A Brain Imaging-Based System and Method for Predicting the Risk of Transition to Mild Cognitive Impairment

Abstract

This invention discloses a system and method for predicting the risk of conversion to mild cognitive impairment (MCI) based on brain imaging, belonging to the field of medical artificial intelligence technology. The system includes: a data preprocessing module for performing voxel-based morphological analysis preprocessing on T1-weighted magnetic resonance imaging (MRI) data of the brain; a brain region feature extraction module for extracting gray matter volume features of each brain region from the preprocessed image based on a fine-grained brain atlas containing at least 90 brain regions; a risk classification model using a trained machine learning algorithm to predict whether a patient belongs to the high-risk or low-risk group for Alzheimer's disease conversion based on the gray matter volume feature vector; and an output module for displaying the prediction results. This invention, by extracting fine-grained brain region features and combining them with machine learning algorithms, achieves automated and high-precision prediction of the risk of MCI patients converting to AD. It has the advantages of objectivity, stability, and high sensitivity, providing an effective auxiliary tool for early clinical screening and individualized intervention.

Description

Technical Field

[0001] This invention relates to the fields of medical data processing and neuroimaging analysis technology, and in particular to a system and method for predicting the risk of conversion to mild cognitive impairment based on brain imaging. Background Technology

[0002] Mild cognitive impairment (MCI) is a transitional stage between normal aging and Alzheimer's disease (AD). MCI patients experience some degree of cognitive decline but have not yet met the diagnostic criteria for dementia. Some patients (MCI converters, MCIc) will progress to AD within a few years, while others (MCI non-converters, MCInc) remain stable or progress slowly. Early and accurate identification of MCIc patients is of great clinical significance for timely intervention, slowing disease progression, and improving patient prognosis.

[0003] Currently, the clinical diagnosis of MCI mainly relies on neuropsychological scale assessments (such as MMSE and MoCA), but this method is highly subjective, easily affected by various factors, and lacks sensitivity in the very early stages of the disease. Structural magnetic resonance imaging (sMRI), as a non-invasive imaging technique, can objectively display microscopic changes in brain structure. Studies have shown that patients with AD and MCI often exhibit gray matter atrophy in specific brain regions (such as the hippocampus and entorhinal cortex). However, traditional qualitative or single-region quantitative analysis is insufficient to provide stable and reliable individualized predictive indicators.

[0004] In recent years, machine learning methods have been introduced into the field of medical image analysis to extract high-dimensional and complex pathological patterns from massive image data. In existing technologies, some studies have attempted to use sMRI features combined with machine learning to diagnose or predict the progression of MCI, but these methods usually have the following shortcomings: (1) the brain regions used are too broad or too few, resulting in insufficient sensitivity to capture subtle changes in gray matter volume; (2) the interpretability, stability and generalization ability of the model need to be improved; (3) there is a lack of a complete and systematic solution that integrates automated preprocessing, feature extraction and high-precision classification models.

[0005] Therefore, there is an urgent need in this field for a system and method that can automatically, objectively and accurately predict the risk of MCI patients turning into AD, in order to make up for the shortcomings of existing clinical diagnostic methods and provide decision support for early individualized intervention. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a system and method for predicting the conversion risk of mild cognitive impairment based on the gray matter volume of multiple brain regions. This system and method can achieve automated and high-precision classification of MCIc and MCInc, and has important clinical application value.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, this invention provides a system for predicting the conversion risk of mild cognitive impairment based on gray matter volume in multiple brain regions. This system includes a data preprocessing module, a brain region feature extraction module, a risk classification model, and an output module. The data preprocessing module processes raw T1-weighted magnetic resonance imaging data, performing standardization and gray matter volume extraction using a voxel-based morphological analysis (VBM) workflow. The brain region feature extraction module extracts the gray matter volume of each brain region as a feature based on a fine-grained brain region atlas (such as the Hammers atlas containing 95 brain regions). The risk classification model employs a trained machine learning algorithm (preferably the K-nearest neighbor algorithm) to classify patients based on the extracted feature vectors. The output module displays the classification results and evaluation indicators.

[0008] Secondly, the present invention provides a method for predicting the risk of conversion of mild cognitive impairment based on gray matter volume of multiple brain regions. The method includes the following steps: acquiring and preprocessing sMRI data; extracting gray matter volume features of multiple brain regions based on fine-grained brain atlases; inputting the features into a pre-trained classification model (such as a KNN model) for risk classification; and outputting the prediction results.

[0009] Thirdly, the present invention provides a storage medium storing a computer program that implements the above-described method.

[0010] Fourthly, the present invention provides an electronic device comprising a processor and executing the above-described program. Beneficial effects

[0011] The present invention has the following beneficial effects: High precision and high sensitivity: By using fine-grained (e.g., 95-area) brain region maps to extract features, compared with traditional coarse-grained region or single-region analysis, it can more comprehensively and sensitively capture gray matter micro-volume changes related to disease progression, thereby significantly improving the accuracy of classification prediction (e.g., test set accuracy can reach over 93%) and AUC value (e.g., over 0.91), and reducing false positive and false negative rates.

[0012] Objectification and Automation: The system has realized an automated processing flow from raw image data to risk prediction results, reducing reliance on human experience and providing objective and repeatable quantitative evaluation indicators.

[0013] High clinical applicability: This invention provides an effective auxiliary tool for early screening and risk stratification of MCI patients, which helps clinicians develop personalized intervention and follow-up strategies and has significant clinical translational potential.

[0014] Good model stability: Experimental results show that the KNN algorithm used exhibits excellent repeatability and stability in this task, which is beneficial for the clinical application and promotion of the model. Attached Figure Description

[0015] ‌ Figure 1 The flowchart illustrates the data preprocessing and feature extraction provided in this embodiment of the invention.

[0016] ‌ Figure 2 For LDA confusion matrix, ROC curve and classification report.

[0017] Figure 3 This provides the confusion matrix, ROC curve, and classification report for KNN.

[0018] Figure 4 This provides the confusion matrix, ROC curve, and classification report for SVM.

[0019] Figure 5 This provides the confusion matrix, ROC curve, and classification report for LDA.

[0020] Figure 6 This provides the confusion matrix, ROC curve, and classification report for KNN.

[0021] Figure 7 This provides the confusion matrix, ROC curve, and classification report for SVM. Detailed Implementation

[0022] The following are specific embodiments of the present invention, which are described in conjunction with the accompanying drawings to further illustrate the technical solutions of the present invention. However, the present invention is not limited to these embodiments.

[0023] Key steps explained: 1) Noise reduction: SANLM is performed 3 times (at the beginning, middle, and end respectively). 2) Registration: Includes affine registration + SPM registration + fast optimization registration 3) Segmentation: SPM initial segmentation → LAS local optimization → AMAP final optimization 4) Surface reconstruction: Independent processing on both sides (accounting for 74% of the total time) 5) Quality Control: The final IQR score affects the reliability of the results (Level B, Medium) Machine Learning In machine learning, confusion matrices and classification metrics are important tools for evaluating the performance of classification models. AUC-ROC curves, on the other hand, are a visual method for evaluating a model's classification ability. In the CAT12 toolbox, T1-weighted MRI images are first preprocessed to remove noise and correct for field inhomogeneities. Then, using prior knowledge from Hammers atlases, combined with information such as voxel signal intensity and location, the brain is segmented. By comparing the features of each voxel with the features of different brain regions in the Hammers atlas, the brain is divided into 95 corresponding brain regions (including cerebrospinal fluid, gray matter, and white matter). Gray matter data is selected for machine learning algorithm analysis.

[0024] Machine learning using VGM in 95 brain regions The following are the basic steps to perform these operations using the preprocessing analysis results: 1. LDA Machine Learning Analysis Based on VGM in Brain Regions A linear discriminant analysis (LDA) model was used to train and evaluate VGM data from 95 brain regions. The model's classification performance was comprehensively measured using metrics such as confusion matrix, ROC curve, and classification report.

[0025] Figure 2 The confusion matrix, ROC curve, and classification report of LDA are displayed.

[0026] Confusion Matrix: The confusion matrix of the LDA model (accuracy: 0.87) presents the specific situation of classification prediction. In the matrix, the values ​​of the intersection regions between the true and predicted labels reflect the prediction hits and misclassifications for each category of sample. Overall, the model performs well in classification prediction for most samples, with a small number of misclassified samples, initially showing that the LDA model has a certain ability to classify and distinguish VGM data from the brain region. TP=7, FN=1, FP=1, TN=6.

[0027] ROC Curve: In the LDA ROC curve, the area under the curve (AUC = 0.96) is close to 1, indicating that the model has excellent classification performance. The trend of the ROC curve reflects the trade-off between the true positive rate and the false positive rate at different thresholds. This curve shape shows that the LDA model trained based on brain region VGM data can effectively distinguish between positive and negative samples, and has high reliability and stability in classification tasks.

[0028] Classification Report: The classification report evaluates the model from multiple dimensions, including precision, recall, F1 score, and number of supporting samples. For label 0.0, the precision is 1.00, recall is 0.88, F1 score is 0.93, and number of supporting samples is 8; for label 1.0, the precision is 0.88, recall is 1.00, F1 score is 0.93, and number of supporting samples is 7. The overall accuracy is 0.93, with precision = 0.93, sensitivity = 0.88, specificity = 0.86, false positive rate = 0.14, and false negative rate = 0.125. The precision, recall, and F1 score of both macro-average and weighted average are also at a high level, further validating the effectiveness of the model in classifying VGM data of brain regions, and the classification performance of samples of different categories is relatively balanced.

[0029] Evaluation of machine learning models based on VGN data from 95 brain regions Using VGMs from 95 brain regions as features, a binary classification machine learning model was constructed (with labels set to 0.0 and 1.0). A confusion matrix was used to visualize the matching relationship between the predictions and the true labels. Precision, recall, and F1 score were calculated using classification reports, and ROC curves were plotted to evaluate the model's discriminative ability.

[0030] Figure 3 It displays the confusion matrix, ROC curve, and classification report of KNN.

[0031] Confusion matrix analysis: There were 8 samples with a true label of 0.0, 7 of which were correctly predicted and 1 was misclassified as 1.0; there were 7 samples with a true label of 1.0, all of which were correctly predicted and misclassified.

[0032] This matrix intuitively reflects the model's bias in recognizing the two types of labels. Class 0.0 has a certain risk of misclassification, while Class 1.0 is accurately identified.

[0033] Analysis Report Interpretation: Precision: 1.00 for class 0.0 (only 1 false positive), 0.88 for class 1.0 (due to 1 false positive in class 0.0); Recall rate: 0.88 for class 0.0 (1 out of 8 cases missed), 1.00 for class 1.0 (all 7 cases were detected); F1 score: 0.93 for both classes, balancing precision and recall; Accuracy: 0.93, 14 out of 15 samples were predicted correctly; The macro average and weighted average are 0.94 and 0.93 respectively, reflecting the overall robustness of the model.

[0034] ROC curve and AUC analysis In the ROC curve (Figure 2), the area under the curve (AUC) reaches 0.91. The AUC is close to 1, indicating that the model has a strong ability to distinguish between the 0.0 and 1.0 brain region VGM samples. The False Positive Rate and True Positive Rate are optimized and balanced, verifying that the model has high discriminative efficiency in brain region classification tasks.

[0035] Performance evaluation of machine learning models based on VGM data from 95 brain regions Using VGM data from 95 brain regions as feature input, a binary classification machine learning model was constructed (classification labels 0.0 and 1.0). A confusion matrix was used to visually represent the matching between the model's predicted labels and the true labels. Precision, recall, and F1 score were calculated using classification reports to quantify the model's classification performance. ROC curves were plotted and the area under the curve (AUC) was calculated to evaluate the model's ability to distinguish between the two classes of samples.

[0036] Figure 4 The confusion matrix, ROC curve, and classification report of SVM are displayed.

[0037] Confusion matrix: All 8 samples with a true label of 0.0 were correctly predicted (corresponding to matrix positions [8,0]); all 7 samples with a true label of 1.0 were also accurately identified (corresponding to matrix positions [0, 7]). This indicates that the model did not misclassify the two types of brain region VGM samples on the current test sample set, and the classification effect was ideal.

[0038] Classification Report: Precision: The precision for both labels is 1.00. For label 0.0, this means that all samples predicted as 0.0 truly belong to class 0.0; similarly for label 1.0, samples predicted as 1.0 are not misclassified.

[0039] Recall: Both classes achieved a recall of 1.00. This means that all samples with a true value of 0.0 were correctly identified, and none of the samples with a true value of 1.0 were missed.

[0040] F1 score: Both classes have an F1 score of 1.00. This metric balances precision and recall, reflecting the excellent classification performance of the model.

[0041] Accuracy: The overall accuracy was 1.00, with all samples correctly classified out of 15 test samples (8 + 7).

[0042] Macro average and weighted average: Both average metrics are 1.00, further validating the model's robust and efficient classification capabilities across different categories and the overall sample.

[0043] ROU curve and AOC In the ROC curve, the area under the curve (AUC) is 1.00. The AUC ranges from 0 to 1; the closer it is to 1, the stronger the model's ability to distinguish between the two classes of samples. Here, the AUC reaches 1.00, indicating that the model can perfectly distinguish between the 0.0 and 1.0 classes of brain region VGM samples. In the trade-off between the False Positive Rate and the True Positive Rate, it achieves an excellent classification performance, with almost no error in distinguishing between the two classes.

[0044] Machine learning was performed on VGM data from 53 out of 95 brain regions. Based on literature review and prior knowledge, priority is given to brain regions known to be susceptible in Alzheimer's disease (AD) or MCI, such as: Medial temporal lobe (hippocampus, entorhinal cortex, amygdala): closely related to memory function, significant atrophy is observed in early AD. Default mode network (DMN) related areas (posterior cingulate cortex, precuneus, medial prefrontal cortex): functional connectivity abnormalities are common in MCI.

[0045] Parts of the parietal and frontal lobes (such as the angular gyrus and prefrontal cortex): involved in executive function and spatial cognition, may be impaired in MCI.

[0046] Significance of gray matter volume differences: Brain regions with significant differences in gray matter volume between the MCIc and MCInc groups were screened using statistical tests.

[0047] 1. LDA machine learning analysis based on brain region VGM data Linear Discriminant Analysis (LDA) is a classic supervised learning method that aims to separate samples of different classes as much as possible after projection by finding the optimal projection direction, thus achieving the classification objective. This experiment primarily evaluates the model performance using the confusion matrix, classification report, ROC curve, and AUC value.

[0048] Figure 5 The confusion matrix, ROC curve, and classification report of LDA are displayed.

[0049] Confusion Matrix Analysis The LDA confusion matrix (accuracy: 0.87) presents the specific classification predictions of the model. In the matrix, Truelabel represents the true class, and Predicted label represents the predicted class. The matrix values ​​show that for a two-class classification task (assuming 0.0 and 1.0 are the two classes), among samples with a true class of 0.0, 7 were correctly predicted, and 1 was misclassified as class 1.0; among samples with a true class of 1.0, 6 were correctly predicted, and 1 was misclassified as class 0.0. This reflects that the model has some misclassifications when classifying samples of the two classes, but the overall number of misclassifications is small, initially showing that the model has good classification ability.

[0050] Analysis Report Interpretation Category 0.0: Precision, recall, and F1 score are all 0.88, supporting 8 samples, indicating that the prediction accuracy and ability to capture real samples are well-balanced when classifying this type of sample.

[0051] Category 1.0: Precision, recall, and F1 score are 0.86, supporting 7 samples. Its classification performance is slightly worse than that of Category 0.0, but its overall performance is still better.

[0052] Overall: The accuracy reached 0.87, meaning that 87% of all samples were correctly classified; the precision, recall, and F1 score of macro average and weighted average were all 0.87, indicating that the model has a relatively balanced classification performance on both types of samples and the overall classification effect is good.

[0053] ROC curve and AUC The LDA ROC curve is plotted with the false positive rate on the horizontal axis and the true positive rate on the vertical axis. The closer the curve is to the upper left corner, the better the model's classification performance. In this experiment, the LDA ROC curve shows an AUC (Area Under Curve) value of 0.89. The closer the AUC value is to 1, the stronger the model's ability to distinguish between positive and negative samples. An AUC value of 0.89 indicates that this LDA model has a good discriminative ability for the two classes of samples based on brain region VGM data, and can effectively identify positive and negative samples.

[0054] Analysis based on KNN machine learning model of VGM data points from 53 brain regions The K-Nearest Neighbors (KNN) algorithm, a classic supervised learning method, determines the class of a current sample by calculating the distance between samples and classifying the K nearest neighbors. This experiment uses a confusion matrix to present the distribution of correct and incorrect classification predictions. Precision, recall, and F1 score from the classification report are used to measure the classification performance at the class level. The overall discriminative ability of the model is evaluated using the ROC curve and the area under the curve (AUC).

[0055] Figure 6 It displays the confusion matrix, ROC curve, and classification report of KNN.

[0056] Confusion matrix: All eight samples with a true class of 0.0 were correctly predicted by the model (predicted as 0.0), demonstrating the model's excellent ability to identify samples of this class. Of the seven samples with a true class of 1.0, six were correctly predicted as 1.0, and only one was misclassified as 0.0. Overall, the model had a low number of misclassifications, initially demonstrating its good classification performance.

[0057] Analysis Report Interpretation For class 0.0: Precision reached 0.89 (meaning that 89% of the samples predicted as class 0.0 were actually class 0.0), recall was 1.00 (all true class 0.0 samples were correctly identified), F1 score was 0.94, and the number of supported samples was 8. This indicates that the model achieved efficient and balanced performance in terms of accurate prediction and comprehensive capture of true samples for this class.

[0058] Category 1.0: Precision 1.00 (all samples predicted as class 1.0 were true class 1.0), Recall 0.86 (86% of true class 1.0 samples were correctly identified), F1 score 0.92, Number of supporting samples 7. Although the recall is slightly lower, the precision is excellent, and the overall classification performance is still quite outstanding.

[0059] Overall performance: The accuracy is 0.93, meaning that about 93% of all samples are correctly classified; the precision, recall, and F1 score of macro average and weighted average are all in the range of 0.93-0.94, reflecting that the model has a balanced overall performance in classifying the two types of samples and has excellent overall classification effect.

[0060] ROC curve and AUC The ROC curve of KNN is plotted with the false positive rate on the horizontal axis and the true positive rate on the vertical axis. The closer the curve is to the upper left corner, the stronger the model's ability to distinguish between positive and negative samples. In this experiment, the area under the ROC curve (AUC) reached 0.96. The AUC value is close to 1, indicating that the model has a very strong discriminative ability for the two classes of samples based on brain region VGM data, and can efficiently distinguish between positive and negative samples, demonstrating excellent classification performance.

[0061] Evaluation of machine learning models based on VGM data from 53 brain regions Using VGM data from 53 brain regions as feature input, a classifier was constructed using a machine learning classification algorithm (the specific algorithm was not specified, but it is speculated to be a supervised classification model) to perform binary classification on the samples.

[0062] Figure 7 The confusion matrix, ROC curve, and classification report of SVM are displayed.

[0063] The confusion matrix shows that all 8 samples with a true label of 0.0 were correctly predicted (predicted label 0.0), with no false positives or false negatives. Of the 7 samples with a true label of 1.0, 6 were correctly predicted (predicted label 1.0), and 1 was misclassified as 0.0 (false negative). The overall misclassification rate is low, and the model has good basic classification ability for both classes of samples.

[0064] Classification Report For class 0.0: Precision 0.89 (89% of samples predicted as 0.0 were actually 0.0), Recall 1.00 (all true 0.0 samples were correctly identified), F1 score 0.94, and number of supporting samples 8. This indicates that the model performs excellently in both accurate identification and comprehensive coverage of samples in this class.

[0065] Category 1.0: Precision 1.00 (all samples predicted as 1.0 were actually 1.0), Recall 0.86 (86% of the true 1.0 samples were correctly identified), F1 score 0.92, number of supporting samples 7. Although the recall is slightly low, the precision is perfect, and the classification performance is still commendable.

[0066] Overall performance: The accuracy reached 0.93 (13 / 15 samples were correctly classified); the precision, recall, and F1 score of macroavg and weighted avg were all approximately 0.93-0.94, indicating that the model has a balanced overall performance on both classes of samples and has excellent overall classification effect.

[0067] ROC curve and AUC The ROC curve trended close to the upper left corner, with an AUC value as high as 0.98. This indicates that the model has an extremely strong ability to distinguish between positive and negative samples. Even with adjustments to the classification threshold, it can maintain a high true positive rate and a low false positive rate, demonstrating robust and excellent classification performance.

[0068] Summary of Clinical Application Significance and Project Value Clinical significance of classification accuracy for brain regions of 1.95 vs. 53: A classification model with 95 brain regions (e.g., KNN accuracy of 93%, AUC=0.891) outperformed one with 53 brain regions, indicating that finer-grained brain region features can more sensitively capture the gray matter volume differences between MCIc (high-risk patients to AD) and MCInc. This helps to identify high-risk patients earlier (predicting AD conversion risk through changes in volume across multiple brain regions before the onset of clinical symptoms), accurately locate abnormal brain regions (such as atrophy patterns in AD-related areas like the hippocampus and medial temporal lobe, providing targets for individualized intervention), and reduce misdiagnosis and missed diagnosis (high classification accuracy, such as LDA with an AUC of 0.96, means lower false positive and false negative rates, avoiding overtreatment of low-risk patients or delayed intervention for high-risk patients).

[0069] This study constructed a high-precision MCIc / MCInc classification model (with an optimal KNN accuracy of 93%) based on gray matter volume features of 95 brain regions, demonstrating the crucial role of fine-grained brain region segmentation in early AD risk prediction. Its clinical significance lies in: To achieve non-invasive and objective early screening, making up for the shortcomings of traditional scales; Provide imaging biomarkers for personalized medicine to guide precision intervention; This lays the research foundation for multimodal AI diagnosis and promotes the development of intelligent assisted diagnosis and treatment of neurodegenerative diseases.

[0070] In the future, it is necessary to expand the sample size to verify generalization and explore the combination of deep learning and multimodal data to further enhance its clinical translational value.

[0071] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. A brain imaging-based system for predicting the risk of conversion to mild cognitive impairment, characterized in that, include: The data preprocessing module is used to receive raw brain T1-weighted magnetic resonance imaging data and preprocess the T1-weighted magnetic resonance imaging data using a voxel-based morphological analysis method, including decranialization, tissue segmentation, spatial normalization and modulation, to extract gray matter volume data. The brain region feature extraction module is connected to the data preprocessing module and is used to extract gray matter volume features of multiple preset brain regions from the standardized gray matter volume image based on the preset brain atlas. A risk classification model is connected to the brain region feature extraction module. The risk classification model is trained to output a prediction result that classifies patients with mild cognitive impairment into high-risk or low-risk Alzheimer's disease based on the gray matter volume features of the multiple preset brain regions. The output module is used to display the prediction results and related classification evaluation indicators.

2. The brain imaging-based system for predicting the risk of conversion to mild cognitive impairment according to claim 1, characterized in that, The preset brain atlas is a fine-grained brain region atlas containing at least 90 brain regions.

3. The system for predicting the risk of conversion to mild cognitive impairment based on brain imaging as described in claim 1, characterized in that, A regional map containing 95 brain regions.

4. The brain imaging-based system for predicting the risk of conversion to mild cognitive impairment according to claim 2, characterized in that, The brain region feature extraction module prioritizes extracting gray matter volume features of relevant brain regions such as the medial temporal lobe, posterior cingulate cortex, precuneus, and medial prefrontal lobe.

5. The brain imaging-based system for predicting the conversion risk of mild cognitive impairment according to any one of claims 2, characterized in that, The risk classification model is constructed using the K-nearest neighbor algorithm, and the AUC value of the K-nearest neighbor algorithm on a test set containing at least 95 brain region features is not less than 0.

91.

6. A method for predicting the risk of conversion to mild cognitive impairment using the prediction system according to any one of claims 1 to 5, characterized in that, Includes the following steps: S1. Obtain raw T1-weighted magnetic resonance imaging (MRI) data of the brain of the individual to be tested; S2. The T1-weighted magnetic resonance imaging data are preprocessed using a voxel-based morphological analysis method to obtain a standardized gray matter volume image; S3. Based on a preset brain atlas, extract gray matter volume features of multiple preset brain regions from the standardized gray matter volume image to form a feature vector; S4. Input the feature vector into a pre-trained risk classification model, and the risk classification model outputs a classification prediction result of whether the individual to be tested belongs to the high-risk or low-risk type of Alzheimer's disease conversion; S5. Output and store the classification prediction results.

7. The method for predicting the risk of conversion to mild cognitive impairment according to claim 6, characterized in that, The training method for the risk classification model includes: Collect a sample dataset with known classification labels, including high-risk and low-risk Alzheimer's disease conversion; Perform steps S1-S3 on the sample data to generate a set of feature vectors for training. The K-nearest neighbors algorithm is used to train the training feature vector set, and the algorithm parameters are optimized until the area under the AUC-ROC curve of the model on the validation set reaches a preset threshold.

Images