A Deep Learning Approach Based on Attention for Predicting AD Transition Probability

Abstract

This invention belongs to the field of data mining technology, specifically relating to a method for predicting the conversion probability of Alzheimer's disease (AD) based on attention-based deep learning. The method includes: acquiring various examination data of the patient to be predicted; dividing the data into follow-up data and baseline data according to follow-up features and baseline features; preprocessing the follow-up data and inputting it into a constructed base model to obtain a follow-up data representation; combining the follow-up data representation with the baseline data and inputting it into a final decision-maker to obtain the probability that the patient will convert from MCI to AD in the next visit. Based on the principle of attention mechanisms, this invention designs a variant of attention for current time-series data prediction scenarios, assigning weights to different features and time points to capture the relationships between features and time points, thereby determining the importance of different features and time points.

Description

Technical Field

[0001] This invention belongs to the field of data mining technology, specifically relating to a method for predicting AD transition probability using deep learning based on attention. Background Technology

[0002] Alzheimer's disease (AD) is a neurodegenerative disease closely related to age. It is estimated that the prevalence is approximately 26.4% in people aged 65-74, increasing to 38.6% in those aged 75-84, and reaching 35.4% in those aged 85 and older. Scientific research indicates that the main pathogenic factors of AD are senile plaques formed by the deposition of β-amyloid protein (Aβ) in the brain and neurofibrillary tangles caused by excessive phosphorylation of Tau protein. The accumulation of these abnormal proteins interferes with normal communication in the nervous system, leading to a gradual decline in cognitive function. Mild cognitive impairment (MCI) is considered a potential early manifestation of AD. Although only about 10%-12% of MCI patients eventually develop AD each year, it affects approximately 22% of people over 65 years of age, with about half of MCI cases caused by AD. In addition, about 22% of people aged 65 or older are in the preclinical AD state, meaning that their cognitive function appears normal, but they already have obvious pathological signs of AD, especially among the elderly aged 90 and older, where this proportion is as high as 50%.

[0003] Given the irreversibility of Alzheimer's disease (AD) and the high cost of care, it is of great significance to use deep learning methods to accurately predict the progression of Alzheimer's disease in its early stages.

[0004] In recent years, many studies have used artificial intelligence methods to study Alzheimer's disease (AD). However, a single modality can only provide information on a specific aspect, and some modalities may not be sensitive enough for early AD detection. Furthermore, baseline non-time-series data cannot predict progression. Therefore, using multimodal time-series data to predict Alzheimer's disease progression has many advantages, as it can overcome the limitations of single-modality and non-time-series data. Multimodal fusion methods can be broadly classified into three categories: early fusion, late fusion, and mid-stage fusion.

[0005] Early fusion, also known as feature-level fusion, refers to fusing data from different modalities in the early stages of model development before feeding it into the model for training. Its advantage is that it can capture low-level correlations between different modalities, but its disadvantages are also obvious: the fused features generated after transformation and scaling of each modality's features usually have high dimensionality, increasing model complexity and computational cost. In 2022, Xu L et al. proposed an early fusion method to predict the progression of Alzheimer's disease using multimodal time-series data. They processed different modalities into potential learning representations, then imputed the data, and finally made predictions, achieving good prediction results. In the same year, El-Sappagh S et al. proposed a two-stage deep learning model to predict the progression of Alzheimer's disease, utilizing neuroimaging data, CS, CSF biomarkers, neuropsychological battery biomarkers, and demographic data to explore the impact of different time series on the results. In the first stage, they performed multi-class predictions for CN, MCI, and AD; in the second stage, they predicted the exact conversion from MCI to AD as a regression task, achieving better results than the comparison model. Deep learning models generally lack interpretability, and interpretability is a particularly important issue in medical settings. This study utilizes attention mechanisms to enhance the interpretability of models.

[0006] Late-stage fusion, also known as decision-level fusion, integrates predictions from different modalities in the later stages of model development. Its advantages include independent processing of each modality, simple model training, and ease of integration. It can easily handle data asynchrony, the entire system can scale with the number of modalities, and each modality's dedicated prediction model can better model that modality. It can also make predictions even when certain modalities are missing from the model input. The disadvantage is that it may not capture the interaction information between different modalities. Huang SC et al. found that in predicting pulmonary embolism using CT scans and EHR data, averaging-based late-stage fusion with regularized DNNs as sub-models outperformed early, mid-stage, and other late-stage fusion strategies. In the field of Alzheimer's disease prediction, Qiu S et al. found that effectively combining MRI data with other non-imaging data to predict AD progression is still relatively limited. They used the NACC dataset to predict Alzheimer's disease progression using different algorithms on MRI scans, MMSE (Mini-Mental State Examination), and LM (Logical Memory) data, and then integrated the results using majority voting, achieving better results than individual models.

[0007] Mid-stage fusion integrates features from different modalities at intermediate layers. This typically involves feature interaction and fusion at intermediate layers of the model, such as through attention mechanisms or shared network layers. Its advantages include capturing intermediate-level correlations between different modalities and better balancing the strengths and weaknesses of early and late-stage fusion. The disadvantage is its complexity in implementation, requiring a well-designed fusion mechanism. Lee G et al. proposed a mid-stage fusion method for predicting Alzheimer's disease, using GRU to predict AD transitions for different modalities separately, and then integrating the results using logistic regression, achieving superior results compared to the baseline.

[0008] In summary, the current limitation of multimodal fusion lies in its neglect of cross-modal interactions and the importance of various modalities in different tasks. Furthermore, the use of time-series data can also lead to suboptimal model results by ignoring the importance of different time points. Summary of the Invention

[0009] To address the aforementioned technical problems, this invention provides a method for predicting AD transition probabilities using attention-based deep learning, comprising:

[0010] S1: Obtain the various examination data of the patient to be predicted, and divide the data into follow-up data and baseline data according to follow-up characteristics and baseline characteristics;

[0011] S2: After preprocessing the follow-up data, it is input into the constructed base model to obtain the follow-up data representation;

[0012] S3: The follow-up data is combined with the baseline data and input into the final decision-maker to obtain the probability that the patient to be predicted will be converted from MCI to AD at the next visit.

[0013] The beneficial effects of this invention are:

[0014] This invention, based on the principle of attention mechanisms, proposes a variant of attention for prediction scenarios involving time series data. This variant can be considered a form of local attention, assigning weights to different features and time points to capture the relationships between features and time points, thus determining the importance of different features and time points. For small sample data, this simplified operation reduces computational complexity, eliminating the need to calculate attention scores during each forward propagation and saving computational costs. Implementation on the ADNI dataset demonstrates that this invention outperforms traditional single machine learning models in terms of accuracy in early AD diagnosis prediction and model generalization ability. This invention boasts high prediction accuracy and practicality. Early diagnosis of Alzheimer's patients helps in timely preparation and preventative measures to slow disease progression and improve patient survival rates, demonstrating promising application prospects. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the structure of the AD conversion probability model predicted by the attention-based deep learning method in this invention;

[0016] Figure 2 This is a schematic diagram of data preprocessing in this invention;

[0017] Figure 3 This is a schematic diagram illustrating the accuracy of the invention in different scenarios;

[0018] Figure 4 This is a schematic diagram of the F2 score for different scenarios in this invention. Detailed Implementation

[0019] The technical solutions of the embodiments 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.

[0020] This invention proposes an attention-based deep learning method for predicting AD transition probabilities, such as... Figure 1 As shown, the method includes: acquiring various examination data of the patient to be predicted, dividing the data according to follow-up characteristics and baseline characteristics, performing a series of preprocessing on the follow-up data, inputting it into the constructed base model to obtain the follow-up data representation, and then combining it with the baseline data to input into the final decision-maker to obtain the probability that the patient to be predicted will be converted from MCI to AD in the next visit;

[0021] The process of building an attention-based deep learning model to predict AD transition probabilities includes:

[0022] S1: Obtain patient's various examination data and classify them according to follow-up characteristics and baseline characteristics.

[0023] We selected a dataset that reflects real-world production environments to predict the probability of AD risk. We chose the ADNI database. The Alzheimer's Disease Neuroimaging Initiative (ADNI) is a longitudinal, multicenter study designed to provide extensive clinical, biomarker, and imaging data across multiple institutions and research organizations to facilitate research on the early diagnosis, tracking, and treatment of Alzheimer's disease. The database includes structural and functional magnetic resonance imaging (MRI and fMRI), positron emission tomography (PET), cerebrospinal fluid biomarkers, demographic data, and clinical and cognitive assessments collected from patients with Alzheimer's disease (AD), mild cognitive impairment (MCI), normal older adults (CN), and other related diseases.

[0024] This invention collected demographic data, cognitive scores, cerebrospinal fluid, and MRI samples from 2419 subjects, from three subgroups: ADNI-1, ADNI-2, and ADNI-3. Demographic data represented baseline characteristics (with no significant longitudinal variation), while the remaining data represented follow-up characteristics. Based on ADNI labels, subjects' clinical status at a given time point was categorized into three types: cognitively normal (CN), mild cognitive impairment (MCI), and Alzheimer's disease (AD). The ADNI dataset was sampled every six months, with the initial examination date designated as the first time, six months later as the second, one year later as the third, and so on, using data from the first six times. Table 1 shows the statistical results of the selected data at baseline.

[0025] Table 1. Data statistics at baseline

[0026]

[0027] S2: Preprocess the dataset to obtain a structured and complete dataset.

[0028] Follow-up characteristics and baseline characteristics are stored in two separate tables. The data table for follow-up characteristics undergoes the following series of processing steps (e.g.) Figure 2 (The remaining data represents 1597 patients with 11096 visits.) For missing data, multiple imputation was used to fill in the gaps. Data tables for baseline characteristics retained patients with the same RID (Relative Individuals Identified) after processing. These data were randomly divided into a 7:3 training set and a test set to predict the conversion from MCI to AD.

[0029] S3: Input the processed follow-up data into the constructed follow-up data model to obtain the data representation.

[0030] Based on the principle of attention mechanisms, we designed a variant of attention for current time series data prediction scenarios. This variant can be considered a form of local attention, assigning weights to different features and time points to capture the relationships between features and time points, thereby determining the importance of different features and time points. For small sample data, this simplified operation can reduce computational complexity, eliminating the need to calculate attention scores during each forward propagation and saving computational costs.

[0031] In the custom attention layer, trainable parameters are learned via backpropagation during training. These weights are initialized to 1, meaning they have no scaling effect on the input data initially, allowing the model to start training from an unbiased state where all inputs are treated equally. Updated through forward and backpropagation, these weights are scaled to varying degrees based on their learned values ​​as training progresses. The loss gradient is updated using an optimizer (Adam), indicating the amount and direction each weight needs to change to minimize the loss. The resulting weighted input is then fed into subsequent steps of model training.

[0032]

[0033] in Indicates weighted input, x t Let α represent the input, β represent the feature weights, and β represent the time weights.

[0034] Recurrent Neural Networks (RNNs) are designed for processing time-series data. Through their internal structure, they effectively capture the relationships between sequences, exhibiting a simple internal structure and excellent performance on short-series tasks. However, they perform poorly on long-series tasks, exhibiting vanishing or exploding gradients. To address this, Long Short-Term Memory (LSTM) networks and Gated Recurrent Units (GRUs) were designed and found to be effective in many areas. GRUs were chosen as the primary component for several advantages. First, they have fewer parameters, accelerating training, while LSTMs require too many parameters, resulting in higher training complexity and a risk of overfitting. Second, the model requires less input data, making GRUs easier to train with a small number of samples. Testing confirmed that GRUs performed better, so GRUs were initially selected as the training unit for the RNN.

[0035] The GRU consists of two parts: an update gate and a reset gate. The calculation method involves concatenating x(t) and h(t-1) with a linear transformation, followed by activation using a sigmoid function. The update gate's value is then applied to h(t-1), representing how much of the information from the previous time step can be utilized. This updated h(t-1) is then used for basic RNN operations: concatenating it with x(t) with a linear transformation, followed by tanh activation, to obtain a new h(t). Finally, the reset gate's value is applied to the new h(t), while (1 - the gate value) is applied to h(t-1). The results are then summed to obtain the final hidden state output h(t). This process means that the reset gate has the ability to reset all previous calculations.

[0036] z t =σ(W z ·[ht-1 ,x t ])

[0037] r t =σ(W r ·[h t-1 ,x t ])

[0038]

[0039] In this process, we combine the attention mechanism with time series learning methods such as GRU to learn the feature representation of the model based on the attention layer, and use the obtained feature representation combined with the baseline data as the input of the final decision-maker.

[0040] S4: Use a multilayer perceptron as the final decision-maker to predict the transition probability.

[0041] In the final prediction, a multilayer perceptron is used to predict the transition from MCI to AD. Information learned from the attention-integrated GRU and concatenated with the cross-sectional (baseline) data are fed into a multilayer perceptron to predict the probability of transitioning from MCI to AD on the next visit.

[0042]

[0043] Where y′ represents the predicted output, W1 and W2 are trainable linear transformation matrices, and c t This represents the context vector of the RNN unit after passing through the attention layer, D represents the baseline data, and b1 and b2 represent the bias vectors.

[0044] S5: Update the parameters of each module according to the loss function to obtain the AD conversion probability model based on attention deep learning method.

[0045] In this invention, the trainable parameters of the RNN unit and the multilayer perceptron are learned by using the binary cross-entropy function. The loss function is used to minimize false negative cases and improve the sensitivity of the model.

[0046]

[0047] In Equation 7, α is a weight set from 0 to 1 for prediction, y is the true range, and y′ is the predicted diagnostic. Based on the proposed custom loss function, all trainable parameters of the RNN, RNN autoencoder, and MLP are updated during model training in backpropagation. All model weights are trained using the Adaptive Moment Estimation (Adam) optimizer, a commonly used optimization algorithm that adaptively adjusts the learning rate for each parameter and typically converges faster than the classic stochastic gradient descent algorithm. The learning rate is set to 0.001. RNN units, batch size, dropout rate, and L2 regularization are tuned hyperparameters.

[0048] S8: Use an attention-based deep learning method to predict the probability model of AD conversion to predict the patient data and obtain the probability that the patient will convert from MCI to AD on the next visit.

[0049] To verify the effectiveness of this invention, comparative experiments were conducted against other traditional machine learning algorithms. This invention selects prediction accuracy and F2 as evaluation parameters:

[0050] 1) Accuracy – ACC

[0051] Accuracy represents the proportion of all correctly predicted samples out of the total sample. The formula for calculating accuracy in binary classification problems is:

[0052]

[0053] Where TP represents the number of correctly predicted positive samples, TN represents the number of correctly predicted negative samples, FP represents the number of incorrectly predicted positive samples, and FN represents the number of incorrectly predicted negative samples.

[0054] 2) F2 score

[0055] The F2 score is a metric that balances precision and recall, but unlike the F1 score, the F2 score places greater emphasis on recall. The specific calculation formula is as follows:

[0056]

[0057] Here, β is a weighting parameter. For the F2 score, β = 2, which means that recall is twice as important as precision.

[0058] In our experiments, we trained and tested the proposed model using longitudinal multimodal data from the ADNI dataset and baseline demographic data (these features are meaningless over time). The training and test sets comprised 70% and 30% of the dataset, respectively. During the experiment setup, the probability of a shift in the next side was predicted in the first two, three, four, five, and six iterations. Comparative experiments were conducted with Support Vector Machines (SVM), Random Forests (RF), and Decision Trees (DT). Since these three algorithms cannot handle longitudinal data, they were trained using the same data as the RNN units to obtain the prediction results.

[0059] Figure 3 , Figure 4 The accuracy of the experiment and the comprehensive index F2 score are shown respectively. It can be seen that both of these indicators of the present invention are better than the selected baseline model. The average accuracy of five predictions can reach 92.2%, and the average F2 score is 84.8%, which is the best performance among all models, and the results are highly competitive.

[0060] This invention addresses the shortcomings of existing methods for early diagnosis and prediction of Alzheimer's disease by using a deep learning method based on attention to predict the probability of AD conversion. The main contributions are as follows:

[0061] 1. By utilizing the concept of attention mechanism, weights are assigned to features and time points respectively. The weights obtained through model training are used in the training process to weight the input data and obtain the importance of different features and time points. Experiments also demonstrate the effectiveness and superiority of the proposed attention layer.

[0062] 2. A method for predicting AD conversion probability based on attention-based multimodal data fusion is proposed. This method combines baseline and follow-up features, considers the complexity of data in real-world production environments, and integrates data from multiple modalities to provide decision support for physicians. Experimental results demonstrate the effectiveness and superiority of this invention.

[0063] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting AD transition probability using attention-based deep learning, characterized in that, include: S1: Obtain the various examination data of the patient to be predicted, and divide the data into follow-up data and baseline data according to follow-up characteristics and baseline characteristics; S2: After preprocessing the follow-up data, it is input into the constructed base model to obtain the follow-up data representation; The base model includes: a dual attention layer and a temporal model; The dual attention layer is used to obtain the feature importance and time step importance of the follow-up data, and represents them with importance weights to obtain a weighted output; The time series model uses a GRU network to obtain the feature representation of the follow-up data by passing the weighted output through the GRU network; The dual attention layer is used to obtain the feature importance and time step importance of the follow-up data, and represents them with importance weights to obtain a weighted output, including: in, Indicates weighted output, x t This represents follow-up data, where α represents feature weights and β represents time weights. S3: The follow-up data is combined with the baseline data and input into the final decision-maker to obtain the probability that the patient to be predicted will be converted from MCI to AD at the next visit; The follow-up data, combined with baseline data, is input into the final decision-maker to obtain the probability that the patient's next visit will result in a conversion from MCI to AD, including: Where y′ represents the predicted output, W1 and W2 are trainable linear transformation matrices, and c t The feature representation of the follow-up data is given by D, which represents the baseline data, and b1 and b2 represent the deviation vectors.

2. The method for predicting AD transition probability using attention-based deep learning according to claim 1, characterized in that, Obtain all examination data of the patient to be predicted, and divide the data into follow-up data and baseline data according to follow-up characteristics and baseline characteristics, including: Various examination data of patients to be predicted were obtained from the ADNI database, including demographic data, cognitive scores, cerebrospinal fluid and MRI samples, from three subgroups: ADNI-1, ADNI-2 and ADNI-3; among them, the demographic data were baseline features with no significant longitudinal changes, and the rest were follow-up features. The ADNI database includes structural and functional magnetic resonance imaging (MRI) and fMRI, positron emission tomography (PET), cerebrospinal fluid biomarkers, demographic data, and clinical and cognitive assessments collected from Alzheimer's disease (AD), mild cognitive impairment (MCI), normal elderly individuals (CN), and other related disease patients.

3. The method for predicting AD transition probability using attention-based deep learning according to claim 1, characterized in that, Preprocessing of follow-up data includes: removing features irrelevant to follow-up data, excluding patient data from a single visit or diagnosis of CN, removing patient data with missing APOE4, and removing features with a missing rate >60%.

Images