Alzheimer's disease long-term prediction and dynamic intervention method based on deep learning

Abstract

The invention discloses an Alzheimer's disease long-term prediction and dynamic intervention method based on deep learning. Accurate dynamic intervention from group statistics to individual dynamics is realized through time sequence alignment, multi-scale time sequence feature extraction, dynamic risk assessment and personalized intervention. Through time sequence alignment and multi-scale time sequence feature extraction, a time-sensitive personalized intervention scheme can be generated. And carrying out continuous time modeling by adopting a cubic Hermite interpolation method and a neural control differential equation, and predicting the long-term risk. By establishing a double-track interaction model, a pathological track and a functional track of a user are analyzed, so that time-varying association among multi-modal data can be dynamically captured. A personalized intervention strategy is provided through reinforcement learning, the intervention strategy is dynamically adjusted in combination with risk reduction amplitude, intervention measure compliance and physiological index change, the effect of short-term behavior change and long-term prediction is balanced, and the Alzheimer's disease is further promoted to be converted from passive treatment to active intervention.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence models, and particularly relates to a long-term prediction and dynamic intervention method for Alzheimer's disease based on deep learning. Background Art

[0002] Alzheimer's disease is a neurological disease that mostly occurs in the elderly population. The clinical manifestations mainly include cognitive decline, mental symptoms and behavioral disorders, and a gradual decline in daily living ability, etc. Most patients and their families believe that the above symptoms are the result of natural aging, resulting in a relatively low early medical treatment rate. As an irreversible and incurable age-related neurodegenerative disease, early prevention and timely intervention are very important.

[0003] Existing Alzheimer's disease prediction models adopt traditional multi-modal fusion methods to solve the differences in structure, format, and distribution of data in different modalities. For example, magnetic resonance imaging is three-dimensional, while genetic data is one-dimensional. This method enables the model to learn the alignment representation between different modalities, thereby effectively integrating these heterogeneous data. The prior art does not consider the temporal dynamic changes between different modalities, that is, there are differences in time scales among multi-source data. For example, heart rate is detected in real time through wearable devices, while magnetic resonance imaging (MRI) is usually obtained once every few months. This shortcoming can lead to short-term fluctuations of high-frequency data and the drowning of long-term trends of low-frequency data, as well as the loss of causal relationships between changes in key biomarkers and behavioral data.

[0004] In addition, existing Alzheimer's disease prevention often gives intervention strategies based on "group statistics". For example, it is analyzed based on group data that increasing social frequency can effectively reduce the risk of Alzheimer's disease, without considering the intervention compliance of users. For example, it is analyzed that enhancing exercise intensity can effectively reduce the risk of Alzheimer's disease, without considering the physiological indicators of users. Moreover, the effect of the intervention measures cannot be tracked in real time, so the intervention measures cannot be dynamically adjusted. Furthermore, existing Alzheimer's disease prevention does not perform long-term prediction of risks. Usually, intervention is carried out only after significant pathological features appear in patients, missing the best prevention window. Summary of the Invention

[0005] In view of the technical problems in the above-mentioned existing technologies, such as not considering the time-scale differences in multi-source data, not considering the user's intervention compliance and giving personalized intervention strategies, not being able to track the effectiveness of intervention measures in real time and make dynamic adjustments, and not being able to achieve long-term prediction of risks, the present technical solution provides a long-term prediction and dynamic intervention method for Alzheimer's disease based on deep learning. By using dynamic time warping (DTW) to perform temporal alignment on multi-modal data with different sampling frequencies, and by cascading the Transformer model and the LSTM model to extract multi-scale temporal features, the data is coordinated and matched in the time dimension to facilitate the formulation of time-sensitive intervention plans; by using the cubic Hermite interpolation method to convert the obtained discrete data into a continuous time series and performing continuous-time modeling through neural controlled differential equations (Neural CDE), long-term risk prediction can be carried out; by establishing a dual-trajectory interaction model to analyze the user's pathological trajectory and functional trajectory, dynamically capture the time-varying correlations between multi-modal data, and provide technical support for generating "personalized" intervention strategies; by using reinforcement learning and reward mechanisms to track the effectiveness of intervention measures in real time, and combining the risk reduction amplitude, the user's intervention compliance and physiological indicators to dynamically and real-time adjust the intervention strategy, promoting the transformation of Alzheimer's disease from "passive intervention" to "active treatment"; effectively solving the above problems.

[0006] The present invention is achieved through the following technical solutions:

[0007] A long-term prediction and dynamic intervention method for Alzheimer's disease based on deep learning, characterized by including the following steps:

[0008] (1) Data preprocessing: Use dynamic time warping DTW to perform temporal alignment on the user's data with different sampling frequencies, and extract multi-scale temporal features through the Transformer model and the LSTM model; coordinate multiple modal data from different sensors, different acquisition frequencies, and different timestamps into a matching state in the time dimension;

[0009] (2) Construction of a dynamic graph neural network model: Use the cubic Hermite interpolation method and neural controlled differential equations to perform continuous-time modeling on discrete data; establish a dual-trajectory interaction model to analyze the user's pathological trajectory and functional trajectory, and dynamically capture the time-varying correlations between multi-modal data; for example, whether increasing the user's exercise intensity can inhibit hippocampal atrophy, so as to achieve precise intervention of "individual dynamics";

[0010] The established dual-trajectory interaction model includes a trajectory encoder, a dynamic graph convolutional layer, and a trajectory interaction gate, processes the input temporal features, updates, transmits, and fuses the node embeddings at each time step, and finally outputs the fused state of all time steps; the specific steps include:

[0011] Initialization: The gated recurrent unit (GRU) is used as the trajectory encoder, and the nn.ModuleList module is used to store two DynamicEdgeConv dynamic graph convolutional layers. An nn.Sequential is used to define a gating module, which includes a fully connected layer nn.Linear and an nn.Sigmoid activation function to control the flow of information in the neural network.

[0012] Forward propagation: For each time step, the trajectory encoder is used to encode the user's pathological and functional characteristics respectively. The node embeddings are updated through the dynamic graph convolutional layer, and the fusion gating value is obtained through the previously defined gating module. Then, the updated features are weighted and fused by the fusion gating value to obtain the fused state. The fused states of each time step are stored in a list, and finally, all time step states are stacked into a tensor using torch.stack and returned.

[0013] (3) Training and evaluation: The data of the last time step of the risk prediction sequence output by the dynamic graph neural network model is used as the input to the policy generator, and the sigmoid activation function is used to output the intervention intensity in terms of motor / cognitive training, providing corresponding intervention strategies. A reward model is established to maximize the cumulative reward through reinforcement learning, dynamically adjust the intervention strategies, balance the changes in short-term behavior and the effects of long-term prediction, and provide the optimal strategy suitable for the user's current state.

[0014] Furthermore, the multi-scale temporal feature extraction in step (1) uses the LSTM model and the Transformer model in series, combining the local perception ability of LSTM and the global modeling ability of Transformer. LSTM is used to extract local features, and then Transformer is used to model the global dependencies to enhance the robustness of feature extraction.

[0015] Furthermore, the Transformer model is based on the self-attention mechanism, processes sequence data in parallel, and captures the global dependencies in the sequence. It is used to process low-frequency data such as the annual amyloid PET deposition rate. The most core formula of the self-attention mechanism used is:

[0016]

[0017] Among them, Q, K, V come from the temporal features of the patient, including: hippocampal volume, Aβ protein level, etc.; P pos is the position encoding matrix used to check the time interval; P geneis the gene prior matrix, which is used to strengthen the attention weights of risk genes; the softmax function normalizes the attention scores of each row to obtain the weight distribution; d is the scaling factor, which is used to stabilize the gradient; is the transpose of the query transformation matrix, W k is the key transformation matrix, W v is the value transformation matrix; the combination of positions with high attention weights hints at key pathological development nodes and assists in clinical decision-making.

[0018] Furthermore, the LSTM model is used to process data such as sleep quality and gait changes collected by wearable devices and high-frequency data such as language fluency scores collected by mobile phone APPs, prevent gradient disappearance or explosion, and capture local features in the sequence. By selecting the hidden state at a certain time step as the local feature representation of the data at that position.

[0019] Furthermore, the dynamic time warping method (DTW) described in step (1) is used for time series alignment to ensure the temporal consistency of multimodal data and prepare for subsequent model training.

[0020] Furthermore, for the pathological trajectory described in step (2), its example metrics include the hippocampal atrophy rate and Aβ deposition, and the time characteristic is low frequency;

[0021] For the functional trajectory, its example metrics are gait stability and MoCA score, and the time characteristic is high frequency;

[0022] The double-trajectory interaction model identifies high-risk conduction paths, quantifies the blocking effects of different intervention measures, identifies the inhibitory effect of functional improvement on the pathological process, generates intervention plans and makes dynamic adjustments.

[0023] Furthermore, an example of the high-risk conduction path is: "APOEε4 → hippocampal atrophy → spatial cognitive decline".

[0024] Furthermore, the neural control differential equation described in step (2) uses the torchcde library to model the continuous time series obtained by the cubic Hermite interpolation method.

[0025] Furthermore, step (2) uses the pytorch library to process continuous time series data.

[0026] Furthermore, in step (2), the input discrete time series data is converted into a continuous time series suitable for processing by the neural control differential equation through cubic Hermite interpolation and derivative estimation of backward differences, so that the neural control differential equation can perform integration and prediction in the continuous time domain.

[0027] Further, the policy generator in step (3) is a policy network including two fully connected layers and a ReLU activation function layer:

[0028] The nn.Sequential module in PyTorch is used to stack multiple neural network layers in sequence. The input of the network will pass through each layer in the order of layer arrangement, and finally the output is obtained;

[0029] The nn.Linear linear layer is used to perform a linear transformation on the input sequence;

[0030] The nn.ReLU activation function layer sets the values less than 0 in the input sequence to 0, and keeps the values greater than 0 unchanged;

[0031] The nn.Linear fully connected layer is used, and the output dimension is 2, corresponding to two different intervention intensities, that is, the intervention intensities in terms of exercise and cognitive training.

[0032] Further, the Sigmoid activation function in step (3) maps the values of the intervention intensities output by the policy network to the interval (0, 1), which is beneficial to more clearly represent the required intervention intensities.

[0033] Further, the reward model in step (3) dynamically adjusts the intervention strategy through reinforcement learning and a reward mechanism, combining the risk reduction amplitude, intervention measure compliance, and physiological index changes, and formulates the intervention strategy by analyzing the connection between short-term behavior changes and long-term prediction effects;

[0034] The state space of reinforcement learning is: the current risk value, intervention measure compliance, physiological index;

[0035] The action space of reinforcement learning is: exercise prescription adjustment and cognitive training recommendation;

[0036] The reward function of reinforcement learning is:

[0037] R = α·ΔRisk + β·Compliance - γ·Cost_(intervention)

[0038] Among them, ΔRisk is the magnitude of risk reduction after intervention, where risk reduction is defined as a positive reward here; Compliance is the compliance of the intervention, that is, the degree to which the user adheres to the intervention measures, such as whether they can adjust the exercise intensity according to the recommendations. The higher the compliance, the better the intervention effect, and the compliance is defined as a positive reward here; Cost_(intervention) is the cost of the intervention, such as economic cost and time cost. The higher the cost, the worse the intervention effect, and the cost of the intervention is defined as a negative reward here; α, β, and γ are weight coefficients, which adjust the importance of each part in the total reward under different application scenarios and priorities.

[0039] Further, in the training and evaluation described in step (3), the training process includes an outer loop and an inner loop. The specific operation method is as follows:

[0040] The outer loop is: ① Traverse the specified number of epochs; ② Reset the environment and obtain the initial state; ③ Initialize the total reward to 0; ④ Print the average reward of the current epoch after the inner loop ends.

[0041] The inner loop is: ① Simulate i intervention cycles; ② Predict the current state to obtain the risk trajectory and formulate an intervention strategy; ③ Apply the intervention strategy to obtain the next state and reward; ④ Accumulate the reward; ⑤ Use the mean square error loss function to calculate the loss, expecting the predicted value of the last time step of the risk trajectory to gradually decrease; ⑥ Backpropagation: Clear the gradients in the optimizer and calculate the gradients of the loss; ⑦ Update the parameters of the model; ⑧ Update the current state to the next state.

[0042] Further, in the model training part of step (3), 5 intervention cycles are simulated, the risk is predicted, and the corresponding intervention strategy is generated.

[0043] Further, the reward mechanism of the reinforcement learning in step (3) is dynamically adjusted based on the magnitude of risk reduction.

[0044] Further, the state space of the reinforcement learning in step (3) is: the current risk value, the compliance of the intervention measure, and the physiological index.

[0045] Further, the action space of the reinforcement learning in step (3) is: exercise prescription adjustment and cognitive training recommendation.

[0046] Further, medical knowledge is introduced for constraint when generating and adjusting the intervention strategy in step (3).

[0047] Beneficial effects

[0048] A method for long-term prediction and dynamic intervention of Alzheimer's disease based on deep learning proposed by the present invention has the following beneficial effects compared with the prior art:

[0049] The present invention aligns multi-modal data of different sampling frequencies of users in time series through Dynamic Time Warping (DTW); uses a cascaded Transformer model and LSTM model for multi-scale time series feature extraction, generates a time-sensitive personalized intervention plan, and improves the robustness of feature extraction; converts the obtained discrete time series into a continuous time series through the cubic Hermite interpolation method; performs continuous-time modeling based on the continuous time series through Neural Controlled Differential Equations (Neural CDE), so as to enable long-term risk prediction; analyzes the pathological trajectory and functional trajectory of users by establishing a dual-trajectory interaction model, dynamically captures the time-varying correlations between multi-modal data, and provides technical support for realizing "personalized intervention"; through reinforcement learning and a reward mechanism, the effect of intervention measures is tracked in real time, that is, whether the current measure can effectively reduce the risk amplitude, and if not, dynamic adjustment is performed, further promoting the transformation of Alzheimer's disease from "passive treatment" to "active intervention". Brief Description of the Drawings

[0050] Figure 1 It is a schematic diagram of the overall process of the present invention.

[0051] Figure 2 It is a training flow chart of the reward model in the present invention. Detailed Embodiments

[0052] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. The described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Without departing from the design concept of the present invention, various variations and improvements made by those of ordinary skill in the art to the technical solutions of the present invention should fall within the protection scope of the present invention.

[0053] Embodiment 1:

[0054] As Figure 1 shown, a long-term prediction and dynamic intervention method for Alzheimer's disease based on deep learning includes the steps:[[]]END]]

[0055] (1) Data preprocessing: Use Dynamic Time Warping (DTW) for time series alignment, and use a Transformer model and an LSTM model for time series feature extraction to coordinate and match various modal data from different sensors, different acquisition frequencies, and different timestamps in the time dimension.

[0056] (11) Dynamic Time Warping (DTW) is used to align time series, and the optimal alignment path between two sequences is found through dynamic programming, so that the cumulative distance after alignment is minimized. The smaller the distance, the higher the similarity. The purpose of DTW is to find the path with the minimum cumulative distance:

[0057] DTW(series1, series2) = C[n - 1][m - 1]

[0058]

[0059] Where 0 ≤ i ≤ n - 1, 0 ≤ j ≤ m - 1; series1 and series2 are two time series with different lengths or speeds, such as physiological signals with different monitoring frequencies (such as electrocardiogram signals, electroencephalogram signals), which can be in the form of a list or an array. Here, let the length of sequence A be n and the length of sequence B be m. D is an n×m distance matrix used to record the distance between each pair of points in the two sequences (such as Euclidean distance). For example, (i, j) records the distance between the i-th point in sequence A and the j-th point in sequence B. Then, starting from the starting point (0, 0) to the ending point (n - 1, m - 1), the minimum cumulative distance C to reach each point is gradually calculated along the shortest path. For example, C[i][j] is the minimum cumulative distance to reach the point (i, j), constructing a cumulative distance matrix. C[n - 1][m - 1] is the minimum cumulative distance to reach the ending point (n - 1, m - 1). Finally, starting from the ending point (n - 1, m - 1), trace back to the starting point (0, 0) in reverse to find an optimal path, that is, the path with the minimum distance.

[0060] Since the smaller the distance, the higher the similarity, the path with the minimum cumulative distance is the best alignment path. Through the DTW algorithm, effective temporal alignment can be performed on multimodal data with different sampling frequencies of users, such as real-time heart rate data and daily exercise intensity.

[0061] (12) Extract temporal features from the temporally aligned data:

[0062] LSTM processes sequence data through a gating mechanism, is suitable for processing short sequences, extracts local features, and is used to process high-frequency physiological signals such as real-time heart rate.

[0063] Transformer is suitable for processing long sequences. Through the self-attention mechanism, it calculates the global time-step correlation weights, captures long-distance dependencies, and is used for extracting comprehensive features by fusing multi-source medical data (such as heart rate, blood oxygen).

[0064] Using Transformer and LSTM in series combines the local perception ability of LSTM and the global modeling ability of Transformer. LSTM is used to extract local features, and then Transformer is used to model global dependencies, enhancing the robustness of feature extraction.

[0065] (13) The Transformer model is set up based on the self-attention mechanism, which can process sequence data in parallel and capture global dependencies in the sequence. This model is used to process low-frequency data such as amyloid PET annual deposition rate. The most core formula of the self-attention mechanism used is:

[0066]

[0067] Q, K, V come from the temporal features of the patient (such as hippocampal volume, Aβ protein level, etc.); W k 、W v 、 are learnable weight matrices. By multiplying the input vector of each encoder with the three weight matrices respectively, the query vector, key vector, and value vector are obtained. Combining the obtained query vector, key vector, and value vector can get three vector matrices Q, K, V. The dot product of the key vector and the query vector obtained from each input vector is calculated to get the corresponding score. d is the scaling factor used to stabilize the gradient. P pos is the position encoding matrix used to check the time interval; P gene is the gene prior matrix used to strengthen the attention weight of risk genes; the softmax function normalizes the attention scores of each row to obtain the weight distribution; multiplying each value vector by the softmax score and summing them up to get the output of the self-attention layer at this position. The combination of positions with high attention weights indicates key pathological development nodes to assist clinical decision-making.

[0068] The Transformer model mainly extracts features through the encoder-decoder. The encoder is responsible for transforming the input sequence into a series of high-dimensional representations, that is, mapping the original features to a high-dimensional space. The decoder generates the final output based on the output of the encoder and the target sequence.

[0069] In the encoder, the multi-head self-attention mechanism can map the input sequence to multiple subspaces respectively, and then calculate the attention weights in each subspace in parallel to capture the temporal correlation patterns in different subspaces. The ReLU activation function is introduced through the feed-forward neural network for non-linear transformation to further extract features and enhance the expression ability of the model. Through the residual network, the details and semantic information in the original input sequence are retained, solving the problems of loss and attenuation that occur in the information transfer of traditional convolutional layers or fully connected layers.

[0070] The decoder calculates the similarity scores between each position and the output of the encoder and the previously generated output, and then uses these scores as weights to weighted sum the output of the encoder and the previously generated output.

[0071] (14) An LSTM model is set up to process high-frequency data such as data from wearable devices (such as sleep quality and gait changes) and language fluency scores collected by mobile phone apps, which can effectively prevent gradient vanishing or explosion.

[0072] Through the control of the forget gate, input gate, and output gate, LSTM can determine which information is forgotten, which new information is added, and which information is output to the hidden state. The hidden state gradually contains the local feature information of the sequence in this process.

[0073] Due to the gated structure of LSTM, it can effectively capture the local features in the sequence. The hidden state of the LSTM layer at a certain time step can be focused on. This hidden state contains the information of the current time step and the previous time steps, and this hidden state can be used as the local feature representation of the data at this position.

[0074] (2) Construction of a dynamic graph neural network model: A dual-trajectory interaction model, namely the pathological trajectory (such as biomarker changes) and the functional trajectory (such as cognitive function changes), is established through a dynamic graph neural network (DGNN), so as to dynamically capture the time-varying correlations between multi-modal data. Neural Controlled Differential Equations (Neural CDE) are used for continuous-time modeling to predict long-term risk trajectories.

[0075] (21) Set the example indicators of the pathological trajectory as the hippocampal atrophy rate and the amount of Aβ deposition. The example indicators of the functional trajectory are gait stability and MoCA score.

[0076] (22) Use the DynamicEdgeConv module in the PyTorch Geometric library to construct a dynamic graph neural network. The nodes of this dynamic graph neural network include pathological nodes and functional nodes:

[0077] Pathological nodes: hippocampal atrophy rate, amyloid PET deposition rate, etc.

[0078] Functional nodes: cognitive score, exercise intensity, sleep quality, etc.

[0079] DynamicEdgeConv can dynamically construct the edges of the graph according to the features of the input data.

[0080] Use the linear layer nn.Linear in Pytorch for linear transformation of the input data. The calculation formula is: y = xW T + b. Where x is the input data, W is the weight matrix, W T is the transpose of the weight matrix, b is the bias vector, and y is the output data.

[0081] The function for calculating edge features is:

[0082]

[0083] where \(W\in\mathbb{R}\) node_dim×2·node_dim , \(b\in\mathbb{R}\) node_dim ; edge_feature is the edge feature, \(X\) i is the source node feature, \(X\) j is the target node feature. \(W\) is a learnable weight matrix used to perform a linear transformation on the combined feature vector. The dimension of this weight matrix is determined by the dimension of the combined feature vector and the dimension of the expected edge feature. \(b\) is a bias vector used to add a constant offset to the result of the linear transformation, and the dimension of this bias vector is the same as the dimension of the expected edge feature.

[0084] During the dynamic edge convolution process, the features of the source node and the target node are concatenated, and the input dimension is twice the node feature dimension, which is \(2\times node\_dim\). After the linear transformation, the output feature dimension is the same as the original feature dimension of the node, which is \(node\_dim\).

[0085] \(\oplus\) represents a feature combination operation, and the combination method is concatenation, that is, the feature vectors of the source node and the target node are concatenated end to end, that is, the low-level features in the encoder and the high-level features in the decoder are concatenated.

[0086]

[0087] where \(concat\_feature\) ij is the concatenated feature vector with dimension \(2\times node\_dim\).

[0088] The KNN algorithm is used, that is, the k-nearest neighbor algorithm. After calculating the feature similarity between all node pairs, the 5 neighbors with the smallest distance are selected for each node, and edges are established between these nodes, that is, \(k = 5\).

[0089] The dynamic graph neural network will update the features of the nodes and the connections of the edges as it evolves over time.

[0090] (23) A dual-trajectory interaction model is constructed using three parts: a trajectory encoder, a dynamic graph convolutional layer, and a trajectory interaction gate. By processing the input temporal features, the node embeddings at each time step are updated, information is transmitted and fused, and finally the fused state of all time steps is output. The specific steps are as follows:

[0091] Initialization: Use the gated recurrent unit (GRU) as the trajectory encoder, and use nn.ModuleList to store two DynamicEdgeConv dynamic graph convolutional layers. Define a gating module using nn.Sequential, which includes a fully connected layer nn.Linear and an nn.Sigmoid activation function to control the flow of information in the neural network.

[0092] Forward propagation: For each time step, use the trajectory encoder to encode the user's pathological and functional features respectively. Update the node embeddings through the dynamic graph convolutional layer and obtain the fusion gating value through the previously defined gating module. Then, weight and fuse the updated features with the fusion gating value to obtain the fused state. Store the fused states of each time step in a list, and finally use torch.stack to stack the states of all time steps into a tensor and return.

[0093] (24) Convert the input discrete time series into a continuous time series suitable for processing by neural control differential equations. For example, the data of the amyloid PET deposition rate and cognitive score of the collected users are both discrete and the data points are uneven. The backward difference method is used to estimate the derivative, and the derivative value is estimated using the data of the current point and the previous point:

[0094]

[0095] where f′(t i ) is the estimated derivative of the current point, f(t i ) is the data of the current point, such as the amyloid PET deposition rate of the user this time. f(t i -1) is the data of the previous point, such as the amyloid PET deposition rate of the user last time. t i is the data point, and t i-1 is the previous data point.

[0096] Use cubic Hermite interpolation method to construct a smooth function on discrete data points, ensuring the continuity of function values and first-order derivatives at data points, which is suitable for processing physiological signals with time-varying rates, such as heart rate variability analysis. Through the values of each data point and the estimated derivative, construct a cubic Hermite interpolation polynomial:

[0097] H(x) = y0·h 00 (t) + y1h 01 (t) + m0h 10 (t) + m1h 11 (t)

[0098]

[0099] Among them, x, x0, and x1 are data points, H(x) is the constructed cubic polynomial, y0 and y1 are the data values of x0 and x1 respectively, m0 and m1 are the derivative values corresponding to x0 and x1 respectively, and h 00 (t), h 01 (t), h 10 (t), h 11 (t) are basis functions, and the values of the corresponding basis functions are obtained by applying boundary conditions to ordinary polynomials.

[0100] Through this formula, the neural control differential equation can be integrated and predicted in a continuous time domain.

[0101] (25) Set the mathematical form of the neural control differential equation as:

[0102] dz(t) = f θ (z(t))dX(t)

[0103] Among them, as can be seen from the above operation method, X(t) is a continuous time series generated by interpolating discrete user data through cubic Hermite interpolation, and f θ is a neural network with parameter θ, z(t) is the risk value at time t, and dz(t) and dX(t) represent the differential forms of z(t) and X(t).

[0104] Through the neural control differential equation, continuous-time modeling of the user's Alzheimer's risk trajectory is realized, thereby realizing long-term prediction of the risk trajectory.

[0105] (26) Perform risk prediction based on the constructed dynamic graph neural network model above; for example, assume that cognitive risk is related to the degree of hippocampal atrophy and exercise intensity, and the corresponding mathematical expression is:

[0106] congitiverisk (t+1)

[0107] = α · atrophy (t) + β · exercise (t) + γ · (atrophy (t) · exercise (t) )

[0108] Among them, congitiverisk (t+1) is the cognitive risk value at time t + 1, atrophy (t) is the degree of hippocampal atrophy at time t, and exercise (t)Let \(I_t\) be the exercise intensity at time \(t\), \(\alpha\) be the linear influence coefficient of hippocampal atrophy on cognitive risk, \(\beta\) be the linear influence intensity of exercise intensity on cognitive risk, and \(\gamma\) be the interaction influence coefficient between hippocampal atrophy and exercise intensity.

[0109] Assume \(\alpha>0\) and \(\beta < 0\). Then, the higher the degree of hippocampal atrophy, the greater the cognitive risk. The greater the exercise intensity, the lower the cognitive risk.

[0110] Assume \(\gamma < 0\). This means that exercise can weaken the negative impact of hippocampal atrophy on cognitive risk.

[0111] Assume \(\gamma>0\). This means that exercise may exacerbate the negative impact of hippocampal atrophy on cognitive risk.

[0112] (3) Training and evaluation: The reward model maximizes the cumulative reward through reinforcement learning to find the optimal intervention strategy, and then dynamically adjusts the intervention strategy to effectively reduce the user's Alzheimer's risk.

[0113] (31) When generating and adjusting the intervention strategy, medical knowledge is introduced for constraint. For example, if the user has coronary heart disease, the training intensity needs to be adjusted accordingly.

[0114] (32) The state space of reinforcement learning is: the current risk value, intervention compliance, and physiological indicators.

[0115] (33) The action space of reinforcement learning is: exercise prescription adjustment and cognitive training recommendation.

[0116] (34) The reward function of reinforcement learning is:

[0117] \(R = \alpha\cdot\Delta Risk+\beta\cdot Compliance-\gamma\cdot Cost_{(intervention)}\)

[0118] where \(\Delta Risk\) is the degree of risk reduction after intervention, and risk reduction is defined as a positive reward.

[0119] Compliance is the compliance of the intervention, that is, the degree to which the user follows the intervention measures. For example, whether the user can adjust the exercise intensity according to the recommendation. The higher the compliance, the better the intervention effect, so compliance is defined as a positive reward.

[0120] \(Cost_{(intervention)}\) is the cost of the intervention, such as economic cost and time cost. The higher the cost, the worse the intervention effect. Therefore, the cost of the intervention here is defined as a negative reward.

[0121] \(\alpha\), \(\beta\), \(\gamma\) are weight coefficients, which are used to adjust the importance of each part in the total reward under different application scenarios and priorities.

[0122] (35) The steps for predicting the risk of Alzheimer's disease and dynamic intervention through the long-term prediction and dynamic intervention system for Alzheimer's disease are as follows:

[0123] ① Obtain multimodal data of different acquisition frequencies and different timestamps of users through methods such as intelligent device collection and medical monitoring. For example, obtain real-time heart rate through a smart watch, and obtain amyloid deposition rate and hippocampal atrophy rate through magnetic resonance imaging.

[0124] ② Align the obtained data in time series through dynamic time warping, extract multi-scale time series features from the aligned data through the Transformer model and the LSTM model, and coordinate and match these data in the time dimension.

[0125] ③ Perform continuous-time modeling on the discrete data through the cubic Hermite interpolation method and the neural control differential equation, so as to perform long-term prediction of the risk.

[0126] ④ Based on the continuous time series, a double-trajectory interaction model is analyzed to analyze the pathological trajectory and functional trajectory of the user, so as to dynamically capture the time-varying correlation between multimodal data. For example, whether increasing the user's exercise intensity can inhibit hippocampal atrophy, so as to achieve precise "individual dynamic" intervention.

[0127] ⑤ Based on the above steps, provide an intervention strategy through reinforcement learning, and dynamically adjust the intervention strategy by combining the risk reduction amplitude, intervention measure compliance, and physiological index changes, balance short-term behavior changes and long-term effects, and provide the optimal strategy suitable for the user's current state.

[0128] (36) According to Figure 2 the training flow chart of the reward model shown, the outer loop is as follows:

[0129] ① Traverse the specified number of epochs; ② Reset the environment and obtain the initial state; ③ Initialize the total reward to 0; ④ After the inner loop ends, print the average reward of the current epoch.

[0130] The inner loop is as follows:

[0131] ① Simulate i intervention cycles; ② Predict the current state, obtain the risk trajectory and formulate an intervention strategy; ③ Apply the intervention strategy to obtain the next state and reward; ④ Accumulate the reward; ⑤ Use the mean square error loss function to calculate the loss, expecting the predicted value of the last time step of the risk trajectory to gradually decrease. ⑥ Backpropagation: Clear the gradients in the optimizer and calculate the gradients of the loss. ⑦ Update the parameters of the model. ⑧ Update the current state to the next state.

[0132] The above has introduced in detail a system for long-term prediction and dynamic intervention of Alzheimer's based on deep learning provided by the examples of the present invention. The description of the above embodiments is only used to help understand the method and its core idea of the present invention; at the same time, for those of ordinary skill in the art, according to the idea of the present invention, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to the present invention.

Claims

1. A long-term prediction and dynamic intervention method for Alzheimer's disease based on deep learning, characterized in that, Including the steps: (1) Data preprocessing: Use Dynamic Time Warping (DTW) to align the data of different sampling frequencies of users in time series, and use Transformer model and LSTM model to extract multi-scale time series features from the data; Harmonize various modal data from different sensors, different acquisition frequencies, and different timestamps into a matching state in the time dimension; (2) Construct a dynamic graph neural network model: Establish a dual-trajectory interaction model through the dynamic graph neural network DGNN, that is, a pathological trajectory such as biomarker changes and a functional trajectory such as cognitive function changes, and dynamically capture the time-varying correlations between multi-modal data; Use the cubic Hermite interpolation method and neural controlled differential equations to perform continuous-time modeling on the discrete data to predict the long-term risk trajectory; The dual-trajectory interaction model is established using a trajectory encoder, a dynamic graph convolutional layer, and a trajectory interaction gate. By processing the input time series features, the node embeddings at each time step are updated, information is transmitted and fused, and finally the state fused at all time steps is output; The specific steps are as follows: Initialization: Use the gated recurrent unit (GRU) as the trajectory encoder, and use the nn.ModuleList module to store two DynamicEdgeConv dynamic graph convolutional layers; Define a gating module using nn.Sequential, which contains a fully connected layer nn.Linear and an nn.Sigmoid activation function to control the flow of information in the neural network; Forward propagation: For each time step, use the trajectory encoder to encode the pathological and functional features of the user respectively. Update the node embeddings through the dynamic graph convolutional layer and obtain the fusion gating value through the previously defined gating module. Then, weight and fuse the updated features through the fusion gating value to obtain the fused state; Store the fused state of each time step in a list, and finally use torch.stack to stack the states of all time steps into a tensor and return; (3) Training and evaluation: Use the data of the last time step of the risk prediction sequence output by the dynamic graph neural network model as the input to the policy generator, and output the intervention intensity in terms of motion / cognitive training through the sigmoid activation function to provide corresponding intervention strategies; Establish a reward model, maximize the cumulative reward through reinforcement learning, dynamically adjust the intervention strategies, balance the changes in short-term behaviors and the effects of long-term predictions, and provide the optimal strategy suitable for the user's current state.

2. The long-term prediction and dynamic intervention method for Alzheimer's disease based on deep learning according to claim 1, wherein: The multi-scale time series feature extraction described in step (1) uses the LSTM model and the Transformer model in series. Combining the local perception ability of LSTM and the global modeling ability of Transformer, use LSTM to extract local features, and then use Transformer to model the global dependencies to enhance the robustness of feature extraction.

3. A method for long-term prediction and dynamic intervention of Alzheimer's disease based on deep learning according to claim 1 or 2, characterized in that: The described Transformer model is based on the self-attention mechanism, processes sequence data in parallel, and captures global dependencies in the sequence. It is used to process low-frequency data such as the annual amyloid PET deposition rate. The most core formula of the self-attention mechanism is: Among them, Q, K, and V come from the temporal features of the patient, including: hippocampal volume, Aβ protein level, etc.; P pos is the position encoding matrix, used to check the time interval; P gene is the gene prior matrix, used to enhance the attention weight of risk genes; the softmax function normalizes the attention scores of each row to obtain the weight distribution; d is the scaling factor, used to stabilize the gradient; is the transpose of the query transformation matrix, W k is the key transformation matrix, W v is the value transformation matrix; the position combination with high attention weight indicates the key pathological development nodes, assisting clinical decision-making; The described LSTM model is used to process high-frequency data such as sleep quality and gait changes collected by wearable devices and language fluency scores collected by mobile phone apps. It prevents gradient vanishing or explosion and captures local features in the sequence. By selecting the hidden state at a certain time step as the local feature representation of the data at that position.

4. A long-term prediction and dynamic intervention method for Alzheimer's disease based on deep learning according to claim 1, characterized in that: For the pathological trajectory described in step (2), its example metrics include the hippocampal atrophy rate and Aβ deposition amount, and the time characteristic is low-frequency; For the functional trajectory, its example metrics are gait stability and MoCA score, and the time characteristic is high-frequency; The described dual-trajectory interaction model identifies high-risk conduction paths, quantifies the blocking effects of different intervention measures, identifies the inhibitory effect of functional improvement on the pathological process, generates intervention plans and makes dynamic adjustments.

5. A long-term prediction and dynamic intervention method for Alzheimer's disease based on deep learning according to claim 4, characterized in that: An example of the high-risk conduction path is: "APOEε4 → hippocampal atrophy → spatial cognitive decline".

6. A long-term prediction and dynamic intervention method for Alzheimer's disease based on deep learning according to claim 1, characterized in that: The neural control differential equation described in step (2) uses the torchcde library to model the continuous time series obtained by the cubic Hermite interpolation method.

7. A long-term prediction and dynamic intervention method for Alzheimer's disease based on deep learning according to claim 1, characterized in that: The policy generator described in step (3) is a policy network containing two fully connected layers and a ReLU activation function layer: Using the nn.Sequential module in PyTorch, multiple neural network layers are stacked in sequence. The input of the network will pass through each layer in the order of layer arrangement, and finally the output is obtained; Using the nn.Linear linear layer to perform a linear transformation on the input sequence; Using the nn.ReLU activation function layer to set the values less than 0 in the input sequence to 0 and keep the values greater than 0 unchanged; Using the nn.Linear fully connected layer, the output dimension is 2, corresponding to two different intervention intensities, namely the intervention intensities in terms of exercise and cognitive training.

8. A long-term prediction and dynamic intervention method for Alzheimer's disease based on deep learning according to claim 1, characterized in that: The Sigmoid activation function described in step (3) maps the values of the intervention intensities output by the policy network to the interval (0, 1), which is beneficial for more clearly representing the required intervention intensities.

9. A long-term prediction and dynamic intervention method for Alzheimer's disease based on deep learning according to claim 1, characterized in that: The reward model described in step (3) dynamically adjusts the intervention strategy through reinforcement learning and the reward mechanism, combining the risk reduction amplitude, intervention measure compliance, and physiological index changes. By analyzing the relationship between short-term behavior changes and long-term prediction effects, intervention strategies are formulated; The state space of reinforcement learning is: the current risk value, intervention measure compliance, physiological index; The action space of reinforcement learning is: exercise prescription adjustment and cognitive training recommendation; The reward function of reinforcement learning is: R = α·ΔRisk + β·Compliance - γ·Cost_(intervention) Among them, ΔRisk is the magnitude of risk reduction after intervention, where risk reduction is defined as a positive reward; Compliance is the compliance of the intervention, that is, the degree to which the user adheres to the intervention measures, such as whether they can adjust the exercise intensity according to the advice; the higher the compliance, the better the intervention effect, and here compliance is defined as a positive reward; Cost_(intervention) is the cost of the intervention, such as economic cost and time cost. The higher the cost, the worse the intervention effect, and here the cost of the intervention is defined as a negative reward; α, β, and γ are weight coefficients, which adjust the importance of each part in the total reward under different application scenarios and priorities.

10. A long-term prediction and dynamic intervention method for Alzheimer's disease based on deep learning according to claim 1, characterized in that: In the training and evaluation described in step (3), the training process includes an outer loop and an inner loop; the specific operation method is as follows: The outer loop is: ① Traverse the specified number of epochs; ② Reset the environment and obtain the initial state; ③ Initialize the total reward to 0; ④ Print the average reward of the current epoch after the inner loop ends; The inner loop is: ① Simulate i intervention cycles; ② Predict the current state to obtain the risk trajectory and formulate an intervention strategy; ③ Apply the intervention strategy to obtain the next state and reward; ④ Accumulate the rewards; ⑤ Calculate the loss using the mean squared error loss function, expecting the predicted value of the last time step of the risk trajectory to gradually decrease; ⑥ Backpropagation: Clear the gradients in the optimizer and calculate the gradients of the loss; ⑦ Update the parameters of the model; ⑧ Update the current state to the next state.

Images