Construction method of brain rehabilitation cognitive training effect prediction model

Abstract

The invention provides a construction method of a brain rehabilitation cognitive training effect prediction model. The method improves the accuracy and individuation level of cognitive training effect prediction. Firstly, state data of a subject is obtained and preprocessed; thirdly, obtaining unified features by adopting an extraction layer, carrying out combined modeling of a spatial topological relation and dynamic time sequence evolution by adopting a fusion layer based on the unified features and generating spatio-temporal features, and adjusting model parameters by adopting an adaptation layer according to baseline features of a subject to ensure personalized prediction; and finally, in a model operation process, triggering online updating through concept drift detection, and generating a confidence interval and an interpretability score of a prediction result based on Bayesian reasoning. According to the method, through state data processing, personalized adaptation, dynamic adaptation and Bayesian reasoning, the precision, stability and interpretability of brain rehabilitation cognitive training effect prediction are remarkably improved, and powerful support is provided for personalized treatment and clinical decision making.

Description

Technical Field

[0001] This invention relates to the field of model building, specifically a method for constructing a model for predicting the effect of cognitive training in brain rehabilitation. Background Art

[0002] Cognitive training for brain rehabilitation, as a treatment method, aims to help patients with brain injuries restore brain function through cognitive training. In recent years, with the continuous development of brain science, neuroscience and artificial intelligence technology, the evaluation and prediction of the effects of cognitive training for brain rehabilitation has gradually become a hot topic in academic and clinical medical research. Traditional techniques assess cognitive status through the subject's condition, thereby better customizing rehabilitation training plans and improving treatment effects. However, traditional techniques mostly rely on a single data modality and usually assume that the training effect is constant, failing to fully consider the individual differences of the subjects and lacking the assessment of the uncertainty of the prediction results.

[0003] Existing technologies combine multiple data sources to evaluate the effectiveness of cognitive training, which can more comprehensively reflect the cognitive state of subjects and improve the accuracy of predictions compared to single-modal data. However, existing technologies have many shortcomings, such as failing to fully consider the individual differences of subjects and usually lacking a mechanism for dynamic adjustment based on individual characteristics. Furthermore, existing models fail to quantify the uncertainty of prediction results and do not provide sufficient interpretability for model results, resulting in a lack of transparency in the practical application of prediction results. Summary of the Invention

[0004] To address the technical problems mentioned in the background section, this invention proposes a method for constructing a predictive model for the effects of cognitive training in brain rehabilitation.

[0005] Therefore, the technical solution adopted by the present invention is as follows: A method for constructing a predictive model for the effectiveness of cognitive training in brain rehabilitation, comprising: S1: Acquire the subject's status data and preprocess the status data.

[0006] S2: Construct a prediction model comprising an extraction layer, a fusion layer, and an adaptation layer; the extraction layer extracts corresponding deep features using different methods based on the modal characteristics of the preprocessed state data, and aligns the deep features using a cross-modal alignment algorithm to obtain unified features; the fusion layer performs joint modeling of spatial topological relationships and dynamic temporal evolution based on the unified features, and finally generates spatiotemporal features; the adaptation layer dynamically adjusts the model parameters based on the baseline characteristics of the subjects.

[0007] S3: Train the prediction model. During the model's operation after training, perform concept drift detection on the newly input state data. When concept drift is detected, trigger online updates and generate confidence intervals and importance scores for the prediction results based on Bayes' theorem.

[0008] Furthermore, the state data includes electroencephalogram (EEG) signals. Near-infrared functional imaging signals Eye movement and kinematic signals and .

[0009] Furthermore, the method for extracting the depth features specifically includes: Based on the modal characteristics of EEG signals in the preprocessed state data, depth features of the EEG signals are extracted using a convolutional neural network. ; Based on the modal characteristics of near-infrared functional imaging signals in the preprocessed state data, depth features of the near-infrared functional imaging signals are extracted using a convolutional neural network combined with a long short-term memory network. ; Based on the modal characteristics of eye movement and kinematic signals in the preprocessed state data, depth features of the eye movement and kinematic signals are extracted using a long short-term memory network. and .

[0010] Furthermore, the specific steps for generating the spatiotemporal features are as follows: 1) Dynamically construct a brain region functional connectivity map, represented as follows: ; in, Indicates the first The node feature matrix of the layer; Indicates the first The node feature matrix of the layer; This indicates the functional connectivity between brain regions. , Represents the adjacency matrix. Represents the identity matrix; Degree matrix; Indicates the first Layer weight matrix; This represents the activation function; through graph convolution operations, it outputs spatial features. ; 2) The temporal features with uniform characteristics are weighted and processed using a temporal attention network, and are represented as follows: ; in, This represents the score at the current time step; A function representing a temporal attention network; This represents the uniform features at the current time step; the score at each time step is converted into weights using the softmax function. The uniform features at each time step are weighted using weights, as shown below: ; in, This represents the temporal features obtained through weighted fusion using a temporal attention network; ultimately, spatiotemporal features are obtained.

[0011] Furthermore, the adaptation layer inputs the subject's baseline features into the multilayer perceptron to obtain adaptation parameters, expressed as follows: ; in, It is a multilayer perceptron network; This represents the adaptation parameters; the personalized adaptation parameters are added to the original model parameters to obtain the adjusted model parameters.

[0012] Furthermore, the prediction model is trained using a multi-task learning framework, defining a short-term prediction branch and a long-term prediction branch, and obtaining short-term and long-term prediction results through the two branches. Define the total loss function as follows: ; in, and They represent the first Short-term and long-term predictive effects for individual subjects, and They represent the first The actual short-term and long-term effects on each subject; This indicates the number of participants.

[0013] Furthermore, based on the KL divergence test method, the concept drift detection is performed, expressed as: ; in, This represents each input data point used in calculating the KL divergence; Represents the probability under the true distribution; This represents the probability under the predicted distribution; when the KL divergence exceeds a set threshold, it is determined that the newly input state data has undergone concept drift, triggering the online update mechanism, as shown below: ; in, This represents the updated model parameters; Indicates the current model parameters; Indicates the learning rate; This represents the current gradient.

[0014] Furthermore, based on Bayes' theorem, and combining the prior distribution and the likelihood function, the posterior distribution of the model parameters is calculated, expressed as: ; in, Represents the likelihood function; Represent the prior distribution; Indicates marginal likelihood; Represents the posterior distribution of the model parameters; This represents the input status data; Represent the prediction result; generate the confidence interval of the prediction result from the posterior distribution; Calculate the importance score of each state data point to the prediction result, expressed as: ; in, Indicates the first The importance score of each state data feature to the prediction result; Indicates the first The gradient of the features of each state data.

[0015] Compared with the prior art, the advantages of the present invention are: 1. The prediction model of the present invention fully integrates multimodal state data through extraction layer, fusion layer and adaptation layer. The extraction layer extracts the deep features of each state data through convolutional neural network (CNN) and long short-term memory network (LSTM). The fusion layer jointly models the model through graph neural network (GCN) and temporal attention network (TAN), which greatly enhances the expressive power and prediction accuracy of the model.

[0016] 2. The adaptation layer of the prediction model of the present invention dynamically adjusts the model parameters according to the baseline characteristics of each subject. This enables the prediction model to take into account the individual differences of the subjects and provide personalized training effect prediction, thereby significantly improving the accuracy and reliability of the prediction.

[0017] 3. This invention introduces a concept drift detection mechanism, which detects changes in data distribution through KL divergence and triggers online updates when concept drift occurs, ensuring that the model can continuously adapt to new data distributions and avoid the decay of model prediction accuracy.

[0018] 4. This invention uses Bayesian inference to generate confidence intervals and importance scores for prediction results, providing uncertainty assessment and interpretability for the model's prediction results, and improving the model's transparency and reliability. This is especially important for clinical applications, as it can provide doctors with more credible and evidence-based prediction results. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart of the method for constructing the prediction model of the present invention; Figure 2 This is a flowchart of step one of the present invention; Figure 3 This is a flowchart of step two of the present invention. Detailed Implementation

[0021] To achieve the above objectives, this invention provides a method for constructing a predictive model for the effect of cognitive training in brain rehabilitation. Please refer to [link to relevant documentation]. Figures 1-3 The method includes: S1: Acquire the subject's state data and preprocess the state data. The purpose of this step is to build a structured and standardized sample set to provide high-quality input for the training of subsequent prediction models; In this embodiment, the subject's electroencephalogram (EEG) signals were acquired using a 32-channel scalp EEG acquisition system (sampling rate 500Hz, electrode layout conforming to the international 10-20 system); the EEG signals are represented as follows: It reflects the electrical activity of brain neuron populations and can capture neural dynamic processes at the millisecond level; The subjects' near-infrared functional imaging signals (FNIRS) were acquired using a 16-channel near-infrared functional imaging device at a sampling rate of 50 Hz, capturing changes in blood oxygen dynamics in the cerebral cortex. The near-infrared functional imaging signals are represented as follows: It reflects the coupling mechanism between neural activity and blood flow and is an important indicator of cognitive load and rehabilitation effect; Eye movement signals were acquired by an eye tracker at a sampling rate of 120 Hz and included fixation point coordinates, saccade path, and pupil diameter; the eye movement signals were represented as follows: Further extraction of features such as fixation duration, saccade amplitude, and pupillary changes can reveal the subject's attention allocation and information processing patterns. Patients with impaired cognitive function often exhibit abnormal fixation patterns. The subjects' kinematic signals were recorded using a wearable inertial sensor at a sampling rate of 120 Hz, including triaxial acceleration and triaxial angular velocity; eye movement signals were represented as follows: This characterizes the subjects' motor execution ability and training compliance; The subject's state data consist of electroencephalogram (EEG) signals, near-infrared functional imaging (NIR) signals, and eye movement and kinematic signals. , is represented as: ; After acquiring the subjects' state data, the data is preprocessed, including noise removal, time alignment, and standardization. First, appropriate noise removal methods are employed for state data of different modalities to ensure signal quality and accuracy. In this embodiment, the electroencephalogram (EEG) signal is processed using Independent Component Analysis (ICA) to remove artifacts from electrooculography (EOG) and electromyography (EMG). ICA removes unwanted artifact signals by decomposing the original EEG signal into multiple independent components. The near-infrared functional imaging signal (fNIRS) signal is processed using wavelet filtering to remove low-frequency baseline drift and high-frequency noise. Wavelet filtering effectively removes irrelevant noise by decomposing the signal into different frequency subbands. The eye movement and kinematic signals are processed for noise removal using median filtering and outlier removal. Median filtering smooths the data and eliminates sharp outliers, ensuring stable and reliable signals. Secondly, since the sampling rate and timestamps of the state data are different, time alignment processing is required to ensure that all state data are analyzed within the same time window. In this embodiment, for EEG signals and near-infrared functional imaging signals, an event-triggered synchronization method is first used to ensure that all data are synchronized at the start and end times of the critical task. Secondly, the dynamic time warping (DTW) method is used to finely align signals with time deviations. DTW calculates the minimum distance between two time series and automatically adjusts their time axes so that state data of different modalities can be compared at the same time point. Finally, to ensure the comparability of state data between different modalities, standardization is required. In this embodiment, the zero-mean unit variance standardization method is used to calculate the mean and standard deviation of the state data for each modality and then standardize it. This ensures that the state data of all modalities can be trained and analyzed under the same standard.

[0022] S2: Construct a prediction model comprising an extraction layer, a fusion layer, and an adaptation layer. The extraction layer extracts corresponding deep features using different methods based on the modal characteristics of the preprocessed state data, and aligns these deep features using a cross-modal alignment algorithm to obtain unified features. The fusion layer, based on the unified features, performs joint modeling of spatial topological relationships and dynamic temporal evolution, ultimately generating spatiotemporal features. The adaptation layer dynamically adjusts the model parameters based on the baseline characteristics of the subjects. In this step, based on the preprocessed state data from step one, a multi-layered prediction model architecture is constructed. This model includes an extraction layer, a fusion layer, and an adaptation layer. The core task of the extraction layer is to extract deep feature representations from the preprocessed state data. Since the characteristics of state data differ for each modality, the extraction layer employs a state data feature extraction method tailored to each modality to obtain representative features and ensure that state data from different modalities can be processed within a unified feature space. Electroencephalogram (EEG) signals contain significant time and frequency domain features, therefore feature extraction is performed using a convolutional neural network (CNN). CNNs possess powerful local pattern capture capabilities, enabling them to identify temporal features and frequency patterns within the signal. The features extracted by a CNN are represented as follows: ; in, This represents the function used for feature extraction in a convolutional neural network (CNN). This represents the deep feature representation of the extracted EEG signal; through convolution operations, the local temporal pattern of the EEG signal can be extracted, and further, the feature dimension can be reduced by pooling layers to retain the most significant feature information; Near-infrared functional imaging signals reflect changes in brain blood oxygen concentration and exhibit temporal characteristics. Therefore, a convolutional neural network (CNN) is used for feature learning of temporal patterns, and a long short-term memory network (LSTM) is used to capture long-term dependencies in order to extract deep features from the near-infrared functional imaging signals. ; Both eye movement and kinematic signals exhibit strong temporal characteristics. Therefore, Long Short-Term Memory (LSTM) networks are used to extract temporal features. LSTM excels at capturing short-term and long-term dependencies in time-series data and can effectively extract depth features from subjects' eye movement and kinematic signals. and Specifically, this includes the subject's gaze patterns and cognitive state, as well as the subject's movement patterns, such as the subject's motor stability and reaction speed. After the extraction layer, the deep feature representation of the subject state data was extracted and recorded as follows: ; To jointly model these different modal deep feature representations, a cross-modal alignment algorithm is used to map all deep feature representations into a unified feature space. In this embodiment, the deep feature representations of different modalities are aligned to a unified feature space using dynamic time warping (DTW) technology, resulting in unified features. This provides a shared feature space for subsequent spatiotemporal modeling and multimodal fusion, enabling joint modeling of deep feature representations of different modalities.

[0023] The core task of the fusion layer is to jointly model the spatial topological relationships and dynamic temporal evolution of brain region functions based on unified features through graph neural networks (GCN) and temporal attention networks (TAN), thereby effectively capturing spatiotemporal information between different modalities. In this way, the fusion layer can not only capture the spatial relationships between brain regions, but also model the temporal changes of multimodal state data, thereby improving the expressive power of the model. The first step in the fusion layer is to use a Graph Convolutional Network (GCN) to model the spatial topological relationships between brain regions. The brain region functional connectivity map is constructed using depth features from EEG and near-infrared functional imaging (NIFI) signals. These depth features are used to calculate the correlation and synchronicity between brain regions, thus building the brain region functional connectivity map. Nodes in the map represent brain regions, and edges represent the functional connectivity strength between brain regions. Through graph convolution operations, GCN can capture the interdependencies between different brain regions. Specifically, The unified features include depth features from EEG and near-infrared functional imaging signals. These features, represented by GCN, are visualized in brain region functional connectivity maps. Aggregate learning is performed on nodes. Represents brain regions, sides Representing functional connectivity between brain regions, GCN extracts spatial features through graph convolution operations, as shown below: ; in, Indicates the first The node feature matrix of a layer, that is, the feature representation of each node in the current layer; Indicates the first The node feature matrix of the layer is the updated node feature representation in the next layer; in the fusion layer, and This represents the feature representation obtained from the extraction layer and the fusion layer, so The deep feature representations of EEG signals and near-infrared functional imaging signals are represented in the unified feature representation obtained from the extraction layer, and these features are then updated in the graph convolution operation; , representing the adjacency matrix of the graph Add the identity matrix The adjacency matrix describes the connections between nodes in the graph, and the identity matrix ensures that each node is also connected to itself; in the fusion layer, The functional connectivity relationships between brain regions are represented by correlation calculations based on depth features of EEG signals and near-infrared functional imaging signals. The degree matrix represents the connectivity of each node. In graph convolution operations, the degree matrix is ​​used to normalize the adjacency matrix to balance the impact of the number of nodes on information propagation. Indicates the first The weight matrix of the layer transforms the node features and is usually optimized through training. In the fusion layer, its role is to adjust the extracted features to ensure the effective fusion of spatial and temporal information. This represents the activation function; through GCN, spatial information of brain region functional connectivity will be effectively extracted, and spatial features will be output. ; The second step of the fusion layer is to use a Temporal Attention Network (TAN) to weight and fuse unified features from different time steps, capturing dynamic changes in the time-series data. Among all the unified features, the deep features of different modalities (EEG, fNIRS, eye tracking, and kinematic signals) can reflect dynamic changes at different levels. The TAN can assign corresponding weights based on the contribution of each time step to the final prediction. The unified feature representation of the current time step is input into the temporal attention network, and a score is output. The importance of the current time step is indicated by the formula: ; in, The function representing the time attention network learns to generate scores associated with each time step; The unified feature representing the current time step, After obtaining the score at each time step, the softmax function is used to convert it into weights. This ensures that the sum of the weights for all time steps is 1, and the model can weight the weights according to the importance of each time step, as shown below: ; in, This indicates the weight of the current time step, i.e., the importance of that time step to the final prediction; This means normalizing the scores over all time steps to ensure that the sum of the weights is 1. The uniform features at each time step are weighted using weights, as follows: ; in, This represents the temporal features after weighted fusion through the Temporal Attention Network (TAN). Through this weighted fusion method, TAN can weight and synthesize the deep feature representations of each time step, thereby highlighting those time periods that have a greater impact on the prediction results and reducing the interference of irrelevant time periods. Finally, spatial and temporal features are dynamically fused through gating units to obtain spatiotemporal features, which will be used for subsequent prediction tasks.

[0024] The task of the adaptation layer is to personalize the model based on the baseline characteristics of the subjects (such as age, years of education, and disease duration), thereby enabling the model to adapt to the characteristics of different subjects and improving the model's accuracy and robustness. The adaptation layer will incorporate the baseline characteristics of the subjects. The data is input into a multilayer perceptron (MLP). Baseline features include the subject's personal information, clinical data, or other factors related to cognitive ability. Through these baseline features, the model can understand the individual differences of each subject and adjust the model parameters accordingly. MLP is a fully connected network consisting of multiple hidden layers and activation functions. It effectively captures the complex relationship between baseline features and model parameters. Specifically, MLP transforms the subject's baseline features into personalized adaptation parameters, which are then applied to the output of the fusion layer. The generation of these adaptation parameters is represented as follows: ; in, It is a multilayer perceptron network that calculates the fitting parameters based on the baseline characteristics of the subjects; This represents the personalized adaptation parameters generated for each subject, reflecting the impact of baseline characteristics on model adjustment; The generated personalized adaptation parameters are added to the original model parameters to obtain the personalized adjusted model parameters. This adjustment dynamically adjusts the output of the fusion layer, enabling the model to be optimized for the specific circumstances of each subject and improve the model's predictive performance at the individual level.

[0025] S3: Train the prediction model. During the model's operation after training, perform concept drift detection on new input state data. When concept drift is detected, trigger online updates and generate confidence intervals and importance scores for the prediction results based on Bayes' theorem. The goal of model training is to learn and predict the short-term and long-term cognitive training effects of subjects through spatiotemporal features. Short-term effects reflect the initial changes after training, while long-term effects reflect the stability and persistence of the training results. Specifically, The model is trained using supervised learning and outputs predictions for short-term and long-term effects. In this embodiment, short-term effects are cognitive improvement scores after one week to one month, and long-term effects are cognitive improvement scores after three to six months. Employing a multi-task learning (MTL) framework, this model simultaneously predicts short-term and long-term effects. Two branches are defined in the model: a short-term prediction branch and a long-term prediction branch. The short-term prediction branch maps spatiotemporal features to short-term effect predictions via a fully connected layer, while the long-term prediction branch maps to long-term effect predictions via another fully connected layer. To jointly optimize short-term and long-term forecasts, a total loss function is defined as follows: ; in, and They represent the first Short-term and long-term predictive effects for individual subjects, and They represent the first The actual short-term and long-term effects on each subject; This represents the number of subjects; by minimizing the loss function, the model parameters are optimized to obtain the best predictive performance. After training, the model can predict short-term and long-term cognitive training effects based on the subjects' input data. The model generates short-term and long-term effects through a forward propagation process as estimates of the cognitive training effect.

[0026] Because cognitive rehabilitation training is influenced by various factors, the distribution of subjects' state data may change over time, during training, and due to individual differences, leading to a decline in predictive effectiveness over long-term use. This step uses KL divergence to dynamically monitor the distribution of subjects' state data, enabling timely detection of changes in data patterns and concept drift. This triggers an online update mechanism to adjust model parameters, thereby improving the model's stability and accuracy in long-term predictions. It significantly enhances the model's adaptability to individual differences and temporal changes, ensuring accurate prediction of cognitive training effects. The distributional difference between the new state data and the training data is calculated using the Kullback-Leibler (KL) divergence test, and expressed as follows: ; in, This represents each data point considered when calculating the KL divergence; it can be a discrete category or a continuous value. This represents the probability under the true distribution, that is, in reality, The probability of taking a certain value; This represents the probability under the predicted distribution, that is, the probability predicted by the model based on the input data. When the KL divergence exceeds the set threshold, the model will determine that the new input state data has undergone concept drift, triggering the online update mechanism so that the model can adapt to the new data distribution. The online update mechanism is implemented through incremental learning, and is represented as follows: ; in, Indicates the current model parameters. Indicates the learning rate; This represents the current gradient, which is the partial derivative of the loss function with respect to the model parameters; this method ensures that the model can be gradually adjusted according to new data distributions to cope with concept drift.

[0027] Bayes' theorem is used to provide a confidence interval for each prediction, offering an assessment of the uncertainty of the prediction results and making the model's output more reliable. Specifically, Based on Bayes' theorem, and combining the prior distribution and the likelihood function, the posterior distribution of the model parameters is calculated, and expressed as: ; in, This represents the likelihood function, which is the expression given the input data. and parameters Prediction results The probability of; This represents the prior distribution, i.e., the initial assumptions about the model parameters; This represents the marginal likelihood, a normalized constant. The posterior distribution of the parameters is represented; the confidence interval of the prediction result is generated from the posterior distribution; in this embodiment, the 95% confidence interval is calculated using the percentile method or the Monte Carlo method, and is expressed as: ; in, This represents the expected value of the predicted value; This represents the width of the confidence interval calculated using the posterior distribution; Finally, the impact of each state data point on the final prediction result is calculated using the posterior distribution. Specifically, the importance of the features of each state data point is evaluated by calculating the gradient or influence degree, as follows: ; in, Indicates the first The importance score of each state data feature to the prediction result; Indicates the first The gradient of features of each state data point; importance scores can explain which input state data features have the greatest impact on the model's final prediction results; After Bayes' theorem and online updates, the model can output predictions of short-term and long-term cognitive training effects, and provide confidence intervals and importance scores for each prediction. Through these outputs, the model can not only provide accurate predictions of cognitive training effects, but also provide decision-makers with detailed information on the credibility of the predictions and the influence of features.

[0028] This invention proposes a method for constructing a predictive model for the effectiveness of cognitive training in brain rehabilitation. The method aims to improve the predictive accuracy of brain rehabilitation training through multimodal data fusion and personalized adaptation. The method includes acquiring the subject's state data and generating a sample set through preprocessing. The core components of the model include an extraction layer, a fusion layer, and an adaptation layer. The extraction layer extracts deep features from different modalities using convolutional neural networks and long short-term memory networks. The fusion layer jointly models spatiotemporal features using graph neural networks and temporal attention networks. The adaptation layer adjusts the model parameters based on the subject's baseline characteristics to ensure personalized prediction. Finally, confidence intervals and importance scores for the prediction results are generated based on Bayesian inference to provide an assessment of the uncertainty and interpretability of the prediction results.

[0029] In summary, this invention significantly improves the accuracy and stability of predicting the effects of cognitive training in brain rehabilitation by processing state data, introducing a personalized adaptation layer, and implementing concept drift detection and online updates. Simultaneously, by providing confidence intervals and interpretability scores for the prediction results through Bayesian inference, it enhances the transparency and credibility of the model, making it particularly suitable for personalized treatment and decision support in clinical medicine.

[0030] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for constructing a predictive model for the effect of cognitive training in brain rehabilitation, characterized in that, The method includes: S1: Acquire the subject's status data and preprocess the status data simultaneously; S2: Construct a prediction model comprising an extraction layer, a fusion layer, and an adaptation layer; the extraction layer extracts corresponding deep features using different methods based on the modal characteristics of the preprocessed state data, and aligns the deep features using a cross-modal alignment algorithm to obtain unified features; the fusion layer performs joint modeling of spatial topological relationships and dynamic temporal evolution based on the unified features, ultimately generating spatiotemporal features; the adaptation layer dynamically adjusts the model parameters based on the baseline characteristics of the subjects; S3: Train the prediction model. During the model's operation after training, perform concept drift detection on the newly input state data. When concept drift is detected, trigger online updates and generate confidence intervals and importance scores for the prediction results based on Bayes' theorem.

2. The method for constructing a brain rehabilitation cognitive training effect prediction model according to claim 1, characterized in that, The status data includes electroencephalogram (EEG) signals. Near-infrared functional imaging signals and eye movement signals With kinematic signals .

3. The method for constructing a brain rehabilitation cognitive training effect prediction model according to claim 2, characterized in that, The specific method for extracting the depth features is as follows: Based on the modal characteristics of EEG signals in the preprocessed state data, depth features of the EEG signals are extracted using a convolutional neural network. ; Based on the modal characteristics of near-infrared functional imaging signals in the preprocessed state data, depth features of the near-infrared functional imaging signals are extracted using a convolutional neural network combined with a long short-term memory network. ; Based on the modal characteristics of eye movement and kinematic signals in the preprocessed state data, depth features of the eye movement signals are extracted using a long short-term memory network. Depth features of kinematic signals .

4. The method for constructing a brain rehabilitation cognitive training effect prediction model according to claim 3, characterized in that, The specific steps for generating the spatiotemporal features are as follows: 1) Dynamically construct a brain region functional connectivity map, represented as follows: ； in, Indicates the first The node feature matrix of the layer; Indicates the first The node feature matrix of the layer; This indicates the functional connectivity between brain regions. , Represents the adjacency matrix. Represents the identity matrix; Degree matrix; Indicates the first Layer weight matrix; This represents the activation function; through graph convolution operations, it outputs spatial features. ; 2) The temporal features with uniform characteristics are weighted and processed using a temporal attention network, and are represented as follows: ； in, This represents the score at the current time step. A function representing a temporal attention network; This represents the uniform features at the current time step; the score at each time step is converted into weights using the softmax function. The uniform features at each time step are weighted using weights, as shown below: ； in, This represents the temporal features after weighted fusion using a temporal attention network.

5. The method for constructing a brain rehabilitation cognitive training effect prediction model according to claim 4, characterized in that, The adaptation layer inputs the subject's baseline features into the multilayer perceptron to obtain adaptation parameters, which are represented as follows: ； in, It is a multilayer perceptron network; This represents the adaptation parameters; the personalized adaptation parameters are added to the original model parameters to obtain the adjusted model parameters.

6. The method for constructing a brain rehabilitation cognitive training effect prediction model according to claim 5, characterized in that, The prediction model is trained using a multi-task learning framework, defining short-term and long-term prediction branches, and obtaining short-term and long-term prediction results through these two branches. Define the total loss function as follows: ； in, and They represent the first Short-term and long-term predictive effects for individual subjects and They represent the first The actual short-term and long-term effects on each subject; This indicates the number of participants.

7. The method for constructing a brain rehabilitation cognitive training effect prediction model according to claim 6, characterized in that, The concept drift detection is performed according to the KL divergence test method, and is expressed as follows: ； in, This represents each input data point used in calculating the KL divergence; Represents the probability under the true distribution; This represents the probability under the predicted distribution; when the KL divergence exceeds a set threshold, it is determined that the newly input state data has undergone concept drift, triggering the online update mechanism, as shown below: ； in, This represents the updated model parameters; Indicates the current model parameters; Indicates the learning rate; This represents the current gradient.

8. The method for constructing a brain rehabilitation cognitive training effect prediction model according to claim 7, characterized in that, According to Bayes' theorem, combining the prior distribution and the likelihood function, the posterior distribution of the model parameters is calculated, expressed as: ； in, Represents the likelihood function; Represent the prior distribution; Indicates marginal likelihood; Represents the posterior distribution of the model parameters; This represents the input status data; Represent the prediction result; generate the confidence interval of the prediction result from the posterior distribution; Calculate the importance score of each state data point to the prediction result, expressed as: ； in, Indicates the first The importance score of each state data feature to the prediction result; Indicates the first The gradient of the features of each state data.

Images