Long-term individual ad risk prediction model and prediction method based on multi-scale deep learning

Abstract

The application belongs to the technical field of brain image computing and medical auxiliary research, and particularly discloses a long-term individual AD risk prediction model and a prediction method based on multi-scale deep learning, which comprises an MRI data acquisition and processing module, an MRI feature extraction module and a risk prediction module. The application adopts the long-term individual AD risk prediction model and the prediction method based on multi-scale deep learning, provides a novel multi-scale model, the model predicts the risk of MCI transforming into AD at the individual level, based on the baseline diagnostic data of the subjects, can obtain the transformation risk once every 6 months within 60 months, thereby effectively tracking the transformation risk of MCI and estimating the transformation time.

Description

Technical Field

[0001] This invention belongs to the field of brain imaging computing and medical assistance research technology, specifically involving a long-term individual AD risk prediction model and prediction method based on multi-scale deep learning. Background Technology

[0002] Alzheimer's disease (AD) is a neurodegenerative disease characterized by progressive impairment of cognitive and memory functions. It is estimated that approximately 152 million people worldwide will be diagnosed with Alzheimer's by 2050. Because the disease is irreversible, treatment remains challenging, with current approaches primarily focused on symptom control and slowing disease progression. Mild cognitive impairment (MCI) is the early stage of Alzheimer's, with approximately 10-15% of MCI patients progressing to Alzheimer's each year. Therefore, early diagnosis and intervention in the MCI stage are crucial for reducing the incidence of Alzheimer's. Predicting whether MCI patients will continue to experience cognitive decline and progress to Alzheimer's is currently a key research focus.

[0003] Currently, research on the conversion of MCI to AD mainly falls into two categories. The first approach treats it as a classification problem, based on machine learning (ML) or deep learning (DL) models, aiming to predict whether a subject will be classified as an AD patient by inputting a feature set at a specific time point. Studies have used plasma phosphorylated tau in combination with other measurable biomarkers, using logistic regression models to predict the progression of cognitive decline patients to AD within 2, 4, or even 6 years. One study constructed a multimodal recurrent neural network to predict the conversion of MCI at multiple time points using longitudinal data. Another study combined interaction effects and a multimodal DL model to predict the long-term progression of AD. These studies focus on classification but do not provide more comprehensive quantitative risk assessments or dynamic disease progression risk analyses.

[0004] The second approach treats AD conversion risk analysis as a survival analysis, with the Cox proportional hazards (CPH) model being the most commonly used. This approach aims to analyze the impact of multiple predetermined variables on patient survival. Two key concepts in survival analysis are event occurrence and survival time. Event occurrence refers to whether MCI (Multiple Organ Dysfunction Infection) converts to AD, while survival time is the time from baseline diagnosis to the last follow-up visit after a patient has converted from pMCI (a type of MCI that will develop into AD) or sMCI (a type of MCI that will not develop into AD) (e.g., loss to follow-up before the conversion event). Studies have quantified the impact of sex, MMSE (Mild Mental Health Examination), and brain age on conversion risk using Cox survival analysis and created a visual scale to provide individual-level risk probabilities. The CPH model allows for long-term and quantitative risk prediction, but it assumes linear constraints and cannot capture the non-linear relationship between input covariates and risk. Furthermore, the input covariates must satisfy the proportional hazards (PH) assumption, which assumes that the hazard ratio remains constant over time, limiting the flexibility to capture changes in hazard ratios over different time periods.

[0005] In recent years, deep learning-based survival analysis methods have emerged, aiming to address the linearity and proportionality constraints of the Cox proportional hazards model. One study constructed a DM-GNN model for personalized cancer prognosis prediction using histopathological images. Another study proposed a MultiSurv model, a deep learning-based end-to-end multimodal discrete-time prediction framework for predicting long-term all-cancer survival. A novel DeepHit model has also been introduced, capturing the joint distribution of survival time and events while simultaneously handling competing risks, ignoring the underlying assumptions of the data. Subsequently, the authors extended the research to dynamic DeepHit for dynamic survival analysis to distinguish relevant competing mortality risks in cystic fibrosis patients. Furthermore, some studies have developed deep learning-based survival analysis models to predict the onset time of Alzheimer's disease (AD) in MCI patients, but these studies only demonstrate population-level survival patterns, neglecting personalized clinical decision-making. Literature on predicting individual AD risk probabilities remains relatively scarce.

[0006] Therefore, there is a need in this field to develop long-term individual AD risk prediction models and methods based on multi-scale deep learning, which can effectively solve the above problems. Summary of the Invention

[0007] The purpose of this invention is to provide a long-term individual AD risk prediction model and method based on multi-scale deep learning. This model predicts the risk of MCI transforming into AD at the individual level. Based on the subject's baseline diagnostic data, it can obtain the transformation risk every 6 months within 60 months, thereby effectively tracking the MCI transformation risk and estimating the transformation time.

[0008] To achieve the above objectives, this invention provides a long-term individual AD risk prediction model based on multi-scale deep learning, including a magnetic resonance imaging (MRI) data acquisition and processing module, an MRI feature extraction module, and a risk prediction module; the MRI feature extraction module consists of four 3D convolutional layers and two 3D fully connected layers.

[0009] Preferably, the MRI data acquisition and processing module acquires baseline T1-weighted structural MRI images from the Alzheimer's Disease Neuroimaging Project (ADNI) public dataset. The original T1-weighted MRI scan images collected from the Alzheimer's Disease Neuroimaging Project (ADNI) public dataset are all in DICOM format and are converted to NIFTI format using Matlab. Preprocessing is performed using the Matlab-based SPM12 and CAT12 brain imaging processing software packages.

[0010] Preferably, the magnetic resonance imaging (MRI) data acquisition and processing module acquires baseline T1-weighted structural MRI images from the Alzheimer's Disease Neuroimaging Project (ADNI) public dataset, including several healthy NC samples, Alzheimer's disease AD samples, stable mild cognitive impairment (sMCI) samples, and progressive mild cognitive impairment (pMCI) samples.

[0011] Each mild cognitive impairment (MCI) sample must include corresponding structural magnetic resonance imaging (sMRI) data and complete clinical information, including age, sex, Mini-Mental State Examination (MMSE) score, and APOE ε4 genotypes A1 and A2.

[0012] For the mild cognitive impairment (MCI) sample, only samples with at least two follow-up records were included.

[0013] Preferably, the risk prediction module consists of two sets of fully connected layers, residual connections, and a multi-head attention mechanism;

[0014] The two sets of fully connected layers are Fully Connected Group 1 FCLayersA and Fully Connected Group 2 FCLayersB;

[0015] FCLayersA contains two fully connected layers with 128 and 256 nodes respectively; FCLayersB contains two fully connected layers with 128 and 64 nodes respectively; the extracted brain structural features, clinical information and gray matter volume constitute multi-scale data, which are passed as input x to FCLayersA.

[0016] Preferably, to capture the complex dependencies between elements in the feature sequence, a multi-head attention mechanism formula is introduced.

[0017]

[0018] Where Q, K, and V represent the features of the query, key, and value, respectively, and d k K represents the embedding dimension. T Indicates the transpose of K;

[0019] Multi-head attention mechanisms calculate attention weights using formulas, initially capturing the potential representations of covariates related to the risk of Alzheimer's disease conversion; FCLayersB will use vector f A (x) is combined with the covariate x and used as input z to directly learn the residuals while maintaining data integrity; the final output vector is f. B (x) is a probability distribution y = [y1, y2, ..., y3]. T ] represents a given subject x with mild cognitive impairment. i When the covariate x is denoted as x, it is transformed into an estimated risk probability of Alzheimer's disease occurring at time t, i.e. Where, x i "e" represents the characteristics of the subject, and "e" represents the occurrence of the event.

[0020] A prediction method based on a multi-scale deep learning long-term individual AD risk prediction model includes the following steps:

[0021] Step S1: The MRI data acquisition and processing module performs craniotomy on the initial brain MRI neuroimages and registers them to the MNI152 standard space.

[0022] Step S2: Based on the Hammers brain atlas template, the image is segmented into gray matter, white matter, and cerebrospinal fluid; through segmentation processing, the corresponding gray matter and white matter segmentation images are obtained, as well as the volume data of gray matter, white matter, and cerebrospinal fluid of 83 brain regions in the Hammers atlas.

[0023] Step S3: Further refine the segmented image by correcting the deviation of intensity non-uniformity and modulating it according to the volume change caused by spatial registration; after preprocessing, images with a quality of less than 80% are excluded.

[0024] Step S4: Smooth the image using an 8mm full-width half-maximum smoothing kernel;

[0025] Step S5: The MRI feature extraction module classifies images of normal cognition and Alzheimer's disease patients using a pre-trained convolutional neural network (NCN) model for healthy individuals.

[0026] Step S6: Fine-tune the images of patients with mild cognitive impairment, and use the output before the last fully connected layer as the extracted brain structural features;

[0027] Step S7: Finally, using gradient-weighted class activation mapping (JEM), the most important decision-making region of the model during the classification process is visualized, namely the brain's region of interest.

[0028] Step S8: The risk prediction module uses the gray matter volume of the brain region and the MRI image features obtained from the MRI data acquisition and processing module and the MRI feature extraction module; combined with clinical information, the three are used as risk variables and input into the risk prediction model.

[0029] Step S9: Predict the specific risk probability that each patient with mild cognitive impairment (MCI) will develop Alzheimer's disease (AD) at every six-month interval over the next five years.

[0030] Preferably, in the magnetic resonance imaging (MRI) data acquisition and processing module, each sample is represented by a triple (x, s, e), where x is a high-dimensional vector representing all features of the sample, s is the survival time, and e is the event occurrence status;

[0031] When e=0, it means that no conversion was observed in the sample before loss to follow-up, and when e=1, it means that the disease was converted to Alzheimer's disease.

[0032] Preferably, the high-dimensional vector x of all features for each sample consists of three types of features, specifically,

[0033] (1) Brain structural features extracted by the MRI feature extraction module;

[0034] (2) Clinical information for each sample, including gender, age, Mini-Mental State Examination (MMSE) score, APOEA1 and APOE A2 gene information;

[0035] (3) Gray matter volumes of 83 brain regions of interest obtained during the data preprocessing stage;

[0036] The sample dataset is D = {(x i ,s i ,e i )} i＝1 The sample of mild cognitive impairment (MCI) used for risk prediction / survival analysis is N.

[0037] Preferably, the five-fold cross-validation method is used to evaluate the performance of the pre-trained convolutional neural network (NCN) model for healthy individuals in classifying Alzheimer's disease (AD) and healthy individuals (NC).

[0038] a. The dataset includes several Alzheimer's disease (AD) samples and healthy human (NC) samples; the model training is carried out for 30 epochs with an initial learning rate of 1e-6, which is dynamically adjusted during the training process, decreasing by 10% every 10 epochs; finally, the final convolutional neural network healthy human NCN model is trained using all Alzheimer's disease (AD) and healthy human (NC) samples, and this model is a pre-trained model.

[0039] b. The output of the first fully connected layer is used as the extracted brain structural features; gradient weighted class activation mapping technology is used to visualize the image regions that contribute most significantly to the classification results by calculating the weighted gradient in the convolutional neural network NCN of healthy people and backpropagating it to the input image.

[0040] c. Using a five-fold cross-validation strategy, grayscale images of stable mild cognitive impairment (sMCI) samples and progressive mild cognitive impairment (pMCI) samples are used to fine-tune the pre-trained convolutional neural network (NCN) of healthy individuals.

[0041] (1) In each fold, the pre-trained model is fine-tuned using the training set partitioned by the fold, and specific features of mild cognitive impairment (MCI) data are captured by adjusting the weights.

[0042] (2) Then input the test set of the fold into the model;

[0043] (3) Extract the fully connected layer before the classification layer and output it as the feature representation of each sample in the test set.

[0044] Preferably, the risk prediction module uses a five-fold cross-validation strategy to train and evaluate model performance;

[0045] a. The inputs to the risk prediction model include a 256-dimensional MRI brain structural feature vector, a 5-dimensional clinical information vector, and an 83-dimensional brain gray matter volume vector, which together form a 344-dimensional multi-scale feature representation.

[0046] b. Each fully connected layer uses the ReLU activation function;

[0047] c. The module is trained for a total of 1500 cycles, with an initial learning rate of 1e-4, which is dynamically adjusted to half of the original learning rate every 500 cycles.

[0048] The present invention employs the above-mentioned long-term individual AD risk prediction model and method based on multi-scale deep learning, and the beneficial effects are as follows:

[0049] The model in this invention can quantify the risk probability of each patient with mild cognitive impairment (MCI) developing Alzheimer's disease (AD) at any time within the next five years, starting from baseline. According to research, there are currently few studies that quantitatively predict the risk of MCI patients developing AD; most studies can only identify high-risk groups. Some studies use the Cox survival analysis model to predict the conversion probability of AD, but the Cox model has linear and proportional constraints, limiting its clinical application. However, the deep learning-based survival analysis model proposed in this invention is not subject to these constraints and is more flexible in application compared to existing methods.

[0050] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0051] Figure 1 This is a model structure diagram of an embodiment of the multi-scale deep learning-based long-term individual AD risk prediction model and prediction method of the present invention; wherein, (a) is a structure diagram of the MRI feature extraction module; and (b) is a structure diagram of the risk prediction module.

[0052] Figure 2 This is a visualization of the region of interest (ROI) in three different sections of an embodiment of the multi-scale deep learning-based long-term individual AD risk prediction model and prediction method of the present invention; wherein, (a) is the axial section; (b) is the sagittal section; and (c) is the coronal section.

[0053] Figure 3 Box plots are shown for four models in the embodiments of the multi-scale deep learning-based long-term individual AD risk prediction model and prediction method of the present invention; where (a) is 6 months; (b) is 12 months; (c) is 18 months; (d) is 24 months; (e) is 30 months; (f) is 36 months; (g) is 42 months; (h) is 48 months; (i) is 54 months; and (j) is 60 months.

[0054] Figure 4 The Kaplan-Meier survival curve is an embodiment of the multi-scale deep learning-based long-term individual AD risk prediction model and prediction method of this invention.

[0055] Figure 5 This is a Bland-Altman plot showing the actual conversion time and predicted conversion time of an MCI individual to AD, based on the multi-scale deep learning long-term individual AD risk prediction model and prediction method of this invention.

[0056] Figure 6 This diagram illustrates the relationship between the actual conversion time and the predicted conversion time in an embodiment of the multi-scale deep learning-based long-term individual AD risk prediction model and method of this invention. Detailed Implementation

[0057] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0058] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0059] Example

[0060] The long-term individual AD risk prediction model and method based on multi-scale deep learning includes a magnetic resonance imaging (MRI) data acquisition and processing module, a magnetic resonance imaging (MRI) feature extraction module, and a risk prediction module.

[0061] The MRI data acquisition and processing module enables the acquisition and processing of brain MRI images, including preprocessed and enhanced data. The module acquires baseline T1-weighted structural MRI images from the Alzheimer's Disease Neuroimaging Project (ADNI) public dataset, including 700 healthy individuals (NC) samples, 766 Alzheimer's disease (AD) samples, 731 stable mild cognitive impairment (sMCI) samples, and 660 progressive mild cognitive impairment (pMCI) samples, as shown in Table 1.

[0062] Table 1. Statistical information of the dataset used in the study.

[0063] NC AD sMCI pMCI Sample size 700 766 731 660 Gender (Male / Female) 382 / 318 374 / 392 436 / 295 375 / 285 age 75.34±6.87 75.02±7.70 74.86±7.57 74.59±7.45 MMSE 29.11±1.12 21.8±4.14 27.88±2.03 26.36±2.22

[0064] Brain imaging software such as SPM and CAT12 were used for data preprocessing and gray matter segmentation of the brain. Eighty-three brain regions were segmented using brain templates to obtain the corresponding gray matter volume for each region. The raw T1-weighted MRI scans collected in the ADNI dataset were all in DICOM format and were converted to NIFTI format using Matlab. Subsequent preprocessing was mainly performed using the Matlab-based SPM12 and CAT12 brain imaging software packages.

[0065] The MRI feature extraction module uses a convolutional neural network (CNN) to extract brain structural features from gray matter images segmented from MRI, such as... Figure 1 As shown in (a), the module consists of four 3D convolutional layers and two 3D fully connected layers.

[0066] The MRI feature extraction module classifies images of patients with normal cognition and Alzheimer's disease using a pre-trained CNN model, then fine-tunes it using images of patients with mild cognitive impairment. Finally, it visualizes the most important decision regions of the model during the classification process, namely the brain regions of interest, through gradient-weighted class activation mapping technology, thereby improving the interpretability of the model.

[0067] Multiple studies have shown that abnormal changes in brain gray matter structure are more significant than those in white matter during cognitive decline. Therefore, pre-processed gray matter images of patients with NC and AD were input into a CNN, and the pre-trained model became an AD-NC classifier. The pre-trained CNN can extract brain structural change features associated with cognitive decline in Alzheimer's disease.

[0068] The risk prediction module implements the function of predicting AD conversion risks, such as... Figure 1 As shown in (b), the model consists of two sets of fully connected layers, residual connections, and a multi-head attention mechanism. The two sets of fully connected layers are Fully Connected Group 1 FCLayersA and Fully Connected Group 2 FCLayersB.

[0069] FCLayersA contains two fully connected layers with 128 and 256 nodes, respectively. FCLayersB contains two fully connected layers with 128 and 64 nodes, respectively. Brain structural features, clinical information, and gray matter volume extracted from MRI constitute multi-scale data, which are passed as input x to FCLayersA.

[0070] To capture the complex dependencies between elements in a feature sequence, a multi-head attention mechanism formula is introduced.

[0071]

[0072] Where Q, K, and V represent the features of the query, key, and value, respectively, and d k K represents the embedding dimension. T This represents the transpose of K.

[0073] Multi-head attention mechanisms calculate attention weights using formulas, thereby initially capturing the potential representations of covariates related to the risk of Alzheimer's disease conversion. FCLayersB will use vector f... A (x) is combined with the covariate x and used as input z to directly learn the residuals while maintaining data integrity. The final output vector is f. B (x) is a probability distribution y = [y1, y2, ..., y3]. T ] represents a given subject x with mild cognitive impairment. i When the covariate x is denoted as x, it is transformed into an estimated risk probability of Alzheimer's disease occurring at time t, i.e. Where, x i "e" represents the characteristics of the subject, and "e" represents the occurrence of the event.

[0074] A prediction method based on a multi-scale deep learning long-term individual AD risk prediction model includes the following steps:

[0075] Step S1: The MRI data acquisition and processing module performs craniotomy on the initial brain MRI neuroimages and registers them to the MNI152 standard space.

[0076] Step S2: Based on the Hammers brain atlas template, segment the image into gray matter, white matter, and cerebrospinal fluid. Through segmentation processing, obtain the corresponding gray matter and white matter segmentation images, as well as the volume data of gray matter, white matter, and cerebrospinal fluid for 83 brain regions in the Hammers atlas.

[0077] Step S3: Further refine the segmented image by correcting for intensity non-uniformity deviations and modulating based on volume changes caused by spatial registration. After preprocessing, images with a quality below 80% will be excluded.

[0078] Step S4: Use an 8mm full-width half-maximum smoothing kernel to smooth the image and improve the signal-to-noise ratio.

[0079] Step S5: The MRI feature extraction module classifies images of patients with normal cognition and Alzheimer's disease using a pre-trained CNN model.

[0080] Step S6: Fine-tune the images of patients with mild cognitive impairment, and use the output before the last fully connected layer as the extracted brain structural features.

[0081] Step S7: Finally, using gradient-weighted class activation mapping (JEM), the most important decision-making region of the model during the classification process is visualized, namely the brain's region of interest.

[0082] Step S8: The risk prediction module utilizes the brain region gray matter volume and MRI image features obtained from the MRI data acquisition and processing module and the MRI feature extraction module. Combined with clinical information, these three elements are input as risk variables into the risk prediction model.

[0083] Step S9: Predict the specific risk probability that each patient with mild cognitive impairment (MCI) will develop into AD at every six-month interval over the next five years, thereby understanding the pathological progression of MCI patients.

[0084] In the MRI data acquisition and processing module, each sample is represented by a triple (x, s, e), where x is a high-dimensional vector representing all features of the sample, s is the survival time, and e is the event occurrence.

[0085] When e=0, it means that no conversion was observed in the sample before loss to follow-up, and when e=1, it means that the disease was converted to Alzheimer's disease.

[0086] The high-dimensional vector x of all features of each sample consists of three types of features, specifically,

[0087] (1) Brain structural features extracted by the MRI feature extraction module.

[0088] (2) Clinical information for each sample, including gender, age, Mini-Mental State Examination (MMSE) score, and APOE A1 and APOE A2 genetic information.

[0089] (3) Gray matter volumes of 83 brain regions of interest obtained during the data preprocessing stage.

[0090] The sample dataset is D = {(x i ,s i ,e i )} i＝1 N represents the total number of MCI samples used for risk prediction / survival analysis.

[0091] The five-fold cross-validation method was used to evaluate the performance of the pre-trained CNN model in AD and NC classification.

[0092] a. The dataset includes 766 AD samples and 700 NC samples. The model was trained for 30 epochs with an initial learning rate of 1e-6, which was dynamically adjusted during training, decreasing by 10% every 10 epochs. Finally, the final CNN model was trained using all the AD and NC samples to obtain stronger classification capabilities. This model is a pre-trained model.

[0093] b. The output of the first fully connected layer is used as the extracted brain structural features. Gradient-weighted class activation mapping is employed to visualize the image regions that contribute most significantly to the classification results by calculating the weighted gradients in the CNN and backpropagating them to the input image.

[0094] c. Employ a five-fold cross-validation strategy, using grayscale images of sMCI and pMCI to fine-tune the pre-trained CNN.

[0095] (1) In order to effectively extract features of all MCI samples and avoid data leakage, in each fold, the pre-trained model is fine-tuned using the training set of the fold partition, and the specific features of the MCI data are captured by adjusting the weights.

[0096] (2) Then input the test set of the fold into the model.

[0097] (3) Extract the fully connected layer before the classification layer and output it as the feature representation of each sample in the test set.

[0098] Since the test set is different for each fold, the feature representation of all MCI samples can be obtained after the five-fold cross-validation is completed.

[0099] The risk prediction module uses a five-fold cross-validation strategy to train and evaluate model performance.

[0100] a. The inputs to the risk prediction model include a 256-dimensional MRI brain structural feature vector, a 5-dimensional clinical information vector, and an 83-dimensional brain gray matter volume vector, which together form a 344-dimensional multi-scale feature representation.

[0101] b. Each fully connected layer uses the ReLU activation function.

[0102] c. The module is trained for a total of 1500 cycles, with an initial learning rate of 1e-4, and is dynamically adjusted to half of the original learning rate every 500 cycles.

[0103] This embodiment uses the Concordance Index (C-index), a commonly used metric in survival analysis, to evaluate the performance of the survival analysis model. This index measures the consistency between the model's predicted survival risk and the actual survival outcome. Specifically, all samples are grouped, with each pair of samples forming a pair. For each pair, if an individual with a longer actual survival time also has a longer predicted survival time, or vice versa, the pair is considered consistent. The C-index value ranges from 0 to 1, with higher values ​​indicating better prediction accuracy.

[0104] In addition, this embodiment also uses the Brier Score (BS) evaluation metric, which measures the predictive performance of the survival model at a specific time point t by evaluating the mean squared distance between the predicted survival probability and the actual survival status. The BS ranges from 0 to 1, with lower values ​​indicating better prediction accuracy. By integrating the BS at different time points, a composite Brier Score (IBS) can be obtained, which is used to evaluate the predictive performance of the survival probability over a certain period of time (5 years in this embodiment).

[0105] The prediction model in this embodiment is evaluated as follows:

[0106] I. Feature Extraction.

[0107] sMRI scans collected from the ADNI dataset were preprocessed to generate GM images and GM volumes for 83 brain regions. This embodiment employs a five-fold cross-validation scheme, training and evaluating a CNN classification model based on GM images of AD and CN. The average accuracy per fold on the test set was 92.91%, indicating excellent performance in distinguishing between AD and NC. After fine-tuning using the MCI training set, the output of the first FC layer was used as the result of the GM feature extraction module. Finally, a 256-dimensional vector representing the brain structural features was obtained for each MCI individual.

[0108] This embodiment uses the Grad-CAM method to visualize regions of interest (ROIs) in three different sections (axial, sagittal, and coronal) of MRI scans to identify brain regions that significantly influence AD ​​and NC classification tasks, such as... Figure 2 As shown in the figure, the visualization results include weighted CAM maps, which highlight regions that play an important role in the classification task, with darker red indicating a greater impact on classification. This embodiment overlays the weighted CAM maps onto the original GM image, combining anatomical information from the MRI scan with the ROIs identified by the model, thus providing a more comprehensive view of the relationship between brain structure and cognitive decline in Alzheimer's disease (AD). Consistent with previous research, this embodiment demonstrates a significant association between cognitive decline in AD and specific brain regions (primarily concentrated in the frontal and temporal lobes). These findings not only validate existing research but also lay a solid foundation for subsequent risk prediction research.

[0109] II. Long-term and quantitative risk prediction.

[0110] To verify the effectiveness of the input feature combination, this embodiment employs a five-fold cross-validation method, using different combinations of three types of features to train and evaluate the multi-scale deep model, as shown in Table 2. A, B, and C represent three types of data: clinical data, gray matter (GM) volume of brain regions, and brain structural features, respectively. The results show that the multi-scale risk prediction model performs best when the three types of data (A, B, and C) are used in combination, with the C-index reaching 0.8881 (lower limit of 95% confidence interval: 0.8709, upper limit: 0.9053) and the IBS reaching 0.0806 (lower limit of 95% confidence interval: 0.0686, upper limit: 0.0927). The results indicate that GM volume contributes the most to the prediction model. When using only this type of data, the C-index is 0.8778 (0.8553, 0.9003) and the IBS is 0.0837 (0.0766, 0.0908), demonstrating the strongest ability to reflect the decline in cognitive ability based on brain region GM volume. Secondly, the extracted brain structural features showed a C-index of 0.8604 (0.8222, 0.8986) and an IBS of 0.0933 (0.0704, 0.1163). Clinical data contributed relatively little to the risk prediction model, with a C-index of 0.7199 (0.6849, 0.7549) and an IBS of 0.1802 (0.1689, 0.1916).

[0111] To verify the effectiveness of the multi-head attention mechanism, this embodiment conducted an experiment, comparing it with a multi-scale model that does not use the attention mechanism (referred to as the noAtt model). The results are shown in Table 2. When using combined ABC data, the C-index of the noAtt model is 0.8789 (95% confidence interval: 0.8616, 0.8962), slightly lower than that of the attention-based multi-scale model, while the IBS is 0.0854 (0.0748, 0.0960), slightly higher. When using other single data or data combinations, the C-index is relatively low and the IBS is relatively high, demonstrating the effectiveness of introducing the multi-head attention mechanism.

[0112] To demonstrate the superiority of the multi-scale model, this embodiment conducted comparative experiments with the Cox proportional hazards model (CPH) and the random survival forest (RSF) model. Among the two prediction methods, the best predictive performance was achieved when all three data types were combined, with C-indices of 0.8162 (0.7892, 0.8433) and 0.8352 (0.8149, 0.8554), and IBS values ​​of 0.1352 (0.1145, 0.1560) and 0.1178 (0.1126, 0.1231), respectively. Regardless of the data combination used for risk prediction, the multi-scale model consistently exhibited a higher C-indice and a lower IBS value, outperforming both the CPH and RSF methods.

[0113] This embodiment uses a combination of three data features to evaluate the risk prediction capabilities of multi-scale models and other models at different time points over a 60-month period. The risk predictions of each model are evaluated using five-fold cross-validation. In each fold validation, the model is trained on the training set and then used to predict the test set, generating the corresponding C-index. Figure 3 As shown, the C-index distributions of the four models at different time points are illustrated, allowing for a comparison of model performance at each time point. The size of the box plot reflects the volatility of model performance; larger boxes indicate greater performance fluctuations. The CPH model generally has high boxes across all time points, indicating significant C-index volatility and poor stability. In contrast, the multi-scale model has smaller boxes across all time points, indicating more consistent performance. The C-index of the multi-scale model ranges from 0.8336 to 0.8992 across time points, demonstrating strong predictive consistency and reliable calibration when using all types of data to predict AD conversion risk over the next 60 months based on baseline data.

[0114] To assess the differences in the C-index among the four models at different time points, this embodiment performed an analysis of variance (ANOVA). At 6 months, the differences between the models were not significant (p>0.05), and the performance of each model was very similar, with significant overlap in the box plots. From 12 months onwards, the performance differences between the models became significant (p<0.05), indicating that in short-term predictions, the differences between the models were not significant, while the deep learning models were more suitable for long-term predictions. Simultaneously, from 12 months onwards, the F-statistic was higher, indicating significant differences between the models, while the differences within groups were smaller.

[0115] This embodiment verifies whether a multi-scale model can effectively predict the occurrence of MCI transformation. The model exhibits the best risk prediction ability when combined with clinical data, GM volume, and brain structural features. Therefore, combined data were used to calculate the predicted risk value for all MCI individuals, and the average predicted risk over 60 months was used as the risk score for each individual. This method comprehensively considers the long-term risk level of MCI individuals over 5 years, rather than being limited to a single risk value. Based on the risk score, instances were divided into high-risk and low-risk groups, with the grouping threshold based on the median risk score. Kaplan-Meier survival curves were plotted to reflect the survival probabilities of the two groups, such as... Figure 4 As shown. Using the log-rank test to compare the high-risk and low-risk groups, the p-value of the multi-scale model was found to be 2.0758 × 10⁻⁶. -243 This indicates a significant difference between the two groups, demonstrating that the model has a strong ability to identify and predict the conversion of MCI patients.

[0116] This embodiment also predicts the specific time of AD conversion based on risk value. For example... Figure 5 The figure shows a comparison between the actual and predicted conversion times of individuals with MCI to AD. If no actual conversion or a predicted no conversion occurs, the predicted conversion time is recorded as 0. Many data points overlap in the figure; points with an overlap of 10 or more are marked. 89.6477% of the data points lie between the upper and lower limits indicated by the dashed lines, at 30.77 and -32.37 respectively, with an average difference of -0.80 between the center lines. This indicates a generally good consistency between the predicted and actual conversion times, demonstrating that the model has satisfactory accuracy and consistency in predicting the time it takes for an MCI individual to convert to AD.

[0117] Furthermore, the density distribution plot provides supplementary evidence supporting the consistency between the actual and predicted conversion times. Two-dimensional kernel density estimation (KDE) can be used to estimate the distribution of data points in a two-dimensional plane. Lighter colors indicate higher data point density. Figure 6As shown, the data points exhibit a clear linear distribution trend, indicating a strong positive linear relationship between the actual conversion time and the predicted conversion time, demonstrating that the model has good predictive performance.

[0118] This embodiment also evaluated the mean absolute error (MAE) between the actual conversion time and the predicted conversion time for different models when combined with various data types, as detailed in Table 2. The results show that the multi-scale model has the strongest predictive performance when the three types of data are used in combination, with an MAE value of 7.3738 (6.7418, 8.0081).

[0119] Table 2 Model Performance Evaluation and Comparison

[0120]

[0121]

[0122] Therefore, the present invention adopts the above-mentioned long-term individual AD risk prediction model and prediction method based on multi-scale deep learning. This model predicts the risk of MCI transforming into AD at the individual level. Based on the baseline diagnostic data of the subjects, it can obtain the transformation risk every 6 months within 60 months, thereby effectively tracking the MCI transformation risk and estimating the transformation time.

[0123] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A prediction method based on a multi-scale deep learning long-term individual AD risk prediction model, characterized in that, Includes the following steps: Step S1: The MRI data acquisition and processing module performs craniotomy on the initial brain MRI neuroimages and registers them to the MNI152 standard space. Step S2: Based on the Hammers brain atlas template, the image is segmented into gray matter, white matter, and cerebrospinal fluid; through segmentation processing, the corresponding gray matter and white matter segmentation images are obtained, as well as the volume data of gray matter, white matter, and cerebrospinal fluid of 83 brain regions in the Hammers atlas. Step S3: Further refine the segmented image by correcting the deviation of intensity non-uniformity and modulating it according to the volume change caused by spatial registration; after preprocessing, images with a quality of less than 80% are excluded. Step S4: Smooth the image using an 8mm full-width half-maximum smoothing kernel; Step S5: The MRI feature extraction module classifies images of normal cognition and Alzheimer's disease patients using a pre-trained convolutional neural network (NCN) model for healthy individuals. Step S6: Fine-tune the images of patients with mild cognitive impairment, and use the output before the last fully connected layer as the extracted brain structural features; Step S7: Finally, using gradient-weighted class activation mapping (JEM), the most important decision-making region of the model during the classification process is visualized, namely the brain's region of interest. Step S8: The risk prediction module uses the gray matter volume of the brain region and the MRI image features obtained from the MRI data acquisition and processing module and the MRI feature extraction module; combined with clinical information, the three are used as risk variables and input into the risk prediction model. Step S9: Predict the specific risk probability that each patient with mild cognitive impairment (MCI) will develop Alzheimer's disease (AD) at every six-month interval over the next five years. The risk prediction module consists of two sets of fully connected layers, residual connections, and a multi-head attention mechanism; The two sets of fully connected layers are Fully Connected Group 1 FCLayersA and Fully Connected Group 2 FCLayersB; FCLayersA contains two fully connected layers with 128 and 256 nodes respectively; FCLayersB contains two fully connected layers with 128 and 64 nodes respectively; extracted brain structural features, clinical information, and gray matter volume constitute multi-scale data, which are used as input. Passed to FCLayersA; Multi-head attention mechanisms calculate attention weights using formulas, initially capturing the potential representations of covariates related to the risk of Alzheimer's disease conversion; FCLayersB will use vector f A (x) is combined with the covariate x and used as input z to directly learn the residuals while maintaining data integrity; the final output vector is f. B (x) is a probability distribution y=[y1,y2,...,y3] T ] represents a given subject x with mild cognitive impairment. i When the covariate x is denoted as x, it is transformed into an estimated risk probability of Alzheimer's disease occurring at time t, i.e. ;where x i "e" represents the characteristics of the subject, and "e" represents the occurrence of the event.

2. The prediction method based on a multi-scale deep learning long-term individual AD risk prediction model according to claim 1, characterized in that: In the MRI data acquisition and processing module, each sample is represented by a triplet (x, s, e), where Let s be a high-dimensional vector representing all features of the sample, where s is the survival time and e is the occurrence of the event. When e=0, it means that no conversion was observed in the sample before loss to follow-up, and when e=1, it means that the disease was converted to Alzheimer's disease.

3. The long-term individual AD risk prediction model and method based on multi-scale deep learning according to claim 2, characterized in that: The high-dimensional vector x of all features of each sample consists of three types of features, specifically for, (1) Brain structural features extracted by the MRI feature extraction module; (2) Clinical information for each sample, including gender, age, Mini-Mental State Examination (MMSE) score, and APOE A1 and APOE A2 genetic information; (3) Gray matter volumes of 83 brain regions of interest obtained during the data preprocessing stage; The sample dataset is D={(x i ,s i ,e i )} i=1 The sample of mild cognitive impairment (MCI) used for risk prediction / survival analysis is N.

4. The prediction method based on a multi-scale deep learning long-term individual AD risk prediction model according to claim 1, characterized in that: The performance of a pre-trained convolutional neural network (NCN) model for healthy individuals in classifying Alzheimer's disease (AD) and healthy individuals (NC) was evaluated using a five-fold cross-validation method. a. The dataset includes several Alzheimer's disease (AD) samples and healthy human (NC) samples; the model training is carried out for 30 epochs with an initial learning rate of 1e-6, which is dynamically adjusted during the training process, decreasing by 10% every 10 epochs; finally, the final convolutional neural network healthy human NCN model is trained using all Alzheimer's disease (AD) and healthy human (NC) samples, and this model is a pre-trained model. b. The output of the first fully connected layer is used as the extracted brain structural features; gradient weighted class activation mapping technology is used to visualize the image regions that contribute most significantly to the classification results by calculating the weighted gradient in the convolutional neural network NCN of healthy people and backpropagating it to the input image. c. Using a five-fold cross-validation strategy, grayscale images of stable mild cognitive impairment (sMCI) samples and progressive mild cognitive impairment (pMCI) samples are used to fine-tune the pre-trained convolutional neural network (NCN) of healthy individuals. (1) In each fold, the pre-trained model is fine-tuned using the training set partitioned by the fold, and specific features of mild cognitive impairment (MCI) data are captured by adjusting the weights; (2) Then input the test set of the fold into the model; (3) Extract the fully connected layer before the classification layer and output it as the feature representation of each sample in the test set.

5. The prediction method based on a multi-scale deep learning long-term individual AD risk prediction model according to claim 1, characterized in that: The risk prediction module uses a five-fold cross-validation strategy to train and evaluate model performance; a. The inputs to the risk prediction model include a 256-dimensional MRI brain structural feature vector, a 5-dimensional clinical information vector, and an 83-dimensional brain gray matter volume vector, which together form a 344-dimensional multi-scale feature representation. b. Each fully connected layer uses the ReLU activation function; c. The module is trained for a total of 1500 cycles, with an initial learning rate of 1e-4, which is dynamically adjusted to half of the original learning rate every 500 cycles.

Images