Early screening multi-modal feature fusion prediction method and system for alzheimer's disease

Abstract

The application discloses an Alzheimer's disease early screening multi-modal feature fusion prediction method and system, and belongs to the technical field of brain disease prediction; the method comprises the following steps: acquiring multi-dimensional data containing cognitive test scores, brain image scanning results and biomarker concentration levels from a patient record database, and obtaining a unified format multi-dimensional data set through standardization processing; adopting a dimension reduction analysis method to extract a key feature vector to capture the covariation relationship between the cognitive test scores and the brain image changes; when the cognitive test score decline amplitude exceeds a preset threshold and the brain image shows atrophy signs, a classification model is constructed through an ensemble learning method to preliminarily classify abnormal signals; and the abnormal signal vector is obtained by fusing the biomarker concentration levels; and the application significantly improves the accuracy and efficiency of early screening of brain diseases such as Alzheimer's disease through precise evaluation of key parts of the human brain, such as the hippocampus.

Description

Technical Field

[0001] This invention relates to the field of brain disease prediction technology, specifically to a multimodal feature fusion prediction method and system for early screening of Alzheimer's disease. Background Technology

[0002] Alzheimer's disease (AD), also known as senile dementia, is the most common and important neurodegenerative disease of unknown cause among the elderly. Its core pathological changes include extracellular senile plaques formed by the deposition of amyloid-β (Aβ) protein, intracellular neurofibrillary tangles formed by hyperphosphorylation of tau protein, and neuronal loss accompanied by glial cell proliferation. With the continued growth of the global aging population, dementia has become the "fourth leading cause of death" among the elderly, after cardiovascular disease, cerebrovascular disease, and cancer, undoubtedly placing enormous pressure on families, society, and the nation.

[0003] Alzheimer's disease has an insidious onset, and its causes and mechanisms are not yet fully understood. Once diagnosed, it is often too late to save patients. Early detection and intervention in AD patients can improve their clinical symptoms, reduce disability and mortality rates, and save on medical expenses.

[0004] In clinical practice, the diagnosis of cognitive impairment relies on the comprehensive analysis of multidimensional data, including cognitive test scores, brain imaging results, and biomarker concentration levels. These heterogeneous data sources differ significantly in format, dimensions, and acquisition criteria, leading to difficulties in data fusion. The relationship between cognitive decline and changes in brain structure is complex, and traditional analytical methods struggle to accurately identify early pathological signals, particularly in capturing subtle covariate relationships.

[0005] Patent CN120413030A discloses a method, medium, and device for predicting the risk of Alzheimer's disease based on feature fusion. The method first acquires sMRI images, FDG-PET images, biomarkers, and intelligence scores of the target subject. Next, the sMRI and FDG-PET image data are input into a three-dimensional convolutional neural network module to extract corresponding first and second image feature maps. Then, these two feature maps are input into a cross-attention fusion module to obtain image fusion features; these features are then input into a U-shaped upsampling network to segment the hippocampus region, obtaining hippocampal features. Subsequently, the image fusion features, hippocampal features, biomarker data, and intelligence score data are merged, and the merged data is input into an extreme learning machine for calculation, thereby obtaining the prediction result of the risk of developing Alzheimer's disease. This method, by leveraging cross-modal data fusion and extracting complementary information from multimodal data, can effectively improve the accuracy of disease risk prediction.

[0006] However, in practice, it cannot meet the complex analytical requirements of multi-level and multi-dimensional risk assessment of cognitive impairment diseases, nor can it achieve refined classification of abnormal signals and accurate quantification of risk levels.

[0007] How to effectively integrate multi-source heterogeneous medical data, accurately identify the correlation patterns between cognitive tests and brain imaging, and establish a precise multi-level risk assessment system has become a key issue that urgently needs to be addressed in the development of early screening technologies for cognitive impairment. Summary of the Invention

[0008] The purpose of this invention is to provide a multimodal feature fusion prediction method and system for early screening of Alzheimer's disease, which can effectively solve the logical correlation problem between multidimensional data fusion and risk prediction.

[0009] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0010] A multimodal feature fusion prediction method for early screening of Alzheimer's disease includes the following steps:

[0011] Step 1: Obtain multidimensional data from the patient record database. The multidimensional data includes cognitive test scores, brain imaging scan results, and biomarker concentration levels. Normalize the multidimensional data using standardization processing to obtain a multidimensional dataset in a unified format.

[0012] Step 2: Based on the unified format multidimensional dataset, use dimensionality reduction analysis to extract key feature vectors. The key feature vectors are used to capture the covariance between cognitive test scores and changes in brain imaging, and the key feature vectors are determined.

[0013] Step 3: If the cognitive test score in the key feature vector decreases by more than a preset threshold and brain imaging shows signs of atrophy, then a classification model is constructed using an ensemble learning method to preliminarily classify the abnormal signal and determine the preliminary category of the abnormal signal. The abnormal signal refers to the deviation index in the cognitive and imaging data.

[0014] Step 4: For the preliminary category of the abnormal signal, obtain the biomarker concentration level as supplementary input, and use a classification fusion method to fuse the preliminary category and the biomarker concentration level to obtain the fused abnormal signal vector;

[0015] Step 5: Calculate the distance between the vectors based on the fused abnormal signal vectors to quantify the correlation strength. If the correlation strength is greater than a preset threshold, it is determined to be a high-risk signal, and a set of high-risk signals is determined.

[0016] Step 6: Based on the high-risk signal set, obtain the multimodal examination combination recommended based on the subject's background information as supplementary input, and fuse the high-risk signal set with the multimodal examination combination to obtain the fused high-risk screening set;

[0017] Step 7: Apply clustering methods to the fused high-risk screening set to group similar signals, and calculate the comprehensive risk score for each group using a weighted average method to obtain the comprehensive risk score for each group;

[0018] Step 8: For each group of comprehensive risk scores, if the score is higher than the preset warning threshold, it is marked as a potential disease risk. The regression prediction method is used to predict the risk progression trend to obtain the final risk identification result.

[0019] In one embodiment of the present invention, step 1 includes: acquiring multidimensional data from a patient database, including cognitive test scores, brain imaging scan results and biomarker concentration levels, to complete the initial data collection and obtain a structured multidimensional data set;

[0020] Based on the collected multidimensional dataset, a standardization processing method is used to normalize the data from different sources, eliminate differences in units, and obtain a multidimensional dataset with a uniform format.

[0021] For a uniformly formatted multidimensional dataset, perform a data integrity check. If data is missing or abnormal, fill it in using a preset interpolation method to determine the multidimensional data matrix with restored integrity.

[0022] We obtained a multidimensional data matrix with restored integrity. For the cognitive test scores and biomarker concentration levels, we used correlation analysis to determine the strength of the association between the data in each dimension and obtained the correlation results between dimensions.

[0023] Based on the correlation results between dimensions, we extracted the concentration levels of biomarkers and brain imaging features that were highly correlated with cognitive test scores to determine the combination of key influencing factors.

[0024] By combining key influencing factors, the support vector machine algorithm is applied to classify the multidimensional data matrix, distinguish patient categories with different health states, and obtain the classified data groups.

[0025] For the classified data groups, generate corresponding feature weight distributions, determine the contribution of each key influencing factor to the classification results, and obtain the final feature importance ranking.

[0026] In one embodiment of the present invention, step 2 includes:

[0027] The unified format multidimensional dataset is obtained, and the cognitive test scores and brain imaging change data are preliminarily sorted using preset data filtering rules to obtain a sorted data subset.

[0028] For the sorted data subset, the high-dimensional data is processed by dimensionality reduction analysis to extract the main feature vectors that can characterize the correlation pattern between cognitive test scores and brain imaging changes, and to determine the set of extracted feature vectors.

[0029] Based on the extracted feature vector set, a data grouping method is used to classify the multidimensional dataset, dividing the data according to different association patterns to obtain the classified data groups;

[0030] For the categorized data groups, if the amount of data in a certain group is lower than a preset threshold, the group is supplemented using data augmentation techniques to obtain supplemented data groups.

[0031] Based on the supplemented data grouping, statistical analysis tools are used to compare the distribution of the main feature vectors within each group, determine the differences in feature vectors between groups, and obtain the difference analysis results.

[0032] Based on the results of the difference analysis, data visualization technology is used to graphically display the feature vector distribution of each group, and the final distribution feature map is determined.

[0033] In one embodiment of the present invention, step 3 includes:

[0034] By extracting relevant data from cognitive tests and brain imaging from key features, an initial dataset was constructed, and preliminary records of score decline and signs of atrophy were completed, resulting in a basic dataset.

[0035] Based on the basic dataset, preset thresholds are applied to filter for signs of score decline and shrinkage. If a score decline exceeding the preset threshold is detected and the image data shows signs of shrinkage, it is marked as a potential abnormal signal, and a subset of data to be classified is determined.

[0036] An ensemble learning approach is used to process the deviation indicators in the subset of data to be classified, construct a classification model, perform preliminary classification of abnormal signals, and obtain classification results.

[0037] Based on the classification results, the correspondence between abnormal signals and the initial categories is analyzed. If the classification results show that the abnormal signal matches a certain initial category, it is classified into the corresponding category to obtain the classified signal set.

[0038] Based on the classified signal set, deviation indicators related to cognitive tests are extracted and compared with the data. If the comparison results show that the deviation indicators are significantly different from historical data, they are marked as high-priority signals to identify key targets for attention.

[0039] By focusing on key targets and combining them with signs of atrophy in brain imaging, detailed feature description records are generated to further archive abnormal signals and obtain the final classification file.

[0040] Based on the final classification files, the mapping relationship between key features and initial categories is organized to complete the full-process classification processing of abnormal signals, resulting in structured classification data storage.

[0041] The preset thresholds are set based on training with historical clinical data. The threshold for the decline in cognitive test scores is set at 15%, which is determined by analyzing the 90th percentile of the score distribution in healthy and early AD groups. The threshold for the association strength is set at 0.7, which is obtained by analyzing the balance point between sensitivity and specificity through ROC curve analysis.

[0042] In one embodiment of the present invention, step 4 includes:

[0043] By obtaining preliminary classification results of abnormal signals from the monitoring system, the signals are pre-classified using a pre-established classification model to obtain initial classification labels.

[0044] Based on the initial classification labels, obtain the corresponding biomarker concentration data, and use the relevant records stored in the database to match and determine the concentration information associated with the classification labels;

[0045] Based on the concentration information obtained from the matching and the initial classification label, the support vector machine algorithm is used to perform feature integration processing to obtain the integrated feature vector;

[0046] If some data in the integrated feature vector deviates from the preset threshold range, the deviating data is standardized to obtain the corrected feature vector.

[0047] By using the corrected feature vectors, we can obtain potential pattern information related to the anomalous signals and determine whether there is an anomalous pattern distribution.

[0048] Based on the abnormal pattern distribution information obtained from the judgment, a pre-set rule base is used to perform pattern comparison to determine the final abnormal signal vector;

[0049] For the final abnormal signal vector, its distribution characteristics in a multi-dimensional space are obtained to obtain the comprehensive classification result of the abnormal signal.

[0050] In one embodiment of the present invention, step 5 includes:

[0051] Abnormal signal data is obtained from the system, and fused vector data is obtained by performing feature extraction and vectorization processing on the abnormal signals.

[0052] For the obtained fused vector data, the Euclidean distance method is used to calculate the distance between each vector to generate a distance matrix;

[0053] Based on the generated distance matrix, the correlation strength between each vector is analyzed and compared using a preset threshold. If the correlation strength is greater than the preset threshold, it is judged as a potentially high-risk signal.

[0054] For the identified potential high-risk signals, signal aggregation processing is performed to generate a set of high-risk signals, and the source information of the relevant signals is marked;

[0055] By performing secondary analysis on the signals in the high-risk signal set, the common characteristics among the signals are obtained, and the core correlation patterns of the high-risk signals are determined.

[0056] Based on the determined core association pattern, the signal set is classified and stored to generate a classified signal subset for subsequent processing.

[0057] The classified signal subset is obtained and prioritized based on the source information to obtain the final high-risk signal processing sequence.

[0058] In one embodiment of the present invention, step 6 includes:

[0059] A set of high-risk signals is obtained from the system. For each signal in the set of high-risk signals, the corresponding background information of the subject is extracted to obtain a background information dataset associated with the signal.

[0060] Based on the obtained background information dataset, a pre-established recommendation model is used to filter out multimodal check combinations that match the background information and determine a list of suitable check combinations.

[0061] For a given list of inspection combinations, the availability status of each inspection combination is obtained. If the availability status meets a preset condition, the inspection combination is bound to the corresponding high-risk signal to obtain a bound signal inspection pair.

[0062] By integrating the data from the bound signal inspection pairs and processing it using a unified data format, the signal data and inspection data are standardized to obtain a standardized comprehensive dataset.

[0063] Based on the standardized comprehensive dataset, common information between signals and inspections is extracted. If the common information reaches a preset threshold, it is judged as a highly correlated data pair, and a highly correlated data subset is obtained.

[0064] For highly relevant subsets of data, obtain the corresponding subject background information, classify and label them, and determine the grouping of the classified screening data.

[0065] By grouping the categorized screening data and combining it with the priority information of the signal set, the data is sorted to obtain the final comprehensive high-risk screening set.

[0066] In one embodiment of the present invention, step 7 includes:

[0067] By performing preliminary data processing on the high-risk screening dataset, structured fusion dataset data is obtained, and the initial dataset processing results are determined.

[0068] Based on the structured results of the fused dataset, a clustering grouping method is used to classify similar signals, resulting in grouped datasets.

[0069] Based on the grouping results data, extract the signal features within each group, summarize the feature summary information for each group, and determine the signal feature set after feature summarization.

[0070] By using the signal feature set after feature summarization, the weighted average of the signal features within each group is calculated to obtain preliminary assessment data of comprehensive risk.

[0071] If the preliminary assessment data of the overall risk exceeds the preset threshold range, the signal feature set will be regrouped to obtain the adjusted grouping result data.

[0072] Based on the adjusted grouping results data, the weighted average is recalculated to determine the final comprehensive risk score;

[0073] The high-risk screening results of each group are ranked based on the final comprehensive risk score to obtain a risk priority sequence.

[0074] In one embodiment of the present invention, step 8 includes:

[0075] Obtain comprehensive risk score data, extract users' health-related indicators from the system, calculate and generate a comprehensive risk score, and determine the basis for preliminary risk assessment;

[0076] If the comprehensive risk score is higher than the preset warning threshold, the comprehensive risk score data is marked as a potential disease risk, and the marked risk dataset is obtained.

[0077] Based on the labeled risk dataset, a regression prediction method is used to analyze the potential disease risk, calculate the risk progression trend, and obtain the predicted value of the progression trend.

[0078] By combining the predicted value of the progress trend with historical health data, we can analyze the possible direction of risk progression and determine the dynamic changes in risk progression.

[0079] If the dynamic changes in the progress of a risk exceed the preset safety range, the risk identification process is triggered to obtain a specific risk category classification.

[0080] Based on risk category classification, relevant health indicators and predictive analysis results are integrated to generate the final risk identification conclusion and determine the priority ranking of users' health status.

[0081] Compared with the prior art, the present invention has the following beneficial effects:

[0082] This invention addresses the problem in business scenarios where traditional methods struggle to effectively integrate multi-source data such as cognitive tests, brain imaging, and biomarkers, leading to inaccurate early high-risk signal identification and imprecise trend prediction. It obtains multidimensional data from a patient record database, standardizes it, extracts key feature vectors to capture covariance relationships, and constructs a classification model to initially categorize abnormal signals if the cognitive score in the vector exceeds a threshold and imaging shows shrinkage. It then fuses biomarker levels to obtain abnormal signal vectors, calculates distances to quantify association strength, and determines the set of high-risk signals. Furthermore, it integrates multimodal examination combinations and clusters them to calculate a comprehensive risk score. If the score exceeds a warning threshold, potential risks are marked, and regression analysis predicts progression trends. This solves the logical connection problem between multidimensional data fusion and risk prediction. This invention improves the accuracy and foresight of cognitive disease screening, enabling efficient early intervention. Attached Figure Description

[0083] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0084] Figure 1 This is a flowchart of the data acquisition and preprocessing process of the present invention.

[0085] Figure 2 This is a flowchart of the feature extraction and key feature vector determination process of the present invention.

[0086] Figure 3 This is a flowchart illustrating the preliminary classification of abnormal signals according to the present invention.

[0087] Figure 4 This is a flowchart of the abnormal signal fusion and high-risk signal determination process of the present invention. Detailed Implementation

[0088] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the embodiments of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0089] The following is in conjunction with the appendix Figures 1-4 The embodiments of the present invention will be described in detail below.

[0090] Example 1:

[0091] This embodiment discloses a multimodal feature fusion prediction method for early screening of Alzheimer's disease, including the following steps:

[0092] Step 1: Obtain multidimensional data from the patient record database. The multidimensional data includes cognitive test scores, brain imaging scan results, and biomarker concentration levels. Normalize the multidimensional data using standardization processing to obtain a multidimensional dataset in a unified format.

[0093] Step 2: Based on the unified format multidimensional dataset, use dimensionality reduction analysis to extract key feature vectors. The key feature vectors are used to capture the covariance between cognitive test scores and changes in brain imaging, and the key feature vectors are determined.

[0094] Step 3: If the cognitive test score in the key feature vector decreases by more than a preset threshold and brain imaging shows signs of atrophy, then a classification model is constructed using an ensemble learning method to preliminarily classify the abnormal signal and determine the preliminary category of the abnormal signal. The abnormal signal refers to the deviation index in the cognitive and imaging data.

[0095] The dimensionality reduction analysis method is principal component analysis (PCA), which retains principal components with 85% variance contribution rate; the ensemble learning method is the random forest algorithm, with 100 trees and a maximum depth of 10, and the input features include cognitive score decline rate and hippocampal volume change rate.

[0096] Step 4: For the preliminary category of the abnormal signal, obtain the biomarker concentration level as supplementary input, and use a classification fusion method to fuse the preliminary category and the biomarker concentration level to obtain the fused abnormal signal vector;

[0097] Step 5: Calculate the distance between the vectors based on the fused abnormal signal vectors to quantify the correlation strength. If the correlation strength is greater than a preset threshold, it is determined to be a high-risk signal, and a set of high-risk signals is determined.

[0098] Step 6: Based on the high-risk signal set, obtain the multimodal examination combination recommended based on the subject's background information as supplementary input, and fuse the high-risk signal set with the multimodal examination combination to obtain the fused high-risk screening set;

[0099] Step 7: Apply clustering methods to the fused high-risk screening set to group similar signals, and calculate the comprehensive risk score for each group using a weighted average method to obtain the comprehensive risk score for each group;

[0100] Step 8: For each group of comprehensive risk scores, if the score is higher than the preset warning threshold, it is marked as a potential disease risk. The regression prediction method is used to predict the risk progression trend to obtain the final risk identification result.

[0101] Furthermore, the deviation index is defined as a cognitive test score that decreases by more than 10% from the baseline or a brain imaging region volume that shrinks by more than 5% from the standard value; the signs of atrophy are determined by a decrease in gray matter density exceeding 2 standard deviations or a decrease in hippocampal volume exceeding 8% in MRI images.

[0102] In the implementation of this invention, all preset thresholds (such as cognitive test score decline threshold, brain image atrophy threshold, association strength threshold, risk score threshold, etc.) are derived from training based on historical clinical data.

[0103] Specifically, this is determined through the following technical process:

[0104] 1. Data preparation: Collect historical clinical data from no fewer than 1,000 confirmed subjects, including cognitive test scores, brain MRI image measurements, and blood biomarker concentrations;

[0105] 2. Feature analysis: Perform distribution statistics for each technical dimension, and calculate the mean, standard deviation, and overlap area between the healthy group and the early AD group;

[0106] 3. Threshold optimization: A grid search combined with cross-validation method is used to test the classification accuracy of different threshold combinations on the validation set;

[0107] 4. Final determination: Select the threshold combination that achieves the highest F1 score on the independent test set as the preset threshold.

[0108] For example, the process of determining the 15% threshold for cognitive test score decline: Analyzing the annual cognitive test changes of 500 healthy subjects and 500 early-stage AD patients, it was found that the median annual decline rate was 3% (IQR: 1-5%) in the healthy group and 18% (IQR: 12-25%) in the early-stage AD group. ROC curve analysis determined that 15% was the optimal classification cutoff point (AUC=0.89).

[0109] Furthermore, step 1 includes:

[0110] By acquiring multidimensional data from the patient database, including cognitive test scores, brain imaging scan results, and biomarker concentration levels, the initial data collection was completed, resulting in a structured multidimensional dataset.

[0111] Based on the collected multidimensional dataset, a standardization processing method is used to normalize the data from different sources, eliminate differences in units, and obtain a multidimensional dataset with a uniform format.

[0112] For a uniformly formatted multidimensional dataset, perform a data integrity check. If data is missing or abnormal, fill it in using a preset interpolation method to determine the multidimensional data matrix with restored integrity.

[0113] We obtained a multidimensional data matrix with restored integrity. For the cognitive test scores and biomarker concentration levels, we used correlation analysis to determine the strength of the association between the data in each dimension and obtained the correlation results between dimensions.

[0114] Based on the correlation results between dimensions, we extracted the concentration levels of biomarkers and brain imaging features that were highly correlated with cognitive test scores to determine the combination of key influencing factors.

[0115] By combining key influencing factors, the support vector machine algorithm is applied to classify the multidimensional data matrix, distinguish patient categories with different health states, and obtain the classified data groups.

[0116] For the classified data groups, generate corresponding feature weight distributions, determine the contribution of each key influencing factor to the classification results, and obtain the final feature importance ranking.

[0117] In practical implementation:

[0118] In some preferred embodiments, during the patient data collection phase, multidimensional information can be extracted from the hospital database, including cognitive test scores such as memory test results, brain imaging data such as hippocampal volume, and biomarker concentrations such as blood amyloid protein levels. For example, assuming a patient has a memory test score of 75, a hippocampal volume of 3.2 cubic centimeters, and an amyloid protein concentration of 150 picograms per milliliter, these data constitute the initial multidimensional data set.

[0119] In one possible implementation, standardization of data with different dimensions can map cognitive test scores, image data, and biomarker concentrations to a range of 0 to 1.

[0120] In some preferred embodiments, a memory test score of 75 points falls within the range of 50 to 100 points across all patients. After normalization, this score may become 0.5, eliminating dimensional differences and facilitating subsequent analysis. This helps improve the stability of model training and prevents excessive influence of a single dimension's data range on the results.

[0121] In some preferred embodiments, if amyloid concentration data for a particular patient is found to be missing during data integrity checks, mean interpolation can be used to fill the gaps with the average concentration of patients in the same group, such as 145 picograms per milliliter. This effectively repairs missing data, ensures the integrity of the multidimensional data matrix, and lays the foundation for subsequent analysis.

[0122] In one possible implementation, correlation analysis can be performed by calculating the correlation coefficient between cognitive test scores and biomarker concentrations. Assuming a correlation coefficient of -0.8 between memory test scores and amyloid concentration indicates a strong negative correlation, meaning higher concentrations correlate with lower cognitive ability. This analysis helps identify key factors significantly impacting cognitive function, improving the specificity of subsequent classification. In some preferred embodiments, after extracting key influencing factors, assuming a high correlation between amyloid concentration and hippocampal volume and cognitive test scores, these can be used as key combinations. When applying the support vector machine algorithm, patients can be categorized into healthy, mild cognitive impairment, and severe impairment groups based on these factors. During classification, the algorithm automatically defines boundaries based on data characteristics, effectively distinguishing different health states and improving diagnostic accuracy.

[0123] In one possible implementation, when generating the feature weight distribution, it is assumed that the weight of amyloid concentration is 0.6 and the weight of hippocampal volume is 0.3, indicating that the former contributes more to the classification results. Such feature importance ranking helps doctors focus on key influencing factors, optimize diagnostic and intervention strategies, and provide data support for subsequent research.

[0124] The processing and analysis of the multidimensional data not only improved the accuracy of classification but also provided a scientific basis for personalized medicine, demonstrating significant technological value.

[0125] Furthermore, step 2 includes:

[0126] The unified format multidimensional dataset is obtained, and the cognitive test scores and brain imaging change data are preliminarily sorted using preset data filtering rules to obtain a sorted data subset.

[0127] For the sorted data subset, the high-dimensional data is processed by dimensionality reduction analysis to extract the main feature vectors that can characterize the correlation pattern between cognitive test scores and brain imaging changes, and to determine the set of extracted feature vectors.

[0128] Based on the extracted feature vector set, a data grouping method is used to classify the multidimensional dataset, dividing the data according to different association patterns to obtain the classified data groups;

[0129] For the categorized data groups, if the amount of data in a certain group is lower than a preset threshold, the group is supplemented using data augmentation techniques to obtain supplemented data groups.

[0130] Based on the supplemented data grouping, statistical analysis tools are used to compare the distribution of the main feature vectors within each group, determine the differences in feature vectors between groups, and obtain the difference analysis results.

[0131] Based on the results of the difference analysis, data visualization technology is used to graphically display the feature vector distribution of each group, and the final distribution feature map is determined.

[0132] In practice:

[0133] In some preferred embodiments, when processing a uniformly formatted cube, preliminary sorting can be performed on cognitive test scores and brain imaging change data. Assume the cognitive test scores extracted from the database range from 0 to 100, and the brain imaging change data includes gray matter volume changes ranging from -5 to 5 cubic centimeters. Using preset filtering rules, records with test scores below 10 or missing imaging data are removed, resulting in a sorted subset of data. This filtering helps ensure data quality and lays the foundation for subsequent analysis.

[0134] In some preferred embodiments, high-dimensional data can be processed using principal component analysis (PCA) during the dimensionality reduction analysis stage. Assuming the original data contains 10 dimensions, such as volume changes in different brain regions and scores on multiple cognitive tests, dimensionality reduction extracts two main feature vectors, representing a comprehensive cognitive ability index and a brain structure change pattern, respectively. This simplifies data complexity while retaining key information.

[0135] In some preferred embodiments, when grouping data for a set of feature vectors, clustering methods can be used to divide the data into three groups based on association patterns, such as a cognitively normal group, a mildly changed group, and a significantly changed group. If a group contains only 20 data points, below a preset threshold of 50, virtual samples are generated to supplement the data using data augmentation techniques.

[0136] In some preferred embodiments, 30 new data points are simulated and generated based on the distribution characteristics of existing data to ensure a balanced amount of grouped data.

[0137] In some preferred embodiments, during the distribution comparison phase, statistical analysis tools are used to analyze the main characteristic vectors of each group. Assuming the mean characteristic vector of the cognitively normal group is 0.8, the mean of the mean vector of the mildly affected group is 0.5, and the mean of the mean vector of the significantly affected group is 0.2, the comparison reveals significant differences between the groups. This helps to reveal the characteristic patterns of different groups and provides a reference for subsequent research.

[0138] In some preferred embodiments, during the data visualization stage, the feature vector distribution can be displayed as a scatter plot. Assuming the horizontal axis represents the comprehensive cognitive ability index and the vertical axis represents the brain structure change pattern, data points in each group are labeled with different colors, such as blue for the normal group, yellow for the mildly changed group, and red for the significantly changed group. The graphical display intuitively reflects the distribution characteristics of each group, facilitating understanding and analysis.

[0139] In some preferred embodiments, for the specific implementation of data augmentation, new samples conforming to the distribution pattern can be generated based on the statistical characteristics of existing data. Assuming the mean of the feature vector of the slightly changed group is 0.5 and the standard deviation is 0.1, the generated new data will fluctuate around this range, such as 0.48 or 0.52. This ensures the reasonableness of the supplementary data and avoids introducing bias.

[0140] In some preferred embodiments, during the specific grouping and classification operations, classification boundaries can be set based on the numerical range of the feature vectors. For example, a cognitive ability composite index greater than 0.7 is classified as the normal group, 0.4 to 0.7 as the mildly affected group, and below 0.4 as the significantly affected group. Clear grouping rules help to quickly segment data and improve processing efficiency.

[0141] Furthermore, step 3 includes:

[0142] By extracting relevant data from cognitive tests and brain imaging from key features, an initial dataset was constructed, and preliminary records of score decline and signs of atrophy were completed, resulting in a basic dataset.

[0143] Based on the basic dataset, preset thresholds are applied to filter for signs of score decline and shrinkage. If a score decline exceeding the preset threshold is detected and the image data shows signs of shrinkage, it is marked as a potential abnormal signal, and a subset of data to be classified is determined.

[0144] An ensemble learning approach is used to process the deviation indicators in the subset of data to be classified, construct a classification model, perform preliminary classification of abnormal signals, and obtain classification results.

[0145] Based on the classification results, the correspondence between abnormal signals and the initial categories is analyzed. If the classification results show that the abnormal signal matches a certain initial category, it is classified into the corresponding category to obtain the classified signal set.

[0146] Based on the classified signal set, deviation indicators related to cognitive tests are extracted and compared with the data. If the comparison results show that the deviation indicators are significantly different from historical data, they are marked as high-priority signals to identify key targets for attention.

[0147] By focusing on key targets and combining them with signs of atrophy in brain imaging, detailed feature description records are generated to further archive abnormal signals and obtain the final classification file.

[0148] Based on the final classification files, the mapping relationship between key features and initial categories is organized to complete the full-process classification processing of abnormal signals, resulting in structured classification data storage.

[0149] In practical implementation:

[0150] In some preferred embodiments, constructing an initial dataset is a crucial step in processing cognitive test and brain imaging data. This initial dataset can be built by collecting data from multiple sources, such as scores obtained through standardized cognitive testing tools, combined with MRI images recording changes in brain structure. For example, assuming a data collection session includes 100 participants, 30 of whom have test scores below average and whose images show signs of brain atrophy, these data will be recorded as a preliminary baseline dataset for subsequent analysis.

[0151] In some preferred embodiments, a preset threshold can be set for screening for signs of score decline and atrophy. For example, a cognitive test score decline of more than 10% and an image atrophy area exceeding 5% are marked as potential abnormal signals. In a particular screening, assuming 15 subjects meet this condition, their data will be categorized into a subset of data to be classified. This helps to quickly identify individuals who may have problems, facilitating further in-depth analysis.

[0152] In some preferred embodiments, when constructing a classification model using ensemble learning methods, multiple algorithms can be combined to process the bias indicators in the data subset. For example, when classifying anomalous signals, the model might comprehensively consider multiple dimensions such as the magnitude of score decline and the location of the atrophic region, ultimately classifying the anomalous signals into three categories: mild, moderate, and severe. If a subject's score drops by 15% and imaging shows slight hippocampal atrophy, they might be classified as a mild anomalous signal. This provides a clear category division for subsequent processing.

[0153] Specifically,

[0154] In some preferred embodiments, when analyzing the correspondence between abnormal signals and initial categories, if a subject's classification result shows moderate abnormality and matches the characteristics of a certain category in historical data, such as hippocampal atrophy accompanied by a decline in memory test scores, then the subject is assigned to the corresponding category, forming a classified signal set. This helps to identify patterns in the data and improve classification accuracy.

[0155] In some preferred embodiments, when extracting bias indicators from the categorized signal set, particular attention can be paid to changes in memory scores in cognitive tests. For example, if a subject's memory score has decreased by 20% compared to historical data and is significantly different from other indicators, it is marked as a high-priority signal. This helps to quickly identify objects that require special attention.

[0156] In some preferred embodiments, when generating feature description records in conjunction with brain imaging, the specific location and extent of atrophy signs can be recorded in detail. For example, if an image of a subject shows a 6% atrophy in the frontal lobe region, this can be combined with the decline in test scores to form a detailed archive. This provides comprehensive data support for subsequent analysis.

[0157] In some preferred embodiments, when organizing the mapping relationship between key features and initial categories, a structured storage framework can be constructed to correspond the classification results of abnormal signals one-to-one with the feature descriptions, such as associating mild abnormalities with atrophy features in specific brain regions. This facilitates data retrieval and long-term tracking, laying the foundation for subsequent research.

[0158] Furthermore, step 4 includes:

[0159] By obtaining preliminary classification results of abnormal signals from the monitoring system, the signals are pre-classified using a pre-established classification model to obtain initial classification labels.

[0160] Based on the initial classification labels, obtain the corresponding biomarker concentration data, and use the relevant records stored in the database to match and determine the concentration information associated with the classification labels;

[0161] Based on the concentration information obtained from the matching and the initial classification label, the support vector machine algorithm is used to perform feature integration processing to obtain the integrated feature vector;

[0162] If some data in the integrated feature vector deviates from the preset threshold range, the deviating data is standardized to obtain the corrected feature vector.

[0163] By using the corrected feature vectors, we can obtain potential pattern information related to the anomalous signals and determine whether there is an anomalous pattern distribution.

[0164] Based on the abnormal pattern distribution information obtained from the judgment, a pre-set rule base is used to perform pattern comparison to determine the final abnormal signal vector;

[0165] For the final abnormal signal vector, its distribution characteristics in a multi-dimensional space are obtained to obtain the comprehensive classification result of the abnormal signal.

[0166] In practice:

[0167] In some preferred embodiments, during the initial classification of anomalous signals, a set of data can be acquired through a monitoring system. For example, a patient's cognitive test score might have decreased by 15%, while brain imaging shows signs of atrophy. Based on a pre-established classification model, the patient is initially categorized as "potentially anomalous." This process involves multi-dimensional analysis of the signals, such as combining historical data on downward trends and imaging features, to ensure the accuracy of the classification labels.

[0168] In some preferred embodiments, when obtaining biomarker concentration data for the initial classification label, it is assumed that the database records a concentration of 2.5 ng / ml for a certain biomarker in the patient, exceeding the normal range of 1.0-2.0 ng / ml. By matching this concentration with the classification label "potential abnormality," its correlation with cognitive decline and visual atrophy is confirmed. The key to this step lies in the comprehensiveness of the database and the rationality of the matching logic, which contributes to the reliability of subsequent analyses.

[0169] In some preferred embodiments, when using the support vector machine algorithm for feature integration, data such as the percentage decrease in cognitive score, the size of the image shrinkage area, and the concentration of biomarkers can be integrated into a single feature vector.

[0170] If, after integration, the biomarker concentration data deviates from a preset threshold range, it is standardized, for example, by proportionally adjusting the concentration values ​​to within a standard range, resulting in a corrected feature vector. This reduces the impact of data bias on subsequent judgments.

[0171] In some preferred embodiments, when obtaining latent pattern information through the corrected feature vector, it is assumed that the analysis reveals that the patient's abnormal signals exhibit a continuously worsening trend in the temporal dimension and are concentrated in specific brain regions in the spatial dimension. This pattern distribution may indicate a specific type of cognitive impairment. By comparing with a preset rule base, such as the "continuous decline + local atrophy" pattern defined in the rule base, the final abnormal signal vector can be further determined. This helps to accurately locate abnormal features.

[0172] In some preferred embodiments, when analyzing the multi-dimensional spatial distribution characteristics of the final abnormal signal vector, it is assumed that its distribution in the cognitive and imaging dimensions is highly similar to other known abnormal cases. The final comprehensive classification result may be categorized as "high-risk group". The classification result can provide a basis for subsequent key monitoring, and the comprehensiveness of the classification is improved through multi-dimensional analysis.

[0173] In some preferred embodiments, the implementation of each technical aspect in the overall process revolves closely around the classification and analysis of anomalous signals. From initial stratification to feature integration, then to pattern comparison and comprehensive classification, each step aims to improve the accuracy of anomalous signal identification. This progressive approach not only refines the classification results but also provides data support for subsequent personalized interventions, significantly improving monitoring efficiency and targeting.

[0174] Furthermore, step 5 includes:

[0175] Abnormal signal data is obtained from the system, and fused vector data is obtained by performing feature extraction and vectorization processing on the abnormal signals.

[0176] For the obtained fused vector data, the Euclidean distance method is used to calculate the distance between each vector to generate a distance matrix;

[0177] Based on the generated distance matrix, the correlation strength between each vector is analyzed and compared using a preset threshold. If the correlation strength is greater than the preset threshold, it is judged as a potentially high-risk signal.

[0178] For the identified potential high-risk signals, signal aggregation processing is performed to generate a set of high-risk signals, and the source information of the relevant signals is marked;

[0179] By performing secondary analysis on the signals in the high-risk signal set, the common characteristics among the signals are obtained, and the core correlation patterns of the high-risk signals are determined.

[0180] Based on the determined core association pattern, the signal set is classified and stored to generate a classified signal subset for subsequent processing.

[0181] The classified signal subset is obtained and prioritized based on the source information to obtain the final high-risk signal processing sequence.

[0182] In practice:

[0183] In some preferred embodiments, feature extraction and vectorization are key steps in processing anomalous signal data. For Alzheimer's disease screening, anomalous signals include abnormal declines in memory test scores and atrophy in specific brain regions. By extracting feature parameters such as the rate of decline in memory test scores and the rate of change in hippocampal volume, a set of feature values ​​is generated, such as a monthly decline rate of 5% in memory scores and an annual decrease rate of 8% in hippocampal volume. These feature values ​​are then converted into vector form, forming fused vector data for subsequent analysis.

[0184] In some preferred embodiments, the Euclidean distance method is often used to measure dissimilarity when calculating the distance between fused vector data. Suppose there are two sets of vector data, representing memory test abnormalities and hippocampal atrophy characteristics of different subjects, respectively. By calculating the distance between the two vectors, a distance matrix can be obtained. The smaller the distance value, the more similar the two signal features are, potentially pointing to the same type of abnormality. This method helps to quickly screen out similar abnormal signals, providing a basis for subsequent analysis.

[0185] In some preferred embodiments, the determination of association strength in the distance matrix analysis relies on a preset threshold. All feature vectors are min-max normalized to the [0,1] interval before distance calculation. Assuming the threshold is set to 0.7, if the Euclidean distance between two vectors is less than 0.7, the association strength is considered high, potentially indicating a high-risk signal. This threshold has been validated using historical data, achieving a classification accuracy of 85% on the test set. This effectively filters out signal data requiring focused attention, reducing the possibility of misjudgment.

[0186] Specifically, for the aggregation and processing of potentially high-risk signals, the signals can be grouped according to their source information. For example, cognitive abnormality signals from the same medical institution can be grouped into one set and their source can be labeled as device A. This facilitates tracing the source of the signals and improves the targeting of subsequent processing.

[0187] When performing secondary analysis on a set of high-risk signals, extracting common features is key. For example, in a group of cognitive abnormality signals, if the memory score decline rate of most signals exceeds 15% and the duration exceeds 10 seconds, this common feature can be considered a core association pattern. This helps to more accurately identify the type of abnormality.

[0188] When storing subsets of signals, they can be categorized based on core correlation patterns. For example, signals with high fluctuation amplitudes can be classified as acute anomalies, while those with lower amplitudes can be classified as chronic anomalies, facilitating quick data retrieval later.

[0189] When prioritizing signals based on source information, the order can be determined by the severity of the signal and the reliability of the source device. For example, acute abnormal signals from core monitoring equipment have the highest priority and are placed at the top of the processing sequence. This prioritization method ensures efficient allocation of processing resources, addressing the most pressing issues first.

[0190] Furthermore, step 6 includes:

[0191] A set of high-risk signals is obtained from the system. For each signal in the set of high-risk signals, the corresponding background information of the subject is extracted to obtain a background information dataset associated with the signal.

[0192] Based on the obtained background information dataset, a pre-established recommendation model is used to filter out multimodal check combinations that match the background information and determine a list of suitable check combinations.

[0193] For a given list of inspection combinations, the availability status of each inspection combination is obtained. If the availability status meets a preset condition, the inspection combination is bound to the corresponding high-risk signal to obtain a bound signal inspection pair.

[0194] By integrating the data from the bound signal inspection pairs and processing it using a unified data format, the signal data and inspection data are standardized to obtain a standardized comprehensive dataset.

[0195] Based on the standardized comprehensive dataset, common information between signals and inspections is extracted. If the common information reaches a preset threshold, it is judged as a highly correlated data pair, and a highly correlated data subset is obtained.

[0196] For highly relevant subsets of data, obtain the corresponding subject background information, classify and label them, and determine the grouping of the classified screening data.

[0197] By grouping the categorized screening data and combining it with the priority information of the signal set, the data is sorted to obtain the final comprehensive high-risk screening set.

[0198] In practice:

[0199] In some preferred embodiments, when processing high-risk signal sets, the extraction of subject background information can involve retrieving signal-related data from the system, such as age, gender, and past medical history. For example, assuming a signal corresponds to a 45-year-old male subject with a history of hypertension, this information will be compiled into a background information dataset for subsequent matching tests. This ensures that the signals are closely linked to the subject's actual condition.

[0200] In some preferred embodiments, when using a recommendation model to screen multimodal examination combinations, suitable examinations can be recommended based on the subject's background information, combined with historical data and expert rules. For example, assuming the aforementioned 45-year-old male has a family history of Alzheimer's disease, the recommendation model might prioritize a combination of PET scan and cerebrospinal fluid examination, as these examinations are highly correlated with Alzheimer's disease-related risk. The screening process can provide more tailored examination plans based on individual differences.

[0201] In some preferred embodiments, the availability of the examination combination can be determined by querying system resources and time schedules to confirm whether the implementation conditions are met. Assuming the cardiac ultrasound equipment is currently available and the appointment time is within the signal processing cycle, the combination can be bound to the corresponding signal to form a signal examination pair. This helps optimize resource allocation and ensure timely examinations.

[0202] In some preferred embodiments, during data integration and standardization, signal data and inspection data can be unified into a structured format. Assuming the signal data contains abnormal frequency values ​​and the inspection data contains equipment parameters, after standardization, both are aligned using timestamps as indexes to form a comprehensive dataset; this facilitates subsequent analysis of the correlation between signals and inspections.

[0203] In some preferred embodiments, when extracting common information and identifying highly correlated data pairs, the matching degree between abnormal signal patterns and examination results can be analyzed. For example, if a signal indicates an abnormal heart rate, and the examination data also suggests arrhythmia, and the two highly overlap in time, if the matching degree exceeds a preset threshold of 80%, it is determined to be a highly correlated data pair, which helps to filter out the signal combinations that truly need attention.

[0204] In some preferred embodiments, when classifying and labeling highly relevant subsets of data, the data can be grouped into different risk levels based on the subjects' background information. For example, based on age and medical history, the data can be divided into high-risk and intermediate-risk groups, with subjects aged 45 years and with a history of hypertension labeled as high-risk. This classification method facilitates subsequent targeted processing.

[0205] In some preferred embodiments, the final ranking of the high-risk screening set can be based on a combination of signal priority and grouping information. Assuming a high-risk group signal has a priority of level 1, the system will prioritize it in the processing sequence, ensuring that the most urgent signals receive a rapid response. This improves processing efficiency and prevents critical signals from being delayed.

[0206] Furthermore, step 7 includes:

[0207] By performing preliminary data processing on the high-risk screening dataset, structured fusion dataset data is obtained, and the initial dataset processing results are determined.

[0208] Based on the structured results of the fused dataset, a clustering grouping method is used to classify similar signals, resulting in grouped datasets.

[0209] Based on the grouping results data, extract the signal features within each group, summarize the feature summary information for each group, and determine the signal feature set after feature summarization.

[0210] By using the signal feature set after feature summarization, the weighted average of the signal features within each group is calculated to obtain preliminary assessment data of comprehensive risk.

[0211] If the preliminary assessment data of the overall risk exceeds the preset threshold range, the signal feature set will be regrouped to obtain the adjusted grouping result data.

[0212] Based on the adjusted grouping results data, the weighted average is recalculated to determine the final comprehensive risk score;

[0213] The high-risk screening results of each group are ranked based on the final comprehensive risk score to obtain a risk priority sequence.

[0214] In practice:

[0215] In some preferred embodiments, preliminary data processing is a key step when processing high-risk screening sets.

[0216] Suppose we are faced with a dataset containing various health risk signals. The goal of data processing is to transform the chaotic signal data into structured, fused aggregate data. A specific implementation could be to first deduplicate and impute missing values, then perform preliminary classification according to signal type, such as grouping abnormal blood pressure and abnormal blood sugar signals separately, forming an initial aggregate processing result. The advantage is that it lays a clear data foundation for subsequent analysis.

[0217] In some preferred embodiments, when using clustering methods to group structured results, classification can be based on signal similarity. Assuming there is risk signal data from 100 subjects, clustering can divide these signals into 5 groups, where signals within each group have high similarity in appearance or risk type. In practice, grouping can be based on the signal's numerical range and the subject's background information; for example, grouping signals of similar age and risk type together helps to more accurately identify risk patterns.

[0218] In some preferred embodiments, when extracting the signal features of each group and summarizing the feature information, it is assumed that a certain group of signals is related to cardiovascular risk. Feature summarization may include information such as average heart rate and blood pressure fluctuation range. The specific implementation method is to statistically analyze the typical indicators of the signals within each group and extract common features to form a signal feature set, which can quickly focus on the core risk points of each group.

[0219] In some preferred embodiments, when calculating the weighted average to obtain preliminary comprehensive risk assessment data, different weights can be assigned to different features. For example, abnormal heart rate has a weight of 0.4, abnormal blood pressure has a weight of 0.3, and other indicators have a weight of 0.3. If the weighted average of a certain group is 0.75, exceeding the preset threshold of 0.6, then secondary grouping is required. In practice, the grouping conditions can be further refined, such as re-dividing by the subject's gender or age group, to obtain adjusted grouping results, which can effectively improve the accuracy of risk assessment.

[0220] In some preferred embodiments, when recalculating the weighted average and determining the final comprehensive risk score, it is assumed that the adjusted weighted average of a certain group decreases to 0.65, which is still higher than the threshold but closer to the actual risk level. The implementation method involves combining the adjusted group data, reallocating weights, and calculating scores. This ensures that the risk assessment results more closely reflect the actual situation.

[0221] In some preferred embodiments, when sorting the high-risk screening results of each group based on the final comprehensive risk score, assuming there are 5 groups of data with scores from highest to lowest of 0.8, 0.75, 0.65, 0.5, and 0.4, the priority sequence is arranged accordingly. Specifically, a sequence list is generated according to the scores from highest to lowest, so that high-risk groups are prioritized during subsequent resource allocation, thereby optimizing resource allocation efficiency and improving the targeting of the screening work.

[0222] Furthermore, step 8 includes:

[0223] Obtain comprehensive risk score data, extract users' health-related indicators from the system, calculate and generate a comprehensive risk score, and determine the basis for preliminary risk assessment;

[0224] For the comprehensive risk score, if its value is higher than the preset warning threshold, the data is marked as a potential disease risk, and the marked risk dataset is obtained;

[0225] Based on the labeled risk dataset, a regression prediction method is used to analyze the potential disease risk, calculate the risk progression trend, and obtain the predicted value of the progression trend.

[0226] By combining the predicted value of the progress trend with historical health data, we can analyze the possible direction of risk progression and determine the dynamic changes in risk progression.

[0227] If the dynamic changes in the progress of a risk exceed the preset safety range, the risk identification process is triggered to obtain a specific risk category classification.

[0228] Based on risk category classification, relevant health indicators and predictive analysis results are integrated to generate the final risk identification conclusion and determine the priority ranking of users' health status.

[0229] In practice:

[0230] In some preferred embodiments, within the business domain of health risk assessment, the acquisition of users' health-related indicators and the generation of comprehensive risk scores can be achieved through multi-dimensional data collection. Assuming the system extracts key indicators such as heart rate, blood pressure, and blood sugar from the user's daily monitoring devices, and combines this with basic information such as the user's age and gender, a comprehensive risk score is generated through weighted calculation. For example, if the heart rate is 85 beats / minute, blood pressure is 130 / 85 mmHg, and blood sugar is 5.5 mmol / L, the system will score these indicators according to preset health standard ranges, ultimately obtaining a comprehensive score, such as 75 points, used to initially determine the level of risk. This ensures the comprehensiveness and objectivity of the assessment.

[0231] In some preferred embodiments, regarding the labeling process for users with a comprehensive risk score exceeding a preset warning threshold (assuming the threshold is 80 points), the system automatically labels a user with a potential disease risk when their score is 85. This not only facilitates subsequent analysis but also helps quickly identify user groups requiring focused attention. The labeled dataset contains basic user information and specific indicators exceeding the threshold, providing data support for further analysis.

[0232] In some preferred embodiments, when using regression prediction methods to analyze potential disease risks, the progression trend of the risk can be predicted based on historical data and the current labeled dataset. For example, assuming the system analyzes a user's blood pressure changes over the past year and finds a gradually increasing trend in their systolic blood pressure, combined with current data, it predicts that the systolic blood pressure may reach 140 mmHg within the next three months, falling into a high-risk range. This can provide important reference for subsequent dynamic change analysis.

[0233] In some preferred embodiments, when analyzing the direction of risk progression by combining historical health data and predicted values, the system may find that a user's elevated blood pressure is related to dietary habits or reduced physical activity. For example, if a user's exercise frequency has decreased from three times a week to once a week over the past six months, and the proportion of high-salt foods in their diet has increased, the system might infer that the risk progression is related to lifestyle, thus providing a basis for subsequent intervention.

[0234] In some preferred embodiments, if the dynamic changes in risk progression exceed the safe range, a risk identification process is triggered, and a specific category can be determined through joint analysis of multiple indicators. For example, assuming both blood pressure and blood sugar remain elevated and the predicted trend worsens, the system may classify it as a cardiovascular disease risk category. This helps to accurately pinpoint the root cause of the problem.

[0235] In some preferred embodiments, when integrating health indicators and prediction results to generate the final risk identification conclusion, the system will synthesize all data and generate a priority ranking of users' health status. For example, if a user is classified as high-risk and their predicted progress is rapid, their priority will be ranked first to allow for timely intervention. This ranking mechanism ensures the rationality of resource allocation and improves overall assessment efficiency.

[0236] It should be noted that in the implementation of this invention, all preset thresholds (such as the cognitive test score decline threshold, brain imaging atrophy threshold, and association strength threshold) are derived based on historical clinical data. Specifically, by collecting a large amount of subject data, statistical methods (such as percentile or ROC curve analysis) are used to determine the optimal threshold to ensure screening accuracy. For example, a 15% cognitive test score decline threshold can be set by analyzing the difference in score distribution between healthy individuals and those with early-stage AD.

[0237] Example 2:

[0238] This embodiment discloses a multimodal feature fusion prediction system for early screening of Alzheimer's disease. The system achieves automated processing and risk assessment of multimodal data through the collaborative work of hardware modules and software algorithms. The multimodal feature fusion prediction method for early screening of Alzheimer's disease specifically includes:

[0239] The data acquisition and preprocessing module is used to acquire multidimensional data from the patient record database. The multidimensional data includes cognitive test scores, brain imaging scan results, and biomarker concentration levels. The multidimensional data is normalized using standardization processing to obtain a multidimensional dataset in a unified format.

[0240] The feature extraction and dimensionality reduction module is used to extract key feature vectors based on the unified format multidimensional dataset using dimensionality reduction analysis methods. The key feature vectors are used to capture the covariance between cognitive test scores and brain image changes, and to determine the key feature vectors.

[0241] An anomaly detection and classification module is used to construct a classification model through an ensemble learning method to preliminarily classify the abnormal signal when the cognitive test score in the key feature vector decreases by more than a preset threshold and the brain imaging shows signs of atrophy, and to determine the preliminary category of the abnormal signal. The abnormal signal refers to the deviation index in the cognitive and imaging data.

[0242] The multi-source data fusion module is used to obtain biomarker concentration levels as supplementary input for the preliminary category of the abnormal signal, and to fuse the preliminary category and the biomarker concentration levels using a classification fusion method to obtain the fused abnormal signal vector.

[0243] The risk assessment and determination module is used to calculate the distance between vectors based on the fused abnormal signal vectors to quantify the correlation strength, and to determine a high-risk signal when the correlation strength is greater than a preset threshold, thereby determining a set of high-risk signals.

[0244] The inspection recommendation and integration module is used to obtain a multimodal inspection combination recommended based on the subject's background information as supplementary input based on the high-risk signal set, and to fuse the high-risk signal set and the multimodal inspection combination to obtain a fused high-risk screening set.

[0245] The risk calculation and grouping module is used to group similar signals by applying a clustering method to the fused high-risk screening set, and to calculate the comprehensive risk score of each group by using a weighted average method.

[0246] The trend prediction and output module is used to mark each group of comprehensive risk scores as potential disease risks when the scores are higher than a preset warning threshold, and to use a regression prediction method to predict the risk progression trend to obtain the final risk identification result.

[0247] The system control center is electrically connected to the data acquisition and preprocessing module, feature extraction and dimensionality reduction module, anomaly detection and classification module, multi-source data fusion module, risk assessment and judgment module, inspection recommendation and integration module, risk calculation and grouping module, and trend prediction and output module, and is used to coordinate the data flow and control signals between the modules.

[0248] In practice, the system architecture consists of the following components:

[0249] The Alzheimer's disease early screening multimodal feature fusion prediction system includes a data acquisition and preprocessing module, a feature extraction and dimensionality reduction module, an anomaly detection and classification module, a multi-source data fusion module, a risk assessment and judgment module, an examination recommendation and integration module, a risk calculation and grouping module, a trend prediction and output module, and a system control center.

[0250] The data acquisition and preprocessing module comprises a database interface unit, a data cleaning unit, and a standardization unit. The database interface unit connects to the patient record database via the hospital information system's API interface, supporting HL7 and DICOM standard protocols. The data cleaning unit employs a rule-based anomaly detection algorithm to identify missing data and outliers. The standardization unit uses a min-max normalization method to uniformly map cognitive test scores, brain imaging data, and biomarker concentrations to the [0,1] interval.

[0251] The feature extraction and dimensionality reduction module integrates principal component analysis algorithm, sets a variance explanation rate threshold of 85%, calculates the eigenvectors of the covariance matrix through eigenvalue decomposition, and retains the principal components with eigenvalues ​​greater than 1 as key feature vectors.

[0252] The anomaly detection and classification module employs a random forest algorithm with 100 decision trees and a maximum depth of 10 layers. Hyperparameters are optimized using five-fold cross-validation. The preset thresholds for cognitive test score decline are set at 15% and for brain imaging atrophy at 10% reduction in hippocampal volume. When a dual anomaly is detected, the module activates the classification process and outputs a preliminary category label for the anomaly signal.

[0253] The multi-source data fusion module uses the support vector machine algorithm, selects the radial basis function as the kernel function, sets the regularization parameter C=1.0, and the kernel function parameter γ='scale'. This module presets the amyloid protein concentration threshold to 200 picograms / mL and the tau protein concentration threshold to 50 picograms / mL.

[0254] The risk assessment and judgment module uses Euclidean distance to calculate the correlation strength between vectors, with a preset correlation strength threshold of 0.7. This module includes a distance matrix generation unit and a threshold comparison unit. When a correlation strength exceeding the threshold is detected, it is automatically marked as a high-risk signal and a set of high-risk signals is generated.

[0255] The examination recommendation and integration module integrates a recommendation engine and a rule base. Based on the subject's age, gender, and family medical history, it selects suitable multimodal examination combinations from the examination item library. This module calculates the matching degree between the examination combination and high-risk signals using a weighted similarity algorithm, generating a standardized comprehensive dataset.

[0256] The risk calculation and grouping module uses the K-means clustering algorithm to group similar signals. Through feature weight allocation and weighted averaging, it generates a comprehensive risk score for each group. This module presets a risk score threshold of 0.6; if the threshold is exceeded, a secondary grouping process is triggered.

[0257] The trend prediction and output module is equipped with a multiple linear regression model and a time series analysis algorithm to predict risk progression trends based on historical health data and current risk scores. This module has a preset warning threshold of 0.8; when a risk score exceeds the warning threshold, it automatically generates risk identification conclusions and priority rankings.

[0258] The system control center employs a multi-threaded architecture to coordinate the workflows of each module, and uses message queues to achieve asynchronous communication between modules, ensuring the real-time performance and stability of the data processing flow. This module integrates system monitoring functions to track the operating status and data processing progress of each module in real time.

[0259] System workflow:

[0260] In practice, the system extracts multidimensional patient data from the hospital database through a data acquisition and preprocessing module. For example, a 65-year-old subject might have a cognitive test score of 72, a brain MRI showing a hippocampal volume of 2.8 cubic centimeters, and blood biomarker testing showing an amyloid protein concentration of 220 picograms per milliliter. The data acquisition and preprocessing module normalizes these heterogeneous data, eliminating dimensional differences.

[0261] The feature extraction and dimensionality reduction module performs dimensionality reduction analysis on the normalized data to extract the main correlation patterns between cognitive test scores and changes in brain imaging. Assuming the extracted feature vectors show a correlation coefficient of 0.75 between cognitive scores and hippocampal volume, this indicates a significant covariance relationship.

[0262] The anomaly detection and classification module detected that the subject's cognitive test score decreased by 18% from the baseline and the hippocampal volume decreased by 12%, both exceeding the preset threshold. The integrated learning classifier was then activated, and the signal was labeled as "moderately abnormal".

[0263] The multi-source data fusion module receives the anomaly classification results and combines them with amyloid protein concentration data to generate a fused feature vector using a support vector machine algorithm. Since the amyloid protein concentration exceeds the 200 ng / mL threshold, the module marks the anomaly signal vector as high priority.

[0264] The risk assessment and judgment module calculates that the Euclidean distance between the vector and the known high-risk pattern is 0.75, which exceeds the preset threshold of 0.7. Therefore, it is judged as a high-risk signal and added to the high-risk signal set.

[0265] The examination recommendation and integration module recommends PET scans and cerebrospinal fluid tests as supplementary examination combinations based on the subject's age and family medical history, and binds and integrates these examination data with high-risk signals.

[0266] The risk calculation and grouping module clusters the fused data and calculates the weighted average risk score of the subject's group as 0.72, which exceeds the threshold of 0.6 but is below the warning threshold of 0.8.

[0267] The trend prediction and output module predicts the risk progression trend of the subject over the next 12 months based on historical data, generates a moderate risk identification conclusion, and recommends a follow-up examination after 6 months.

[0268] The system control center monitors the entire data processing flow to ensure that all modules work together and ultimately generate a structured risk assessment report for clinicians to refer to.

[0269] Through multi-module collaboration and algorithm integration, this system has achieved automation, precision, and standardization in early screening for Alzheimer's disease, significantly improving screening efficiency and accuracy.

[0270] In this invention, early screening for Alzheimer's disease is broken down into basic technical units such as multimodal data acquisition, feature extraction, anomaly detection, and risk assessment. Through multimodal data fusion (cognitive test scores, brain imaging results, and biomarker concentration levels), key feature vectors are extracted using standardization and dimensionality reduction analysis methods to accurately capture the covariance between cognitive tests and changes in brain imaging. This achieves refined classification of abnormal signals, significantly reducing the false positive rate and improving the detection rate of early pathological signals. The correlation strength is quantified by calculating the Euclidean distance between fused vectors, and high-risk signals are automatically identified based on a preset threshold (e.g., correlation strength > 0.7). A comprehensive risk score (threshold 0.6-0.8) is generated using clustering and weighted averaging methods, making the risk assessment process entirely data-driven, avoiding subjective bias, and reducing risk grading errors. More importantly, this invention recommends multimodal examination combinations based on the subject's background information (e.g., age, family history) and generates personalized screening sets through data binding and standardization, improving examination efficiency. Simultaneously, a regression prediction model predicts risk progression trends, providing dynamic evidence for intervention strategies and shortening the diagnostic cycle.

[0271] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0272] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multimodal feature fusion prediction method for early screening of Alzheimer's disease, characterized in that, Includes the following steps: Step 1: Obtain multidimensional data from the patient record database. The multidimensional data includes cognitive test scores, brain imaging scan results, and biomarker concentration levels. Normalize the multidimensional data using standardization processing to obtain a multidimensional dataset in a unified format. Step 2: Based on the unified format multidimensional dataset, use dimensionality reduction analysis to extract key feature vectors. The key feature vectors are used to capture the covariance between cognitive test scores and changes in brain imaging, and the key feature vectors are determined. Step 3: If the cognitive test score in the key feature vector decreases by more than a preset threshold and brain imaging shows signs of atrophy, a classification model is constructed using an ensemble learning method to preliminarily classify the abnormal signal and determine its preliminary category. The abnormal signal refers to a deviation index in the cognitive and imaging data. The deviation index is defined as a cognitive test score decreasing by more than 10% from the baseline or a brain imaging region shrinking by more than 5% from the standard value. Step 4: For the preliminary category of the abnormal signal, obtain the biomarker concentration level as supplementary input, and use a classification fusion method to fuse the preliminary category and the biomarker concentration level to obtain the fused abnormal signal vector; Step 5: Calculate the distance between the vectors based on the fused abnormal signal vectors to quantify the correlation strength. If the correlation strength is greater than a preset threshold, it is determined to be a high-risk signal, and a set of high-risk signals is determined. Step 6: Based on the high-risk signal set, obtain the multimodal examination combination recommended based on the subject's background information as supplementary input, and fuse the high-risk signal set with the multimodal examination combination to obtain the fused high-risk screening set; Step 7: Apply clustering methods to the fused high-risk screening set to group similar signals, and calculate the comprehensive risk score for each group using a weighted average method to obtain the comprehensive risk score for each group; Step 8: For each group of comprehensive risk scores, if the comprehensive risk score of each group is higher than the preset warning threshold, it is marked as a potential disease risk. The regression prediction method is used to predict the risk progression trend to obtain the final risk identification result.

2. The multimodal feature fusion prediction method for early screening of Alzheimer's disease according to claim 1, characterized in that, Step 1 includes: By acquiring multidimensional data from the patient database, including cognitive test scores, brain imaging scan results, and biomarker concentration levels, the initial data collection was completed, resulting in a structured multidimensional dataset. Based on the collected multidimensional dataset, a standardization processing method is used to normalize the data from different sources, eliminate differences in units, and obtain a multidimensional dataset with a uniform format. For a uniformly formatted multidimensional dataset, perform a data integrity check. If data is missing or abnormal, fill it in using a preset interpolation method to determine the multidimensional data matrix with restored integrity. We obtained a multidimensional data matrix with restored integrity. For the cognitive test scores and biomarker concentration levels, we used correlation analysis to determine the strength of the association between the data in each dimension and obtained the correlation results between dimensions. Based on the correlation results between dimensions, we extracted the concentration levels of biomarkers and brain imaging features that were highly correlated with cognitive test scores to determine the combination of key influencing factors. By combining key influencing factors, the support vector machine algorithm is applied to classify the multidimensional data matrix, distinguish patient categories with different health states, and obtain the classified data groups. For the classified data groups, generate corresponding feature weight distributions, determine the contribution of each key influencing factor to the classification results, and obtain the final feature importance ranking.

3. The multimodal feature fusion prediction method for early screening of Alzheimer's disease according to claim 1, characterized in that, Step 2 includes: The unified format multidimensional dataset is obtained, and the cognitive test scores and brain imaging change data are preliminarily sorted using preset data filtering rules to obtain a sorted data subset. For the sorted data subset, the high-dimensional data is processed by dimensionality reduction analysis to extract the main feature vectors that can characterize the correlation pattern between cognitive test scores and brain imaging changes, and to determine the set of extracted feature vectors. Based on the extracted feature vector set, a data grouping method is used to classify the multidimensional dataset, dividing the data according to different association patterns to obtain the classified data groups; For the categorized data groups, if the amount of data in a certain group is lower than a preset threshold, the group is supplemented using data augmentation techniques to obtain supplemented data groups. Based on the supplemented data grouping, statistical analysis tools are used to compare the distribution of the main feature vectors within each group, determine the differences in feature vectors between groups, and obtain the difference analysis results. Based on the results of the difference analysis, data visualization technology is used to graphically display the feature vector distribution of each group, and the final distribution feature map is determined.

4. The multimodal feature fusion prediction method for early screening of Alzheimer's disease according to claim 1, characterized in that, Step 3 includes: By extracting relevant data from cognitive tests and brain imaging from key features, an initial dataset was constructed, and preliminary records of score decline and signs of atrophy were completed, resulting in a basic dataset. Based on the basic dataset, preset thresholds are applied to filter for signs of score decline and shrinkage. If a score decline exceeding the preset threshold is detected and the image data shows signs of shrinkage, it is marked as a potential abnormal signal, and a subset of data to be classified is determined. An ensemble learning approach is used to process the deviation indicators in the subset of data to be classified, construct a classification model, perform preliminary classification of abnormal signals, and obtain classification results. Based on the classification results, the correspondence between abnormal signals and the initial categories is analyzed. If the classification results show that the abnormal signal matches a certain initial category, it is classified into the corresponding category to obtain the classified signal set. Based on the classified signal set, deviation indicators related to cognitive tests are extracted and compared with the data. If the comparison results show that the deviation indicators are significantly different from historical data, they are marked as high-priority signals to identify key targets for attention. By focusing on key targets and combining them with signs of atrophy in brain imaging, detailed feature description records are generated to further archive abnormal signals and obtain the final classification file. Based on the final classification files, the mapping relationship between key features and initial categories is organized to complete the full-process classification processing of abnormal signals, resulting in structured classification data storage.

5. The multimodal feature fusion prediction method for early screening of Alzheimer's disease according to claim 1, characterized in that, Step 4 includes: By obtaining preliminary classification results of abnormal signals from the monitoring system, the signals are pre-classified using a pre-established classification model to obtain initial classification labels. Based on the initial classification labels, obtain the corresponding biomarker concentration data, and use the relevant records stored in the database to match and determine the concentration information associated with the classification labels; Based on the concentration information obtained from the matching and the initial classification label, the support vector machine algorithm is used to perform feature integration processing to obtain the integrated feature vector; If some data in the integrated feature vector deviates from the preset threshold range, the deviating data is standardized to obtain the corrected feature vector. By using the corrected feature vectors, we can obtain potential pattern information related to the anomalous signals and determine whether there is an anomalous pattern distribution. Based on the abnormal pattern distribution information obtained from the judgment, a pre-set rule base is used to perform pattern comparison to determine the final abnormal signal vector; For the final abnormal signal vector, its distribution characteristics in a multi-dimensional space are obtained to obtain the comprehensive classification result of the abnormal signal.

6. The multimodal feature fusion prediction method for early screening of Alzheimer's disease according to claim 1, characterized in that, Step 5 includes: Abnormal signal data is obtained from the system, and fused vector data is obtained by performing feature extraction and vectorization processing on the abnormal signals. For the obtained fused vector data, the Euclidean distance method is used to calculate the distance between each vector to generate a distance matrix; Based on the generated distance matrix, the correlation strength between each vector is analyzed and compared using a preset threshold. If the correlation strength is greater than the preset threshold, it is judged as a potentially high-risk signal. For the identified potential high-risk signals, signal aggregation processing is performed to generate a set of high-risk signals, and the source information of the relevant signals is marked; By performing secondary analysis on the signals in the high-risk signal set, the common characteristics among the signals are obtained, and the core correlation patterns of the high-risk signals are determined. Based on the determined core association pattern, the signal set is classified and stored to generate a classified signal subset for subsequent processing. The classified signal subset is obtained and prioritized based on the source information to obtain the final high-risk signal processing sequence.

7. The multimodal feature fusion prediction method for early screening of Alzheimer's disease according to claim 1, characterized in that, Step 6 includes: A set of high-risk signals is obtained from the system. For each signal in the set of high-risk signals, the corresponding background information of the subject is extracted to obtain a background information dataset associated with the signal. Based on the obtained background information dataset, a pre-established recommendation model is used to filter out multimodal check combinations that match the background information and determine a list of suitable check combinations. For a given list of inspection combinations, the availability status of each inspection combination is obtained. If the availability status meets a preset condition, the inspection combination is bound to the corresponding high-risk signal to obtain a bound signal inspection pair. By integrating the data from the bound signal inspection pairs and processing it using a unified data format, the signal data and inspection data are standardized to obtain a standardized comprehensive dataset. Based on the standardized comprehensive dataset, common information between signals and inspections is extracted. If the common information reaches a preset threshold, it is judged as a highly correlated data pair, and a highly correlated data subset is obtained. For highly relevant subsets of data, obtain the corresponding subject background information, classify and label them, and determine the grouping of the classified screening data. By grouping the categorized screening data and combining it with the priority information of the signal set, the data is sorted to obtain the final comprehensive high-risk screening set.

8. The multimodal feature fusion prediction method for early screening of Alzheimer's disease according to claim 1, characterized in that, Step 7 includes: By performing preliminary data processing on the high-risk screening dataset, structured fusion dataset data is obtained, and the initial dataset processing results are determined. Based on the structured results of the fused dataset, a clustering grouping method is used to classify similar signals, resulting in grouped datasets. Based on the grouping results data, extract the signal features within each group, summarize the feature summary information for each group, and determine the signal feature set after feature summarization. By using the signal feature set after feature summarization, the weighted average of the signal features within each group is calculated to obtain preliminary assessment data of comprehensive risk. If the preliminary assessment data of the overall risk exceeds the preset threshold range, the signal feature set will be regrouped to obtain the adjusted grouping result data. Based on the adjusted grouping results data, the weighted average is recalculated to determine the final comprehensive risk score; The high-risk screening results of each group are ranked based on the final comprehensive risk score to obtain a risk priority sequence.

9. The multimodal feature fusion prediction method for early screening of Alzheimer's disease according to claim 1, characterized in that, Step 8 includes: Obtain comprehensive risk score data, extract users' health-related indicators from the system, calculate and generate a comprehensive risk score, and determine the basis for preliminary risk assessment; If the comprehensive risk score is higher than the preset warning threshold, the comprehensive risk score data is marked as a potential disease risk, and the marked risk dataset is obtained. Based on the labeled risk dataset, a regression prediction method is used to analyze the potential disease risk, calculate the risk progression trend, and obtain the predicted value of the progression trend. By combining the predicted value of the progress trend with historical health data, we can analyze the possible direction of risk progression and determine the dynamic changes in risk progression. If the dynamic changes in the progress of a risk exceed the preset safety range, the risk identification process is triggered to obtain a specific risk category classification. Based on risk category classification, relevant health indicators and predictive analysis results are integrated to generate the final risk identification conclusion and determine the priority ranking of users' health status.

10. A multimodal feature fusion prediction system for early screening of Alzheimer's disease, used to execute the multimodal feature fusion prediction method for early screening of Alzheimer's disease as described in claim 1, characterized in that, Specifically, it includes: The data acquisition and preprocessing module is used to acquire multidimensional data from the patient record database. The multidimensional data includes cognitive test scores, brain imaging scan results, and biomarker concentration levels. The multidimensional data is normalized using standardization processing to obtain a multidimensional dataset in a unified format. The feature extraction and dimensionality reduction module is used to extract key feature vectors based on the unified format multidimensional dataset using dimensionality reduction analysis methods. The key feature vectors are used to capture the covariance between cognitive test scores and brain image changes, and to determine the key feature vectors. An anomaly detection and classification module is used to construct a classification model through an ensemble learning method to preliminarily classify the abnormal signal when the cognitive test score in the key feature vector decreases by more than a preset threshold and the brain imaging shows signs of atrophy, and to determine the preliminary category of the abnormal signal. The abnormal signal refers to the deviation index in the cognitive and imaging data. The multi-source data fusion module is used to obtain biomarker concentration levels as supplementary input for the preliminary category of the abnormal signal, and to fuse the preliminary category and the biomarker concentration levels using a classification fusion method to obtain the fused abnormal signal vector. The risk assessment and determination module is used to calculate the distance between vectors based on the fused abnormal signal vectors to quantify the correlation strength, and to determine a high-risk signal when the correlation strength is greater than a preset threshold, thereby determining a set of high-risk signals. The inspection recommendation and integration module is used to obtain a multimodal inspection combination recommended based on the subject's background information as supplementary input based on the high-risk signal set, and to fuse the high-risk signal set and the multimodal inspection combination to obtain a fused high-risk screening set. The risk calculation and grouping module is used to group similar signals by applying a clustering method to the fused high-risk screening set, and to calculate the comprehensive risk score of each group by using a weighted average method. The trend prediction and output module is used to mark each group of comprehensive risk scores as potential disease risks when the scores are higher than a preset warning threshold, and to use a regression prediction method to predict the risk progression trend to obtain the final risk identification result. The system control center is electrically connected to the data acquisition and preprocessing module, feature extraction and dimensionality reduction module, anomaly detection and classification module, multi-source data fusion module, risk assessment and judgment module, inspection recommendation and integration module, risk calculation and grouping module, and trend prediction and output module, and is used to coordinate the data flow and control signals between the modules.

Images