Alzheimer's disease prediction system based on health data and prediction method thereof

Abstract

This invention provides an Alzheimer's disease prediction system and method based on health data, including the following steps: Step 1, acquisition and preprocessing of multimodal health data; Step 2, feature screening and analysis; Step 3, prediction model construction and optimization: selecting logistic regression, random forest, XGBoost and LightGBM algorithms to construct prediction models; Step 4, model evaluation and interpretation; Step 5, model deployment: deploying the best-performing model through a convenient access interface to achieve early risk screening for Alzheimer's disease; High prediction accuracy: integrating multimodal health data to comprehensively characterize AD pathological features, constructing and optimizing various machine learning models, among which the random forest model achieves an accuracy of 0.9535 and an AUC value of 0.9491, the XGBoost and LightGBM models have similar performance and are significantly better than the traditional logistic regression model, with low false negative and false positive rates.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and medical prediction technology, specifically to an Alzheimer's disease prediction system and method based on health data. Background Technology

[0002] Alzheimer's disease (AD) is an irreversible neurodegenerative disease with no effective cure, and its global incidence continues to rise. Early screening and diagnosis are crucial to slowing disease progression, but early symptoms of AD are often subtle, and public awareness of the disease is insufficient. Early symptoms are easily mistaken for age-related causes, leading to delays in intervention.

[0003] Existing AD diagnosis and screening technologies have significant limitations: clinical "gold standard" diagnostic methods (such as cerebrospinal fluid testing and PET scans) are highly accurate, but they are invasive or costly, making them difficult to widely adopt, and by the time a diagnosis is made, most patients have already progressed to the middle or late stages of the disease; traditional symptom assessment and cognitive scale screening are easy to promote, but they rely on the subjective judgment of specialists, lacking precision and failing to achieve large-scale, dynamic early risk screening.

[0004] In recent years, computer-aided diagnostic technology has been gradually applied to Alzheimer's disease (AD) screening. Some studies have proposed diagnostic methods based on single-modality data (such as structural magnetic resonance imaging), but these methods have failed to form a complete evaluation system, cannot fully characterize the pathological features of AD, and the one-sidedness of information leads to limited predictive accuracy. Other studies have constructed relevant predictive models, but they suffer from fragmented data collection, weak model generalization ability, and a lack of convenient interfaces for ordinary users and medical staff, making it difficult to achieve widespread adoption of the technology. At the same time, existing statistical methods cannot fully capture the complex nonlinear relationships between multiple factors when processing high-dimensional, multi-source health data, resulting in model performance that is difficult to meet the needs of clinical practice.

[0005] In summary, there is an urgent need in the field of early screening for Alzheimer's disease for a precise, low-cost, easy-to-use, and widely applicable technological solution to address the shortcomings of existing methods in terms of accuracy, accessibility, and practicality. Summary of the Invention

[0006] This invention provides an Alzheimer's disease prediction system and method based on health data, aiming to solve the problems mentioned in the background art.

[0007] An Alzheimer's disease prediction method based on multimodal health data includes the following steps: Step 1: Acquisition and preprocessing of multimodal health data: Acquire multimodal health data including demographic characteristics, lifestyle indicators, clinical history, physiological and biochemical indicators, cognitive function assessment and symptom manifestations. After data cleaning and integrity checks, the data is randomly divided into training set and test set at a ratio of 8:2. Step 2, Feature Screening and Analysis: Comparisons were performed between groups for all features. For continuous variables, the two independent samples t-test was used, and for categorical variables, the χ² test was used (significance level α=0.05). Features that were significantly associated with Alzheimer's disease (P<0.05) were screened. Step 3: Predictive Model Construction and Optimization: Select logistic regression, random forest, XGBoost and LightGBM algorithms to construct the predictive model, and use a nested cross-validation strategy to optimize the hyperparameters. The inner loop performs 5-fold cross-validation to search for the optimal hyperparameter combination, and the outer loop evaluates the model's generalization performance. The optimization goal is to maximize the recall rate while taking other evaluation metrics into account. Step 4: Model Evaluation and Interpretation: Accuracy, recall, precision, AUC, and F1 score are used as evaluation metrics. ROC curves, PR curves, and confusion matrices are used to verify model performance. Gini importance and SHAP tools are used to quantify the contribution of features to the prediction results and reveal the association between features and disease outcomes. Step 5, Model Deployment: Deploy the best-performing model through a convenient access interface to achieve early risk screening for Alzheimer's disease.

[0008] Preferably, the multimodal health data specifically includes: demographic characteristics such as age, gender, body type, and education level; lifestyle indicators such as physical activity duration, diet quality score, smoking and drinking status; clinical history such as family history of Alzheimer's disease, cardiovascular disease, and diabetes; physiological and biochemical indicators such as systolic blood pressure, diastolic blood pressure, total cholesterol, and low-density lipoprotein cholesterol; cognitive function assessment data such as Mini-Mental State Examination (MMSE) score and Activities of Daily Living (ADL) score; and symptom data such as memory complaints, disorientation, and personality changes.

[0009] Preferably, the data preprocessing further includes missing value imputation and outlier handling. Continuous data is imputed using the mean or median, and categorical data is imputed using the mode. Extreme outliers are identified and removed using statistical methods.

[0010] Preferably, the significant correlation features include MMSE score, functional assessment score, activities of daily living score, memory complaints, behavioral problems, sleep quality, and high-density lipoprotein cholesterol level.

[0011] Preferably, the hyperparameter optimization covers key parameters such as regularization coefficient, maximum tree depth, learning rate, and subsample ratio.

[0012] Preferably, the model evaluation further includes performance validation using average precision (AP) in the context of class imbalance.

[0013] Preferably, the SHAP tool can reveal the nonlinear relationship between key features and Alzheimer's disease outcomes, as well as the interaction effects between features.

[0014] Preferably, the convenient access interface is in the form of a webpage, and the deployment process includes website design, backend development, functional testing, filing deployment, and maintenance support.

[0015] An Alzheimer's disease prediction system based on multimodal health data includes: Data acquisition and preprocessing module: used to acquire multimodal health data, perform data cleaning, integrity checks, missing value handling, outlier handling, and dataset partitioning; Feature filtering module: Filters features that are significantly associated with Alzheimer's disease through statistical tests; Model building and optimization module: Builds various machine learning models and optimizes hyperparameters using nested cross-validation; Model Evaluation and Interpretation Module: Evaluates model performance based on multi-dimensional indicators, and performs feature importance analysis and model interpretation using Gini importance and SHAP tools; Deployment module: Provides a convenient web-based access interface to complete model deployment and application. Among them, each module realizes data transmission through the data interface. The output of the data acquisition and preprocessing module serves as the input of the feature selection module, and the output of the feature selection module is passed to the model construction and optimization module.

[0016] Preferably, the model building and optimization module supports adaptive model updates, periodically collects new data and re-evaluates model performance, and triggers model updates when performance deteriorates or feature distribution changes.

[0017] Due to the adoption of the above scheme, the beneficial effects of the present invention are: high prediction accuracy: integrating multimodal health data, comprehensively characterizing AD pathological features, constructing and optimizing various machine learning models, among which the random forest model has an accuracy of 0.9535 and an AUC value of 0.9491, the XGBoost and LightGBM models have similar performance, significantly outperforming the traditional logistic regression model, and both the missed diagnosis rate and the misdiagnosis rate are low.

[0018] Scientific feature selection: Features that are significantly associated with AD are selected through statistical tests to reduce redundant information interference, while identifying core predictive features (MMSE score, functional evaluation score, etc.) to improve model efficiency and reliability.

[0019] High interpretability: By integrating Gini importance with the SHAP tool, the contribution of features to prediction results is quantified, the association mechanism between features and diseases is revealed, the "black box" problem is solved, and clinicians' trust and acceptance of the model are improved.

[0020] Wide accessibility: The model is deployed via a web page, requiring no professional equipment or complex operations. The general public can perform self-testing, and primary healthcare institutions can conduct large-scale screening, effectively solving the problem of technology popularization.

[0021] Highly adaptable: It supports adaptive model updates, can cope with dynamic changes in health data characteristics, maintains the long-term accuracy of the model through regular evaluation and updates, and meets the needs of long-term clinical applications. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the overall process of the prediction method of the present invention; Figure 2 A heatmap showing the significance level distribution of continuous variables related to AD; Figure 3 Comparison of intergroup distributions of AD-related categorical variables Figure 1 ; Figure 4 Comparison of intergroup distributions of AD-related categorical variables Figure 2 ; Figure 5 Comparison of intergroup distributions of AD-related categorical variables Figure 3 ; Figure 6 Comparison of intergroup distributions of AD-related categorical variables Figure 4 ; Figure 7 A comparison chart of ROC curves for the classification performance of each model; Figure 8 This is a performance evaluation graph for models based on precision-recall. Figure 9 The confusion matrix of the classification results of each model on the test set; Figure 10 This is a graph showing a multi-model comparative analysis based on feature importance. Figure 11 This is a diagram illustrating the importance and influence direction of features in the three-model system based on SHAP values. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0024] like Figure 1-11As shown, an Alzheimer's disease prediction method based on multimodal health data includes the following steps: Step 1: Acquisition and preprocessing of multimodal health data: Acquire multimodal health data including demographic characteristics, lifestyle indicators, clinical history, physiological and biochemical indicators, cognitive function assessment and symptom manifestations. After data cleaning and integrity checks, the data is randomly divided into training set and test set at a ratio of 8:2. Step 2, Feature Screening and Analysis: Comparisons were performed between groups for all features. For continuous variables, the two independent samples t-test was used, and for categorical variables, the χ² test was used (significance level α=0.05). Features that were significantly associated with Alzheimer's disease (P<0.05) were screened. Step 3: Predictive Model Construction and Optimization: Select logistic regression, random forest, XGBoost and LightGBM algorithms to construct the predictive model, and use a nested cross-validation strategy to optimize the hyperparameters. The inner loop performs 5-fold cross-validation to search for the optimal hyperparameter combination, and the outer loop evaluates the model's generalization performance. The optimization goal is to maximize the recall rate while taking other evaluation metrics into account. Step 4: Model Evaluation and Interpretation: Accuracy, recall, precision, AUC, and F1 score are used as evaluation metrics. ROC curves, PR curves, and confusion matrices are used to verify model performance. Gini importance and SHAP tools are used to quantify the contribution of features to the prediction results and reveal the association between features and disease outcomes. Step 5, Model Deployment: Deploy the best-performing model through a convenient access interface to achieve early risk screening for Alzheimer's disease.

[0025] The multimodal health data specifically includes: demographic characteristics such as age, gender, body type, and education level; lifestyle indicators such as physical activity duration, diet quality score, smoking and drinking status; clinical history such as family history of Alzheimer's disease, cardiovascular disease, and diabetes; physiological and biochemical indicators such as systolic blood pressure, diastolic blood pressure, total cholesterol, and low-density lipoprotein cholesterol; cognitive function assessment data such as Mini-Mental State Examination (MMSE) score and Activities of Daily Living (ADL) score; and symptom data such as memory complaints, disorientation, and personality changes.

[0026] The data preprocessing also includes missing value imputation and outlier handling. Continuous data is imputed with the mean or median, and categorical data is imputed with the mode. Extreme outliers are identified and removed using statistical methods.

[0027] The significant associated features include MMSE score, functional assessment score, activities of daily living score, memory complaints, behavioral problems, sleep quality, and high-density lipoprotein cholesterol level.

[0028] The hyperparameter optimization covers key parameters such as regularization coefficient, maximum tree depth, learning rate, and subsample ratio.

[0029] The model evaluation also includes performance validation using average precision (AP) in the context of class imbalance.

[0030] The SHAP tool can reveal the nonlinear relationship between key features and Alzheimer's disease outcomes, as well as the interaction effects between features.

[0031] The convenient access interface is in web page format, and the deployment process includes website design, backend development, functional testing, filing deployment, and maintenance support.

[0032] An Alzheimer's disease prediction system based on multimodal health data includes: Data acquisition and preprocessing module: used to acquire multimodal health data, perform data cleaning, integrity checks, missing value handling, outlier handling, and dataset partitioning; Feature filtering module: Filters features that are significantly associated with Alzheimer's disease through statistical tests; Model building and optimization module: Builds various machine learning models and optimizes hyperparameters using nested cross-validation; Model Evaluation and Interpretation Module: Evaluates model performance based on multi-dimensional indicators, and performs feature importance analysis and model interpretation using Gini importance and SHAP tools; Deployment module: Provides a convenient web-based access interface to complete model deployment and application. The model building and optimization module supports adaptive model updates, periodically collecting new data and re-evaluating model performance. Model updates are triggered when performance degrades or feature distribution changes. Data transmission between modules is achieved through data interfaces. The output of the data acquisition and preprocessing module serves as the input to the feature selection module, and the output of the feature selection module is then passed to the model building and optimization module. In this embodiment, the Alzheimer's Disease Dataset, released by Rabie on the Kaggle platform, is used. This dataset is a publicly synthesized, multi-dimensional collection of clinical data, licensed under the CCBY 4.0 license, allowing its use for academic and research purposes with appropriate citation. Currently, this dataset has received relatively few citations in academic literature, thus possessing high original analytical value. The dataset contains 2149 patient records, with each sample encompassing 33 attributes. It systematically integrates variables from multiple dimensions, including demographic characteristics, lifestyle indicators, clinical history, physiological and biochemical indicators, cognitive function assessments, and symptom presentations. The dataset includes demographic data such as age (60-90 years), sex, body type, and education level; lifestyle habits such as physical activity duration (0-10 hours per week), dietary quality score (0-10 points), and smoking and alcohol consumption status, reflecting the potential impact of individual behavioral patterns on cognitive health; clinical history including family history of Alzheimer's disease, cardiovascular disease, diabetes, depression, and head trauma; physiological and biochemical indicators including systolic blood pressure (90-180 mmHg), diastolic blood pressure, total cholesterol (150-300 mg / dL), and LDL cholesterol; cognitive and functional assessments represented by the Mini-Mental State Examination (MMSE, 0-30 points) and Activities of Daily Living (ADL) scores; and symptom manifestations encompassing various behavioral and cognitive symptoms such as memory complaints, disorientation, and personality changes. This dataset is characterized by comprehensive variables, diverse types, and sufficient sample size, providing a solid data foundation for constructing a systematic and reliable Alzheimer's disease risk prediction model.

[0033] Experimental Methods: This study focuses on the construction and mechanism exploration of Alzheimer's disease prediction models driven by health data. The overall technical approach covers four main stages: data acquisition and cleaning, feature analysis and screening, multi-model construction and hyperparameter optimization, and model evaluation and interpretability analysis. The specific process is as follows: First, the raw health dataset is obtained from the public data platform Kaggle and subjected to quality checks and preprocessing. Then, the correlation between features and AD is initially evaluated through statistical tests and visualization methods. Based on this, multiple machine learning models are constructed, and hyperparameters are tuned using nested cross-validation. Finally, the model performance is validated on independent test sets, and key risk factors and their biological significance are explored in depth using various interpretability methods.

[0034] Data preprocessing involves performing integrity checks and preliminary cleaning on the raw data. The data is then randomly divided into a training set and a test set at an 8:2 ratio. The training set contains 1719 samples, and the test set contains 430 samples to ensure the reliability of model training and evaluation.

[0035] Feature analysis and visualization were performed to initially understand the distribution differences of each feature between the AD group and the non-AD group. Inter-group comparisons were conducted for all features: continuous variables were analyzed using a two-sample t-test, and categorical variables were analyzed using a chi-square test, with the significance level set at α=0.05. Based on this, a heatmap with the "p-value" as the core indicator was created to visually demonstrate the strength and significance of the association between each feature and the AD diagnosis, providing a statistical basis for subsequent feature selection.

[0036] The prediction model was constructed and hyperparameters optimized using four widely used machine learning algorithms: Logistic Regression, Random Forest, LightGBM, and XGBoost. To fully utilize the performance of each model and avoid overfitting, a nested cross-validation strategy was employed for hyperparameter tuning. The inner loop performed 5-fold cross-validation to search for the optimal hyperparameter combination, while the outer loop evaluated the model's generalization performance. The tuning objective was to maximize recall while also considering accuracy, F1 score, and AUC. The optimization process covered key parameters such as regularization coefficient, maximum tree depth, and learning rate.

[0037] Model evaluation and interpretability analysis: Model performance was evaluated using metrics such as accuracy, precision, recall, F1 score, and AUC. Comprehensive validation was performed on the test set using ROC curves, PR curves, and confusion matrices. To further analyze AD risk factors and prediction mechanisms, this study integrates multiple interpretability analysis methods: Firstly, it utilizes Gini importance and Shapley Additive Explanations (SHAP) provided by random forest and gradient boosting tree models to quantify the contribution of each feature to AD risk from different dimensions; secondly, it employs visualization tools such as SHAP summary diagrams to reveal the nonlinear relationship between key features and AD outcomes, as well as the interaction effects between features.

[0038] Dataset feature analysis and univariate screening: The statistical results of the potential association between each feature in the dataset and Alzheimer's disease are shown in Table 1. Several features showed highly significant statistical differences between the AD group and the non-AD group (P<0.05). Among them, the differences in cognitive and functional assessment indicators were the most prominent, including MMSE score (P=4.00×10⁻³²) and functional assessment score (P=5.72×10⁻³²). 70 Daily living activities score (P=6.02×10⁻) 57 ), etc. At the same time, several clinical symptoms and behavioral manifestations, such as memory complaints (P=1.53×10⁻), were also observed. 45 ) and behavioral problems (P=4.73×10⁻²) 5The correlation was also very strong. Furthermore, high-density lipoprotein cholesterol levels (P=0.050) and sleep quality (P=0.009) also showed some inter-group differences. Based on log... 10 Feature heatmaps drawn from (P-value) Figure 2 These salient features are visually represented by stacked bar charts of representative categorical features. Figure 3-6 This further reveals the distribution pattern of the key categorical variables between the two groups.

[0039] Table 1 Comparison of baseline data between the two groups of patients Note: Measurement data in the table are expressed as means, and count data are expressed as number of cases (percentage) (simplified as " / " in the table); comparisons between groups were performed using t-tests (measurement) and χ² tests (counting), with the corresponding statistics being t and χ². The p-value is the significance level, and p < 0.05 indicates that the difference is statistically significant.

[0040] in, Figure 3-6 The horizontal axis represents different clinical characteristics and demographic variables, where 0 and 1 (or 0, 1, 2, and 3) represent different categories of each variable; the vertical axis represents the number of people. Orange bars represent the number of people without the disease for each characteristic, and blue bars represent the number of people with the disease. From left to right, the distribution of characteristics such as gender, smoking, family history of Alzheimer's disease, cardiovascular disease, diabetes, depression, history of head injury, hypertension, memory complaints, behavioral problems, confusion, disorientation, personality changes, difficulty completing tasks, forgetfulness, physical constitution, and education level is shown.

[0041] A comparison and evaluation of the prediction performance of multiple models was conducted. Based on the given models, cross-validation was used to obtain evaluation metrics such as accuracy, recall, precision, AUC, and F1 score, which are listed in Table 2. The Logistic Regression model performed poorly across all evaluation metrics, with an accuracy of 0.7767, a recall of 0.9079, a precision of 0.6273, an AUC of 0.8862, and an F1 score of 0.7419, exhibiting a high recall but low precision. The Random Forest model performed best among all models, with high performance across all metrics. Its accuracy reached 0.9535, its recall was 0.9276, its precision was 0.9400, its AUC was 0.9491, and its F1 score was 0.9338. The XGBoost model performs similarly to the Random Forest model, exhibiting stable and excellent performance across all metrics. Its accuracy is 0.9488, recall 0.9211, precision 0.9333, AUC 0.9493, and F1 score 0.9272. The LightGBM model's performance is highly similar to XGBoost, also achieving 0.9488 accuracy, 0.9145 recall, 0.9392 precision, 0.9479 AUC, and an F1 score of 0.9267. While it slightly surpasses XGBoost in precision, it falls slightly short in recall, AUC, and F1 score. However, these differences are relatively small, and overall classification performance remains at a high level.

[0042] Table 2 Comparison of Model Prediction Performance Metrics Based on Test Set From the ROC curve ( Figure 7 As can be seen, the curves for LightGBM, RandomForest, and XGBoost are all significantly closer to the upper left corner, with AUC values ​​exceeding 0.94, demonstrating excellent discriminative ability. LightGBM has the highest AUC (0.9479), followed closely by the other two; while LogisticRegression (AUC = 0.8862) performs relatively weakly. (PR curves) Figure 8 Further evaluation of the models in the context of class imbalance showed that LightGBM still had the highest mean precision (AP=0.9452), and the other ensemble models also all had a precision above 0.92, significantly outperforming LogisticRegression (AP=0.8249). Confusion matrix ( Figure 9The results show that LightGBM, XGBoost, and RandomForest can effectively identify most positive and negative samples on the test set with a small number of misclassifications, especially RandomForest and XGBoost, which only misclassified 20 cases.

[0043] Feature selection and model interpretability analysis: The key feature sets ultimately selected by each model are shown in Table 3. Analysis revealed that MMSE scores, functional assessment scores, memory complaints, behavioral problems, and activities of daily living scores were adopted by all four models, constituting the core feature set for AD prediction. Regarding model specificity, Logistic Regression relies more on static features such as demographic attributes and medical history; while the superior tree model extensively incorporates lifestyle factors (such as exercise, alcohol consumption, and sleep) and physiological metabolic indicators (such as blood pressure and blood lipids). A bar chart based on feature importance (Gini importance) is presented. Figure 10 Analysis shows that in high-performing ensemble learning models, the aforementioned cognitive and functional features rank among the most important. However, there are also differences in feature emphasis among models; for example, the LightGBM model assigns higher weights to indicators such as diet quality and triglyceride levels.

[0044] Table 3 Key features of AD prediction for four machine learning algorithms SHAP interpretability analysis results ( Figure 11 The study further quantified the direction and strength of the influence of features on the prediction results. All ensemble models consistently confirmed that a decrease in the values ​​of cognitive and functional features (such as MMSE, functional assessment, and ADL) significantly increases the risk of AD. Furthermore, different models revealed their unique feature focuses: XGBoost paid additional attention to the history of head injury; LightGBM broadly identified physiological indicators such as blood lipids, blood pressure, and BMI, as well as lifestyle factors such as diet and sleep; RandomForest highlighted the protective effect of weekly exercise volume. SHAP analysis also showed that the influence distribution of some features (such as memory complaints) was relatively discrete, suggesting that their effects may be modulated by other factors.

[0045] This invention achieves accurate prediction of Alzheimer's disease by integrating multimodal health data and scientific feature screening, combined with an optimized machine learning model. It also has strong interpretability and broad accessibility, providing an efficient and reliable technical solution for early screening of Alzheimer's disease, and has significant clinical value and promising prospects for promotion.

[0046] The above description of the embodiments is intended to enable those skilled in the art to understand and use the present invention. It will be apparent to those skilled in the art that various modifications can be made to these embodiments, and the general principles described herein can be applied to other embodiments without inventive effort. Therefore, the present invention is not limited to the above embodiments. Improvements and modifications made by those skilled in the art based on the principles of the present invention without departing from the scope of the invention should be within the protection scope of the present invention. The above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. 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 method for predicting Alzheimer's disease based on multimodal health data, characterized in that, Includes the following steps: Step 1: Acquisition and preprocessing of multimodal health data: Acquire multimodal health data including demographic characteristics, lifestyle indicators, clinical history, physiological and biochemical indicators, cognitive function assessment and symptom manifestations. After data cleaning and integrity checks, the data is randomly divided into training set and test set at a ratio of 8:

2. Step 2, Feature Screening and Analysis: Comparisons were performed between groups for all features. For continuous variables, the two independent samples t-test was used, and for categorical variables, the χ² test was used. The significance level was set at α=0.

05. Features that were significantly associated with Alzheimer's disease were screened out, with a p-value < 0.

05. Step 3: Predictive Model Construction and Optimization: Select logistic regression, random forest, XGBoost and LightGBM algorithms to construct the predictive model, and use a nested cross-validation strategy to optimize the hyperparameters. The inner loop performs 5-fold cross-validation to search for the optimal hyperparameter combination, and the outer loop evaluates the model's generalization performance. The optimization goal is to maximize the recall rate while taking other evaluation metrics into account. Step 4: Model Evaluation and Interpretation: Accuracy, recall, precision, AUC, and F1 score are used as evaluation metrics. ROC curves, PR curves, and confusion matrices are used to verify model performance. Gini importance and SHAP tools are used to quantify the contribution of features to the prediction results and reveal the association between features and disease outcomes. Step 5, Model Deployment: Deploy the best-performing model through a convenient access interface to achieve early risk screening for Alzheimer's disease.

2. The prediction method according to claim 1, characterized in that, The multimodal health data specifically includes: demographic characteristics such as age, gender, body type, and education level; lifestyle indicators such as physical activity duration, diet quality score, smoking and drinking status; clinical history such as family history of Alzheimer's disease, cardiovascular disease, and diabetes; physiological and biochemical indicators such as systolic blood pressure, diastolic blood pressure, total cholesterol, and low-density lipoprotein cholesterol; cognitive function assessment data such as Mini-Mental State Examination (MMSE) score and Activities of Daily Living (ADL) score; and symptom data such as memory complaints, disorientation, and personality changes.

3. The prediction method according to claim 1, characterized in that, The data preprocessing also includes missing value imputation and outlier handling. Continuous data is imputed with the mean or median, and categorical data is imputed with the mode. Extreme outliers are identified and removed using statistical methods.

4. The prediction method according to claim 1, characterized in that, The significant associated features include MMSE score, functional assessment score, activities of daily living score, memory complaints, behavioral problems, sleep quality, and high-density lipoprotein cholesterol level.

5. The prediction method according to claim 1, characterized in that, The hyperparameter optimization covers key parameters such as regularization coefficient, maximum tree depth, learning rate, and subsample ratio.

6. The prediction method according to claim 1, characterized in that, The model evaluation also includes performance validation using average precision (AP) in the context of class imbalance.

7. The prediction method according to claim 1, characterized in that, The SHAP tool can reveal the nonlinear relationship between key features and Alzheimer's disease outcomes, as well as the interaction effects between features.

8. The prediction method according to claim 1, characterized in that, The convenient access interface is in web page format, and the deployment process includes website design, backend development, functional testing, filing deployment, and maintenance support.

9. An Alzheimer's disease prediction system based on multimodal health data, characterized in that, include: Data acquisition and preprocessing module: used to acquire multimodal health data, perform data cleaning, integrity checks, missing value handling, outlier handling, and dataset partitioning; Feature filtering module: Filters features that are significantly associated with Alzheimer's disease through statistical tests; Model building and optimization module: Builds various machine learning models and optimizes hyperparameters using nested cross-validation; Model Evaluation and Interpretation Module: Evaluates model performance based on multi-dimensional indicators, and performs feature importance analysis and model interpretation using Gini importance and SHAP tools; Deployment module: Provides a convenient web-based access interface to complete model deployment and application. Each module transmits data through a data interface. The output of the data acquisition and preprocessing module serves as the input of the feature selection module, and the output of the feature selection module is passed to the model construction and optimization module.

10. The prediction system according to claim 9, characterized in that, The model building and optimization module supports adaptive model updates, regularly collects new data and re-evaluates model performance, and triggers model updates when performance deteriorates or feature distribution changes.

Images