Vascular cognitive impairment prediction method, computer program product and terminal

Abstract

The invention discloses a vascular cognitive impairment prediction method, a computer program product and a terminal, and belongs to the field of deep learning. Feature information of medical time sequence data of a patient is extracted based on a time sequence data processing model; compressed feature representation of the feature information is learned through the information bottleneck; and performing prediction according to the compressed feature representation to obtain a prediction result of the disease progress. According to the method, the time sequence data processing model is used for capturing time dependence in the time sequence data, features related to the disease progression of dementia are effectively extracted from clinical data, then the features are screened by adopting information bottleneck, the most important features for classification prediction are identified, redundant and irrelevant information is eliminated, and the accuracy of classification prediction is improved. Therefore, the accuracy of classification prediction is improved. Furthermore, in order to simulate and generate various possible future data paths and reflect the uncertainty in the data, future process data is generated by using a diffusion model after feature extraction.

Description

Technical Field

[0001] This invention relates to the field of deep learning, and more particularly to a method for predicting vascular cognitive impairment, a computer program product, and a terminal. Background Technology

[0002] Vascular cognitive impairment (VCI) originates from cerebrovascular diseases and is associated with vascular risk factors (such as hypertension, diabetes, and hyperlipidemia). It is a large syndrome ranging from subjective cognitive decline and mild cognitive impairment to dementia. Studies have shown that VCI has become the second leading cause of cognitive impairment in the elderly, second only to Alzheimer's disease. Early dementia often manifests only as memory loss, easily overlooked by patients and their families. Most cases are diagnosed only when the condition has progressed to an advanced stage. However, there is no radical cure for late-stage dementia; treatment primarily focuses on interventions targeting modifiable factors. Furthermore, research indicates that most patients with mild cognitive impairment can alleviate symptoms and slow disease progression, or even recover, through early cognitive training, exercise training, and medication. Therefore, early screening for vascular dementia is crucial. Current screening methods for cognitive impairment include standardized scale assessments, cerebrospinal fluid examination, and functional magnetic resonance imaging (fMRI). These methods have limitations in community screening, being expensive and potentially requiring qualified medical professionals. Therefore, a simple method suitable for large-scale screening would be more useful for the early screening of vascular dementia.

[0003] In the medical field, artificial intelligence technology is revolutionizing disease prediction and management, including diagnosis, prediction, treatment, and management. Currently, the main methods for disease prediction can be broadly categorized into three types: classical regression methods, machine learning methods, and deep learning methods. Traditional disease risk prediction is primarily based on the Cox proportional hazards regression model (Cox model) and logistic regression model. The Cox model has been widely used in medical research and is the most frequently applied multifactor regression analysis method in traditional survival analysis and risk prediction. In recent years, feature selection and supervised learning modeling methods in the field of machine learning have also been increasingly used for disease prediction problems. Machine learning methods offer advantages in improving model interpretability and predictive performance. Deep learning has brought disruptive changes to multiple fields such as image recognition, speech recognition, and natural language understanding in recent years. In the analysis of electronic medical record data, some research has also utilized deep learning methods to build disease risk prediction models, such as CNN or RNN models.

[0004] However, machine learning and classical regression methods suffer from limitations in disease prediction, including handling complexity, large-scale data processing capabilities, weak generalization ability, and poor dynamic data processing capabilities. This is primarily because machine learning and classical regression methods typically assume linear or simple nonlinear relationships between data, while the complexity of medical data (especially when involving large amounts of interaction and high-dimensional data) often exceeds the processing capabilities of these models. Dementia is a progressive disease whose symptoms and effects gradually worsen over time. Time-series data can capture this gradual process, providing continuous information about disease progression. However, machine learning models and traditional regression methods have limitations in processing time-series data, while deep learning models have proven particularly effective in capturing and analyzing complex patterns in time-series data. Currently, there is considerable research on deep learning for imaging data analysis, but research on exploring cognitive impairment using clinical variables that change over time is still limited. Summary of the Invention

[0005] The purpose of this invention is to overcome the problems of the prior art and provide a method for predicting vascular cognitive impairment, a computer program product, and a terminal.

[0006] The objective of this invention is achieved through the following technical solution: a method for predicting vascular cognitive impairment, comprising the following steps:

[0007] Preprocessing of time-series data includes one or more of the following: data grouping, missing value handling, outlier detection, data shaping, and normalization; the time-series data consists of patients' medical records and / or imaging data.

[0008] The feature information of time series data is extracted based on the time series data processing model. The time series data processing model can be any one or more of the following: temporal convolutional network, long short-term memory network, gated recurrent unit, and Transformer model.

[0009] Learn compressed feature representations of feature information by addressing information bottlenecks;

[0010] Predictions are made based on compressed feature representations to obtain predictions of disease progression.

[0011] In one example, missing value handling includes:

[0012] The diffusion model was trained using time-series data without missing information.

[0013] Noisy data is generated by gradually adding noise to the feature vector at each time step through the noise prediction network in the diffusion model.

[0014] By using a denoising network in the diffusion model to predict and remove noise at each time step, the feature data is gradually restored, thereby achieving missing value imputation.

[0015] In one example, extracting feature information from temporal data using a temporal convolutional network includes the following steps:

[0016] Constructing a temporal convolutional network: The temporal convolutional network includes a sequentially connected input layer, hidden layer, and output layer. The hidden layer can be one or more layers. The hidden layer includes a convolutional block, which includes a sequentially connected dilated convolutional layer, a weight normalization layer, an activation function layer, and a dropout layer. The dilated convolutional layer in each convolutional block has residual connections.

[0017] Feature extraction: First, the feature information of the time series data is captured through dilated convolutional layers. Then, the weights in the dilated convolutional layers are normalized through weight normalization layers. Next, non-linear processing is performed through activation function layers. Finally, regularization is performed through Dropout layers. The above feature extraction process is repeated and superimposed, and residual processing is performed. Finally, the feature information of the time series data is output.

[0018] In one example, the dilated convolutional layer captures long-term, low-frequency feature information of temporal data based on dynamic convolution, including the following steps:

[0019] For each time point, multiple Q and K vectors are generated, where Q is represented by the features of the current time point and K is represented by the features of historical time points;

[0020] The similarity between vectors Q and K is calculated, and then normalized to obtain the attention weights. An analysis of variance is performed on the attention weights. If the variance is greater than a threshold, the first convolutional kernel is used to perform the analysis of variance on the attention weights; otherwise, the second convolutional kernel is used to perform the analysis of variance on the attention weights. The first convolutional kernel is smaller than the second convolutional kernel.

[0021] In one example, before the step of learning a compressed feature representation of feature information through an information bottleneck, a step of generating future process data is also included:

[0022] Train a diffusion model using the feature information of time series data;

[0023] Noisy data is generated by gradually adding noise to the feature vector at each time step through the noise prediction network in the diffusion model.

[0024] By using a denoising network in the diffusion model to predict and remove noise at each time step, feature data is gradually recovered, thereby generating future process data.

[0025] In one example, a compressed feature representation of future process data is learned through an information bottleneck, including:

[0026] Future process data x tnThe relevant information between the two variables, Y and Y, representing the actual dementia situation data, is defined as mutual information I(X; Y), where X represents the input set of future process data.

[0027] By minimizing Excluding irrelevant data that does not contribute to the prediction of Y for x tn The information in ' is compressed to obtain Capture with x tn The relevant features are used to obtain the filtered features; among them It is the minimum sufficient statistic of X relative to Y, and β represents the compression parameter.

[0028] In one example, the method further includes:

[0029] The loss of the information bottleneck module and the loss of the prediction network are trained together and adjusted with weight factors to obtain the overall loss.

[0030] It should be further noted that the technical features corresponding to the above examples can be combined or substituted to form new technical solutions.

[0031] The present invention also includes a computer program product comprising a computer program that, when executed by a processor, implements the steps of the method for predicting vascular cognitive impairment formed by any or a combination of the above examples.

[0032] The present invention also includes a storage medium storing computer instructions thereon, which, when executed, perform the steps of the vascular cognitive impairment prediction method formed by any or more of the above examples.

[0033] The present invention also includes a terminal comprising a memory and a processor, the memory storing computer instructions executable on the processor, wherein the processor, when executing the computer instructions, performs the steps of the vascular cognitive impairment prediction method formed by any or more of the above examples.

[0034] Compared with the prior art, the beneficial effects of the present invention are:

[0035] 1. In one example, the present invention utilizes a time series data processing model to capture the time dependence in time series data, effectively extracts features related to the progression of dementia from clinical data, and then uses information bottlenecks to filter the features, identify the features most important for classification prediction, and eliminate redundant and irrelevant information, thereby improving the accuracy of classification prediction.

[0036] 2. In one example, the diffusion model is used for missing value imputation, which can generate imputed data that is highly consistent with the distribution of the original data, ensuring the authenticity and consistency of the data. It can also preserve the dynamic characteristics of the patient's condition over time, enabling the predictive model to capture the true trend of disease development.

[0037] 3. In one example, a temporal convolutional neural network is used to capture the temporal dependence in time series data, identify markers of vascular dementia that progress over time, track the relationship between long-term progression of case data and eventual onset of the disease, and improve the accuracy of feature extraction.

[0038] 4. In one example, a smaller dynamic convolution kernel is used to capture short-term, high-frequency features, while a larger dynamic convolution kernel is used to capture long-term, low-frequency features, thereby extracting more comprehensive and accurate feature information and improving computational efficiency.

[0039] 5. In one example, in order to simulate the generation of multiple possible future data paths and reflect the uncertainty in the data, a diffusion model is used to generate future process data after feature extraction.

[0040] 6. In one example, the loss of classification prediction and the loss of the information bottleneck module are trained together and adjusted by weight factors. This allows the network to learn the latent representation corresponding to the real dementia classification result while obtaining the classification prediction result, thus obtaining feature information useful for the dementia classification prediction task. Attached Figure Description

[0041] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. The accompanying drawings are provided to provide a further understanding of the present application and constitute a part of the present application. The same reference numerals are used in these drawings to denote the same or similar parts. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation of the present application.

[0042] Figure 1 A flowchart illustrating a method provided as an example of the present invention;

[0043] Figure 2 This is a schematic diagram of a temporal convolutional network provided as an example of the present invention;

[0044] Figure 3 The following is a structural diagram of a prediction system provided as an example of the present invention. Detailed Implementation

[0045] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] In the description of this invention, it should be noted that the directions or positional relationships indicated by terms such as "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer" are based on the directions or positional relationships shown in the accompanying drawings. They are used only for the convenience of describing this invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention. Furthermore, the use of ordinal numbers (e.g., "first and second," "first to fourth," etc.) is for distinguishing objects and is not limited to this order, and should not be construed as indicating or implying relative importance.

[0047] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "joining" should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0048] Furthermore, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0049] In one example, such as Figure 1 As shown, a method for predicting vascular cognitive impairment includes the following steps:

[0050] S1: Preprocess the time series data, including one or more of the following: data grouping, missing value handling, outlier detection, data shaping, and normalization; the time series data is the patient's medical record data and / or imaging data.

[0051] S2: Extract feature information of time series data based on time series data processing model. The time series data processing model can be any one or more of the following: temporal convolutional network, long short-term memory network, gated recurrent unit, and Transformer model.

[0052] S3: Learn compressed feature representations of feature information by addressing information bottlenecks;

[0053] S4: Based on the compressed feature representation, a prediction is made to obtain the predicted results of the disease process, including the speed of disease development, the severity of the disease, and the future state.

[0054] In step S1, the patient's medical record data includes the patient's medical records, medical history, medication records, vital signs, physiological signal data (such as electrocardiogram data), etc.; the imaging data are medical images, such as X-rays, CT scans, MRI images, etc.

[0055] For data preprocessing, data grouping involves excluding volunteers diagnosed with dementia before the baseline assessment date in the dataset. Grouping is performed using hospital inpatient records or death records. After grouping, volunteers are labeled y = {0, 1, 2}, where y = 0 represents normal, y = 1 represents vascular dementia, and y = 2 represents other types of dementia. Next, data with more than 30% missing values ​​are deleted. Missing values ​​are replaced using constant imputation or model imputation. Optionally, a diffusion model is used to handle missing values. The diffusion model is first trained on data without missing values, and the knowledge gained from the training is used for backdiffusion to impute missing values. Outlier detection uses quartiles to identify outliers, and each outlier is replaced with the previous non-missing value. Since the follow-up assessment time point may vary from volunteer to volunteer, data normalization and shaping are required. A fixed time window length is determined, using four months as an example. Feature data for each volunteer is aggregated according to the time window length, and the average value within each time period is calculated as the data within that time window. After data shaping, the feature data for each volunteer is denoted as X = (x... t0 x t1 , ......, x tn ), where ti represents the i-th time window, and each volunteer has tn data points after data shaping. Finally, the min-max normalization method is used to scale all features to the interval [0, 1].

[0056] In step S2, when two or more models are used to extract feature information, the feature information output by each model can be fused to improve the accuracy of feature information extraction.

[0057] In step S3, although the data information is fused based on the time series data processing model, redundant features still exist. Therefore, this invention uses information bottleneck to filter features, identify the most important features for classification prediction, and eliminate redundant and irrelevant information.

[0058] This invention utilizes a time-series data processing model to capture the time dependence in time-series data and identify vascular dementia markers that progress over time, such as changes in a certain biomarker that are closely related to the progression of vascular dementia. This effectively extracts features related to the progression of dementia from clinical data. Then, information bottlenecks are used to filter the features, identify the features most important for classification and prediction, and eliminate redundant and irrelevant information, thereby ensuring the accuracy of classification and prediction.

[0059] In one example, missing values ​​are inevitable during follow-up. To fill in high-quality missing value data, missing value processing is performed, including:

[0060] The diffusion model was trained using time-series data without missing information.

[0061] Noisy data is generated by gradually adding noise to the feature vector at each time step through the noise prediction network in the diffusion model.

[0062] By using a denoising network in the diffusion model to predict and remove noise at each time step, the feature data is gradually restored, thereby achieving missing value imputation.

[0063] The diffusion model can be an SSSD model; specifically, the diffusion model is trained with complete data, and noise is gradually added to the feature vector at each time step t to generate noisy data x. t , in It is a parameter for controlling noise intensity, ∈ t It is noise that follows a normal distribution. α t =1-β t , parameter β t This represents the noise intensity used to control each time step t. βt gradually increases, and the proportion of noise data gradually increases during the forward diffusion process; as the time step changes, β... t As the noise level increases, more and more noise data is introduced during the forward diffusion process. Then, a noise prediction network is trained to predict the noise level at each time step t. θ (x t loss function (t) Mean squared error measures the difference between the predicted noise and the actual added noise. The network is trained to minimize the prediction error, while the reverse denoising process involves using the denoising network to predict and remove noise at each time step t, gradually recovering the feature data. Where Z is random noise distributed in a standard normal pattern, σ t The noise standard deviation is given at time step t, and missing values ​​are filled in through this process.

[0064] In one example, a Temporal Convolutional Network (TCN) is used to extract feature information from temporal data, including the following steps:

[0065] Build as Figure 2 The temporal convolutional network shown comprises a sequentially connected input layer, hidden layers, and an output layer. This example has two hidden layers. Each hidden layer consists of a convolutional block, which includes a sequentially connected dilated convolutional layer, a weight normalization layer, an activation function layer, and a dropout layer. The dilated convolutional layers within each block have residual connections. The dilated convolutional layer introduces an appropriate dilation factor *d* and kernel size. A smaller kernel size is better suited for capturing short-term, high-frequency features, while a larger kernel size can capture longer-term, low-frequency features. The weight normalization layer normalizes the weights in the convolutional layers, making the optimization process more stable and faster. The activation function layer introduces a non-linear ReLU activation function, enabling the network to learn and represent more complex patterns and relationships. The dropout layer temporarily removes some neurons from the network during training, preventing the model from becoming overly sensitive to specific training data. Residual connections allow the network's input to "skip" one or more layers, merging with the output of that layer through addition.

[0066] Feature extraction: First, the feature information of the time series data is captured through dilated convolutional layers. Then, the weights in the dilated convolutional layers are normalized through weight normalization layers. Next, non-linear processing is performed through activation function layers. Finally, regularization is performed through Dropout layers. The above feature extraction process is repeated and superimposed, and residual processing is performed. Finally, the feature information of the time series data is output.

[0067] Temporal convolutional neural network models study feature data that changes over time. It can be seen that the value of each layer in the *ti*-th time window depends only on the values ​​of the previous layer in the *ti*-th, *ti-1*, *ti-2*, ... time windows. This is a characteristic of causal convolution. Furthermore, the extraction of information from the previous layer by each layer can be leapfrog, growing exponentially by a size *d*. This is a characteristic of dilated convolution. Using dilated convolution requires padding with zeros at each layer, with a padding size of (k-1)d. Each information node in a dilated convolutional layer can fuse information from the initial input feature data within the range of (K-1)d-1, where K is the kernel size and d is the dilation factor. In X... ti Convolution at point is defined as F is the filter, F = (f1, f2, ..., f...). k ), X ti This represents the feature data in the ti-th time window. K represents the time delay or number of historical steps considered by the filter; w represents the weight parameter; s represents the different time delay steps of the filter; and finally, the output layer outputs the feature data X. new =(x t0 x t1 , ......, x tn At this time, X new (x tn It has already absorbed the feature information from the previous time steps. Through multiple operations, it captures complex time dependencies and extracts data feature information.

[0068] In one example, to improve data processing efficiency and enhance the model's responsiveness to abnormal dynamics and sudden events, this invention, in the dilated convolutional layer, borrows the idea of ​​multi-head attention mechanism, and improves the originally fixed convolutional kernel to implement dynamic convolution, in order to capture long-term, low-frequency feature information and short-term, high-frequency feature information of time-series data, including the following steps:

[0069] For each time point, multiple Q and K vectors are generated, where Q is represented by the features of the current time point and K is represented by the features of historical time points;

[0070] The similarity between vectors Q and K is calculated, and then normalized to obtain the attention weights. An analysis of variance is performed on the attention weights. If the variance is greater than a threshold, the first convolutional kernel (smaller convolutional kernel) is used to perform the analysis of variance on the attention weights; otherwise, the second convolutional kernel (larger convolutional kernel) is used to perform the analysis of variance on the attention weights. The first convolutional kernel is smaller than the second convolutional kernel.

[0071] Specifically, multiple Q and K vectors are generated for each time node. Q is represented by the features of the current time node, and K is represented by the features of historical nodes X. ti-1 X ti-2 X ti-3 X ti-4 Express it as Q = X ti-1 K1 = X ti-1 K2 = X ti-2 K3 = X ti-3 K4 = X ti-4 The similarity between vectors Q and K is calculated, and then the attention weight S is obtained by softmax normalization. i S i =softmax(QK) iNext, an analysis of variance is performed on the attention weights. Larger variance indicates that the weights are concentrated at a few time points, requiring a smaller convolutional kernel. Smaller variance indicates that the data weights are widely distributed and the data changes relatively smoothly, thus requiring a larger convolutional kernel. Then, an activation function is added to simulate a non-linear process. A weight normalization layer addresses the gradient explosion problem, and a Dropout layer prevents overfitting. This process is repeated and stacked, and residual connections are then used to construct a temporal convolutional neural network model.

[0072] In one example, before the step of learning a compressed feature representation of feature information through an information bottleneck, a step of generating future process data is also included:

[0073] Train a diffusion model using the feature information of time series data;

[0074] Noisy data is generated by gradually adding noise to the feature vector at each time step through the noise prediction network in the diffusion model.

[0075] By using a denoising network in the diffusion model to predict and remove noise at each time step, feature data is gradually recovered, thereby generating future process data.

[0076] Specifically, with X new =(x t0 x t1 , ......, x tn-1 The diffusion model is trained by progressively adding noise to the feature vector at each time step t, generating noisy data x. t , in It is a parameter for controlling noise intensity, ∈ t It is noise that follows a normal distribution. α t =1-β t , parameter β t β represents the noise intensity used to control each time step t. t As the noise data gradually increases during the forward diffusion process, its proportion gradually increases; as the time step changes, β... t As the noise level increases, more and more noise data is introduced during the forward diffusion process. Then, a noise prediction network is trained to predict the noise level at each time step t. θ (x t loss function (t) Mean squared error measures the difference between the predicted noise and the actual added noise. The network is trained to minimize the prediction error, while the reverse denoising process involves using the denoising network to predict and remove noise at each time step t, gradually recovering the feature data. Where Z is random noise distributed in a standard normal pattern, σ tThe noise standard deviation at time step t is used to generate future process data, with X as the input. new (x tn X is calculated using a noise predictor. new (x tn The noise data is used to generate future process data x based on the derived formula. tn ′.

[0077] In one example, a compressed feature representation of future process data is learned through an information bottleneck, including:

[0078] Future process data x tn The relevant information between two variables, X and Y, in the actual dementia data is defined as mutual information, (X; Y), and it is assumed that x... tn There is a statistical correlation between Y and Y, and X represents the input set of future process data;

[0079] In this case, Y implicitly determines x. tn Relevant and irrelevant features in ' are minimized Excluding irrelevant data that does not contribute to the prediction of Y for x tn The information in ' is compressed to obtain Capture with x tn The relevant features are used to obtain the filtered features; among them It is the minimum sufficient statistic of X relative to Y, and β represents the compression parameter.

[0080] Specifically, because the feature information of time series still contains some features irrelevant to the classification task, this invention inputs the feature data processed by the temporal convolutional neural network model into the information bottleneck module for feature filtering. The principle of the information bottleneck is to compress the information in X by excluding irrelevant parts that do not contribute to the prediction of Y, and to use the relevant parts of X relative to Y... express, It is the minimum sufficient statistic of X with respect to Y, assuming a Markov chain. Under constraints Minimize Therefore, the optimization objective of the information bottleneck is to minimize it: β is a trade-off parameter. The formula for calculating mutual information is: p(x|y) represents the joint marginal probability density function of random variables X and Y; p(x) represents the marginal probability density function of random variable X; p(x, y) represents the joint probability density function. Each volunteer has data from tn time windows, therefore the optimization objective of this invention can be expressed as: λ ti This represents the weighting parameter.

[0081] In the calculation of mutual information, this invention utilizes the nonnegativity of KL divergence and Bayesian transform approximation to calculate mutual information. The final loss function LI for the information bottleneck module is obtained as follows:

[0082]

[0083] Where N is the number of cases, E ∈ E ∈t1 q represents the expected value; q represents the conditional probability distribution. In a given compressed representation Under the given conditions, the negative log-likelihood of the conditional probability distribution q of the input y; β represents the trade-off parameter used to balance the contributions of different terms in the loss function; Given the original data x, we obtain the compressed representation. The conditional probability; express The prior distribution is then determined. Finally, the filtered feature vectors are input into a fully connected layer (classification prediction network) to obtain the disease progression prediction result.

[0084] In one example, the method also includes:

[0085] The loss of the information bottleneck module and the loss of the prediction network are trained together and adjusted with weight factors to obtain the overall loss.

[0086] Specifically, the overall loss expression is as follows:

[0087]

[0088] in, H represents the loss for classification prediction; H represents the entropy function; α represents the weighting parameter of the loss function.

[0089] A vascular cognitive impairment prediction system based on a temporal convolutional network and information bottleneck is constructed by training according to the loss function. Figure 3 As shown, the system includes a sequentially connected temporal convolutional network, a diffusion model, an information bottleneck module, and a fully connected layer. By using this system to predict patient data, the prediction results of disease progression can be obtained, which can greatly promote the research of deep learning in the early screening task of vascular dementia.

[0090] Combining the above examples, a preferred example method of the present invention is obtained, which includes the following steps:

[0091] S1': Preprocess the time series data, including data grouping, missing value handling, outlier detection, data shaping, and normalization.

[0092] S2': Extract feature information of time series data based on time series data processing model. The time series data processing model can be any one or more of the following: temporal convolutional network, long short-term memory network, gated recurrent unit, and Transformer model.

[0093] S3': Generates future process data based on the feature information of time series data using a diffusion model;

[0094] S4': Learn compressed feature representations of future process data by leveraging information bottlenecks;

[0095] S5': Based on the compressed feature representation, a prediction is made to obtain the predicted results of the disease process, including the speed of disease development, the severity of the disease, and the future state.

[0096] This invention proposes a method for predicting the progression of vascular cognitive impairment based on temporal convolutional networks and a diffusion model. It utilizes temporal convolutional neural networks to capture the temporal dependencies in data, tracking the relationship between long-term progression of case data and the eventual onset of the disease, thereby achieving feature extraction. Since disease progression prediction involves the analysis of data over a long period, to improve model stability and efficiency and more accurately capture sudden events and abnormal dynamics, this invention adds a process to the temporal convolutional neural network model that dynamically updates the convolutional kernel size based on the drastic changes in data. This improves processing efficiency while enhancing the model's responsiveness to abnormal dynamics and sudden events. The extracted features are then input as the initial input to the diffusion model. By progressively adding noise, the data is transformed from its original distribution to a high-noise distribution. Starting with completely noisy data, noise is gradually removed to generate future disease progression data, thus achieving disease progression prediction. However, because medical data typically contains a large number of features, and these features may exhibit high correlation, multicollinearity can affect the model's stability and predictive ability. Therefore, this invention utilizes information bottlenecks to screen features closely related to the disease before performing disease progression prediction, improving the accuracy of the model's predictions. The screened features are then input into a fully connected layer to output the predicted disease progression.

[0097] Furthermore, since vascular cognitive impairment is a long-term developing disease, missing values ​​are inevitable in the dataset. Traditional missing value imputation methods (such as mean imputation and interpolation) may not be able to fully utilize the time dependence and complex patterns of the data. However, the diffusion model combined with the temporal convolutional network can capture the complex time dependence and dynamic changes in time series data, so that the imputed data can reflect the long-term trend of the time series. This invention can capture and generate high-quality missing parts through forward diffusion and backward denoising processes. Therefore, in the data preprocessing stage, this invention also introduces the diffusion model to perform missing value imputation on the data, thereby restoring the data and providing reliable basic data information for subsequent classification prediction.

[0098] An example of the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the vascular cognitive impairment prediction method formed by any or a combination of the above examples. The processor may be a single-core or multi-core central processing unit or a specific integrated circuit, or one or more integrated circuits configured to implement the present invention.

[0099] An example of the present invention also provides a storage medium having the same inventive concept as a vascular cognitive impairment prediction method formed by any or more of the above examples, wherein computer instructions are stored thereon, which, when executed, perform the steps of the vascular cognitive impairment prediction method formed by any or more of the above examples.

[0100] Based on this understanding, the technical solution of this embodiment, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0101] An example of the present invention also provides a terminal having the same inventive concept as any or a combination of examples corresponding to the above-described method for predicting vascular cognitive impairment, including a memory and a processor. The memory stores computer instructions executable on the processor, and the processor executes the steps of the above-described method for predicting vascular cognitive impairment when executing the computer instructions. The processor may be a single-core or multi-core central processing unit or a specific integrated circuit, or one or more integrated circuits configured to implement the present invention.

[0102] In one example, the terminal, i.e., the electronic device, is manifested in the form of a general-purpose computing device. The components of the electronic device may include, but are not limited to: at least one processing unit (processor) mentioned above, at least one storage unit mentioned above, and a bus connecting different system components (including storage units and processing units).

[0103] The storage unit stores program code that can be executed by the processing unit to perform the steps described in the "Exemplary Methods" section of this specification according to various exemplary embodiments of the present invention. For example, the processing unit can perform one of the above-described methods for predicting vascular cognitive impairment.

[0104] The storage unit may include a readable medium in the form of a volatile storage unit, such as a random access memory (RAM) 3201 and / or a cache storage unit, and may further include a read-only memory (ROM).

[0105] The storage unit may also include a program / utility having a set (at least one) of program modules, including but not limited to: an operating system, one or more application programs, other program modules, and program data, each or some combination of these examples may include an implementation of a network environment.

[0106] A bus can represent one or more of several types of bus structures, including a memory cell bus or memory cell controller, a peripheral bus, a graphics acceleration port, a processing unit, or a local bus that uses any of the various bus structures.

[0107] The electronic device can also communicate with one or more external devices (e.g., keyboards, pointing devices, Bluetooth devices, etc.), one or more devices that enable a user to interact with the electronic device, and / or any device that enables the electronic device to communicate with one or more other computing devices (e.g., routers, modems, etc.). This communication can be achieved through input / output (I / O) interfaces. Furthermore, the electronic device can communicate with one or more networks (e.g., local area networks (LANs), wide area networks (WANs), and / or public networks, such as the Internet) via a network adapter. The network adapter communicates with other modules of the electronic device via a bus. It should be understood that other hardware and / or software modules can be used in conjunction with the electronic device, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0108] Through the above description, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solution according to this exemplary embodiment can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, terminal device, or network device, etc.) to execute the method of the exemplary embodiment of this application.

[0109] The above detailed embodiments are a description of the present invention. It should not be considered that the specific embodiments of the present invention are limited to these descriptions. For those skilled in the art, several simple deductions and substitutions can be made without departing from the concept of the present invention, and all of these should be considered to fall within the protection scope of the present invention.

Claims

1. A method for predicting vascular cognitive impairment, characterized in that, Includes the following steps: Preprocessing of time-series data includes one or more of the following: data grouping, missing value handling, outlier detection, data shaping, and normalization; the time-series data consists of patients' medical records and / or imaging data. The feature information of time series data is extracted based on the time series data processing model. The time series data processing model can be any one or more of the following: temporal convolutional network, long short-term memory network, gated recurrent unit, and Transformer model. Learn compressed feature representations of feature information by addressing information bottlenecks; Predictions are made based on compressed feature representations to obtain predictions of disease progression.

2. The method for predicting vascular cognitive impairment according to claim 1, characterized in that, Missing value handling includes: The diffusion model was trained using time-series data without missing information. Noisy data is generated by gradually adding noise to the feature vector at each time step through the noise prediction network in the diffusion model. By using a denoising network in the diffusion model to predict and remove noise at each time step, the feature data is gradually restored, thereby achieving missing value imputation.

3. The method for predicting vascular cognitive impairment according to claim 1, characterized in that, When extracting feature information from temporal data using a temporal convolutional network, the following steps are included: Constructing a temporal convolutional network: The temporal convolutional network includes a sequentially connected input layer, hidden layer, and output layer. The hidden layer can be one or more layers. The hidden layer includes a convolutional block, which includes a sequentially connected dilated convolutional layer, a weight normalization layer, an activation function layer, and a dropout layer. The dilated convolutional layer in each convolutional block has residual connections. Feature extraction: First, the feature information of the time series data is captured through dilated convolutional layers. Then, the weights in the dilated convolutional layers are normalized through weight normalization layers. Next, non-linear processing is performed through activation function layers. Finally, regularization is performed through Dropout layers. The above feature extraction process is repeated and superimposed, and residual processing is performed. Finally, the feature information of the time series data is output.

4. The method for predicting vascular cognitive impairment according to claim 3, characterized in that, The dilated convolutional layer captures long-term, low-frequency feature information of time-series data based on dynamic convolution, including the following steps: For each time point, multiple Q and K vectors are generated, where Q is represented by the features of the current time point and K is represented by the features of historical time points; The similarity between vectors Q and K is calculated, and then normalized to obtain the attention weights. An analysis of variance is performed on the attention weights. If the variance is greater than a threshold, the first convolutional kernel is used to perform the analysis of variance on the attention weights; otherwise, the second convolutional kernel is used to perform the analysis of variance on the attention weights. The first convolutional kernel is smaller than the second convolutional kernel.

5. The method for predicting vascular cognitive impairment according to claim 1, characterized in that, Before the step of learning the compressed feature representation of feature information through the information bottleneck, there is also a step of generating future process data: Train a diffusion model using the feature information of time series data; Noisy data is generated by gradually adding noise to the feature vector at each time step through the noise prediction network in the diffusion model. By using a denoising network in the diffusion model to predict and remove noise at each time step, feature data is gradually recovered, thereby generating future process data.

6. The method for predicting vascular cognitive impairment according to claim 5, characterized in that, Learning compressed feature representations of future process data through information bottlenecks includes: Future process data x tn The relevant information between the two variables, Y and Y, representing the actual dementia situation data, is defined as mutual information I(X; Y), where X represents the input set of future process data. By minimizing Excluding irrelevant data that does not contribute to the prediction of Y for x tn The information in ' is compressed to obtain Capture with x tn The relevant features are used to obtain the filtered features; among them It is the minimum sufficient statistic of X relative to Y, and β represents the compression parameter.

7. The method for predicting vascular cognitive impairment according to claim 1, characterized in that, The method further includes: The loss of the information bottleneck module and the loss of the prediction network are trained together and adjusted with weight factors to obtain the overall loss.

8. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of a method for predicting vascular cognitive impairment as described in any one of claims 1-7.

9. A terminal comprising a memory and a processor, wherein the memory stores computer instructions executable on the processor, characterized in that, When the processor executes the computer instructions, it performs the steps of the method for predicting vascular cognitive impairment as described in any one of claims 1-7.

Images