Fusion multi-dimensional index of female alzheimer's disease early risk prediction method
By constructing a female-specific indicator system and XGBoost model, combined with SHAP analysis, a low-cost, highly interpretable early risk prediction system for Alzheimer's disease was developed. This system addresses the issues of low accuracy in female AD screening and difficulty in adapting to primary healthcare institutions, achieving efficient and accurate early risk warning and health management.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 赵文慧
- Filing Date
- 2026-02-25
- Publication Date
- 2026-05-29
AI Technical Summary
Existing Alzheimer's disease screening and prediction models lack specific designs for women, resulting in low identification accuracy, a lack of early screening systems, high costs and difficulty in popularizing traditional diagnostic methods, and a lack of threshold standards for women-specific biomarkers, making it impossible for primary healthcare institutions to achieve effective screening.
We have constructed a core indicator system that includes indicators specific to women, adopted the XGBoost prediction model, and combined it with SHAP value analysis to develop a low-cost, highly interpretable early risk prediction system that is suitable for use in primary healthcare institutions and provides personalized health management recommendations.
It improves the accuracy of early risk prediction for Alzheimer's disease in women, reduces testing costs, adapts to primary healthcare institutions, provides personalized health management, reduces the burden on families and society, and promotes the upgrading of the health industry.
Abstract
Description
Technical Field
[0001] This invention relates to the fields of Alzheimer's disease screening and artificial intelligence medical technology, specifically a method for predicting the early risk of Alzheimer's disease in women by integrating multi-dimensional indicators. It is applicable to intelligent assessment of the early risk of Alzheimer's disease, identification of high-risk groups, and provision of health management suggestions for women over 45 years of age in menopause and middle-aged and elderly groups. It can be implemented in primary medical institutions, physical examination institutions, health care centers, and other scenarios. Background Technology
[0002] Alzheimer's disease (AD) is an insidious degenerative disease of the central nervous system and the main type of dementia. In my country, the disease affects a large number of people over 60 years of age, imposing a heavy medical, care, and economic burden on families and society. It has also become the fifth leading cause of death in my country. Clinical studies have confirmed that the prevalence of AD is significantly higher in women than in men. From 1992 to 2021, the rate of increase in the standardized prevalence of AD in elderly women was higher than that in men. From 2022 to 2030, the projected increase in the standardized incidence rate in women is expected to far exceed that in men. The disease burden (DALY) of Alzheimer's disease and other dementias has consistently been higher in women than in men, and its growth rate over the past decade has also been ahead of men.
[0003] The core contributing factors to the high incidence of Alzheimer's disease (AD) in women are highly correlated with their physiological characteristics. The sharp drop in estrogen levels and the rise in follicle-stimulating hormone (FSH) during perimenopause can trigger core pathological changes in AD, such as abnormal phosphorylation of tau protein and deposition of amyloid-β protein, through mechanisms such as activation of the C / EBPβ / AEP signaling pathway. Simultaneously, factors such as high fasting blood glucose and obstructive sleep apnea further increase the risk of disease in women. Furthermore, animal studies abroad have also confirmed that menopause exacerbates metabolic and inflammatory abnormalities, which are highly consistent with the pathological manifestations of AD in postmenopausal women.
[0004] The current field of Alzheimer's disease screening faces significant challenges: First, existing screening and prediction models are mostly gender-neutral, neglecting women's specific physiological risk factors and resulting in insufficient accuracy in identifying women, with a lack of dedicated AD testing systems for women. Second, AD has an insidious onset, with clinical diagnosis often occurring in the middle or late stages, and there are currently no effective drugs. Early warning is the only effective path for disease prevention and control, but early screening systems are severely lacking at present. Third, there is a lack of specific standards for female AD biomarker thresholds, and general models do not incorporate women-specific risk factors such as menopause, reproductive history, and hormone therapy history, significantly reducing predictive accuracy. Fourth, traditional diagnostic gold standards are invasive, expensive, and complex to operate, making them unavailable in primary healthcare institutions due to high costs and operational difficulties, leading to significant burdens on family care and a continuous increase in national healthcare and elderly care expenditures.
[0005] In existing technologies, testing solutions from international pharmaceutical giants are expensive, have limited assessment dimensions, rely on large laboratories, and have poor accessibility. While general screening tools are widely used and low-cost, their gender-neutral design results in a 15%-20% lower accuracy rate for identifying Alzheimer's disease (AD) in women. Products from emerging technology companies often lack sufficient clinical validation, ignore gender differences, and have weak interpretability. Meanwhile, the rapid iteration and cost reduction of liquid biopsy technology have provided an industrial foundation for the standardized and low-cost detection of AD-related biomarkers. Combined with the explicit requirements for early detection, early diagnosis, and early intervention in the National Action Plan for Addressing Alzheimer's Disease (2024-2030), developing an early risk prediction system for AD specifically for women has become an urgent industry need. Summary of the Invention
[0006] 3.1 Purpose of the invention:
[0007] This invention addresses the shortcomings of existing technologies and industry pain points by providing a method for predicting the early risk of Alzheimer's disease in women that integrates multi-dimensional indicators. Its core objective is:
[0008] (1) Construct a core indicator system that includes indicators specific to women, accurately cover the specific risk factors of AD in women, and fill the gap in the indicator system for AD risk assessment specific to women;
[0009] (2) Develop a female-specific XGBoost prediction model to achieve an AUC≥0.9 for early risk prediction of AD in women, thereby improving the accuracy of AD risk prediction in women;
[0010] (3) Through AI technology, we can achieve low-cost, large-scale, and standardized early risk screening for AD in women, adapt to the use scenarios of primary healthcare institutions, reduce the medical burden on society and families, and contribute to the national strategies of Healthy China and active aging.
[0011] 3.2 Technical Solution:
[0012] To achieve the aforementioned objectives, the present invention employs a multi-dimensional indicator-integrated early risk prediction system for Alzheimer's disease in women. This system includes an indicator system construction module, a data preprocessing module, a female-specific XGBoost prediction model module, a model interpretation module, a primary care adaptation tool module, and a health management suggestion output module. These modules work together to accurately predict and output the early risk of Alzheimer's disease in women. The specific technical solution is as follows:
[0013] (1) Indicator System Construction Module: This module is used to construct a core indicator system for women with AD, which is divided into 3 women-specific indicators (menopausal age, history of hormone replacement therapy, number of births), 2 basic biomarker indicators (p-tau217 plasma concentration, GFAP plasma concentration), and 5 clinical / lifestyle indicators (MMSE simplified score, PSQI sleep quality, exercise frequency, dietary structure, history of hypertension). The indicators with non-zero coefficients are screened by using the LASSO regression model combined with 10-fold cross-validation. At the same time, a graph theory indicator association network is constructed. The node association strength and degree centrality are used to calculate the Pearson correlation coefficient to identify the core nodes, verify the rationality of the indicator system, and clarify the quantitative collection standards, detection methods and data sources of each indicator.
[0014] (2) Data preprocessing module: This module is used to perform quality control and standardization on the collected female sample data. The data sources include female samples from the ADNI public database and female samples collected offline in the community. Missing values are filled by median / mode and outliers are removed by the 3σ principle to control the error rate ≤1%. The samples are divided into training set and test set in a 7:3 ratio, and domestic samples are selected as external validation set. Z-score standardization is performed on continuous variables and one-hot encoding is performed on categorical variables. At the same time, three core interaction items are constructed, namely menopausal age × p-tau217, hormone treatment history × GFAP, and number of births × MMSE score, which are included in the model training dimension.
[0015] (3) Female-specific XGBoost prediction model module: This module is the core prediction unit of the system. It uses the XGBoost binary classification algorithm to build a female AD-specific prediction model. The basic parameters are set as learning rate η=0.1, maximum depth maxdepth=6, subsample ratio subsample=0.8, and objective function is binary:logistic. The parameters are tuned by grid search combined with Bayesian optimization. The learning rate ranges from 0.01 to 0.1 and the maximum depth ranges from 4 to 8. An early stopping mechanism of early_stopping_rounds=50 is set to avoid model overfitting. The model performance indicators reach AUC≥0.90 and accuracy≥88% on the test set. The performance improvement on the external validation set compared with the general model is ≥10%.
[0016] (4) Model interpretation module: This module uses the SHAP value analysis method to calculate the contribution of each indicator and interaction item to the prediction results, and generates SHAP.summary_plot and risk factor heat map to clarify the core role of female-specific indicators in risk prediction, thereby improving the interpretability of the model and the trust of grassroots medical staff.
[0017] (5) Grassroots Adaptation Tool Module: This module performs lightweight processing on the XGBoost prediction model, removes features with a contribution of <5%, and reduces the model feature dimension to 13 dimensions (10 basic indicators + 3 core interaction items); extracts the optimal parameters of the model, and writes the risk probability calculation formula: Risk probability = 1 / [1+e^-(β0+∑βⱼxⱼ+∑γ k x_int k ]; Develop Excel calculation templates and Web / mini-program front-end prototypes to achieve integrated operation of indicator input, automatic calculation of interactive items, risk probability output, and risk level (low / medium / high) determination. The operation time is ≤5 minutes / case and the detection cost is ≤500 yuan / time.
[0018] (6) Health management advice output module: Based on the risk level output by the model, this module combines abnormalities in women-specific indicators and clinical / lifestyle indicators to generate personalized health management advice, covering four major modules: hormone regulation, dietary conditioning, cognitive training, and sleep improvement. It also provides specialist referral guidance for high-risk users.
[0019] 3.3 Beneficial effects:
[0020] Compared with the prior art, the present invention has the following significant advantages:
[0021] (1) Innovative perspective, filling the gap in the industry: This invention is a rare AD prediction technology in China that focuses on female gender differences. It is the first to construct a core indicator system containing three female-specific indicators, breaking through the limitations of the traditional model's "gender-independent design". It accurately captures specific risk factors such as women's perimenopause and reproductive history, filling the gap in the field of female-specific AD risk prediction.
[0022] (2) High prediction accuracy and strong interpretability: The XGBoost binary classification algorithm combined with the core interaction item design is adopted to achieve an AUC≥0.9 for early risk prediction of AD in women, which is ≥10% higher than the performance of general models. At the same time, SHAP value analysis and graph theory index association network are introduced to clarify the contribution of each index to the prediction results, which solves the "black box" problem of traditional AI medical models and improves clinical acceptance.
[0023] (3) Adaptable to primary care and highly accessible: Through the development of lightweight models and primary care adaptation tools, standardized operation, controllable cost, and simplified process are achieved. The operation time is ≤5 minutes / case, the detection cost is ≤500 yuan / time, and there is no need to rely on complex imaging equipment. It can be implemented in primary medical institutions such as community health service centers and county-level medical consortia, and can achieve large-scale popularization.
[0024] (4) Full-process closed loop with strong service: A full-process service closed loop of "indicator collection - risk prediction - result interpretation - health management - referral guidance" has been constructed. It not only realizes early risk warning of AD, but also provides users with personalized and tiered health management suggestions, realizing the extension from "screening" to "intervention" and enhancing the practical application value of the system.
[0025] (5) Low cost and high efficiency with significant social value: Using finger-prick blood to collect biomarkers and combining them with questionnaire surveys to obtain clinical / lifestyle indicators, the detection cost is significantly reduced compared to traditional invasive detection methods; through early warning and intervention, the occurrence and development of AD in women can be effectively delayed, reducing the pressure on family care and national medical and elderly care expenditures, while promoting the extension of the elderly health industry to the "prevention end" and promoting industrial transformation and upgrading. Detailed Implementation
[0026] The following detailed description of the multi-dimensional indicator-integrated early risk prediction system for Alzheimer's disease in women, based on specific implementation steps, is provided. The scope of protection of this invention is not limited to the following embodiments.
[0027] Example 1:
[0028] This embodiment provides a detailed implementation process for a multi-dimensional indicator-integrated early risk prediction system for Alzheimer's disease in women, including the following steps:
[0029] (1) Indicator collection and quality control
[0030] Data was collected from female subjects aged 45 and above using an indicator system construction module, following quantitative collection standards. Menopausal age was accurate to ±6 months. Hormone therapy history included duration, dosage, and discontinuation time. Reproductive history included live births and miscarriages. For p-tau217 plasma concentration, intra-assay CV was ≤5%, inter-assay CV was ≤8%, and the limit of detection for GFAP plasma concentration was ≤0.01 ng / mL. MMSE simplified score ranged from 0-30, and PSQI sleep quality score ranged from 0-21 (>7 was considered abnormal). Data sources included female samples from the ADNI database and samples collected offline in the community. Quality control was performed using a data preprocessing module, filling missing values for continuous variables with the median and missing values for categorical variables with the mode. Outliers were removed using the 3σ principle, controlling the error rate to ≤1%.
[0031] (2) Data standardization and interaction item construction
[0032] The quality-controlled sample data was divided into training and testing sets in a 7:3 ratio. 200 female samples aged 55-85 from a domestic community were selected as the external validation set. Z-score standardization was performed on continuous variables (such as age and p-tau217 concentration), and one-hot encoding was performed on categorical variables (such as history of hypertension and exercise frequency). Three core interaction terms were constructed: menopausal age × p-tau217, history of hormone therapy × GFAP, and number of births × MMSE score, which were included in the model training dimensions.
[0033] (3) Model training and optimization
[0034] Using the female-specific XGBoost prediction model module, an XGBoost binary classification model was initialized with the following basic parameters: learning rate η=0.1, maximum depth=6, subsample ratio=0.8, and objective function: binary:logistic. Parameter tuning was performed using grid search (learning rates 0.01, 0.05, 0.1; maximum depths 4, 6, 8) combined with Bayesian optimization. An early stopping mechanism of early_stopping_rounds=50 was set, and the optimal parameter combination was selected through 10-fold cross-validation. The final model achieved an AUC of 0.92 and an accuracy of 90% on the test set, and a 12% performance improvement compared to the general model on the external validation set.
[0035] (4) Model explanation and visualization
[0036] Using the SHAP value analysis method in the model interpretation module, the contribution of each indicator and interaction term was calculated, and a risk factor heatmap was generated. The results showed that menopausal age × p-tau217 was the core influencing factor for predicting the risk of AD in women, with a contribution of 25%, followed by PSQI sleep quality and history of hypertension. This verified the rationality of the indicator system and the interpretability of the model.
[0037] (5) Development and application of grassroots adaptation tools
[0038] The model was lightweighted using a basic adaptation tool module, removing features with a contribution of less than 5% and reducing the model's feature dimension to 13. The optimal parameters β0, β₱, and γ were extracted from the model. k We developed a risk probability calculation formula, created an Excel calculation template, added input format validation and drop-down selection options, and enabled automatic calculation of interactive items, output of risk probability, and determination of risk level after indicator input (low risk: risk probability <30%, medium risk: 30% ≤ risk probability <60%, high risk: risk probability ≥60%). A trial was conducted on 150 samples in two community health service centers. The average operation time was 3.5 minutes per sample, the average testing cost was 400 yuan per test, and the prediction accuracy reached 89%, meeting the needs of primary healthcare institutions.
[0039] (6) Output of health management recommendations
[0040] Based on the risk level output by the model, personalized health management recommendations are generated for the subjects through the health management recommendation output module: For low-risk subjects, basic lifestyle adjustment recommendations are provided, such as exercising more than 3 times a week, optimizing dietary structure, and improving sleep; for medium-risk subjects, targeted interventions are provided based on abnormal indicators, such as providing hormone regulation guidance for menopausal women and sleep improvement programs for those with sleep disorders; for high-risk subjects, an in-depth risk analysis report is generated, and neurology specialist referral guidance is provided, while a follow-up reminder mechanism is established.
[0041] Example 2:
[0042] This embodiment provides an implementation method for the system in a health checkup institution. Targeting women aged 45 and above undergoing health checkups, the predictive system of this invention is incorporated into a "Women's Brain Health Special Screening Package." The specific implementation steps are as follows:
[0043] (1) The physical examination institution provides finger-prick blood collection and questionnaire survey for female examinees, collects 10 core indicators in the indicator system, and enters the Web prototype of the grassroots adaptation tool through the online port;
[0044] (2) The Web prototype automatically completes data quality control, standardization processing and risk probability calculation, and outputs risk level and basic prediction report in real time;
[0045] (3) Provide tiered services for those undergoing physical examinations. Provide personalized health management advice for those with low / medium risk, and provide in-depth interpretation for those with high risk in conjunction with neurologists, as well as value-added services such as subsequent cognitive training and intervention guidance.
[0046] (4) The medical examination institution will desensitize the screening data and feed it back to the system backend for continuous iteration and optimization of the model to improve the prediction accuracy.
[0047] The early risk prediction system for Alzheimer's disease in women that integrates multi-dimensional indicators can be flexibly adapted to different application scenarios, with the form of grassroots-level tools (Web / mini-program, customized version for connecting to electronic medical record systems). It can also be implemented through a dual-track model of public welfare and commerce, providing free initial screening for AD risk for middle-aged and elderly women at the grassroots level, and then providing paid in-depth screening and intervention services for medium- and high-risk users, achieving a win-win situation for both commercial and social value.
[0048] The above description is only a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solution of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for predicting the early risk of Alzheimer's disease in women by integrating multi-dimensional indicators, characterized in that, Includes the following steps: (1) Construct a core indicator system for women with Alzheimer's disease. This system includes 3 women-specific indicators, 2 basic biomarker indicators and 5 clinical / lifestyle indicators. The women-specific indicators are menopausal age, history of hormone replacement therapy and number of births. The basic biomarker indicators are p-tau217 plasma concentration and GFAP plasma concentration. The clinical / lifestyle indicators are MMSE simplified score, PSQI sleep quality, exercise frequency, dietary structure and history of hypertension. (2) After collecting the above-mentioned indicator data of the target population of women aged 45 and above and performing quality control processing, the data were standardized and coded. At the same time, three core interactive items of female characteristics were constructed: menopausal age × p-tau217, hormone therapy history × GFAP, and number of births × MMSE score. (3) Input the processed index data and core interaction terms into the XGBoost binary classification prediction model to complete model training and optimization, so that the model has an AUC ≥ 0.9 and an accuracy ≥ 88% for predicting the early risk of Alzheimer's disease in women; (4) Input the indicator data of the person to be predicted into the trained model after processing in step 2), and output the early risk probability of Alzheimer's disease and the corresponding risk level.
2. The method for predicting the early risk of Alzheimer's disease in women according to claim 1, characterized in that, The core indicator system specifically for Alzheimer's disease in women was constructed in the following way: Candidate indicators were extracted through core literature review and consultation with clinical experts, and indicators with Pearson correlation coefficient <0.3 and those difficult to collect at the grassroots level were eliminated to obtain core candidate indicators; a graph theory indicator association network was constructed with core candidate indicators and Alzheimer's disease onset status as nodes, and the edge weights and degree centrality were calculated using Pearson correlation coefficient to identify core nodes; Using Alzheimer's disease incidence status as the dependent variable and core candidate indicators as independent variables, a LASSO regression model was constructed. The optimal penalty coefficient was determined through 10-fold cross-validation, and indicators with non-zero coefficients were retained to form a standardized core indicator system.
3. The method for predicting the early risk of Alzheimer's disease in women according to claim 1, characterized in that... When training and tuning the XGBoost binary classification prediction model, the learning rate η was set to 0.01-0.1 and the maximum depth to 4-8. Grid search combined with Bayesian optimization was used for parameter tuning. An early stopping mechanism of early_stopping_rounds=50 was set to control overfitting. After the model training was completed, the contribution of each indicator and core interaction term to the prediction results was analyzed using SHAP values. Features with a contribution of <5% were removed to achieve model lightweighting. The lightweight model still maintained stable prediction performance when the indicator collection missing rate was ≤10%.
4. The method for predicting the early risk of Alzheimer's disease in women according to claim 1, characterized in that, For the aforementioned core indicators, female-specific quantitative collection standards were established: menopausal age was accurate to ±6 months; surgical menopause was marked with menopause type and time; hormone replacement therapy history was recorded with duration, dosage, and discontinuation time, with a time accuracy of 0.5 years; the number of births was counted as the sum of live births and miscarriages; the intra-assay CV for p-tau217 plasma concentration was ≤5%, and the inter-assay CV was ≤8%; the limit of detection for GFAP plasma concentration was ≤0.01 ng / mL.
5. A system for predicting the early risk of Alzheimer's disease in women, implementing the method of any one of claims 1-4, characterized in that, The system comprises an indicator acquisition module, a data processing module, a model training module, and a risk prediction module, which are connected in sequence. The indicator acquisition module is adapted to primary healthcare scenarios and is used to collect core indicator data specific to women over 45 years of age according to women-specific quantitative collection standards, with a single collection time of ≤5 minutes. The data processing module is used to perform quality control, standardize and encode the collected data, and construct three core interactive items of female characteristics. The model training module is used to complete the training, optimization, and lightweight processing of the XGBoost binary classification prediction model. The risk prediction module is used to input the data to be predicted and output the early risk probability and risk level of Alzheimer's disease. The system can be embedded into commonly used office software in primary healthcare for lightweight deployment.