Remote assessment system for mild cognitive impairment based on meta-learning and hybrid experts

Abstract

The present application provides a mild cognitive impairment remote evaluation system based on meta-learning and mixed experts, belonging to the technical field of artificial intelligence, intelligent medical treatment and remote health evaluation, comprising: a multi-modal behavior feature encoding module, outputting a multi-modal disease-specific behavior knowledge vector; a meta-learning pathology representation calculation module, used for converting the multi-modal disease-specific behavior knowledge vector into pathological meta-features corresponding to pathological mechanisms, and combining to form a static pathological meta-feature matrix; a pathological mechanism synergy relationship modeling module, used for taking each pathological meta-feature as a node, learning the synergy relationship between different pathological mechanisms, and outputting a static pathological synergy relationship matrix; a static memory layer, jointly constructed by the static pathological meta-feature matrix and the static pathological synergy relationship matrix; a dynamic joint decision module, weighting and fusing the output results of activated experts to obtain a comprehensive decision result; and a risk output module, used for generating a mild cognitive impairment risk probability according to the comprehensive decision result.

Description

Technical Field

[0001] This invention belongs to the fields of artificial intelligence, smart healthcare and remote health assessment technology, and specifically relates to a remote assessment system for mild cognitive impairment based on meta-learning and hybrid experts for remote screening of mild cognitive impairment. Background Technology

[0002] Mild cognitive impairment (MCI) is a critical transitional stage between normal aging and Alzheimer's disease, and a core window for early screening and intervention of cognitive degenerative diseases. While MCI symptoms are mild and unclear, timely identification and continuous monitoring can effectively slow disease progression and significantly reduce long-term care costs for families and society. Current clinical MCI identification primarily relies on brain imaging, biomarker detection, and cognitive scale assessments. While these methods possess a certain degree of accuracy, they suffer from drawbacks such as high equipment costs, complex procedures, and difficulty in high-frequency repetition, making them unsuitable for large-scale screening needs in families, communities, and grassroots settings. Furthermore, manual scale assessments are highly subjective and heavily reliant on professional personnel, further limiting their effectiveness in widespread adoption.

[0003] In recent years, voice, text, task response, and interaction behavior patterns have been proven to reflect dynamic changes in cognitive function, providing a feasible path for building a low-threshold, remote intelligent assessment system. However, existing related technical solutions still have significant shortcomings:

[0004] First, modality utilization is insufficient. Some existing cognitive impairment screening programs rely solely on single speech or text data. For example, typical speech feature-based screening methods usually only extract speech prosodic features such as speech rate, fundamental frequency, and pauses to build a classification model. They do not incorporate key behavioral information such as task execution trajectory, interaction sequence, and operation error patterns into a unified analysis framework. Therefore, it is difficult to comprehensively characterize the complex abnormalities of mild cognitive impairment in different cognitive domains.

[0005] Second, knowledge representation is isolated. Some existing multimodal cognitive screening solutions use end-to-end deep learning models, which directly feed multimodal inputs into a unified classifier and learn features only through data fitting. They do not explicitly introduce pathological mechanisms such as hippocampal memory coding disorders, impaired prefrontal executive function, and language semantic disorders, resulting in models lacking medical semantic support and having insufficient generalization ability and clinical interpretability.

[0006] Third, the decision-making mechanism is rigid. Existing mainstream deep learning models usually use a single fixed network structure to process all subjects. For example, schemes based on unified classification architectures such as ResNet and Transformer use the same reasoning path for different subjects. They cannot dynamically adjust the reasoning strategy according to individual cognitive impairment patterns and are difficult to adapt to the significant heterogeneity of mild cognitive impairment.

[0007] Fourth, there is insufficient interpretability. Existing solutions can usually only output high and low risk results. Even if some technologies use ex post-hoc interpretation methods such as SHAP and LIME, they mainly remain at the level of feature contribution, making it difficult to reconstruct the pathological reasoning path and clarify the source of anomalies. This fails to provide sufficient support for review, examination, and clinical trust in medical scenarios. Summary of the Invention

[0008] Existing intelligent assessment schemes for mild cognitive impairment generally suffer from problems such as insufficient modality utilization, isolated pathological knowledge representation, rigid decision-making mechanisms, and insufficient interpretability, making it difficult to achieve accurate representation and individualized assessment of multiple pathological mechanisms under remote, low-threshold conditions. To address the shortcomings of existing technologies, this invention provides a remote assessment system for mild cognitive impairment based on meta-learning and hybrid experts. A multimodal behavioral feature encoding module performs unified representation learning on multimodal data generated by subjects in remote digital cognitive tasks, including speech, text, behavioral trajectories, task performance, and covariates, outputting a multimodal disease-specific behavioral knowledge vector. A meta-learning pathological representation calculation module associates this multimodal disease-specific behavioral knowledge vector with multiple pre-set dominant pathological mechanisms, using multiple pre-trained pathological meta-learners to extract pathological meta-features corresponding to each pathological mechanism, combining them to form a static pathological meta-feature matrix. A pathological mechanism collaborative relationship modeling module uses each pathological meta-feature as a node, employing a graph attention network to learn and quantify the collaboration and dependence between different pathological mechanisms. The system generates a static pathological synergy matrix based on competitive relationships. This static matrix, along with the static pathological meta-feature matrix, forms a static memory layer, organizing scattered pathological mechanism knowledge into a structured knowledge layer containing mechanism features and their relationships. During the reasoning phase, the dynamic joint decision-making module calculates the matching degree between the subject's current multimodal disease-specific behavioral knowledge vector and each pathological meta-feature in the static memory layer. Using a sparse-gated Top-K strategy, it adaptively activates fewer than all pathological mechanism experts to participate in joint reasoning. The outputs of the activated experts are then weighted and fused to obtain a comprehensive decision result, ultimately generating the probability of mild cognitive impairment risk. The system also outputs a structured traceability report, recording key intermediate information such as the set of activated experts, gating weights, pathological mechanism synergy weights, and pathological activation values, providing a traceable pathological mechanism explanation for the evaluation results. In the preferred scheme, the pathological meta-learner adopts a prototype network architecture and introduces a supervised comparative learning mechanism and a historical feature library during the training phase to enhance the discriminativeness and stability of pathological meta-features. At the same time, an atypical pattern processing flow is set up. When the highest matching score is lower than a preset threshold, the corresponding sample is marked as awaiting expert review instead of being forcibly classified, ensuring the safety and reliability of the system in practical applications.

[0009] The specific technical solution adopted by this invention to solve its technical problem is as follows:

[0010] A remote assessment system for mild cognitive impairment based on meta-learning and hybrid experts includes:

[0011] The multimodal behavioral feature encoding module is used to perform unified representation learning on the multimodal behavioral data of the subjects and output multimodal disease-specific behavioral knowledge vectors;

[0012] The meta-learning pathological representation calculation module includes multiple pre-trained pathological meta-learners, each corresponding to a preset dominant pathological mechanism, used to convert the multimodal disease-specific behavioral knowledge vector into pathological meta-features of the corresponding pathological mechanism, and combine them to form a static pathological meta-feature matrix.

[0013] The pathological mechanism synergy modeling module is used to learn the synergy between different pathological mechanisms using each pathological element feature as a node, and outputs a static pathological synergy matrix.

[0014] The static memory layer is constructed jointly by the static pathological element feature matrix and the static pathological synergistic relationship matrix.

[0015] The dynamic joint decision-making module includes multiple pathological mechanism experts that correspond one-to-one with the preset dominant pathological mechanism. It is used to adaptively activate less than all of the pathological mechanism experts by means of a sparse gating Top-K strategy based on the matching degree between the multimodal disease behavior knowledge vector and the pathological element features in the static memory layer, and to perform weighted fusion on the output results of the activated experts to obtain a comprehensive decision result.

[0016] The risk output module is used to generate the risk probability of mild cognitive impairment based on the comprehensive decision results.

[0017] Furthermore, the system also includes a structured source tracing report generation module, which records key intermediate information during the model inference process and outputs a structured source tracing report containing the risk probability and its pathological mechanism explanation.

[0018] Furthermore, the pathological meta-learner employs a prototype network architecture. During meta-learning training, a support set and a query set are constructed for each pathological mechanism subtask. The prototype vector corresponding to the pathological mechanism is calculated using samples from the support set. Then, a loss function is constructed based on the distance or similarity between the query set samples and the prototype vector to train the corresponding pathological meta-learner. After training, each pathological meta-learner can map the input multimodal disease-specific behavioral knowledge vector to a pathological meta-feature space aligned with the corresponding pathological mechanism prototype.

[0019] Furthermore, the training labels for the pathological meta-learner are generated through at least one of the following methods: generating initial labels based on cognitive task rules, automatically labeling based on behavioral feature statistical rules, generating pseudo-labels based on unsupervised clustering, and correcting based on manual labeling. In specific implementation, initial labels can first be generated for each subject according to the preset mapping relationship between remote digital cognitive task types and pathological mechanisms. Then, the initial labels are automatically corrected based on the behavioral feature statistical thresholds of each task. For samples whose labels still cannot be determined, a clustering algorithm is used to group the multimodal behavioral features and assign pseudo-labels. Finally, a small number of samples are extracted for manual labeling by clinical experts to correct the rules and clustering parameters.

[0020] Furthermore, during training, the meta-learning pathological representation calculation module employs a supervised contrastive learning mechanism to train the pathological meta-learner. This optimizes pathological meta-features through intra-class aggregation of similar pathological mechanism samples and inter-class separation of dissimilar pathological mechanism samples. The module also incorporates a historical feature library, including both current batch and historical batch pathological meta-features within the contrastive learning scope. This historical feature library is updated using a first-in, first-out (FIFO) strategy. By introducing a historical feature library for cross-batch feature reuse, the problem of insufficient positive samples under small-batch training conditions can be effectively alleviated, enhancing the consistency and generalization ability of pathological meta-features in a larger sample space.

[0021] Furthermore, the pathological mechanism synergy modeling module employs a graph attention network. It calculates and normalizes the attention coefficients between nodes to obtain synergy weights. Based on these synergy weights, it performs weighted fusion of neighboring node features to update node representations. The synergy weights of all node pairs constitute the static pathological synergy matrix. Specifically, multiple pre-defined dominant pathological mechanisms are used as nodes in the graph structure. The initial feature of each node is its corresponding pathological meta-feature representation. For any two pathological mechanism nodes, the attention coefficients between them are calculated using a learnable linear transformation and attention weight vector. After normalization, the synergy weights representing the strength of the synergistic influence between pathological mechanisms are obtained. Through weighted fusion and updating of neighboring node features, node features not only retain their own pathological representation information but also incorporate contextual synergy information related to their pathological mechanisms.

[0022] Furthermore, the dynamic joint decision-making module is configured to trigger an atypical mode processing flow when the highest value among all matching scores is lower than a preset threshold, marking the corresponding sample as awaiting expert review or outputting a prompt message indicating that the current mode does not cover it. By setting the atypical mode processing flow, the forced classification of atypical or noisy samples that do not match any existing pathological mechanism can be avoided, ensuring the safety and reliability of the system in practical applications.

[0023] Furthermore, the number of the pre-defined dominant pathological mechanisms is greater than or equal to two, including one or more of the following: memory encoding disorder, memory consolidation or retrieval disorder, executive function disorder, language and semantic disorder, visuospatial disorder, and attention disorder. Each dominant pathological mechanism corresponds to a pathological meta-learner and a pathological mechanism expert network, thereby decoupling the multiple possible pathological mechanisms of mild cognitive impairment into independent computable knowledge units.

[0024] Furthermore, the multimodal behavioral data includes speech data, text data, behavioral trajectory data, task performance data, and covariate data. This multimodal behavioral data originates from remote digital cognitive tasks performed by the subjects, including one or more of the following: speech description tasks, delayed image recall tasks, reaction time judgment tasks, path tracing tasks, and sequential digit manipulation tasks. Specifically, speech description tasks correspond to language semantic features, delayed image recall tasks correspond to memory-related features, reaction time judgment tasks correspond to attention and executive function-related features, and path tracing tasks correspond to visuospatial processing-related features. The multimodal behavioral data undergoes preprocessing before encoding. This preprocessing includes noise reduction, silence removal, human voice detection, and adaptive gain control for speech data; word segmentation and semantic cleaning for text data; smoothing and outlier removal for behavioral trajectory data; and standardization for covariate data. Through the structured mapping relationship between the above tasks and pathological mechanisms, the system can associate behavioral data from different sources with corresponding pathological knowledge, providing a structured input foundation for subsequent pathological meta-feature extraction and dynamic joint reasoning.

[0025] Furthermore, in the static memory layer, the static pathological synergy relation matrix is ​​stored together with the static pathological meta-feature matrix as an adjacency matrix of a graph structure. Thus, the static memory layer not only stores the independent feature representations of each pathological mechanism, but also stores the synergistic, dependent, or competitive relationships between mechanisms in the form of a graph structure, constituting a pathological knowledge layer containing structured relational knowledge, providing a knowledge foundation for the dynamic joint decision-making module to perform relation-based reasoning.

[0026] Compared to existing technologies, this invention and its preferred solutions offer at least the following advantages: They achieve comprehensive integration and utilization of multimodal behavioral data, enabling a more complete characterization of the abnormal manifestations of mild cognitive impairment in different cognitive domains; they explicitly integrate clinical pathology knowledge into the model architecture, providing clear medical semantic support for feature representations and enhancing the model's generalization ability; they employ a dynamic expert scheduling mechanism, which adaptively adjusts the inference path based on individual cognitive impairment patterns, adapting to the significant heterogeneity of mild cognitive impairment; they achieve interpretability and traceability of assessment results through structured source reports, providing effective support for result review and clinical decision-making in medical settings; and the overall solution adopts a non-invasive remote assessment method, eliminating the need for expensive professional equipment and adapting to large-scale screening needs in various scenarios such as families, communities, and grassroots levels. Attached Figure Description

[0027] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0028] Figure 1 This is a system principle and architecture block diagram of an embodiment of the present invention. Detailed Implementation

[0029] To make the features and advantages of the present invention more apparent and understandable, specific embodiments are described below in detail:

[0030] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used in this specification have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0031] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0032] This invention provides a remote assessment system for mild cognitive impairment based on meta-learning and hybrid experts. Based on this system, users can input multimodal data, including the subject's voice, text, behavioral trajectories, task performance, and covariates, through remote digital cognitive tasks. This input data is then processed sequentially through a data preprocessing module, a multimodal behavioral feature encoding module, a meta-learning pathological representation calculation module, a pathological mechanism collaborative relationship modeling module, a static memory layer construction module, and a dynamic joint decision-making module. This process comprehensively processes the input multimodal behavioral data, outputting a mild cognitive impairment risk probability and a structured source tracing report. This enables remote, non-invasive, and interpretable auxiliary assessment of mild cognitive impairment, allowing for intelligent identification of mild cognitive impairment risk. Simultaneously, the structured source tracing report generation module records key intermediate results during the model's inference process and outputs the mild cognitive impairment risk probability and its interpretation.

[0033] The remote assessment system for mild cognitive impairment based on meta-learning and hybrid experts provided in this invention specifically includes the following modules:

[0034] The data acquisition module is used to collect multimodal behavioral data of subjects during remote digital cognitive tasks;

[0035] The data preprocessing module, which communicates with the data acquisition module, is used to clean, standardize, and perform quality control on multimodal behavioral data.

[0036] The multimodal behavior feature encoding module communicates with the data preprocessing module and is used to perform unified representation learning on the preprocessed multimodal behavior data, outputting multimodal disease-specific behavior knowledge vectors.

[0037] The meta-learning pathological representation calculation module communicates with the multimodal behavior feature encoding module. It is used to divide the multimodal disease-specific behavior knowledge vector into multiple pathological mechanism sub-tasks, train multiple pathological meta-learners, and output a static pathological meta-feature matrix.

[0038] The pathological mechanism synergy modeling module is connected to the meta-learning pathological characterization calculation module to learn the synergy between different pathological mechanisms and output a static pathological synergy matrix.

[0039] The static memory layer construction module communicates with the meta-learning pathological characterization calculation module and the pathological mechanism collaborative relationship modeling module, and is used to jointly construct the static memory layer from the static pathological meta-feature matrix and the static pathological collaborative relationship matrix.

[0040] The dynamic joint decision-making module communicates with the static memory layer construction module and is used to adaptively select the most relevant pathological mechanism experts to participate in joint reasoning based on the degree of matching between the subject's current behavioral state and the static memory layer.

[0041] The risk output module, which communicates with the dynamic joint decision-making module, is used to generate the risk probability of mild cognitive impairment.

[0042] The structured source tracing report generation module communicates with the dynamic joint decision-making module and the risk output module to record key intermediate information during the model inference process and output evaluation results and their interpretation reports.

[0043] In the aforementioned system, multimodal behavioral data specifically includes speech data, text data, behavioral trajectory data, task performance data, and covariate data; remote digital cognitive tasks include one or more of the following: speech description task, delayed image recall task, reaction time judgment task, path tracking task, and sequential digit manipulation task. Different remote digital cognitive tasks correspond to different cognitive domains and pathological mechanisms. Specifically, the speech description task is used to represent language semantic-related features, the delayed image recall task is used to represent memory-related features, the reaction time judgment task is used to represent attention and executive function-related features, and the path tracking task is used to represent visuospatial processing-related features.

[0044] The specific processing logic of the data preprocessing module is as follows: noise reduction and silence removal are performed on the speech data, and voice detection and adaptive gain control can be further performed; word segmentation, semantic cleaning and anomaly correction are performed on the text data; smoothing and outlier removal are performed on the behavioral trajectory data, and one or more of the following methods can be used: moving average, exponential moving average and outlier detection; standardization is performed on the covariate data.

[0045] The multimodal behavior feature encoding module is used to fuse and encode speech, text, behavior trajectory and covariate features to obtain multimodal disease-specific behavior knowledge vectors.

[0046] The specific implementation logic of the meta-learning pathological representation calculation module is as follows: Multiple pathological mechanism sub-tasks are constructed based on a set of pre-defined dominant pathological mechanisms. The number of dominant pathological mechanisms is K and K≥2. Dominant pathological mechanisms include one or more of the following: memory encoding impairment, memory consolidation or retrieval impairment, executive function impairment, language and semantic impairment, visuospatial impairment, and attention impairment. A pathological meta-learner is trained for each pathological mechanism sub-task, outputting the corresponding pathological meta-feature representation and combining them to form a static pathological meta-feature matrix. Pathological mechanism sub-task labels are generated through at least one of the following methods: generating initial labels based on cognitive task rules, automatically labeling based on behavioral feature statistical rules, generating pseudo-labels based on unsupervised clustering, and correcting based on a small amount of manual labeling. During the pre-training stage, this module employs a supervised comparative learning mechanism to improve the discriminativeness and stability of pathological meta-features. Furthermore, it combines a historical feature bank mechanism to achieve cross-batch sample expansion and information reuse. The length of the historical feature bank is a fixed value, preferably 512–2048, and it is updated using a first-in-first-out strategy to include both current batch features and historical batch features within the scope of supervised comparative learning.

[0047] The specific implementation logic of the pathological mechanism synergy modeling module is as follows: learn the synergy between different pathological mechanisms through graph attention mechanism to generate a static pathological synergy matrix; and further, perform weighted summation of the features of neighboring nodes based on the synergy weight to update the node feature representation.

[0048] The dynamic joint decision-making module comprises a dynamic decision scheduling layer, a sparse gating layer, and a joint decision-making layer. The sparse gating layer retains relevant pathological mechanism experts to participate in joint reasoning through a Top-K selection strategy. The number of pathological mechanism experts retained in the sparse gating layer is K', where K' < K, preferably 2 or 3. When the maximum matching score is lower than a preset threshold θ, an atypical pattern processing procedure is triggered. The threshold θ is preferably in the range of 0.2 to 0.5. The atypical pattern processing procedure means that when the matching score between the subject's behavior pattern and any existing pathological mechanism is lower than the threshold θ, the system marks the sample as awaiting expert review or outputs a prompt message that is not covered by the current pattern, without forcibly classifying it into any known pathological mechanism, thereby avoiding misclassification of atypical or noisy samples.

[0049] The risk output module generates the risk probability of mild cognitive impairment by weighted and fused the output results of experts on activated pathological mechanisms.

[0050] The report output by the structured source tracing report generation module includes at least one or more of the following: behavioral state coding, pathological mechanism relevance weight, pathological mechanism synergy weight, activation expert set, joint prediction weight, pathological activation value, and risk probability. The report may also include one or more of the following: dominant pathological mechanism name, inter-mechanism synergy strength, key behavioral evidence summary, and historical result comparison information. The report is output in JSON, table, or visualization panel format. This report can be used for physician review, historical record management, and subsequent follow-up reference.

[0051] This system is suitable for home self-testing, community screening, outpatient initial screening, and primary care follow-up scenarios.

[0052] Based on the above systems, such as Figure 1 As shown, this embodiment fully implements the solution of the present invention through the following steps:

[0053] S1. Collect multimodal behavioral data of subjects during remote digital cognitive tasks.

[0054] In this embodiment, multimodal behavioral data includes voice data, text data, behavioral trajectory data, task performance data, and covariate data.

[0055] Remote digital cognitive tasks include one or more of the following: speech description tasks, delayed image recall tasks, reaction time judgment tasks, path tracing tasks, and sequential digit manipulation tasks. Different remote digital cognitive tasks correspond to different cognitive domains and pathological mechanisms: speech description tasks collect language behavior features such as speech rate, pauses, and semantic coherence, corresponding to pathological mechanisms related to language and semantic disorders; delayed image recall tasks collect behavioral performance during encoding, consolidation, and retrieval, corresponding to pathological mechanisms related to memory encoding and retrieval disorders; reaction time judgment tasks collect reaction delay, error patterns, and attention switching features, corresponding to pathological mechanisms related to attention disorders and executive function disorders; and path tracing tasks collect path deviation, trajectory stability, and visuospatial processing features, corresponding to pathological mechanisms related to visuospatial disorders. Through this task-mechanism mapping relationship, the system can structurally associate remote behavioral data with pathological knowledge.

[0056] S2. The multimodal behavioral data is cleaned, standardized, and quality controlled through the data preprocessing module.

[0057] The specific processing steps include: noise reduction and silent segment removal for speech data, and further implementation of human voice detection and adaptive gain control. Human voice detection identifies and removes silent segments and background noise segments, retaining only valid speech content. Adaptive gain control dynamically adjusts the volume to reduce the impact of differences in the remote acquisition environment. Text data undergoes word segmentation, semantic cleaning, and anomaly correction. Behavioral trajectory data is smoothed and anomaly removal is performed. One or more of the following methods can be used: moving average, exponential moving average, and anomaly detection, to reduce trajectory jitter caused by equipment errors and remove data points with significantly abnormal speed, acceleration, or trajectory patterns. Covariate data is standardized.

[0058] S3. The preprocessed multimodal behavior data is subjected to unified representation learning through the multimodal behavior feature encoding module, and multimodal disease-specific behavior knowledge vectors are output.

[0059] The multimodal behavior feature encoding module maps speech features, text features, behavioral trajectory features, and covariate features from different sources and in different forms to a unified semantic space. The information from each modality is then fused using the following formula to obtain the multimodal disease-specific behavior knowledge vector F:

[0060]

[0061] in, It is a non-linear activation function; , , , These are the linear transformation weight matrices for speech, text, behavioral trajectory, and covariate modalities, respectively. , , , These are single-modal feature vectors for speech, text, behavioral trajectory, and covariate modality, respectively. This serves as a bias term. This knowledge vector not only preserves anomalous information across different modalities but also reflects the contextual relationships between various behavioral manifestations, providing a foundational input for subsequent pathological characterization calculations.

[0062] As a preferred implementation in this embodiment, the multimodal behavior feature encoding module adopts a Transformer-based shared encoder architecture. Specifically, for the speech modality, 1024-dimensional speech temporal features are first extracted using a pre-trained Wav2Vec2.0 base model, and then a fixed-length speech feature vector is obtained through average pooling. For the text modality, 768-dimensional text semantic features are extracted using a pre-trained BERT-base model. For the behavior trajectory modality, 256-dimensional trajectory dynamic features are extracted using a one-layer one-dimensional convolutional neural network (with a kernel size of 3 and a stride of 1) combined with a two-layer long short-term memory network (LSTM). For the covariate modality, a 128-dimensional feature vector is mapped through a one-layer fully connected layer. After concatenating the above four single-modal features, they are input into a two-layer Transformer encoder for unified representation learning. The Transformer encoder has 8 attention heads, a hidden layer dimension of 512, a feedforward network dimension of 2048, and uses GELU as the activation function. Finally, a 512-dimensional multimodal disease-specific behavior knowledge vector F is output. Those skilled in the art can make reasonable adjustments within the range of 1-4 layers of Transformer encoders, 4-16 attention heads, and 256-1024 dimensions of hidden layers, depending on the actual data scale and computing resources.

[0063] S4. Through the meta-learning pathological representation calculation module, the multimodal disease-specific behavioral knowledge vector is divided into multiple pathological mechanism sub-tasks, and multiple pathological meta-learners are trained to output a static pathological meta-feature matrix.

[0064] The meta-learning pathological representation computation module does not model all samples uniformly. Instead, it divides the multimodal disease-specific behavioral knowledge vector into multiple pathological mechanism sub-tasks based on a preset pathological mechanism. Each meta-learner learns only for a single dominant pathological mechanism, and extracts the pathological meta-feature representation corresponding to the i-th pathological mechanism using the following formula. And finally construct the static pathological element feature matrix M:

[0065]

[0066] in, is the pathological feature vector corresponding to the i-th pathological mechanism; This is the learner for the i-th pathological element; For multimodal disease-specific behavioral knowledge vectors; The total number of dominant pathological mechanisms, and This design enables the model to transform from "a single classifier" to a structured knowledge representation of "multiple pathological mechanism experts".

[0067] This step specifically includes the following sub-steps:

[0068] ① Construction of Pathological Mechanism Subtasks: Based on the pre-defined K dominant pathological mechanisms associated with mild cognitive impairment, K pathological mechanism subtasks are constructed. These dominant pathological mechanisms include, but are not limited to, memory encoding disorders, memory consolidation / retrieval disorders, executive function disorders, language and semantic disorders, visuospatial disorders, and attention disorders. The goal of each pathological mechanism subtask is to learn a representation function that can determine whether its corresponding pathological mechanism is dominant in the subject's behavioral data.

[0069] ② Generation of Pathological Mechanism Subtask Labels: Pathological mechanism subtask labels are generated through at least one of the following methods: generating initial labels based on cognitive task rules, automatically labeling based on behavioral feature statistical rules, grouping multimodal behavioral features using unsupervised clustering methods and generating pseudo-labels, and correcting with a small number of manually labeled samples to form semi-supervised training data. In a preferred embodiment, pathological mechanism subtask labels are generated jointly in the following order: generating initial labels for each subject based on the preset mapping relationship between remote digital cognitive task types and pathological mechanisms; automatically correcting the initial labels based on the behavioral feature statistical thresholds of each task (e.g., speech rate below X, reaction time coefficient of variation above Y); for samples where labels cannot be determined after the above steps, using the K-Means clustering algorithm to group multimodal behavioral features, aligning the cluster centers with typical pathological patterns, and assigning pseudo-labels; randomly selecting 5%–10% of the samples for manual labeling by clinical experts to correct rules and clustering parameters.

[0070] ③ Meta-learner training process: For each pathological mechanism subtask, a support set and a query set are constructed based on multimodal disease-specific behavioral knowledge vectors, and the corresponding pathological meta-learner is trained using a meta-learning framework. During training, the prototype vector or central representation corresponding to the pathological mechanism is first calculated using the support set samples, and then a loss function is constructed based on the distance or similarity between the query set samples and the prototype vector to update the pathological meta-learner parameters. In a preferred embodiment, the meta-learning pathological representation calculation module uses the supervised contrastive learning loss function SupCon Loss instead of the traditional triplet loss during the pre-training stage, and its expression is:

[0071]

[0072] in, The supervised contrastive learning loss for the nth sample; This represents the number of positive samples that belong to the same pathological mechanism category as sample n. This is the set of positive samples that belong to the same pathological mechanism category as sample n; , , They are sample n, positive sample p, and comparison sample, respectively. Normalized pathological features; the operator "·" is for vector dot product operation; This is a temperature coefficient used to adjust the distribution of similarity. It is the set of all comparable samples excluding sample n itself.

[0073] ④ Generation of static pathological meta-feature matrix: After training, for any new subject sample, its multimodal disease-specific behavioral knowledge vector F is input into all K pathological meta-learners. Each pathological meta-learner outputs the pathological meta-feature representation of the sample for the corresponding pathological mechanism. ,in This is a high-dimensional feature vector extracted by the corresponding pathomorphic learner and aligned with the prototype representation of the pathological mechanism. Finally, all pathomorphic feature representations are combined to form a static pathomorphic feature matrix:

[0074]

[0075] In a preferred implementation of this embodiment, each pathological meta-learner in the meta-learning pathological representation computation module is implemented using a 3-layer multilayer perceptron (MLP) architecture. Specifically, the first fully connected layer maps the input 512-dimensional multimodal disease-specific behavioral knowledge vector to 256-dimensional features, the second fully connected layer maps the 256-dimensional features to 128-dimensional features, and the third fully connected layer outputs a 64-dimensional pathological meta-feature vector. ReLU activation is used across all layers, and a Dropout layer (with a dropout rate of 0.2) is added to prevent overfitting. During prototype network training, the support set size for each pathological mechanism subtask is set to 5, and the query set size is set to 15. Euclidean distance is used to calculate the similarity between query samples and prototype vectors, and supervised contrastive learning loss is used for joint optimization. The length of the historical feature library is preferably set to 1024, and the temperature coefficient is [not specified]. The preferred setting is 0.07.

[0076] Furthermore, the meta-learning pathological characterization calculation module incorporates a historical feature database mechanism to achieve cross-batch sample expansion and information reuse. The length of the historical feature database is a fixed value, preferably 512–2048, and it is updated using a first-in-first-out (FIFO) strategy. After each round of training, the pathological meta-features of the current batch are added to the tail of the queue, and the earliest features to enter the queue are removed according to the FIFO strategy. During training, the features of the current batch and the historical features cached in the queue are included in the supervised comparative learning scope to expand the effective sample set, alleviate the problem of insufficient positive samples under small-batch training conditions, and enhance the consistency and generalization ability of pathological meta-features in a larger sample space.

[0077] S5. Through the pathological mechanism synergy modeling module, learn the synergistic relationships between different pathological mechanisms and output a static pathological synergy matrix.

[0078] Since mild cognitive impairment is not driven by a single pathological mechanism but is the result of the combined effects of multiple pathological mechanisms, the system further constructs a pathological mechanism synergy matrix A after obtaining the static pathological meta-feature matrix M. Through graph attention, the system can learn the influence strength between different pathological mechanisms, quantifying "which mechanisms are easily co-activated" and "which mechanisms have competitive relationships." This step specifically includes the following sub-steps:

[0079] ① Construction of the pathological mechanism relationship graph: K pathological mechanisms are used as K nodes in the graph structure, where the initial features of each node are its corresponding static pathological element feature vectors. .

[0080] ② Calculation of attention coefficient between nodes: For any two pathological mechanism nodes i and j, calculate the attention coefficient of node j to node i. The importance of pathological mechanism j to pathological mechanism i is characterized by the following formula:

[0081]

[0082] in, denoted as the original attention coefficient of pathological mechanism j to pathological mechanism i; The LeakyReLU nonlinear activation function is used. This represents the weight vector of a single-layer feedforward neural network. A shared linear transformation matrix used to unify the dimensions of pathological meta-features; , ...

[0083] ③ Collaborative relationship weight normalization: normalization of attention coefficients Normalization was performed to obtain the weights of the synergistic relationship of pathological mechanisms. The calculation formula is as follows:

[0084]

[0085] in, The normalized weight of the synergistic relationship of pathological mechanisms represents the degree of synergistic influence of pathological mechanism j on pathological mechanism i within the neighborhood of pathological mechanism node i. This is the softmax normalization operation performed along dimension j.

[0086] ④ Node feature update: based on collaborative relationship weights The feature representation of node i is updated by weighted summation of the features of neighboring nodes. The update formula is as follows:

[0087]

[0088] in, This is a feature representation of the updated pathological mechanism i; It is a non-linear activation function; Let i be the set of neighboring nodes of the pathological mechanism node i; For shared linear transformation matrices; Let be the initial meta-feature vector of pathological mechanism j. Through this update process, the features of pathological mechanism node i not only retain its own pathological characterization information, but also integrate contextual collaborative information with its related pathological mechanisms.

[0089] ⑤ Generation of static pathological synergy matrix: weighting the synergy relationships between all pathological mechanism nodes. By combining these elements, a static pathological synergy matrix A is formed, where the elements in matrix A are... , representing the strength of the synergistic influence of pathological mechanism j on pathological mechanism i. The static pathological synergy matrix is ​​used to describe the synergistic, dependent, or competitive relationships between different pathological mechanisms, and together with the static pathological meta-feature matrix, it constitutes the static memory layer required for subsequent dynamic joint decision-making.

[0090] In a preferred implementation of this embodiment, the pathological mechanism synergy modeling module is implemented using a single-layer graph attention network (GAT). Specifically, the GAT has 4 attention heads, each with an output dimension of 32. The updated node features are obtained through concatenation, and the activation function is LeakyReLU (with a negative slope of 0.2). Residual connections are added to enhance gradient propagation. When the number of dominant pathological mechanisms K=6, the generated static pathological synergy matrix A is a 6×6 square matrix. Those skilled in the art can reasonably adjust the GAT layer count to 1-3 layers and the number of attention heads to 2-8 layers, depending on the complexity of the pathological mechanisms.

[0091] S6. By constructing a static memory layer module, the static pathological meta-feature matrix and the static pathological synergistic relationship matrix are jointly constructed into a static memory layer.

[0092] The static memory layer comprises two core components: a pathological meta-feature matrix M, which stores the feature representations of each pathological mechanism; and a pathological synergy matrix A, which describes the strength of associations between different pathological mechanisms. The static memory layer, consisting of the static pathological meta-feature matrix M and the static pathological synergy matrix A, not only stores the features of individual mechanisms but also the relationships between mechanisms, providing a knowledge base for subsequent dynamic joint decision-making.

[0093] S7. Through the dynamic joint decision-making module, based on the degree of matching between the subject's current behavioral state and the static memory layer, the most relevant pathological mechanism experts are adaptively selected to participate in joint reasoning.

[0094] The dynamic joint decision-making module comprises a dynamic decision scheduling layer, a sparse gating layer, and a joint decision-making layer. It adaptively invokes the most relevant pathological mechanism experts to participate in reasoning based on the subject's current behavioral state, rather than having all experts participate in the calculation simultaneously. This step specifically includes the following sub-steps:

[0095] ① Dynamic Decision Scheduling Layer: Based on the subject's current multimodal disease-specific behavioral knowledge vector F, a learnable query network is used. Generate a behavior state query vector q, and then calculate the relationship between the query vector q and each pathological meta-feature in the static memory layer. Matching score between The matching score is used to characterize the degree of correlation between the current subject and the corresponding pathological mechanism, and its expression is:

[0096]

[0097]

[0098] in, The query vector represents the subject's current behavioral state; It is a learnable query network used to map multimodal disease-specific behavioral knowledge vectors to query vectors; For multimodal disease-specific behavioral knowledge vectors; The matching score between the subject's behavioral state and the i-th pathological mechanism; Let be the meta-feature vector of the i-th pathological mechanism; This is a vector dot product operation.

[0099] ②Sparse gated layer: based on matching score All pathological mechanism experts were ranked, and a Top-K selection strategy was used to retain the experts with the highest matching scores. One pathological mechanism expert, among whom Preferably, Choose 2 or 3. For the selected set of experts... Furthermore, softmax normalization is applied to the corresponding matching scores to obtain normalized gating weights. Its expression is:

[0100]

[0101] in, The contribution weight of the i-th activated pathology mechanism expert in this joint inference; To perform softmax normalization only on the matching scores within the set of activated experts; A collection of expert indexes on activated pathological mechanisms.

[0102] Furthermore, a matching score threshold can be set. When the highest matching score falls below a preset threshold, the "Other / Atypical Mode" processing flow is triggered. The preferred value range is 0.2 to 0.5. The atypical pattern processing procedure refers to: when the matching score between the subject's behavioral pattern and any existing pathological mechanism is below the threshold... When this happens, the system marks the sample as awaiting expert review or outputs a prompt message indicating that it is not covered by the current model, instead of forcibly classifying it into any known pathological mechanism, thereby avoiding misclassification of atypical or noisy samples.

[0103] ③ Joint decision-making layer: Only activate the index set The pathological mechanism expert network is used, and the multimodal disease-specific behavioral knowledge vector F of the current subject is input into these activated expert networks. The output results corresponding to the experts of each pathological mechanism were obtained. Its expression is:

[0104]

[0105] in, This represents the output of the i-th pathological mechanism expert network. For the i-th pathological mechanism expert network; This is a knowledge vector for multimodal disease-specific behaviors.

[0106] ④ Weighted fusion of expert outputs: The output results of experts on activated pathological mechanisms According to the corresponding normalized gating weights After weighted fusion, the final comprehensive decision vector P is obtained, and its expression is:

[0107]

[0108] in, This serves as the final comprehensive decision vector, representing the overall decision-making outcome of the current subject under the combined influence of multiple related pathological mechanisms, and acting as the basic input for generating the risk probability of mild cognitive impairment in the future. Let be the gating weight for the i-th activated expert; A collection of expert indexes on activated pathological mechanisms.

[0109] As a preferred implementation of this embodiment, the query network in the dynamic joint decision-making module A two-layer multilayer perceptron is used to map the input 512-dimensional multimodal disease-specific behavioral knowledge vector F to a 64-dimensional behavioral state query vector q. ReLU activation is used between layers. Each pathological mechanism expert network... All models employ a three-layer multilayer perceptron architecture. The first fully connected layer maps the 512-dimensional input to 256-dimensional features, the second fully connected layer maps the 256-dimensional features to 128-dimensional features, and the third fully connected layer outputs a one-dimensional log-odds probability of pathological mechanism risk. ReLU activation is used between layers, and a BatchNorm layer is added for normalization. Top-K selection parameters are used for the sparse gating layers. The preferred setting is 2, which is the matching score threshold. The preferred setting is 0.3.

[0110] S8. Generate the risk probability of mild cognitive impairment based on the output of the selected pathology mechanism expert.

[0111] The risk output module obtains the risk log odds by weighting and fusing the output results of the activated pathological mechanism experts; then, the risk log odds are mapped to the [0,1] interval by the Sigmoid function to obtain the risk probability p of mild cognitive impairment.

[0112] S9. Record key intermediate information during the model inference process through the structured traceability report generation module, and output the evaluation results and their interpretation report.

[0113] The structured source tracing report should include at least one or more of the following: behavioral state coding, pathological mechanism relevance weights, pathological mechanism synergy weights, activation expert set, joint prediction weights, pathological activation values, and risk probabilities. It may also further include one or more of the following: the name of the dominant pathological mechanism, inter-mechanism synergy strength, key behavioral evidence summaries, and historical outcome comparison information. The report should be output in JSON, table, or visualization panel format. This report can be used for physician review, historical record management, and subsequent follow-up reference.

[0114] Two further test and verification examples are provided below:

[0115] Example 1: Evaluation effect verification based on clinical simulation dataset

[0116] In one specific embodiment, a clinical simulation dataset is constructed, comprising samples of mild cognitive impairment and normal control samples. The dataset includes multimodal behavioral data of subjects in digital cognitive tasks such as verbal description, delayed image recall, reaction time judgment, and path tracking, and is labeled by clinical experts based on neuropsychological scale results and pathological knowledge tags. Preferably, the pathological mechanisms can be defined as categories such as memory encoding disorders, executive function disorders, language and semantic disorders, and attention disorders, and each sample is assigned a dominant pathological mechanism label or multi-label annotation information.

[0117] In this embodiment, the multimodal behavioral feature encoding module is implemented using a Transformer or an equivalent shared encoding structure; the meta-learning pathological representation module is pre-trained using a prototype network; the pathological mechanism collaborative relationship modeling module is implemented using a graph attention network; and the dynamic joint decision-making module includes a query network, a sparse gating layer, and an expert network. During training, the dataset can be divided into a training set, a validation set, and a test set, and the model can be trained using cross-entropy loss, prototype distance loss, or a combination thereof. Preferably, the model training hyperparameters are set as follows: the learning rate is... The batch size is 16–64, and the number of training rounds is 50–200. The meta-learner is trained using the Adam optimizer, and the number of layers in the graph attention network is preferably 1–3.

[0118] In a preferred embodiment, the system of the present invention was validated based on a clinical simulation dataset. Experimental results show that, in the task of identifying the risk of mild cognitive impairment, the system of the present invention outperforms control schemes that use only a single voice modality input or a unified classifier structure in terms of accuracy, recall, AUC, or other statistical indicators. Furthermore, the present invention can output activation expert sets, pathological mechanism co-weights, and structured source tracing reports, thereby improving the interpretability of results while ensuring assessment performance.

[0119] Example 2: Explainable Output Example

[0120] In another specific embodiment, for a given subject sample, the system outputs a structured source tracing report simultaneously with the probability of mild cognitive impairment risk. This report may display:

[0121] (1) A collection of experts on activated pathological mechanisms;

[0122] (2) Gating weights for each activated expert;

[0123] (3) The synergistic weights related to activating experts in the pathological mechanism synergistic relationship matrix;

[0124] (4) Pathological activation values ​​output by each expert;

[0125] (5) Final risk probability.

[0126] In a preferred embodiment, the structured source tracing report is output in JSON, tabular, or visual panel format, and includes at least the following fields: subject ID, task type, behavioral feature summary, set of activated experts, gating weights of each expert, synergistic weights of pathological mechanisms, pathological activation value, final risk probability, and comparison with historical results. For example, the system can output, "This high-risk judgment is mainly driven by the memory coding disorder mechanism and the executive function disorder mechanism, with the memory coding disorder mechanism having the highest gating weight and the executive function disorder mechanism exhibiting a high synergistic influence." In this way, doctors can not only obtain the risk result but also view the source of the corresponding pathological mechanism and its synergistic relationship, thereby improving the interpretability and credibility of the model in medical scenarios.

[0127] In summary, compared with existing technologies, the present invention achieves multimodal behavioral data collection through remote digital cognitive tasks, extending mild cognitive impairment screening from centralized in-hospital testing to family, community, and grassroots scenarios, thereby improving screening coverage and ease of application. Through meta-learning-based collaborative modeling of pathological representations and mechanisms, static pathological knowledge is transformed into callable and computable internal knowledge units, enhancing the ability to represent pathological mechanisms. By employing a hybrid expert and sparse gating strategy, dynamic expert invocation is performed based on the current performance of different subjects, improving adaptability to complex and heterogeneous samples. Finally, structured source reporting ensures that results are analyzable, traceable, and verifiable, enhancing the transparency and credibility of assessment results in healthcare settings.

[0128] Furthermore, a supervised comparative learning mechanism is introduced into the meta-learning pathological representation calculation stage. This mechanism aggregates samples of similar pathological mechanisms within a class and separates samples of dissimilar pathological mechanisms between classes. It also combines this with a historical feature database to achieve cross-batch sample expansion and reuse of historical features, thereby improving the stability, clarity, and discriminative power of pathological meta-features. This approach provides a higher-quality static pathological representation foundation for subsequent modeling of collaborative relationships among pathological mechanisms and dynamic joint decision-making by hybrid experts.

[0129] It should be noted that the computer-based data processing system provided by this invention outputs results only for the auxiliary assessment of cognitive impairment risk and is not directly used for disease diagnosis or treatment decisions.

[0130] Those skilled in the art will understand that embodiments of this application can be provided as systems, apparatus, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code.

[0131] This application is described with reference to flowchart illustrations and / or block diagrams of systems, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0132] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0133] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0134] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A remote assessment system for mild cognitive impairment based on meta-learning and hybrid experts, characterized in that, include: The multimodal behavioral feature encoding module is used to perform unified representation learning on the multimodal behavioral data of the subjects and output multimodal disease-specific behavioral knowledge vectors; The meta-learning pathological representation calculation module includes multiple pre-trained pathological meta-learners, each corresponding to a preset dominant pathological mechanism, used to convert the multimodal disease-specific behavioral knowledge vector into pathological meta-features of the corresponding pathological mechanism, and combine them to form a static pathological meta-feature matrix. The pathological mechanism synergy modeling module is used to learn the synergy between different pathological mechanisms using each pathological element feature as a node, and outputs a static pathological synergy matrix. The static memory layer is constructed jointly by the static pathological element feature matrix and the static pathological synergistic relationship matrix. The dynamic joint decision-making module includes multiple pathological mechanism experts that correspond one-to-one with the preset dominant pathological mechanism. It is used to adaptively activate less than all of the pathological mechanism experts by means of a sparse gating Top-K strategy based on the matching degree between the multimodal disease behavior knowledge vector and the pathological element features in the static memory layer, and to perform weighted fusion on the output results of the activated experts to obtain a comprehensive decision result. The risk output module is used to generate the risk probability of mild cognitive impairment based on the comprehensive decision results.

2. The remote assessment system for mild cognitive impairment based on meta-learning and hybrid experts according to claim 1, characterized in that: It also includes a structured source tracing report generation module, which records key intermediate information during the model reasoning process and outputs a structured source tracing report containing the risk probability and the explanation of its pathological mechanism.

3. The remote assessment system for mild cognitive impairment based on meta-learning and hybrid experts according to claim 1, characterized in that: The pathological meta-learner adopts a prototype network architecture.

4. The remote assessment system for mild cognitive impairment based on meta-learning and hybrid experts according to claim 1, characterized in that: The training labels for the pathological meta-learner are generated in at least one of the following ways: generating initial labels based on cognitive task rules, automatically labeling based on behavioral feature statistical rules, generating pseudo-labels based on unsupervised clustering, or correcting based on manual labeling.

5. The remote assessment system for mild cognitive impairment based on meta-learning and hybrid experts according to claim 1, characterized in that: During training, the meta-learning pathological representation calculation module employs a supervised comparative learning mechanism to train the pathological meta-learner. It optimizes pathological meta-features by aggregating samples of the same type of pathological mechanism within each class and separating samples of different types of pathological mechanisms between each class. The meta-learning pathological representation calculation module also incorporates a historical feature library, including the current batch of pathological meta-features and historical batches of pathological meta-features into the comparative learning scope. The historical feature library is updated using a first-in, first-out (FIFO) strategy.

6. The remote assessment system for mild cognitive impairment based on meta-learning and hybrid experts according to claim 1, characterized in that: The pathological mechanism synergy modeling module adopts a graph attention network. It calculates and normalizes the attention coefficients between nodes to obtain the synergy weights. Based on the synergy weights, it performs weighted fusion of the features of neighboring nodes to update the node representation. The synergy weights of all node pairs constitute the static pathological synergy matrix.

7. The remote assessment system for mild cognitive impairment based on meta-learning and hybrid experts according to claim 1, characterized in that: The dynamic joint decision-making module is also configured to: when the highest value among all matching scores is lower than a preset threshold, trigger the atypical mode processing flow, mark the corresponding sample as pending expert review or output a prompt message that is not covered by the current mode.

8. The remote assessment system for mild cognitive impairment based on meta-learning and hybrid experts according to claim 1, characterized in that: The number of the pre-defined dominant pathological mechanisms is greater than or equal to 2, including one or more of the following: memory encoding disorder, memory consolidation or retrieval disorder, executive function disorder, language and semantic disorder, visuospatial disorder, and attention disorder.

9. The remote assessment system for mild cognitive impairment based on meta-learning and hybrid experts according to claim 1, characterized in that: The multimodal behavioral data includes speech data, text data, behavioral trajectory data, task performance data, and covariate data. The multimodal behavioral data originates from remote digital cognitive tasks performed by the subjects. These remote digital cognitive tasks include one or more of the following: speech description task, delayed image recall task, reaction time judgment task, path tracking task, and sequential digit manipulation task. Among these, the speech description task corresponds to language semantic-related features, the delayed image recall task corresponds to memory-related features, the reaction time judgment task corresponds to attention and executive function-related features, and the path tracking task corresponds to visuospatial processing-related features. The multimodal behavioral data undergoes preprocessing before encoding. The preprocessing includes noise reduction, silence removal, human voice detection, and adaptive gain control for speech data; word segmentation and semantic cleaning for text data; smoothing and outlier removal for behavioral trajectory data; and standardization for covariate data.

10. The remote assessment system for mild cognitive impairment based on meta-learning and hybrid experts according to claim 1, characterized in that: In the static memory layer, the static pathological synergy relation matrix is ​​stored together with the static pathological meta-feature matrix as an adjacency matrix of a graph structure.

Images