Alzheimer's disease condition score prediction method

Abstract

The invention relates to the technical field of condition prediction, and particularly discloses a condition score prediction method for Alzheimer's disease, which comprises the following steps: S1, acquiring longitudinal multi-time-point MRI image data and corresponding clinical score data of a target subject; s2, processing to obtain a feature matrix of the brain region and brain region data and clinical scores at different time points; s3, performing sparse representation on the feature matrix by using a dictionary learning algorithm to obtain a sparse representation matrix; s4, constructing an SLSTM, taking the sparse representation matrixes of the baseline moment and the M06 and M12 time points as input, and training a model to predict clinical scores of subsequent time points M18, M24 and M36; according to the method, sparse representation is performed on MRI image data through dictionary learning, redundant features are removed, features highly related to disease change are extracted, and a long-term dependency relationship in a longitudinal time sequence is captured in combination with a double-layer stacked LSTM structure, so that the model can more accurately predict clinical scores of future time points.

Description

Technical Field

[0001] The present invention belongs to the technical field of disease prediction, and in particular relates to a method for predicting Alzheimer's disease disease score. Background Art

[0002] Alzheimer's disease (AD), as a complex neurodegenerative disease, is characterized by the gradual loss and functional decline of brain nerve cells. Its fundamental pathological mechanism is still the focus of scientific exploration, and the treatment after diagnosis faces the dilemma of being unable to cure it. The course of AD usually starts quietly from mild cognitive impairment (MCI) and gradually deteriorates to a comprehensive impact on daily life ability. Early symptoms such as amnesia are often underestimated or ignored until the disease deteriorates significantly, and then it attracts attention. At this time, the treatment options are relatively limited. With the acceleration of global population aging, the number of dementia cases has risen sharply. It is estimated that by 2050, the number of dementia patients in the world will exceed 150 million, of which Alzheimer's patients will account for half of the country, which will put great pressure on the social and economic system. On the economic level, the cost of treatment and care for AD has become a burden that cannot be ignored. In 2018, the global expenditure on this has exceeded the trillion US dollar mark, and it is predicted that this figure will double to a staggering 2 trillion US dollars by 2030. What is more serious is the lack of effective prevention and treatment methods for AD worldwide. Coupled with its incurability, patient care and drug development have become a heavy burden on society.

[0003] In China, one of the countries with the fastest growth rates of AD, this challenge is particularly urgent. However, what is worrying is that there are still many misunderstandings about AD among the public and even medical professionals, making it difficult for patients to be accurately identified and effectively intervened in the early stages of the disease. At present, it is still challenging to distinguish between normal aging and AD prodromal states, and early intervention methods are limited, which further exacerbates social attention to the AD issue. Therefore, improving public awareness and strengthening research on early diagnosis and intervention strategies have become important issues that need to be urgently addressed by all sectors of society.

[0004] In the clinical assessment process of Alzheimer's disease, doctors rely on comprehensive neuroimaging technology (covering MRI, fMRI, etc.) and standardized dementia assessment scales (such as MMSE, ADAS-Cog, CDR) to comprehensively examine the condition. The core of this process is the accurate acquisition of clinical scores, which must be performed in person by professionally qualified neurologists. However, the current reality facing China is the significant contradiction between the surge in the number of patients and the scarcity of professional physician resources, which makes efficient and accurate assessment work face huge challenges. In addition, the process of collecting clinical scale scores is not easy. It requires doctors to invest a lot of time in detailed assessments, and the patient's cooperation and compliance often become key factors affecting the quality of the score, further increasing the difficulty of the score.

[0005] In this regard, the inventors propose a method for predicting Alzheimer's disease symptom score to solve the above problems. Summary of the invention

[0006] The purpose of the present invention is to provide a method for predicting Alzheimer's disease condition score to solve the problems raised in the above background technology.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] A method for predicting Alzheimer's disease condition score, comprising:

[0009] S1. Obtain longitudinal multi-time-point MRI imaging data and corresponding clinical score data of target subjects from a public database;

[0010] S2, performing standardization processing on the MRI image data to extract the characteristic matrix of each brain region;

[0011] Preprocessing the clinical score data to obtain brain region data and clinical scores of each subject at different time points;

[0012] S3, using a dictionary learning algorithm to perform sparse representation on the feature matrix of the MRI image data at each time point to obtain a sparse representation matrix;

[0013] By optimizing the dictionary matrix and sparse representation matrix, important features related to the learning task are extracted, redundant information is removed, and model complexity is reduced;

[0014] S4. Construct a long short-term memory network model SLSTM with a double-layer stacking structure to capture the dependencies of time series and the correlation between brain regions in longitudinal MRI data.

[0015] Saving the sparse representation matrices at different time points, using the sparse representation matrices at baseline and at time points M06 and M12 as input, training the model to predict the clinical scores at subsequent time points M18, M24, and M36;

[0016] S5. The Pearson correlation coefficient R and mean absolute error MAE are used to evaluate the prediction results. The closer the R value is to 1 and the smaller the MAE value is, the better the model prediction performance is.

[0017] Compare the prediction performance based on different feature selections and model structures to determine the optimal model structure.

[0018] Preferably, the MRI image information includes data at a baseline time and multiple subsequent time points.

[0019] Preferably, the clinical scoring data include the Mini-Mental State Examination (MMSE), the Alzheimer's Disease Assessment Scale-Cognitive Part (ADAS-Cog), the Clinical Dementia Rating Scale-SOB, and the Clinical Dementia Rating Scale-Global Score (CDR-GLOB).

[0020] Preferably, the process of obtaining the sparse representation matrix by the dictionary learning algorithm is:

[0021] S21, set the longitudinal time point MRI image data as a matrix represents the d-dimensional MRI data of m subjects at time point t, then the data of baseline, month 06, month 12, month 18, month 24, and month 36 are represented by {X 1 , X 2 , X 3 , X 4 , X 5};

[0022] S22, using a dictionary learning algorithm to remove redundant features irrelevant to the current learning task, and converting the data into a sparse representation. The dictionary learning is solved by a variable alternating optimization method. After fixing the dictionary matrix, the sparse representation is updated, and then the sparse representation is fixed to update the dictionary matrix until convergence;

[0023] S23. The sparse representation of the dictionary matrix and the sample at each time point is obtained by using the K-singular value decomposition method with a column-by-column update strategy. The size of the dictionary is controlled by setting the size of the dictionary vocabulary k, thereby affecting the sparsity, and obtaining the sparse representation matrix.

[0024] Preferably, the dictionary learning algorithm takes the baseline time point as an example, assuming that its image data set is X 1 ={x 1 ,x 2 ,···,x i ,···x m},in The formula is as follows:

[0025]

[0026] Where k is the vocabulary size of the dictionary, is the dictionary matrix, For x i sparse representation of ;

[0027] For the above formula, the variable alternation optimization method is used to solve it:

[0028] First, the first step is to fix the dictionary B;

[0029] The second step is to use a iTo update the dictionary B with the initial value, equation (1.1) can be written as

[0030]

[0031] in ||·|| F is the Frobenius norm of the matrix.

[0032] Preferably, the K-singular value decomposition method of the column-by-column update strategy is used to solve the above formula (1.2) to obtain:

[0033]

[0034] Let b i represents the i-th column of the dictionary matrix B, a i represents the i-th row of the sparse matrix A. When updating the i-th column of the dictionary, the other columns are fixed, so E i =(X 1 -∑ j≠i b j a j ) is fixed, and a i Only non-zero elements are retained, E i Only keep b i and a i The product of the non-zero elements of , and then perform singular value decomposition, initialize B and iterate repeatedly, and finally obtain the dictionary B in the baseline state and the sparse representation A of the sample.

[0035] Preferably, the implementation process of the SLSTM is:

[0036] S41, forget gate f t Decide what information to discard from the cell state; the formula is as follows:

[0037] f t =σ(W f ·[h t-1 ,x t ]+b f )(1.5)

[0038] Among them, W f is the weight matrix, x t is the input information at time t. In this method, the input information is x t The sparse representation of a i ,h t-1 is the hidden state at time t-1, σ is the activation function sigmoid, and its value range is (0,1).

[0039] S42, input gate i t Decide what information should be added to the cell state, using ht-1 and x t To decide which information to update. Then use h t-1 and x t Get new candidate cell information through the tanh layer This information may be updated to the cell information.

[0040] i t =σ(W i ·[h t-1 ,x t ]+b i )=σ(W ih h t-1 +W ix x t ]+b i )(1.6)

[0041]

[0042] S43, memory unit update, where f t Decision C t The degree of retention, i t Determine the value of this update, update as follows

[0043]

[0044] S43, finally the output gate, which determines the output information through the sigmoid layer,

[0045] o t =σ(W o ·[h t-1 ,x t ]+b o ) (1.9)

[0046] S44, activate the cell state and dot product output through the tanh layer to obtain the final hidden state output,

[0047] h t =o t *tanh(C t ) (1.10)

[0048] S44, after obtaining the hidden state output, the output y is obtained through the softmax function t

[0049] y t =softmax(W y ·h t )

[0050] A double-layer stacked LSTM is used to pass the output of one LSTM layer as input to the next LSTM layer. Each LSTM layer has its own hidden state and memory unit. By connecting two LSTM layers, information transmission and shared connection can be achieved. The two stacked LSTM layers can enhance the expressiveness of the model, allowing the model to better capture long-term dependencies in sequence data.

[0051] Preferably, the model prediction ability is determined by calculating the Pearson correlation coefficient R and the mean absolute error MAE between the true score and the predicted score, which are defined as follows:

[0052]

[0053]

[0054] Where Y represents the actual clinical score, Indicates the predicted clinical score cov(·) is the covariance, σ(·) is the standard deviation, and the smaller the calculated MAE is, the better, and the closer R is to 1, the better.

[0055] Compared with the prior art, the present invention has the following beneficial effects:

[0056] (1) The present invention uses dictionary learning to perform sparse representation of MRI image data, remove redundant features, and extract features that are highly correlated with disease changes, effectively reducing the interference of noise on model prediction. Combined with a double-layer stacked LSTM structure, it captures the long-term dependencies in the longitudinal time series, enabling the model to more accurately predict clinical scores at future time points.

[0057] (2) The dictionary learning algorithm of the present invention reduces the feature dimension and optimizes the data representation through sparse representation, thereby significantly reducing the computational complexity of model training and reasoning. This not only improves the efficiency of model training, but also makes real-time prediction possible in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 The present invention is a flow chart of a method for predicting Alzheimer's disease condition score. DETAILED DESCRIPTION

[0059] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0060] Embodiment 1:

[0061] See also Figure 1 As shown, a method for predicting the condition score of Alzheimer's disease comprises:

[0062] S1. Obtain longitudinal multi-time-point MRI imaging data and corresponding clinical score data of target subjects from a public database;

[0063] S2, performing standardization processing on the MRI image data to extract the characteristic matrix of each brain region;

[0064] Preprocessing the clinical score data to obtain brain region data and clinical scores of each subject at different time points;

[0065] S3, using a dictionary learning algorithm to perform sparse representation on the feature matrix of the MRI image data at each time point to obtain a sparse representation matrix;

[0066] By optimizing the dictionary matrix and sparse representation matrix, important features related to the learning task are extracted, redundant information is removed, and model complexity is reduced;

[0067] S4. Construct a long short-term memory network model SLSTM with a double-layer stacking structure to capture the dependencies of time series and the correlation between brain regions in longitudinal MRI data.

[0068] Saving the sparse representation matrices at different time points, using the sparse representation matrices at baseline and at time points M06 and M12 as input, training the model to predict the clinical scores at subsequent time points M18, M24, and M36;

[0069] S5. The Pearson correlation coefficient R and mean absolute error MAE are used to evaluate the prediction results. The closer the R value is to 1 and the smaller the MAE value is, the better the model prediction performance is.

[0070] Compare the prediction performance based on different feature selections and model structures to determine the optimal model structure.

[0071] As can be seen from the above, dictionary learning is used to sparsely represent MRI image data, remove redundant features, and extract features that are highly correlated with disease changes, effectively reducing the interference of noise on model prediction. Combined with a double-layer stacked LSTM structure, the long-term dependencies in the longitudinal time series are captured, allowing the model to more accurately predict clinical scores at future time points;

[0072] The dictionary learning algorithm uses sparse representation to reduce feature dimensions and optimize data representation, thereby significantly reducing the computational complexity of model training and reasoning. This not only improves model training efficiency, but also makes real-time predictions possible in practical applications.

[0073] Embodiment 2:

[0074] refer to Figure 1 As shown, the MRI image information includes data at a baseline time and multiple subsequent time points.

[0075] Specifically, the clinical scoring data include the Mini-Mental State Examination (MMSE), the Alzheimer's Disease Assessment Scale-Cognitive Part (ADAS-Cog), the Clinical Dementia Rating Scale-SOB, and the Clinical Dementia Rating Scale-GLOB.

[0076] Specifically, the process of obtaining the sparse representation matrix by the dictionary learning algorithm is as follows:

[0077] S21, set the longitudinal time point MRI image data as a matrix represents the d-dimensional MRI data of m subjects at time point t, then the data of baseline, month 06, month 12, month 18, month 24, and month 36 are represented by {X 1 , X 2 , X 3 , X 4 , X 5};

[0078] S22, using a dictionary learning algorithm to remove redundant features irrelevant to the current learning task, and converting the data into a sparse representation. The dictionary learning is solved by a variable alternating optimization method. After fixing the dictionary matrix, the sparse representation is updated, and then the sparse representation is fixed to update the dictionary matrix until convergence;

[0079] S23. The sparse representation of the dictionary matrix and the sample at each time point is obtained by using the K-singular value decomposition method with a column-by-column update strategy. The size of the dictionary is controlled by setting the size of the dictionary vocabulary k, thereby affecting the sparsity, and obtaining the sparse representation matrix.

[0080] Specifically, the dictionary learning algorithm takes the baseline time point as an example, assuming that its image data set is X 1 ={x 1 ,x 2 ,···,x i ,···x m},in The formula is as follows:

[0081]

[0082] Where k is the vocabulary size of the dictionary, is the dictionary matrix, For x i sparse representation of ;

[0083] For the above formula, the variable alternation optimization method is used to solve it:

[0084] First, the first step is to fix the dictionary B;

[0085] The second step is to use a i To update the dictionary B with the initial value, equation (1.1) can be written as

[0086]

[0087] in ||·|| F is the Frobenius norm of the matrix.

[0088] Specifically, the K-singular value decomposition method of the column-by-column update strategy is used to solve the above formula (1.2) to obtain:

[0089]

[0090] Let b i represents the i-th column of the dictionary matrix B, a i represents the i-th row of the sparse matrix A. When updating the i-th column of the dictionary, the other columns are fixed, so E i =(X 1 -∑ j≠i b j a j ) is fixed, and a i Only non-zero elements are retained, E i Only keep b i and a i The product of the non-zero elements of , and then perform singular value decomposition, initialize B and iterate repeatedly, and finally obtain the dictionary B in the baseline state and the sparse representation A of the sample.

[0091] Specifically, the implementation process of the SLSTM is as follows:

[0092] S41, forget gate f t Decide what information to discard from the cell state; the formula is as follows:

[0093] f t =σ(W f ·[h t-1 ,x t ]+b f ) (1.5)

[0094] Among them, W f is the weight matrix, x t is the input information at time t. In this method, the input information is x t The sparse representation of a i ,h t-1is the hidden state at time t-1, σ is the activation function sigmoid, and its value range is (0,1).

[0095] S42, input gate i t Decide what information should be added to the cell state, using h t-1 and x t To decide which information to update. Then use h t-1 and x t Get new candidate cell information through the tanh layer This information may be updated to the cell information.

[0096] i t =σ(W i ·[h t-1 ,x t ]+b i )=σ(W ih h t-1 +W ix x t ]+b i )(1.6)

[0097]

[0098] S43, memory unit update, where f t Decision C t The degree of retention, i t Determine the value of this update, update as follows

[0099]

[0100] S43, finally the output gate, which determines the output information through the sigmoid layer,

[0101] o t =σ(W o ·[h t-1 ,x t ]+b o ) (1.9)

[0102] S44, activate the cell state and dot product output through the tanh layer to obtain the final hidden state output,

[0103] h t =o t *tanh(C t ) (1.10)

[0104] S44, after obtaining the hidden state output, the output y is obtained through the softmax function t

[0105] yt =softmax(W y ·h t )

[0106] A double-layer stacked LSTM is used to pass the output of one LSTM layer as input to the next LSTM layer. Each LSTM layer has its own hidden state and memory unit. By connecting two LSTM layers, information transmission and shared connection can be achieved. The two stacked LSTM layers can enhance the expressiveness of the model, so that the model can better capture the long-term dependencies in the sequence data. All data at the three time points of baseline, M06, and M12 are subjected to dictionary learning to remove redundant features irrelevant to the current learning task. At the same time, the data is converted into a suitable sparse representation form, and then the imaging data of all subjects at the three time points are sparsely represented. t As input, it is sent to the stacked LSTM network for learning to predict the scores at future time points (M18, M24, M36).

[0107] Specifically, the Pearson correlation coefficient R and the mean absolute error MAE between the true score and the predicted score are calculated to determine the model prediction ability, which is defined as follows:

[0108]

[0109]

[0110] Where Y represents the actual clinical score, Indicates the predicted clinical score cov(·) is the covariance, σ(·) is the standard deviation, and the smaller the calculated MAE is, the better, and the closer R is to 1, the better.

[0111] As can be seen above, by extracting features from the data at baseline, 6 months, and 12 months, the model can better learn the laws of disease progression. The double-layer stacked LSTM structure further enhances the model's ability to express features of data at different time points and improves its ability to predict new data;

[0112] Predicting clinical scores (such as MMSE, ADAS-Cog, CDR-SOB) at future time points can help identify potential trends in disease changes and provide support for personalized medical decisions. For example, for patients with a clear downward trend in scores, clinicians can adjust treatment plans in advance to delay disease progression;

[0113] By introducing the Pearson correlation coefficient (R) and mean absolute error (MAE) to evaluate the prediction effect, this method can comprehensively measure the model performance. R close to 1 indicates that the prediction result is highly correlated with the true value, and a small MAE indicates a low prediction error, ensuring the reliability of the model in different task scenarios;

[0114] This method is not only applicable to the clinical score prediction of Alzheimer's disease (AD), but can also be extended to the condition monitoring and progression prediction scenarios of other chronic diseases. By changing the input features and target variables, it can be applied to a variety of medical data analysis tasks.

[0115] Embodiment three:

[0116] This design is specifically applied to the field of Alzheimer's disease (AD) assessment. Alzheimer's disease is a neurodegenerative disease that seriously affects the quality of life of patients. Its disease assessment and prediction are of great significance to disease management. Using the above method, a model for predicting the patient's future disease score (such as MMSE, ADAS-Cog, CDR-SOB, etc.) is constructed and verified through longitudinal MRI imaging data and clinical score data to provide a more accurate assessment of disease progression.

[0117] Specifically, longitudinal multi-time point data of 805 participants were downloaded from the public database (Alzheimer's Disease Neuroimaging Initiative) ADNI, including MRI imaging information and clinical scores at month 06 (M06), month 12 (M12), month 18 (M18), month 24 (M24) and month 36 (M36). It is particularly noteworthy that the baseline MRI data set included a complete set of 805 samples, which were divided into 186 patients with Alzheimer's disease (AD), 393 individuals with mild cognitive impairment (MCI), and 226 healthy controls (Normal Control, NC) according to diagnostic groups. In addition, in order to comprehensively evaluate the changes in the participants' cognitive status, we also obtained four widely recognized clinical rating scales from the participants, namely the Mini-Mental State Examination (MMSE), the Alzheimer's Disease Assessment Scale-Cognitive Part (ADAS-Cog), the Clinical Dementia Rating Scale Total Score (CDR-SOB), and the Clinical Dementia Rating Scale Global Score (CDR-GLOB). These scales are used by professional physicians and provide a quantitative and scientific basis for the evaluation of each subject's cognitive function. The dataset used in this study comes from the ADNI public database, including baseline MRI t1-weighted (T1w) data and four clinical scores (ADAS-Cog, CDR-GLOB, CDR-SOB, MMSE). At the same time, the subjects' follow-up MRI and scale scores at 06 months (M06), 12 months (M12), 18 months (M18), 24 months (M24), and 36 months (M36) were obtained. The method proposed in this article is as follows Figure 1As shown in the figure, first, MRI image data and four clinical scores at six time points (baseline, M06, M12, M18, M24, M36) are obtained, and data preprocessing is performed to obtain brain region data and clinical scores of each subject at different time points; then, matrix sparseness is performed through traditional dictionary learning (dictionary learning, DL). Through sparse representation, not only can the learning task be simplified, but also key data features can be extracted; considering the mutual correlation between brain regions and the correlation of image data at longitudinal time points, a stacked long short-term memory network (Long Short Term Memory, LSTM) is used for regression training, and fitting is performed through loss function to achieve clinical score prediction at future time points. The overall prediction score method includes two parts: dictionary learning (DL) and regression prediction of stacked long short-term memory network (SLSTM), so the proposed method is referred to as DLSLSTM model.

[0118] Further, S1: Data Acquisition

[0119] Further selected from the above ADNI database (Alzheimer's Disease Neuroimaging Initiative):

[0120] Number of subjects: 100 (50 AD patients, 30 mild cognitive impairment (MCI) patients, 20 healthy controls (HC))

[0121] MRI imaging data time points: baseline (BL), 6 months (M06), 12 months (M12), 18 months (M18), 24 months (M24), 36 months (M36)

[0122] Corresponding clinical scoring data:

[0123] Mini-Mental State Examination (MMSE)

[0124] Alzheimer's Disease Assessment Scale-Cognitive (ADAS-Cog)

[0125] Clinical Dementia Rating Scale-SOB (CDR-SOB)

[0126] S2: Data preprocessing

[0127] MRI image standardization:

[0128] The grayscale of MRI images was normalized.

[0129] Extract key brain region features (such as hippocampus, cortical thickness, etc.) to form a brain region feature matrix at each time point.

[0130] Clinical scoring standardization:

[0131] Handle missing values ​​(using linear interpolation).

[0132] Normalize the score range to fit the model input.

[0133] S3: Feature extraction (dictionary learning)

[0134] Using sparse representation method:

[0135] initialization:

[0136] Determine the shape of the MRI feature matrix at each time point: md (m is the number of subjects, d is the feature dimension).

[0137] Initialize dictionary matrix and the sparse representation matrix k is the dictionary vocabulary size (set k=64).

[0138] Iterative Optimization:

[0139] Through the variable alternation optimization method, fix B to update A, and then fix A to update B until convergence.

[0140] The K-singular value decomposition method using a column-by-column update strategy accelerates convergence.

[0141] Finally, the sparse representation matrix A at each time point is generated t , which can be used as input for subsequent models.

[0142] S4: SLSTM model construction and training

[0143] Model structure:

[0144] Two-layer stacked LSTM: input layer (sparse representation matrix A{BL}, A{M06}, A{M12}) → first layer LSTM → second layer LSTM → output layer (predicted clinical scores Y{M18}, Y{M24}, Y{M36}).

[0145] Number of hidden units in each LSTM layer: 128.

[0146] Training process:

[0147] Loss function: Mean Squared Error (MSE).

[0148] Optimizer: Adam, learning rate is set to 0.001.

[0149] Batch size: 32, number of training rounds: 100.

[0150] Data split: 70% for training, 20% for validation, and 10% for testing.

[0151] S5: Result evaluation

[0152] The Pearson correlation coefficient (R) and mean absolute error (MAE) were used to evaluate the model performance:

[0153] Experimental results:

[0154] MMSE Prediction:

[0155] R=0.91

[0156] MAE = 1.23

[0157] ADAS-Cog predictions:

[0158] R=0.88

[0159] MAE = 2.15

[0160] CDR-SOB prediction:

[0161] R=0.85

[0162] MAE=0.98

[0163] Comparative analysis:

[0164] Removing the dictionary learning module: R decreases by about 8% and MAE increases by about 15%.

[0165] Using a single-layer LSTM: R drops by about 5%, and MAE increases by about 10%.

[0166] From the above, we can see that this method significantly improves the prediction accuracy of clinical scores at future time points, especially after removing redundant features through dictionary learning, the model complexity is reduced and the prediction effect is improved; the experimental results show that the Pearson correlation coefficient (R) of the model reaches 0.91, and the prediction error (MAE) is significantly lower than that of the traditional method.

[0167] This model can provide an effective tool for monitoring the condition of AD patients and assist doctors in optimizing treatment plans.

[0168] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for predicting Alzheimer's disease condition score, characterized in that: include: S1. Obtain longitudinal multi-time-point MRI imaging data and corresponding clinical score data of target subjects from a public database; S2, performing standardization processing on the MRI image data to extract the characteristic matrix of each brain region; Preprocessing the clinical score data to obtain brain region data and clinical scores of each subject at different time points; S3, using a dictionary learning algorithm to perform sparse representation on the feature matrix of the MRI image data at each time point to obtain a sparse representation matrix; By optimizing the dictionary matrix and sparse representation matrix, important features related to the learning task are extracted, redundant information is removed, and model complexity is reduced; S4. Construct a long short-term memory network model SLSTM with a double-layer stacking structure to capture the dependencies of time series and the correlation between brain regions in longitudinal MRI data. Saving the sparse representation matrices at different time points, using the sparse representation matrices at baseline and at time points M06 and M12 as input, training the model to predict the clinical scores at subsequent time points M18, M24, and M36; S5. The Pearson correlation coefficient R and mean absolute error MAE are used to evaluate the prediction results. The closer the R value is to 1 and the smaller the MAE value is, the better the model prediction performance is. Compare the prediction performance based on different feature selections and model structures to determine the optimal model structure.

2. A method for predicting Alzheimer's disease condition score according to claim 1, characterized in that: The MRI image information includes data at a baseline time and subsequent multiple time points.

3. The method for predicting Alzheimer's disease condition score according to claim 1, characterized in that: The clinical scoring data include Mini-Mental State Examination (MMSE), Alzheimer's Disease Assessment Scale-Cognitive Part (ADAS-Cog), Clinical Dementia Rating Scale-SOB and Clinical Dementia Rating Scale-GLOB.

4. A method for predicting Alzheimer's disease condition score according to claim 1, characterized in that: The process of obtaining the sparse representation matrix by the dictionary learning algorithm is as follows: S21, set the longitudinal time point MRI image data as a matrix represents the d-dimensional MRI data of m subjects at time point t, then the data of baseline, month 06, month 12, month 18, month 24, and month 36 are represented by {X1, X2, X3, X4, X5}; S22, using a dictionary learning algorithm to remove redundant features irrelevant to the current learning task, and converting the data into a sparse representation. The dictionary learning is solved by a variable alternating optimization method. After fixing the dictionary matrix, the sparse representation is updated, and then the sparse representation is fixed to update the dictionary matrix until convergence; S23. The sparse representation of the dictionary matrix and the sample at each time point is obtained by using the K-singular value decomposition method with a column-by-column update strategy. The size of the dictionary is controlled by setting the size of the dictionary vocabulary k, thereby affecting the sparsity, and obtaining the sparse representation matrix.

5. A method for predicting Alzheimer's disease condition score according to claim 4, characterized in that: The dictionary learning algorithm takes the baseline time point as an example, assuming that its image data set is X1 = {x1, x2, ···, x i ,···x m },in The formula is as follows: Where k is the vocabulary size of the dictionary, is the dictionary matrix, For x i sparse representation of ; For the above formula, the variable alternation optimization method is used to solve it: First, the first step is to fix the dictionary B; The second step is to use a i To update the dictionary B with the initial value, equation (1.1) can be written as in ||·|| F is the Frobenius norm of the matrix.

6. A method for predicting Alzheimer's disease condition score according to claim 4, characterized in that: The K-singular value decomposition method of the column-by-column update strategy is used to solve the above equation (1.2) to obtain: Let b i represents the i-th column of the dictionary matrix B, a i represents the i-th row of the sparse matrix A. When updating the i-th column of the dictionary, the other columns are fixed, so E i =(X1-∑ j≠i b j a j ) is fixed, and a i Only non-zero elements are retained, E i Only keep b i and a i The product of the non-zero elements of , and then perform singular value decomposition, initialize B and iterate repeatedly, and finally obtain the dictionary B in the baseline state and the sparse representation A of the sample.

7. A method for predicting Alzheimer's disease condition score according to claim 4, characterized in that: The implementation process of the SLSTM is as follows: S41, forget gate f t Decide what information to discard from the cell state; the formula is as follows: f t =σ(W f ·[h t-1 ,x t ]+b f )(1.5) Among them, W f is the weight matrix, x t is the input information at time t, and the input information is x t The sparse representation of a i ,h t-1 is the hidden state at time t-1, σ is the activation function sigmoid, and its value range is (0,1). S42, input gate i t Decide what information should be added to the cell state, using h t-1 and x t To decide which information to update. Then use h t-1 and x t Get new candidate cell information through the tanh layer This information may be updated to the cell information. i t =σ(W i ·[h t-1 ,x t ]+b i )=σ(W ih h t-1 +W ix x t ]+b i )(1.6) S43, memory unit update, where f t Decision C t The degree of retention, i t Determine the value of this update, update as follows S43, output gate, determines the output information through the sigmoid layer, the t =σ(W o ·[h t-1 ,x t ]+b o )(1.9) S44, activate the cell state and dot product output through the tanh layer to obtain the final hidden state output, h t =o t *tanh(C t )(1.10) S44, after obtaining the hidden state output, the output y is obtained through the softmax function t y t =softmax(W y ·h t ) A double-layer stacked LSTM is used to pass the output of one LSTM layer as input to the next LSTM layer. Each LSTM layer has its own hidden state and memory unit. By connecting two LSTM layers, information transmission and shared connection can be achieved. The two stacked LSTM layers can enhance the expressiveness of the model, allowing the model to better capture long-term dependencies in sequence data.

8. The method for predicting Alzheimer's disease condition score according to claim 1, characterized in that: The model prediction ability is judged by calculating the Pearson correlation coefficient R and the mean absolute error MAE between the true score and the predicted score, which are defined as follows: Where Y represents the actual clinical score, Indicates the predicted clinical score cov(·) is the covariance, σ(·) is the standard deviation, and the smaller the calculated MAE is, the better, and the closer R is to 1, the better.

Images