay window and obtaining the time-series data sequence is as follows:

[0092] Preset time delay window length The value is determined based on the time scale of the cognitive task and the physiological signal response characteristics, and the range is [value range missing]. In this embodiment, it is set This value is based on the academic research conclusions of cognitive neuroscience, referring to the time window theory of cognitive processing in "Temporal Windows as Logistical Basis for Cognitive Processing" (Institutional Knowledge Base of Institute of Psychology, Chinese Academy of Sciences), and combined with the neural temporal representation research in "Neural Representation of Time across Complementary Reference Frames" (MPG.PuRe 2025). 0.5s can completely cover the core association cycle (0.4-0.6s) of "stimulus-physiology-behavior" in cognitive tasks. It avoids missing key signals due to an excessively short window and prevents noise from being introduced due to an excessively long window. It is a conventional fitting value for multimodal temporal association analysis.

[0093] The aligned multimodal data is divided into continuous time segments along the time dimension, with the length of each time segment equal to a preset time delay window. Forming a time window sequence ( (Total number of time windows).

[0094] The nodes of the dynamic heterogeneous graph represent physiological and behavioral variables, and the total number of nodes is given by... Then the first Nodes ( ) in the Time window The time series data sequence inside is ,in For the first Each node Inner The values ​​of each sampling point ( (for sampling point index) This is the sampling rate of the modal data corresponding to this node.

[0095] S22.2 Calculate the correlation coefficient of time series data sequences at different nodes under the time delay window;

[0096] Specifically, the specific steps for calculating the correlation coefficient of node time-series data sequences are as follows:

[0097] For any two nodes and ( ), in the Time window Within, calculate the time-series data sequences of both. and correlation coefficient In this embodiment, the Pearson correlation coefficient is used, and the calculation formula is as follows:

[0098] ;

[0099] in, (Time window) (Total number of sampling points within) for The mean, for The mean;

[0100] Correlation coefficient The value range is [-1, 1]. The larger the absolute value, the stronger the correlation between the two nodes within that time window.

[0101] S22.3. Normalize the correlation coefficients of all node pairs to obtain the initial values ​​of the time delay weights of the edges in the graph. ;in, , All are node indices in a dynamic heterogeneous graph;

[0102] Specifically, correlation coefficient normalization and initial value calculation of time delay weights.

[0103] For all node pairs In the Time window Correlation coefficient within Normalization is performed using the Min-Max normalization method, mapping the values ​​to the [0,1] interval to obtain the initial values ​​for the delay weights. (correspond In the The specific values ​​for each time window are calculated using the following formula:

[0104] ;

[0105] in, For all node pairs in The minimum value of the internal correlation coefficient. For all node pairs in The maximum value of the internal correlation coefficient;

[0106] Indicates the first Within a time window, from node Pointing to node The initial values ​​of the time delay weights of the directed edges. , All are node indices in a dynamic heterogeneous graph. ), representing the nodes within this time window For nodes The initial strength of potential interaction relationships.

[0107] S22.4, will As the initial weights of the directed edges in the dynamic heterogeneous graph, a sequence of time graphs that evolves with the progress of the task is generated.

[0108] Specifically, the generation of dynamic heterogeneous graph time series sequences involves the following steps:

[0109] For each time window Using all physiological and behavioral variables as nodes, Construct a static heterogeneous graph with initial weights for directed edges. ,in ( For the first (nodes). (directed edge) Corresponding node To the node Potential interaction relationships);

[0110] Concatenate all the static heterogeneous graphs corresponding to the time windows in chronological order to form a sequence of time-series graphs that evolve with the task progress. This sequence fully preserves the dynamic changes in the potential interaction relationships between nodes within different time windows.

[0111] In this embodiment, the differentiable causal discovery module 23 introduces a differentiable causal discovery mechanism to process the time series graph sequence and jointly optimize the existence probability and causal direction of edges in the graph. The process by which the differentiable causal discovery module 23 optimizes the existence probability and causal direction of graph edges based on the time series graph sequence by constructing a joint optimization objective function includes the following steps:

[0112] S23.1 Determine the predicted value of the node time series data at the target time, and the set of parent nodes of the node corresponding to the target time, to clarify the basis for causal relationship fitting;

[0113] Specifically, the implementation of determining the node prediction value and the set of parent nodes is as follows:

[0114] Target time definition: Let the current processing time window be... The target time is The first Each sampling time ( ), denoted as ;

[0115] node At the target time The true value is Predicted value It is predicted by combining its historical sampling data and the historical data of its parent node set;

[0116] Parent node set definition: Let node The set of parent nodes is , indicating a node At the target time The set of nodes whose values ​​have potential causal influence. (excluding nodes) (Itself), the initial set of parent nodes is composed of dynamic heterogeneous graphs Medium weight nodes The threshold value of 0.3 is used for preliminary screening of strongly correlated nodes. The value of 0.3 is based on the classic classification standard for correlation coefficient strength in "Applied Multivariate Statistical Analysis" (Johnson & Wichern), and also aligns with the research conclusion of "Screening and Evaluation of Non-volatile Decision Traits for Oviposition Resistance / Susceptibility in Oriented Fruit Fly" (Chinese Journal of Agricultural Science, 2023)—when the absolute value of the correlation coefficient is ≥0.3, the variables have a statistically significant effective association, which can eliminate random noise and retain core features. It is a common dividing line between "weak correlation" and "moderate correlation" in the field of feature screening and is widely used in multi-domain filtering tasks.

[0117] S23.2 Construct a prediction error term to characterize the fitting accuracy of causal relationships, ensuring the accuracy of causal inference;

[0118] Specifically, the construction of the prediction error term is implemented as follows:

[0119] The prediction error term is used to characterize the fitting accuracy of causal relationships, based on nodes. The true value Compared with the predicted value The deviation is constructed using the following formula:

[0120] ;

[0121] in, , For nodes To the node Causal weights (to be optimized). For nodes For nodes The causal delay (determined by the time offset when the correlation coefficient is at its maximum in S22.2). For nodes The bias term; This represents the total number of time windows for the dynamic heterogeneous graph time series sequence. The total number of sampling points within a single time window. The total number of nodes in the dynamic heterogeneous graph;

[0122] The smaller the value, the closer the prediction result based on the set of parent nodes is to the true value, and the higher the fitting accuracy of the causal relationship.

[0123] S23.3, Add a sparsity regularization term to constrain the redundancy of causal relationships;

[0124] Specifically, the steps for constructing the sparse regularization term are as follows:

[0125] The sparsity regularization term is used to constrain the redundancy of causal relationships and avoid too many spurious causal edges. L1 regularization is used, and the formula is:

[0126] ;

[0127] in, is the regularization coefficient, with a value range of [0.001, 0.01]. In this embodiment, it is set as follows: This value originates from the academic consensus in the field of sparse causal discovery. Referring to research findings in "Sparse Causal Discovery in MultivariateTime Series" (ar5iv 2025) and "A Sparsity Principle for Partially Observable Causal Representation Learning" (arXiv 2024), the typical parameter range for sparse regularization in causal models is 0.001-0.01. 0.005 falls in the middle of this range, balancing model sparsity with the preservation of causal relationships, which is consistent with... The parameter optimization logic based on 1,2-norm regularization has been proven effective in dynamic Granger causal learning tasks.

[0128] Introducing L1 regularization can make most causal weights The value approaches 0, retaining only a few significant causal edges, which aligns with the sparsity characteristic of causal relationships in cognitive decline.

[0129] S23.4 Integrate the prediction error term and the sparsity regularization term to form an improved Granger causal loss function LGranger, and simultaneously optimize the existence probability and causal direction of edges in the graph based on the improved Granger causal loss function LGranger.

[0130] Specifically, the construction and optimization of the improved Granger causal loss function are as follows:

[0131] Integrating prediction error terms With sparsity regularization term and embedding neurophysiological prior constraints This leads to an improved Granger causality loss function. The formula is:

[0132] ;

[0133] in, These are neurophysiological prior constraints used to ensure the biological plausibility of causal relationships. For example, based on common sense in neuroscience, eye-tracking signal nodes do not directly drive EEG signal nodes; therefore, they constrain the causal weights of such node pairs. The formula is (Ω represents the set of node pairs that do not conform to common sense in neurophysiology);

[0134] Joint optimization objective: Employ an end-to-end structure search algorithm based on Gumbel-Sigmoid to simultaneously optimize the existence probability of edges in the graph. The probability of the existence of an edge is related to the causal direction. Calculation, where For the Sigmoid function, For random noise that follows a Gumbel distribution, The temperature parameter (initially 1.0, decaying to 0.1 with iteration);

[0135] Optimization process: Minimize using the stochastic gradient descent (SGD) algorithm The number of iterations was set to 1000, the learning rate was set to 0.001, and the final convergence yielded a biologically plausible temporal causal weight matrix. and the probability matrix of the existence of edges ,when At that time, it is determined that there is a slave node. To the node The causal edge, the causal direction is → .

[0136] In this embodiment, the cognitive decline causal chain modeling module 24, based on the output of the differentiable causal discovery module 23, identifies driving decline patterns, generates an individualized cognitive function decline causal chain map, establishes a personalized baseline, and calculates the longitudinal deviation of the causal path strength. The process by which the cognitive decline causal chain modeling module 24 calculates the longitudinal deviation of the causal path strength based on the personalized baseline after generating the cognitive decline causal chain map includes the following steps:

[0137] S24.1 Based on the initial detection data, determine the first... Initial strength value of causal path , for the first Establish personalized baselines based on causal paths;

[0138] Specifically, the steps for establishing a personalized baseline are as follows:

[0139] Causal path definition: The first in the causal chain diagram of individualized cognitive function decline causal path ( , The total number of causal paths (where is the total number of causal paths) consists of continuous causal edges, denoted as . ,in path length ( );

[0140] Initial intensity value calculation: based on the temporal causal weight matrix corresponding to the first detection data. Calculate the first Initial strength value of causal path The formula is:

[0141] ;

[0142] in, For the initial detection data, from node To the node Causal weights;

[0143] Personalized baseline: As the first The personalized baseline of the causal path is denoted as This is used for subsequent comparative analysis of the changing trends of path intensity.

[0144] S24.2, Obtain the first The first test Current strength value of the causal path Capture cross-cycle change data;

[0145] Specifically, obtain the current intensity value. The specific steps are as follows:

[0146] No. Second test ( , The number of tests, the first test After completion, the temporal causal weight matrix corresponding to this detection is obtained through the differentiable causal discovery module 23. ;

[0147] Using the same formula as S24.1, calculate the first... The causal path in the first Current intensity value in this test:

[0148] ;

[0149] in, For the first In the second detection data, from node To the node Causal weights;

[0150] Directly reflects the first During the second test, the first The intensity of causal influence of each causal path.

[0151] S24.3, based on and Calculate the first Longitudinal deviation of causal paths .

[0152] Specifically, calculate the longitudinal deviation. The specific steps are as follows:

[0153] Longitudinal deviation Used to characterize the The formula for the degree of change in the strength of a causal path relative to a personalized baseline is:

[0154] ;

[0155] in, The range of values ​​is , The larger the value, the more significant the change in the strength of the causal path;

[0156] like ( To avoid deviation from the threshold, this embodiment sets If the value is within the specified range, the causal path is determined to be a driving path of cognitive decline and included in the core path set of the individualized cognitive decline causal chain map. This value conforms to the clinical criteria for cognitive impairment assessment, refers to the diagnostic criteria for mild cognitive impairment (MCI) in the "Chinese Guidelines for the Diagnosis and Treatment of Dementia and Cognitive Impairment (2025 Edition)," and aligns with the clinical consensus of the Australian Dementia Society webinar (2025)—a 20% change in cognitive indicators can effectively distinguish between "physiological decline" and "pathological deviation," cover measurement errors (approximately ±10%), and accurately identify substantial decline trends; it is a commonly used threshold for early cognitive impairment screening.

[0157] In this embodiment, the sequence prediction and output module 25 uses an LSTM-attention network to model the multi-turn causal path strength sequence, outputting an individualized cognitive function decline causal chain map and key driving paths to the personalized intervention strategy generation unit 3, while also outputting a standardized clinical report. The sequence prediction and output module 25, based on the multi-turn causal path strength sequence, utilizes an LSTM-attention network to calculate the attention weights at each time step to achieve sequence modeling, including the following steps:

[0158] S25.1, Obtaining the multi-round causal path strength sequence after inputting it into the LSTM network to generate the first... Step Hidden State and context vectors representing global features of the sequence. This forms the basis for constructing a two-dimensional input of local temporal features and global features;

[0159] Specifically, to obtain the first Step Hidden State and context vector The specific steps are as follows:

[0160] Definition of multi-round causal path strength sequence: Assume that multiple rounds of causal path strength sequence are performed. The first test (including the first test), the The intensity sequence of the causal path is The intensity sequences of all causal paths constitute the input matrix. (dimension is) );

[0161] LSTM network input: Input matrix Input the data into the LSTM network in the detection order, at each time step... The input is The Column (corresponding to all causal paths in the first column) The intensity of the second detection, the dimension is ×1); The hidden layer dimension of the LSTM network is set to 64, the activation functions of the forget gate, input gate, and output gate are all Sigmoid functions, and the cell state update uses the tanh function;

[0162] No. Step Hidden State LSTM network processing After the intensity sequence of the second detection, output the first... Step Hidden State (Dimensions are 64×1) Including the Local temporal characteristics of the strength of all causal paths during the second detection;

[0163] Context vector For all Hidden state Perform average pooling to obtain a context vector representing the global features of the sequence. (Dimensions are 64×1), the formula is:

[0164] .

[0165] S25.2, based on and The association mapping relationship is calculated for the first Attention score for each time step;

[0166] Specifically, the steps for calculating attention scores are as follows:

[0167] Based on hidden state With context vector The association mapping relationship is calculated. Attention score at each time step The formula is:

[0168] ;

[0169] in, (Dimensions are 1×64) (1×64 dimension) is the weight matrix. (1×1 dimension) is the bias term. For activation functions;

[0170] Attention score Used to measure the The correlation strength between local features and global features at each time step. The larger the value, the greater the contribution of the intensity data of that time step to sequence modeling.

[0171] S25.3. Normalize the attention scores for all time steps to obtain the first... Attention weights for each time step This enables the modeling of multi-round causal path strength sequences.

[0172] Specifically, attention weights The specific operations for computation and sequence modeling are as follows:

[0173] Attention weight :

[0174] Attention score for all time steps Normalization is performed, and the softmax function is used to obtain the first... Attention weights for each time step The formula is:

[0175] ;

[0176] in:

[0177] Indicates the total number of detections (including the first detection and the length of the corresponding multi-round causal path strength sequence).

[0178] Indicates the time step index (value range is...) , corresponding to the (the timing step corresponding to the next detection)

[0179] The time step traversal variable (range and...) represents the time step traversal variable in the summation process. Consistent);

[0180] This represents the natural exponential function, used to map attention scores to non-negative values;

[0181] The value range of is [0,1], and This ensures the normalization property of the weights, that is, the sum of the attention weights of all time steps is 1;

[0182] Sequence modeling: Incorporating attention weights Hidden state corresponding to the timing step Weighted summation yields the sequence modeling result that integrates local and global features. The formula is:

[0183] ;

[0184] Output results: Based on modeling results Filter out attention weights The causal path corresponding to (0.2 as an empirical threshold) is used as the key driving path. Combined with the individualized cognitive function decline causal chain map, a standardized clinical report is generated, which includes the trend of path strength changes, analysis of driving decline patterns, and intervention suggestions. This report is then simultaneously output to the personalized intervention strategy generation unit 3. The value of 0.2 is based on academic research on attention feature selection. According to the algorithm logic of "Sequential Attention for Feature Selection" (ResearchGate 2022), 0.2 is a general dividing value that distinguishes between "core weights" and "secondary weights". Features with weights ≥ 0.2 contribute more than 80% to the model output. This approach can control computational complexity while focusing on key causal paths and has been widely validated in feature selection tasks using efficient attention mechanisms.

[0185] Personalized intervention strategy generation unit 3 receives the individualized cognitive function decline causal chain map and key driving path output by the central AI processing and analysis unit 2, generates a personalized cognitive intervention strategy that targets and blocks the decline path, and synchronizes the generated personalized cognitive intervention strategy to the effect tracking and feedback closed loop unit 4.

[0186] In this embodiment, the personalized intervention strategy generation unit 3 includes an intervention input parsing module 31, a target pathway blocking strategy generation module 32, and an intervention strategy synchronization module 33, wherein:

[0187] The intervention input analysis module 31 receives the individualized cognitive function decline causal chain map and key driving paths output by the central AI processing and analysis unit 2, and analyzes the path associations and core features of the key driving paths in the causal chain map.

[0188] Specifically, the intervention input parsing module 31 receives the individualized cognitive function decline causal chain graph (GraphML format) and key driving path list output by the central AI processing and analysis unit 2, as follows:

[0189] Graph structured analysis: Extracting the modal types (eye movement / EEG / behavioral) and core feature indicators of nodes in the graph, clarifying the start-end association and causal strength of causal edges. Timely delay We will analyze the complete chain of each key driving path and associate it with the corresponding cognitive function dimensions (such as EEG-attention, eye movement-visual search, and behavior-executive function).

[0190] Core Feature Extraction: Extracting the initial intensity value of each key driving path. Current intensity value and longitudinal deviation ,according to Determine intervention priority by sorting in descending order; identify the core sensitive features of the starting drive node of the pathway (such as EEG nodes). Wave power ratio and fixation stability of eye movement nodes are used as indicators for monitoring intervention effects.

[0191] Output: A structured parsing report, including a node-feature mapping table, a path-cognitive dimension association table, an intervention priority ranking table, and a list of sensitive monitoring indicators.

[0192] The targeted pathway blocking strategy generation module 32 generates a personalized cognitive intervention strategy that targets and blocks the decline pathway based on the analysis results of the intervention input analysis module 31.

[0193] Specifically, the targeted pathway blocking strategy generation module 32 generates intervention strategies based on the analysis report of the intervention input analysis module 31, following the logic of "path type - intervention goal - task selection - personalized configuration". The specific implementation is as follows:

[0194] Intervention target setting: For each key driving path, the core objectives are "to block the abnormal causal transmission, weaken the abnormal activation of driving nodes, and optimize the function of target nodes";

[0195] Task-based selection:

[0196] Physiological-behavioral pathway: Select targeted physiological regulation training + behavioral optimization training (such as gaze control training + response inhibition training);

[0197] Multimodal sequential approach: Select cross-modal collaborative training + single-modal targeted training (such as neurofeedback training + vision-decision collaborative training).

[0198] Physiological-physiological pathway: Select rhythm regulation training (such as auditory / visual stimulation training based on neural rhythms);

[0199] Personalized configuration: Training duration, frequency, and initial difficulty are based on individual baselines. Path priority ( (The larger the value, the higher the training frequency) setting; adjust the rules to reflect changes in sensitive indicators and causal strength after intervention. Based on changes, dynamically adapt to individual intervention effects;

[0200] Output: A personalized intervention package, including intervention goals, task list, parameter configuration, monitoring indicators, and adjustment rules.

[0201] The intervention strategy synchronization module 33 synchronizes the personalized cognitive intervention strategy generated by the target path blocking strategy generation module 32 to the effect tracking and feedback closed-loop unit 4.

[0202] Specifically, the intervention strategy synchronization module 33 is responsible for synchronizing the personalized intervention plan package to the effect tracking and feedback closed-loop unit 4, and the specific implementation is as follows:

[0203] Format standardization: The intervention plan package is converted into a structured JSON format, including intervention strategy ID, user identifier, associated path ID, task details, monitoring indicators and synchronization timestamp, to ensure that the receiving end can parse it directly;

[0204] Secure transmission: Data is transmitted using HTTPS protocol combined with encryption mechanisms to prevent tampering or leakage;

[0205] Synchronization Guarantee: Set up a reception confirmation and retry mechanism to automatically retry if no confirmation signal is received; record synchronization logs (including synchronization time, status, and verification results) to support subsequent traceability;

[0206] Output: The standardized personalized intervention strategy is synchronized to the effect tracking and feedback closed loop unit 4 to complete the accurate delivery of the intervention strategy.

[0207] The effect tracking and feedback closed-loop unit 4 receives personalized cognitive intervention strategies synchronized by the personalized intervention strategy generation unit 3, dynamically monitors the implementation process and effects of the personalized cognitive intervention strategy, collects multimodal physiological and behavioral data after intervention, and transmits the collected post-intervention multimodal data back to the central AI processing and analysis unit 2. The central AI processing and analysis unit 2 iterates the cognitive decline causal chain map and modeling results based on the transmitted post-intervention multimodal data, and feeds back the iterated results to the personalized intervention strategy generation unit 3, which then corrects the intervention strategy parameters.

[0208] In this embodiment, the effect tracking and feedback closed-loop unit 4 includes an intervention strategy receiving module 41, an intervention process monitoring module 42, a feedback data acquisition module 43, and a data feedback and iteration connection module 44, wherein:

[0209] Intervention strategy receiving module 41 receives personalized cognitive intervention strategies synchronized by personalized intervention strategy generation unit 3;

[0210] Specifically, the core function of the intervention strategy receiving module 41 is to accurately receive the personalized cognitive intervention strategy synchronized by the personalized intervention strategy generation unit 3, which is implemented as follows:

[0211] Receiving mechanism: It adopts the communication protocol (HTTPS protocol) that matches the intervention strategy synchronization module 33 to receive personalized intervention strategy data in structured JSON format;

[0212] Data verification: The received data is verified for completeness and format. The core fields (intervention strategy ID, user unique identifier, associated critical path ID, intervention task list, and monitoring indicator list) are checked to see if they are complete and if the format conforms to the preset specifications. If the verification passes, the data is stored in the local cache. If the verification fails, a retransmission request is returned to the personalized intervention strategy generation unit 3.

[0213] Data parsing and storage: Parse JSON format data, extract key information such as intervention task type, implementation parameters (duration, frequency), and core monitoring indicators, and store it in an index based on "user identifier-strategy ID" to provide directly callable strategy basis for subsequent intervention process monitoring module 42.

[0214] The intervention process monitoring module 42 monitors the implementation process of the personalized cognitive intervention strategy and the corresponding physiological and behavioral changes based on the personalized cognitive intervention strategy obtained by the intervention strategy receiving module 41.

[0215] Specifically, the intervention process monitoring module 42 dynamically monitors the intervention implementation process and individual physiological and behavioral responses based on the intervention strategy parsed by the intervention strategy receiving module 41, as follows:

[0216] Implementation process monitoring: Monitor the compliance of intervention execution according to the task type and parameters in the intervention strategy, including whether the intervention is implemented according to the preset duration and frequency, whether the task difficulty is adjusted according to the rules, and record the completion status of intervention implementation (such as "completion rate" and "number of interruptions").

[0217] Dynamic monitoring of physiological behavior: based on clearly defined core monitoring indicators (such as EEG nodes) according to intervention strategies. The data includes wave power ratio, fixation stability of eye movement nodes, and reaction time of behavioral nodes. Physiological and behavioral data changes during the intervention process are captured in real time using non-invasive monitoring devices (consistent with the type of acquisition device in multimodal data perception and acquisition unit 1). The data is recorded in segments of 100ms-500ms time windows to ensure that the data is synchronized with the timing of the intervention process.

[0218] Anomaly warning: If the intervention implementation deviates from the preset parameters (such as the duration of a single intervention is less than 80% of the preset value) or if the physiological behavior data shows extreme abnormalities (such as excessive EEG signal noise), a real-time warning will be triggered, prompting the intervention implementer (such as medical staff or individuals) to adjust the implementation method or investigate equipment problems.

[0219] Feedback data acquisition module 43 collects post-intervention multimodal physiological and behavioral data within the monitoring scope of intervention process monitoring module 42;

[0220] Specifically, the feedback data acquisition module 43, based on the monitoring range of the intervention process monitoring module 42, collects multimodal physiological and behavioral data after the intervention, providing data support for the iterative analysis of the central AI processing and analysis unit 2. The specific implementation is as follows:

[0221] Data collection timing: Data collection should be initiated according to the cycle set by the intervention strategy (e.g., immediate collection after a single intervention, or centralized collection after the end of a weekly intervention cycle).

[0222] The data collected is consistent with the data collection dimensions of the multimodal data perception and acquisition unit 1, including eye movement signals (gaze coordinates, gaze duration), electroencephalogram (EEG signals (potentials in specified leads, power percentage of characteristic frequency bands), and behavioral data (reaction time, task accuracy), to ensure the continuity of data dimensions.

[0223] Data preprocessing: The collected raw data is standardized and preprocessed, including noise reduction (bandpass filtering of EEG signals from 5-30Hz and removal of abnormal values ​​from eye movement signals), timestamp marking (adding a uniform timestamp to each data point based on the intervention end time), and format conversion (converting to a format consistent with time-series tokenized multimodal data) to ensure that the data can be directly parsed by the central AI processing and analysis unit 2.

[0224] The data feedback and iteration connection module 44 feeds back the multimodal data collected by the feedback data acquisition module 43 after intervention to the central AI processing and analysis unit 2, and receives the cognitive decline causal chain map and modeling results after the central AI processing and analysis unit 2 iterates based on the fed-back data, and synchronizes them to the personalized intervention strategy generation unit 3.

[0225] Specifically, the data feedback and iteration connection module 44 is responsible for the feedback of data and the connection of iteration results, constructing a closed-loop link of "collection-analysis-correction", which is implemented as follows:

[0226] Feedback data transmission: The multimodal data after intervention, which is preprocessed by the feedback data collection module 43, is associated and marked according to "user identifier-strategy ID-collection timestamp" and transmitted back to the central AI processing and analysis unit 2 through an encrypted transmission mechanism (the same encryption method as the intervention strategy synchronization module 33). The transmission is accompanied by a record of the intervention implementation status (completion rate, abnormal situations).

[0227] Iteration result reception: Receive the results of the central AI processing and analysis unit 2 after iteration based on the feedback data, including the updated individualized cognitive function decline causal chain map, key driving path adjustment results, and updated values ​​of longitudinal deviation of causal path strength;

[0228] Iterative Integration: The iteration results are synchronized to the personalized intervention strategy generation unit 3 in a structured format, with the "iteration version number" and "core adjustment point" (such as "causal strength of key driving path A") marked during synchronization. "Reducing by 15%" provides a clear basis for the target pathway blocking strategy generation module 32 to modify the intervention strategy parameters;

[0229] Closed-loop logging: Records data feedback status, iteration result reception status, and time nodes for synchronization to personalized intervention strategy generation unit 3. Log files are established according to "user identifier - iteration cycle" to support subsequent traceability and closed-loop effect evaluation.

[0230] Those skilled in the art will understand that the process of implementing all or part of the steps of the above embodiments can be carried out by hardware or by a program instructing the relevant hardware.

[0231] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An integrated platform for early screening and intervention of cognitive impairment driven by artificial intelligence, characterized in that: include: Multimodal data perception and acquisition unit (1) is based on eye-tracking and EEG synchronous recording technology to collect synchronous eye-tracking signals, EEG signals and behavioral task data streams of cognitive impairment risk groups, associate key events of cognitive tasks to generate timestamps, complete the perception and acquisition of time-marked multimodal physiological and behavioral data, and transmit the output time-marked multimodal data to the central AI processing and analysis unit (2). The central AI processing and analysis unit (2) receives time-stamped multimodal data transmitted by the multimodal data perception and acquisition unit (1), and completes cross-modal temporal alignment of eye-tracking, EEG and behavioral data streams with the key event timestamps of the cognitive task as anchor points; constructs a dynamic heterogeneous graph structure, maps physiological and behavioral variables as graph nodes, initializes directed edges with time delay weights to represent the potential interaction relationships between variables, and generates a time-series graph sequence that evolves with the task progress; introduces a differentiable causal discovery mechanism, and uses an end-to-end structure search algorithm based on Gumbel-Sigmoid to jointly optimize the existence probability of edges in the graph. With the causal direction, an improved Granger causal loss function and neurophysiological prior constraints are embedded to automatically learn a biologically reasonable temporal causal weight matrix; based on the temporal causal weight matrix, the driving decline pattern is identified, and an individualized cognitive function decline causal chain map is generated. A personalized baseline is established based on the first detection data, and the longitudinal deviation of the intensity of each causal path in the subsequent detection data is calculated; the LSTM-attention network is used to model the multi-round causal path intensity sequence, and the individualized cognitive function decline causal chain map and key driving paths are output to the personalized intervention strategy generation unit (3), while a standardized clinical report is output at the same time; Personalized intervention strategy generation unit (3) receives the individualized cognitive function decline causal chain map and key driving path output by the central AI processing and analysis unit (2), generates a personalized cognitive intervention strategy that targets and blocks the decline path, and synchronizes the generated personalized cognitive intervention strategy to the effect tracking and feedback closed loop unit (4). The effect tracking and feedback closed-loop unit (4) receives the personalized cognitive intervention strategy synchronized by the personalized intervention strategy generation unit (3), dynamically monitors the implementation process and effect of the personalized cognitive intervention strategy, collects multimodal physiological and behavioral data after intervention, and transmits the collected multimodal data after intervention back to the central AI processing and analysis unit (2); the central AI processing and analysis unit (2) iterates the cognitive decline causal chain map and modeling results based on the transmitted multimodal data after intervention, and feeds back the iterated results to the personalized intervention strategy generation unit (3), which then corrects the intervention strategy parameters.

2. The AI-driven integrated platform for early screening and intervention of cognitive impairment according to claim 1, characterized in that, The multimodal data sensing and acquisition unit (1) includes a signal synchronization acquisition module (11), a task event timestamp association module (12), and a data format standardization module (13), wherein: The signal synchronization acquisition module (11) is based on eye-tracking and electroencephalogram (EEG) synchronization recording technology to simultaneously acquire eye-tracking signals, EEG signals and behavioral task data streams of people at risk of cognitive impairment. The task event timestamp association module (12) associates key events of cognitive tasks and generates corresponding timestamp markers for the collected eye movement signals, EEG signals and behavioral task data streams; The data format standardization module (13) performs format standardization processing on the eye movement signal, EEG signal and behavioral task data stream after the timestamp is marked, and transmits the processed time-marked multimodal physiological and behavioral data to the central AI processing and analysis unit (2).

3. The AI-driven integrated platform for early screening and intervention of cognitive impairment according to claim 2, characterized in that, The central AI processing and analysis unit (2) includes a cross-modal temporal alignment module (21), a dynamic heterogeneous graph construction module (22), a differentiable causal discovery module (23), a cognitive decay causal chain modeling module (24), and a sequence prediction and output module (25), wherein: The cross-modal temporal alignment module (21) receives the temporally marked multimodal data transmitted by the multimodal data perception and acquisition unit (1) and completes the cross-modal temporal alignment of eye movement, EEG and behavioral data streams with the key event timestamps of cognitive tasks as anchor points. The dynamic heterogeneous graph construction module (22) constructs a dynamic heterogeneous graph structure based on the aligned multimodal data, maps physiological and behavioral variables to graph nodes, initializes directed edges with time delay weights to represent the potential interaction relationships between variables, and generates a time sequence graph sequence that evolves with the task progress. The differentiable causal discovery module (23) introduces a differentiable causal discovery mechanism to process the time series graph sequence and jointly optimize the existence probability of edges and causal direction in the graph. The cognitive decline causal chain modeling module (24) identifies driving decline patterns based on the output of the differentiable causal discovery module (23), generates an individualized cognitive function decline causal chain map, establishes a personalized baseline, and calculates the longitudinal deviation of the causal path strength. The sequence prediction and output module (25) uses an LSTM-attention network to model the intensity sequence of multi-round causal paths, outputs an individualized cognitive function decline causal chain map and key driving paths to the personalized intervention strategy generation unit (3), and outputs a standardized clinical report.

4. The AI-driven integrated platform for early screening and intervention of cognitive impairment according to claim 3, characterized in that, The cross-modal temporal alignment module (21) includes a cognitive task anchor benchmark determination submodule and a multimodal data timestamp calibration submodule, wherein: The cognitive task anchor benchmark determination submodule extracts the triggering time of stimulus presentation and response instructions in the cognitive task, and uses the triggering time as the benchmark time point for multimodal data alignment; The multimodal data timestamp calibration submodule matches the sampling times of eye-tracking signals, electroencephalogram signals, and behavioral task data streams with the reference time points. After calibration, the time deviation of each modality data does not exceed 1 / 2 of the corresponding data sampling period.

5. The AI-driven integrated platform for early screening and intervention of cognitive impairment according to claim 4, characterized in that, The dynamic heterogeneous graph construction module (22) performs the following steps when constructing the dynamic heterogeneous graph: initial value calculation of directed edge time delay weights and time series graph sequence generation. S22.1 Determine the preset time delay window and obtain the time series data sequence corresponding to the graph node; S22.2 Calculate the correlation coefficient of time-series data sequences at different nodes under the time delay window; S22.

3. Normalize the correlation coefficients of all node pairs to obtain the initial values ​​of the time delay weights of the edges in the graph. ;in, , All are node indices in a dynamic heterogeneous graph; S22.4, will As the initial weights of the directed edges in the dynamic heterogeneous graph, a sequence of time graphs that evolves with the progress of the task is generated.

6. The AI-driven integrated platform for early screening and intervention of cognitive impairment according to claim 5, characterized in that, The differentiable causal discovery module (23), based on the time series graph sequence, optimizes the probability of graph edge existence and causal direction by constructing a joint optimization objective function, including the following steps: S23.1 Determine the predicted value of the node time series data at the target time, and the set of parent nodes of the node corresponding to the target time, to clarify the basis for causal relationship fitting; S23.2 Construct a prediction error term to characterize the fitting accuracy of causal relationships, ensuring the accuracy of causal inference; S23.3, Add a sparsity regularization term to constrain the redundancy of causal relationships; S23.4 Integrate the prediction error term and the sparsity regularization term to form an improved Granger causal loss function LGranger, and simultaneously optimize the existence probability and causal direction of edges in the graph based on the improved Granger causal loss function LGranger.

7. The AI-driven integrated platform for early screening and intervention of cognitive impairment according to claim 6, characterized in that, The cognitive decline causal chain modeling module (24), after generating the cognitive decline causal chain map, calculates the longitudinal deviation of the causal path strength based on the personalized baseline in the following steps: S24.1 Based on the initial detection data, determine the first... Initial strength value of causal path , for the first Establish personalized baselines based on causal paths; S24.2, Obtain the first The first test Current strength value of the causal path Capture cross-cycle change data; S24.3, based on and Calculate the first Longitudinal deviation of causal paths .

8. The AI-driven integrated platform for early screening and intervention of cognitive impairment according to claim 7, characterized in that, The sequence prediction and output module (25) uses an LSTM-attention network to calculate the attention weights of time steps to achieve sequence modeling based on a multi-round causal path strength sequence. The process includes the following steps: S25.1, Obtaining the multi-round causal path strength sequence after inputting it into the LSTM network to generate the first... Step Hidden State and context vectors representing global features of the sequence. This forms the basis for constructing a two-dimensional input of local temporal features and global features; S25.2, based on and The association mapping relationship is calculated for the first Attention score for each time step; S25.

3. Normalize the attention scores for all time steps to obtain the first... Attention weights for each time step This enables the modeling of multi-round causal path strength sequences.

9. The AI-driven integrated platform for early screening and intervention of cognitive impairment according to claim 8, characterized in that, The personalized intervention strategy generation unit (3) includes an intervention input parsing module (31), a target pathway blocking strategy generation module (32), and an intervention strategy synchronization module (33), wherein: The intervention input analysis module (31) receives the individualized cognitive function decline causal chain map and key driving path output by the central AI processing and analysis unit (2), and analyzes the path association relationship and core features of the key driving path in the causal chain map; The targeted path blocking strategy generation module (32) generates a personalized cognitive intervention strategy for targeting and blocking the decline path based on the analysis results of the intervention input analysis module (31). The intervention strategy synchronization module (33) synchronizes the personalized cognitive intervention strategy generated by the target path blocking strategy generation module (32) to the effect tracking and feedback closed-loop unit (4).

10. The AI-driven integrated platform for early screening and intervention of cognitive impairment according to claim 9, characterized in that, The effect tracking and feedback closed-loop unit (4) includes an intervention strategy receiving module (41), an intervention process monitoring module (42), a feedback data acquisition module (43), and a data feedback and iteration connection module (44), wherein: The intervention strategy receiving module (41) receives the personalized cognitive intervention strategy synchronized by the personalized intervention strategy generation unit (3); The intervention process monitoring module (42) monitors the implementation process of the personalized cognitive intervention strategy and the corresponding physiological behavior changes based on the personalized cognitive intervention strategy obtained by the intervention strategy receiving module (41). The feedback data acquisition module (43) collects post-intervention multimodal physiological and behavioral data within the monitoring scope of the intervention process monitoring module (42); The data feedback and iteration connection module (44) feeds back the multimodal data after intervention collected by the feedback data acquisition module (43) to the central AI processing and analysis unit (2), and receives the cognitive decline causal chain map and modeling results after the iteration of the feedback data from the central AI processing and analysis unit (2), and synchronizes them to the personalized intervention strategy generation unit (3).

Images