A dementia risk prediction method and system for a home and hospital hierarchical scenario

Abstract

The present application relates to a kind of dementia risk prediction method and system for family and hospital hierarchical scene.The method includes: obtaining prospective cohort data, is divided into the family end feature set matched with family scene, the hospital end feature set matched with hospital scene;Two kinds of feature sets are respectively subjected to feature screening and redundancy processing, obtain the family end core feature set applicable to family scene, the hospital end core feature set applicable to hospital scene;Based on feature importance ranking construction sequential forward inclusion mechanism, respectively determine family scene and hospital scene input feature;According to the scene model selection instruction of user, call target model to carry out risk prediction, obtain dementia risk estimation result.Can make risk prediction and data availability match, realize the continuous application path of family end and hospital end, make prediction probability more close to real event rate.

Description

Technical Field

[0001] This invention relates to the field of medical artificial intelligence and data-driven modeling technology, and in particular to a method and system for predicting dementia risk based on machine learning for different data acquisition scenarios. Background Technology

[0002] Dementia is a progressive neurodegenerative syndrome whose morbid and economic burden is continuously increasing globally. Extensive clinical and research evidence indicates that the pathophysiological changes in dementia can occur years or even two decades before clinical symptoms appear. Accurately identifying high-risk individuals at this subclinical stage will provide a crucial intervention window for lifestyle interventions and the implementation of emerging disease-modifying therapies (DMTs), significantly contributing to delaying or even halting disease progression.

[0003] In practical applications of dementia risk prediction, the availability and feature distribution of data vary significantly across different scenarios, forming two typical application scenarios: In the family / community scenario, the available information mainly consists of self-reported questionnaires, self-tested vital signs, lifestyle data, and socioeconomic data from research subjects. This type of data is inexpensive to acquire and easy to promote, but the information dimensions are relatively coarse and the accuracy is limited. In the hospital / clinical scenario, information such as routine test indicators, precise physiological measurement data, and performance on some cognitive tasks can be obtained through professional medical means. The data dimensions are richer and the accuracy is higher, but the acquisition process depends on professional resources, is more costly, and is difficult to popularize on a large scale. These different scenarios not only differ in data acquisition conditions, but their feature distributions also show significant shifts, making it difficult for models trained based on a single scenario to be directly transferred and applied in another scenario.

[0004] Existing technologies related to dementia risk prediction mainly include traditional risk scoring / statistical models and single-scenario machine learning models. There are also some models that only complete offline modeling without conducting systematic clinical applicability assessments. However, traditional risk scoring / statistical models (such as CAIDE and ANU-ADRI) typically use a fixed and small set of variables to build the model. While they are easily generalized, their overall predictive ability is limited, and the availability of variables and applicable scenarios are significantly constrained, making them unsuitable for different application scenarios. Single-scenario machine learning models are mostly based on specific data sources or specific clinical scenarios, making them highly dependent on the scenario. The core variables they rely on are often unavailable at the home level or difficult to collect using standardized criteria, hindering cross-scenario deployment and application. Models lacking systematic calibration and clinical applicability assessments only emphasize the model's discriminative power (AUC), paying insufficient attention to calibration effects and thresholding decision indicator optimization. They also cannot support dementia risk estimation over multiple time spans, limiting their practical clinical application value.

[0005] Therefore, traditional dementia risk prediction techniques often suffer from incomplete predictions and poor applicability due to poor cross-scenario adaptability, significant redundancy of high-dimensional multi-domain features, insufficient practicality, and a lack of visualization and interpretation loops. There is an urgent need for a dementia risk prediction method and system that can effectively model across different data acquisition scenarios, mitigate the impact of feature distribution differences, and improve the model's cross-scenario applicability. Summary of the Invention

[0006] Based on this, in order to solve the above-mentioned technical problems, a method and system for predicting dementia risk in stratified scenarios of home and hospital is provided. By constructing a feature modeling mechanism with scenario constraints and a scenario-based collaborative prediction system, the feature distribution offset under different data acquisition conditions is effectively solved, the model's cross-scenario generalization ability and prediction stability are improved, the risk prediction can be matched with data availability, and a continuous application path can be realized between the home and hospital ends, making the predicted probability closer to the actual event rate.

[0007] A method for predicting dementia risk in stratified family and hospital settings, the method comprising:

[0008] We acquire forward-looking cohort data and divide the scenarios based on data acquisition costs, availability constraints, and differences in feature distribution under different scenarios. The scenarios are divided into a home-side feature set that matches the home scenario and a hospital-side feature set that matches the hospital scenario.

[0009] The home-side feature set and the hospital-side feature set are respectively subjected to feature filtering and correlation redundancy removal processing to obtain a home-side core feature set suitable for home scenarios and a hospital-side core feature set suitable for hospital scenarios.

[0010] For the core feature sets of the home and hospital, a forward inclusion mechanism based on feature importance ranking is used to iteratively train the model during the gradual inclusion of features. The feature inclusion termination condition is determined based on the marginal gain change of the model performance index, thereby determining the input features for the home scenario and the input features for the hospital scenario, respectively.

[0011] A family-based dementia risk prediction model is trained based on the input features from the family end, and a hospital-based dementia risk prediction model is trained based on the input features from the hospital end, thus completing the construction of a scenario-based collaborative dual-model prediction system for family and hospital-based hierarchical scenarios.

[0012] Obtain the user's scenario model selection instruction, call the target model from the family-based dementia risk prediction model and the hospital-based dementia risk prediction model to perform risk assessment, and output the individual dementia risk prediction result.

[0013] In one embodiment, prospective queue data is acquired, and scenario segmentation is performed based on data acquisition cost, availability constraints, and differences in feature distribution under different scenarios. The segmentation is divided into a home-side feature set matching the home scenario and a hospital-side feature set matching the hospital scenario, including:

[0014] We acquired prospective cohort data, with follow-up of new-onset all-cause dementia as the outcome measure. Individuals with no baseline history of dementia diagnosis and follow-up information were included to form a database.

[0015] Extract all-dimensional candidate features related to dementia risk from the database to form an all-dimensional candidate feature pool;

[0016] Based on the data acquisition cost and availability standards in real-world application scenarios, as well as the differences in feature distribution under different scenarios, the features in the full-dimensional candidate feature pool are divided into a home-side feature set that matches the home scenario and a hospital-side feature set that matches the hospital scenario.

[0017] In one embodiment, feature filtering and relevance redundancy removal are performed on the home-side feature set and the hospital-side feature set, respectively, to obtain a home-side core feature set suitable for home scenarios and a hospital-side core feature set suitable for hospital scenarios, including:

[0018] Variables with a missing proportion exceeding a preset threshold in the family-side feature set and hospital-side feature set are removed respectively. Categorical variables are one-hot encoded and missing values ​​are treated as independent categories. Missing indicator features are generated for continuous variables and normalized.

[0019] Gradient boosting tree models are trained on the processed family-side feature set and hospital-side feature set respectively. The importance scores of each feature are calculated and sorted in descending order. Based on the features sorted in descending order, candidate feature sets for the family side and hospital side are selected.

[0020] The correlation coefficients between each feature pair in the candidate feature set for the home end and the candidate feature set for the hospital end are calculated respectively. Based on the correlation coefficients, redundancy removal is performed to obtain the core feature set for the home end suitable for the home scenario and the core feature set for the hospital end suitable for the hospital scenario.

[0021] In one embodiment, for the core feature sets of the home and hospital scenarios, a forward inclusion mechanism based on feature importance ranking is used. During the gradual feature inclusion process, model iterative training is performed, and the feature inclusion termination condition is determined based on the marginal gain change of the model performance index. This determines the input features for the home scenario and the input features for the hospital scenario, respectively, including:

[0022] Each feature in the family-side core feature set and the hospital-side core feature set is sorted in descending order according to its importance score to form a family-side feature inclusion order list and a hospital-side feature inclusion order list.

[0023] Gradient boosting tree model was uniformly selected as the training model, the number of features included was set to 0, and the initial discriminative power AUC value of the model was recorded.

[0024] According to the family-side feature inclusion order list and the hospital-side feature inclusion order list, individual features are progressively input into the gradient boosting tree model for repeated training operations. Each time a feature is added, the model is retrained and the area under the receiver operating characteristic (AUC) curve is calculated based on cross-validation. When, after at least two consecutive feature additions, the increase in the AUC curve in two adjacent rounds is not greater than a preset threshold, feature inclusion is stopped, and the final family-side input features and hospital-side input features are determined.

[0025] In one embodiment, a family-based dementia risk prediction model is trained based on the family-based input features, and a hospital-based dementia risk prediction model is trained based on the hospital-based input features, including:

[0026] The family-side training dataset and the hospital-side training dataset are determined based on the aforementioned family-side input features and hospital-side input features, respectively.

[0027] On the aforementioned family-based training dataset and hospital-based training dataset, a preliminary family-based model and a preliminary hospital-based model were trained using the gradient boosting tree algorithm, respectively. The model parameters were then optimized through cross-validation to obtain the final family-based dementia risk prediction model and hospital-based dementia risk prediction model.

[0028] In one embodiment, the method further includes:

[0029] Determine the model evaluation indicators, and use the model evaluation indicators to evaluate the performance of the family-based dementia risk prediction model and the hospital-based dementia risk prediction model, respectively, and obtain the evaluation results;

[0030] Based on the evaluation results, the parameters of the family-based dementia risk prediction model and the hospital-based dementia risk prediction model are adjusted.

[0031] In one embodiment, the user's scenario model selection instruction is obtained, and the target model is called from the family-based dementia risk prediction model and the hospital-based dementia risk prediction model to perform risk assessment, and the individual dementia risk prediction result is output, including:

[0032] Establish a mapping relationship between model type and calling interface for the family-based dementia risk prediction model and the hospital-based dementia risk prediction model, respectively;

[0033] Obtain the user's scene model selection instruction, extract the target model type from the scene model selection instruction, and find the corresponding calling interface based on the mapping relationship to call the target model;

[0034] Determine the input feature vector and input the input feature vector into the target model for risk prediction.

[0035] In one embodiment, the output of an individual's dementia risk prediction results includes:

[0036] The target model outputs a summary of risk probabilities across multiple time spans.

[0037] The risk probability summary is adjusted based on different predictive factors to output individual dementia risk prediction results and generate explanatory visualizations of risk changes.

[0038] A dementia risk prediction system for stratified scenarios involving families and hospitals, the system comprising:

[0039] The data acquisition and segmentation module is used to acquire prospective cohort data and segment the scenarios based on data acquisition costs, availability constraints, and differences in feature distribution under different scenarios, dividing them into a home-side feature set that matches the home scenario and a hospital-side feature set that matches the hospital scenario.

[0040] The data preprocessing module is used to perform feature filtering and correlation redundancy removal on the home-side feature set and the hospital-side feature set respectively, to obtain a home-side core feature set suitable for home scenarios and a hospital-side core feature set suitable for hospital scenarios.

[0041] The input feature determination module is used to perform iterative model training during the gradual feature inclusion process for the core feature sets of the home end and the core feature sets of the hospital end, based on the sequential forward inclusion mechanism of feature importance ranking, and to determine the feature inclusion termination condition based on the marginal gain change of the model performance index, thereby determining the input features of the home scene and the input features of the hospital scene respectively.

[0042] The model training module is used to train a family-side dementia risk prediction model based on the family-side input features and a hospital-side dementia risk prediction model based on the hospital-side input features, thereby completing the construction of a scenario-based collaborative dual-model prediction system for family and hospital-level scenarios.

[0043] The risk prediction module is used to obtain the user's scenario model selection instruction, call the target model from the family-side dementia risk prediction model and the hospital-side dementia risk prediction model to perform risk assessment, and output the individual dementia risk prediction result.

[0044] This invention introduces a feature partitioning and modeling mechanism with scene constraints, constructing feature subsets and prediction models for different data acquisition conditions, effectively alleviating the performance degradation problem caused by feature distribution differences between different scenarios; through sequential forward inclusion and feature redundancy removal mechanisms, it achieves optimized selection of input features, improving model stability and generalization ability; by constructing a scenario-based collaborative dual-model prediction system, the model can be flexibly invoked in different application scenarios, thereby reducing feature acquisition costs while improving prediction accuracy and enhancing the model's continuous application capability in multiple scenarios. Attached Figure Description

[0045] Figure 1 This is an application environment diagram of a dementia risk prediction method for a family and hospital stratification scenario in one embodiment;

[0046] Figure 2 This is a flowchart illustrating a dementia risk prediction method for a family- and hospital-based stratified scenario in one embodiment.

[0047] Figure 3 This is a schematic diagram of the computer device interface when selecting a model in one embodiment.

[0048] Figure 4 This is a schematic diagram illustrating the results of dementia risk estimation in one embodiment and its visualization.

[0049] Figure 5 This is a block diagram of a dementia risk prediction system for a family and hospital stratification scenario in one embodiment;

[0050] Figure 6 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0052] The dementia risk prediction method for stratified family and hospital settings provided in this application can be applied to, for example... Figure 1 The application environment shown. For example... Figure 1As shown, the application environment includes computer device 110. Computer device 110 can acquire prospective cohort data and, based on data acquisition costs, availability constraints, and differences in feature distribution under different scenarios, divide the data into a home-side feature set matching the home scenario and a hospital-side feature set matching the hospital scenario. Computer device 110 can perform feature filtering and relevance redundancy removal processing on the home-side and hospital-side feature sets respectively, obtaining a home-side core feature set suitable for the home scenario and a hospital-side core feature set suitable for the hospital scenario. Computer device 110 can then, based on a sequential forward inclusion mechanism ranked by feature importance, gradually incorporate features from the home-side and hospital-side core feature sets. The model undergoes iterative training, and features are included as termination conditions based on changes in the marginal gain of model performance indicators, thereby determining the input features for the home scenario and the hospital scenario, respectively. Computer device 110 can train a home-based dementia risk prediction model based on the home-based input features and a hospital-based dementia risk prediction model based on the hospital-based input features, thus completing the construction of a scenario-based collaborative dual-model prediction system for tiered scenarios in homes and hospitals. Computer device 110 can obtain the user's scenario model selection instructions, call the target model from the home-based dementia risk prediction model and the hospital-based dementia risk prediction model for risk assessment, and output the individual dementia risk prediction result. The computer device 110 can be, but is not limited to, various personal computers, laptops, smartphones, robots, unmanned aerial vehicles, tablets, etc.

[0053] In one embodiment, such as Figure 2 As shown, a method for predicting dementia risk in stratified scenarios of home and hospital is provided, including the following steps:

[0054] Step 202: Obtain prospective cohort data and divide the scenarios based on data acquisition cost, availability constraints, and differences in feature distribution under different scenarios, dividing them into a home-side feature set that matches the home scenario and a hospital-side feature set that matches the hospital scenario.

[0055] Computer devices can select prospective cohort data, include research individuals with baseline dementia who have follow-up information, use new-onset all-cause dementia during follow-up as the outcome, and divide candidate features into family feature sets and hospital feature sets according to the availability of real-world data.

[0056] In one embodiment, a dementia risk prediction method for stratified scenarios in families and hospitals may further include a data collection and segmentation process. The specific process includes: acquiring prospective cohort data, using follow-up of newly diagnosed all-cause dementia as the outcome indicator, including individuals with no baseline history of dementia diagnosis and follow-up information to form a database; extracting full-dimensional candidate features related to dementia risk from the database to form a full-dimensional candidate feature pool; and, based on the data acquisition cost and availability standards in real-world application scenarios and the differences in feature distribution under different scenarios, dividing the features in the full-dimensional candidate feature pool into a family-side feature set matching the family scenario and a hospital-side feature set matching the hospital scenario.

[0057] Computer equipment can preferentially select standardized prospective population cohort data with large sample sizes, long-term follow-up, and multi-dimensional information recording. In this embodiment, the computer equipment can preferably use UK Biobank cohort data, which must meet the core requirements of complete baseline information, sufficient follow-up period, and clear labeling of outcome events. It can cover multi-dimensional data such as demographics, lifestyle, clinical examination, cognitive assessment, and disease outcomes, providing a sufficient data foundation for the construction of dual-scenario feature sets.

[0058] Next, the computer equipment can use new-onset all-cause dementia as the core outcome indicator to rigorously define the research subjects in the cohort data, screen out the research population that meets the modeling requirements, and standardize and calibrate the new-onset all-cause dementia outcome in the screened dataset to ensure the accuracy and consistency of the outcome events. Then, based on the screened prospective cohort dataset, all multi-dimensional candidate features related to dementia risk are extracted from the original data, and the extracted dimensions cover the core related areas of dementia risk prediction.

[0059] Computer devices can be categorized based on their actual data acquisition capabilities in different real-world scenarios, combined with dimensions such as acquisition cost, operational difficulty, and dependence on professional resources. This allows for precise segmentation of features in the full-dimensional candidate feature pool into a home feature set and a hospital feature set. The home feature set includes self-reported / self-tested information, lifestyle information, socioeconomic information, medical history information, and basic physical examination information. The hospital feature set, in addition to the home feature set, incorporates routine clinical laboratory indicators, physiological measurement data, and short cognitive task performance data.

[0060] Step 204: Perform feature filtering and correlation redundancy removal on the home-side feature set and the hospital-side feature set respectively to obtain the home-side core feature set suitable for home scenarios and the hospital-side core feature set suitable for hospital scenarios.

[0061] After obtaining the feature sets from home and hospital, computer devices can extract a core feature set that combines high predictive value, low redundancy, and scenario adaptability through feature importance quantification screening and correlation redundancy elimination operations.

[0062] Specifically, in one embodiment, a dementia risk prediction method for stratified scenarios in families and hospitals may further include data processing, filtering, and redundancy removal processes. The specific processes include: removing variables with a missing proportion exceeding a preset threshold from both the family-side and hospital-side feature sets; performing one-hot encoding on categorical variables and treating missing values ​​as independent categories; generating missing indicator features for continuous variables and performing normalization; training gradient boosting tree models on the processed family-side and hospital-side feature sets respectively; calculating the importance score of each feature and sorting them in descending order; selecting candidate feature sets for families and hospitals based on the descending-ordered features; calculating the correlation coefficient between feature pairs in both the family-side and hospital-side candidate feature sets respectively; and performing redundancy removal based on the correlation coefficient to obtain a core feature set for families suitable for family scenarios and a core feature set for hospitals suitable for hospital scenarios.

[0063] Computer equipment can preprocess the feature sets from home and hospitals by performing missing data filtering, categorical variable processing, continuous variable processing, and dimensionality control of multi-select category fields. This allows the home and hospital feature sets, which have undergone standardized preprocessing such as missing data filtering, variable encoding, and normalization, to be treated as independent processing objects. The system performs format verification and outlier checks on the feature data, ensuring that the data quality of both feature sets meets requirements. Specifically, missing data filtering removes variables with a missing percentage exceeding a threshold (e.g., 80%) to reduce interference from noise and unusable variables; categorical variable processing involves one-hot encoding of categorical variables, treating missing data as an independent category so the model can learn information about the "missing data itself"; continuous variable processing adds missing indicator features (1 = missing, 0 = not missing) to continuous variables and normalizes them (e.g., scaling to [0,1] using min–max) to improve model convergence efficiency and feature comparability; and dimensionality control of multi-select category fields involves field-level aggregation of the multi-column features generated by expanding the multi-select category fields, and can use dimensionality reduction to extract the main components, reducing dimensionality inflation and enhancing robustness.

[0064] Next, the computer equipment can take the processed family-side and hospital-side feature sets as inputs, respectively, and use newly diagnosed all-cause dementia as the outcome label to independently train gradient boosting tree models. Based on the feature splitting gain and feature usage frequency output during model training, the comprehensive importance score of each feature is calculated, and the features in both feature sets are ranked from highest to lowest score. To improve cross-system transferability, the computer equipment can combine engineering availability and clinical applicability constraints to verify each ranked feature, eliminating features that are difficult to reproduce across systems, rely on complex algorithm derivation, have excessively high acquisition costs in the given scenario, or have no practical clinical significance. Among the retained features, the top 50 features are selected based on their importance scores to form the Top 50 candidate feature sets for the family side and the Top 50 candidate feature sets for the hospital side, achieving initial feature simplification. The gradient boosting tree model can employ LightGBM.

[0065] For the candidate feature sets after initial screening, computer equipment can identify highly redundant feature clusters through correlation calculations, remove redundant features, and obtain non-redundant feature sets for each scenario. Specifically, the computer equipment can pair features in the candidate Top 50 feature sets for home and hospital scenarios, calculate the Pearson correlation coefficient between each pair of features to quantify the degree of linear correlation between features; a unified correlation judgment threshold (preferably |r|>0.6) is set as the standard to identify highly correlated feature pairs and highly correlated feature clusters formed by multiple highly correlated features in the feature sets. The computer equipment can independently analyze each highly correlated feature pair / feature cluster, retaining only one representative feature within the cluster and removing the remaining redundant features; independent features without high correlation in the feature sets are directly retained, ultimately forming the non-redundant Top 30 core feature sets for home and hospital scenarios, respectively.

[0066] Step 206: For the core feature sets of the home and hospital, a forward inclusion mechanism based on feature importance ranking is used to iteratively train the model during the gradual inclusion of features. The feature inclusion termination condition is determined based on the marginal gain change of the model performance index, thereby determining the input features for the home scenario and the input features for the hospital scenario, respectively.

[0067] Based on the existing core feature sets for home and hospital use, computer devices can use the model discriminant power (AUC) gain saturation as the core criterion to accurately select the most valuable features for predicting dementia risk from the core feature sets through sequential forward inclusion and successive training and verification, and determine the final input features.

[0068] In one embodiment, a dementia risk prediction method for a family- and hospital-based stratified scenario may further include a process for determining input features. This process includes: sorting each feature in the family-side core feature set and the hospital-side core feature set in descending order of importance score to form a family-side feature inclusion order list and a hospital-side feature inclusion order list; uniformly selecting a gradient boosting tree model as the training model, setting the number of features included to 0 and recording the initial AUC base value of the model; progressively inputting individual features into the gradient boosting tree model for repeated training operations according to the family-side and hospital-side feature inclusion order lists, retraining the model after each feature addition and calculating the area under the receiver operating characteristic (AUC) curve based on cross-validation; stopping feature inclusion when the increase in the AUC curve in two adjacent rounds is no greater than a preset threshold after at least two consecutive feature additions, and determining the final family-side and hospital-side input features.

[0069] Within the feature range of the core feature sets for home and hospital terminals that have undergone redundancy removal, the computer equipment sorts various features in descending order based on the importance scores calculated in the previous stage, forming a feature inclusion order list. At the same time, unified basic parameters for model training are determined, a gradient boosting tree model is selected as the training model, new-onset all-cause dementia is used as the outcome label for model training, cross-validation is used as the model performance verification method, and the standard for model training and evaluation is unified.

[0070] Specifically, the computer device can set the number of features included to 0, and the initial discriminative power (AUC) of the model is recorded as the base value. Simultaneously, the AUC gain saturation criterion is set as follows: after adding a single feature, if the AUC obtained from model cross-validation is less than or equal to a preset threshold, and this condition is met for two consecutive additions of features, then the AUC gain is considered to be approaching saturation, and feature inclusion is stopped. Next, according to the feature inclusion order in descending order of importance, single-feature progressive inclusion and model retraining are performed for both the home and hospital sides. The AUC gain after each round of feature addition is verified one by one. When the AUC gain of two consecutive additions of features is less than or equal to the preset threshold, the model's discriminative power is considered to be approaching saturation. At this point, the number of features included in the previous round is the optimal number of features for this scenario. Based on the optimal number of features, the top N features are selected from the feature inclusion order list for the corresponding scenario to form the final input feature set for that scenario, i.e., the final input features for the home side and the final input features for the hospital side.

[0071] Step 208: A family-based dementia risk prediction model is trained based on family input features, and a hospital-based dementia risk prediction model is trained based on hospital input features, thus completing the construction of a scenario-based collaborative dual-model prediction system for family and hospital-based hierarchical scenarios.

[0072] The computer device can add features step by step in order of importance and repeat the training to observe the change in discriminative power (AUC) as the number of features increases. When the AUC gain approaches saturation, the final number of features (e.g., the Top 10 predictors for each feature) is selected to obtain the family-based dementia risk prediction model: using low-burden family predictors as input for community / family early screening; and the hospital-based dementia risk prediction model: using routine clinical hospital predictors as input for fine assessment of the hospital environment.

[0073] In one embodiment, a dementia risk prediction method for a family- and hospital-based stratified scenario may further include a model training process, which specifically includes: determining a family-based training dataset and a hospital-based training dataset based on family-based input features and hospital-based input features, respectively; training a preliminary family-based model and a preliminary hospital-based model on the family-based training dataset and the hospital-based training dataset, respectively, and optimizing the model parameters through cross-validation to obtain the final family-based dementia risk prediction model and the hospital-based dementia risk prediction model.

[0074] Computer devices can train corresponding models based on the Top 30 feature sets for home and hospital use respectively. Using the determined final input features for home and hospital use as core inputs, and new-onset all-cause dementia as the outcome label, the models are trained independently in different scenarios, following the same training framework, and the gradient boosting tree algorithm is used to complete the construction of dementia risk prediction models for home and hospital use.

[0075] Specifically, computer equipment can extract feature data that perfectly matches the final input features at the home and hospital ends, respectively. Combined with the corresponding new-onset all-cause dementia outcomes, this forms training datasets for the home and hospital ends. Next, a gradient boosting tree model (e.g., LightGBM) is uniformly selected as the core training algorithm. Independent model training frameworks are initialized for both the home and hospital ends. General basic hyperparameters are set based on the gradient boosting tree algorithm, including learning rate, decision tree depth, number of leaf nodes, number of iterations, and sample sampling ratio. Multiple rounds of training and iteration are performed on the home and hospital models respectively. Through multiple iterations of the gradient boosting tree, the mapping relationship between features and dementia outcomes is fitted, resulting in preliminary models for both the home and hospital ends. A pre-set validation set is then substituted into the preliminary models, and the discriminant power (AUC), accuracy, sensitivity, and specificity of the two models on the validation set are calculated to preliminarily assess the basic performance of the models. Parameter tuning is then performed to obtain the dementia risk prediction models for both the home and hospital ends.

[0076] In one embodiment, a dementia risk prediction method for a family- and hospital-based stratified scenario may further include a model evaluation process, which specifically includes: determining model evaluation indicators, and using the model evaluation indicators to evaluate the performance of the family-based dementia risk prediction model and the hospital-based dementia risk prediction model, respectively, to obtain evaluation results; and adjusting the parameters of the family-based dementia risk prediction model and the hospital-based dementia risk prediction model based on the evaluation results.

[0077] Computer devices can use cross-validation (e.g., 50% fold) to obtain internal validation performance estimates, use ROC curves and AUC to evaluate discriminative power, and can cover multiple time spans (e.g., 5 years, 10 years, full follow-up).

[0078] In this embodiment, the computer device can group predicted risks into decimal places and compare predicted risks with observed event rates; it can also use the Hosmer-Lemeshow goodness-of-fit test to evaluate consistency; it can determine probability thresholds by maximizing the Youden index and calculate threshold-dependent indicators such as accuracy, sensitivity, specificity, precision, and F1, providing an operational standard for actual risk stratification.

[0079] Step 210: Obtain the user's scenario model selection instruction, call the target model from the family-based dementia risk prediction model and the hospital-based dementia risk prediction model to conduct risk assessment, and output the individual dementia risk prediction result.

[0080] In one embodiment, a dementia risk prediction method for stratified scenarios in families and hospitals may further include a process of user selection of scenario models. The specific process includes: establishing a mapping relationship between model types and calling interfaces for family-side dementia risk prediction models and hospital-side dementia risk prediction models respectively; obtaining the user's scenario model selection instruction, extracting the target model type from the scenario model selection instruction, and finding the corresponding calling interface based on the mapping relationship to call the target model; determining the input feature vector, and inputting the input feature vector into the target model for risk prediction.

[0081] Computer devices can deploy the trained home-based dementia risk prediction model and hospital-based dementia risk prediction model to the backend node, build a unified model calling framework, assign scenario model identifiers to home-based and hospital-based models, and thus establish a mapping relationship table of model identifier-model type-call interface.

[0082] like Figure 3As shown, the computer device's interface can include a scene model selection entry point, providing model selection options and a command submission trigger button to facilitate user input. Once the user completes the scene model selection and clicks the submit button, the user's scene model selection command is captured in real time. This command can include information such as the selected model type, command submission time, and user operation identifier. Next, the computer device extracts the model type identifier information to determine whether the user has selected a home-based or hospital-based dementia risk prediction model. It then calls the target model via an interface and uses the target model to perform risk prediction. Specifically, the user can choose either a home-based or hospital-based dementia risk prediction model as the target model, inputting 10 corresponding prediction features to receive a real-time individual dementia risk estimate.

[0083] In one embodiment, a dementia risk prediction method for a family and hospital stratified scenario may also include a result presentation process, which includes: outputting a risk probability summary over multiple time spans through a target model; adjusting the risk probability summary based on different predictive factors; outputting individual dementia risk prediction results; and generating interpretive visualizations of risk changes.

[0084] In this embodiment, the computer device can provide summaries across multiple time spans (e.g., 5 years, 10 years, or more than 10 years) and visual displays, such as... Figure 4 As shown. Specifically, it can also provide age-based explanatory visualizations, that is, under the condition that other predictors are fixed, it shows the changes in risk output under different age inputs, which helps users understand the trend of risk changing with age, and can be used for explanation and communication.

[0085] In one embodiment, a dementia risk prediction method for stratified family and hospital settings is provided. This method constructs a stratified dementia risk prediction model based on prospective cohort data from the UK Biobank, and implements interactive, individualized risk estimation. The specific implementation process is as follows:

[0086] We acquired UK Biobank cohort data to construct a dataset of research subjects, and used newly diagnosed all-cause dementia as the outcome label for model training.

[0087] Based on the availability of real-world data, the candidate features in the dataset are divided into two feature sets: one for families and one for hospitals.

[0088] Data preprocessing was carried out uniformly on the two feature sets, variables with a missing ratio of more than 80% were removed, categorical variables were one-hot encoded and missing values ​​were treated as independent categories, missing indicator features were added to continuous variables and normalization was completed.

[0089] The importance of each feature was calculated using the gradient boosting tree algorithm in the feature sets of home and hospital scenarios respectively, and the Top 50 features in each scenario were selected by combining engineering and clinical conversion constraints.

[0090] Pearson correlation analysis was performed on the top 50 features of each scenario. Features with an absolute correlation coefficient greater than 0.6 were grouped into highly correlated clusters. Within each cluster, representative features were retained according to feature importance to obtain the non-redundant top 30 features of each scenario.

[0091] The top 30 features for each scenario are sequentially fed into training and validation in descending order of feature importance. The final top 10 predicted features are determined based on the model's discriminative power AUC gain saturation state. The home model and hospital model are trained to obtain the dementia risk prediction model.

[0092] Five-fold cross-validation was used to comprehensively evaluate the two models, including calculating the AUC of the ROC curve to evaluate the model's discriminative power, evaluating the overall performance with the AP index, evaluating the model's calibration with the decimal calibration curve and the Hosmer-Lemeshow goodness-of-fit test, determining the risk stratification threshold by maximizing the Youden index, and calculating threshold-dependent indicators such as output accuracy, sensitivity, and specificity.

[0093] The trained and validated home and hospital models are packaged into an interactive risk estimation tool. Users can select the corresponding model according to the actual scenario, input the 10 predictive features of the model, and the tool will output the individual dementia risk estimation results and visualization content in real time. It also provides explanatory graphs of risk changes with age and scenario, as well as risk summary information for multiple time spans such as 5 years, 10 years, and the full follow-up period.

[0094] In this embodiment, a two-tiered model—a family-based dementia risk prediction model and a hospital-based dementia risk prediction model—matches risk prediction with data availability, enabling a continuous application path of low-cost universal screening at the family level and high-precision stratification at the hospital level. High discriminative power is achieved using only 10 predictive features. Both the family-based and hospital-based dementia risk prediction models exhibit strong AUC and remain robust across multiple time spans. Importance ranking and correlation redundancy removal mechanisms reduce collinearity and redundant information, improving model stability and interpretability, and enhancing transferability to different application environments. Calibration analysis and HL tests make the predicted probabilities closer to the actual event rates, facilitating individualized absolute risk expression and decision support. Real-time input / output and visual interpretation are supported, improving usability in non-specialist scenarios and promoting early identification and prevention strategy development.

[0095] It should be understood that although the steps in the flowchart above are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowchart above may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0096] In one embodiment, such as Figure 5 As shown, a dementia risk prediction system for stratified scenarios in families and hospitals is provided, including: a data acquisition and segmentation module 510, a data preprocessing module 520, an input feature determination module 530, a model training module 540, and a risk prediction module 550, wherein:

[0097] The data acquisition and segmentation module 510 is used to acquire prospective cohort data and segment the scenarios based on data acquisition cost, availability constraints and differences in feature distribution under different scenarios, into a home-side feature set that matches the home scenario and a hospital-side feature set that matches the hospital scenario.

[0098] The data preprocessing module 520 is used to perform feature filtering and correlation redundancy removal on the home-side feature set and the hospital-side feature set respectively, to obtain the home-side core feature set suitable for home scenarios and the hospital-side core feature set suitable for hospital scenarios.

[0099] The input feature determination module 530 is used to perform iterative model training during the gradual feature inclusion process of the core feature sets of the home end and the core feature sets of the hospital end based on the sequential forward inclusion mechanism of feature importance ranking, and determine the feature inclusion termination condition based on the marginal gain change of the model performance index, thereby determining the input features of the home scene and the input features of the hospital scene respectively.

[0100] The model training module 540 is used to train a family-based dementia risk prediction model based on family input features and a hospital-based dementia risk prediction model based on hospital input features, thus completing the construction of a scenario-based collaborative dual-model prediction system for family and hospital-based hierarchical scenarios.

[0101] The risk prediction module 550 is used to obtain the user's scenario model selection instruction, call the target model from the family-side dementia risk prediction model and the hospital-side dementia risk prediction model to conduct risk assessment, and output the individual dementia risk prediction results.

[0102] In one embodiment, the data acquisition and segmentation module 510 is further used to acquire prospective cohort data, with follow-up of newly diagnosed all-cause dementia as the outcome indicator, and to include individuals with no baseline history of dementia diagnosis and follow-up information to form a database; extract all-dimensional candidate features related to dementia risk from the database to form an all-dimensional candidate feature pool; and, based on the data acquisition cost and availability standards in real-world application scenarios and the differences in feature distribution under different scenarios, divide the features in the all-dimensional candidate feature pool into a family-side feature set that matches the family scenario and a hospital-side feature set that matches the hospital scenario.

[0103] In one embodiment, the data preprocessing module 520 is further configured to remove variables whose missing proportions exceed a preset threshold from the family-side feature set and the hospital-side feature set, respectively; perform one-hot encoding on categorical variables and treat missing values ​​as independent categories; generate missing indicator features for continuous variables and perform normalization processing; train gradient boosting tree models on the processed family-side feature set and hospital-side feature set respectively; calculate the importance score of each feature and sort them in descending order; select candidate feature sets for the family-side and hospital-side based on the features sorted in descending order; calculate the correlation coefficient between each feature pair in the candidate feature sets for the family-side and hospital-side candidate feature sets respectively; perform redundancy removal processing based on the correlation coefficient to obtain the core feature set for the family-side suitable for the family scenario and the core feature set for the hospital-side suitable for the hospital scenario.

[0104] In one embodiment, the input feature determination module 530 is further configured to sort each feature in the family-side core feature set and the hospital-side core feature set in descending order of importance score, forming a family-side feature inclusion order list and a hospital-side feature inclusion order list; uniformly select the gradient boosting tree model as the training model, set the number of features included to 0, and record the initial discriminative power AUC base value of the model; according to the family-side feature inclusion order list and the hospital-side feature inclusion order list, input individual features step by step into the gradient boosting tree model for repeated training operations, retraining the model after each feature is added and calculating the area under the receiver operating characteristic curve (AUC) based on cross-validation; when the increase in the AUC of the receiver operating characteristic curve in two adjacent rounds is not greater than a preset threshold after at least two consecutive feature additions, feature inclusion is stopped, and the final family-side input features and hospital-side input features are determined.

[0105] In one embodiment, the model training module 540 is further configured to determine the family-side training dataset and the hospital-side training dataset based on the family-side input features and the hospital-side input features, respectively; and to train the family-side preliminary model and the hospital-side preliminary model using the gradient boosting tree algorithm on the family-side training dataset and the hospital-side training dataset, respectively, and to optimize the model parameters through cross-validation to obtain the final family-side dementia risk prediction model and the hospital-side dementia risk prediction model.

[0106] In one embodiment, the model training module 540 is further configured to determine model evaluation metrics, and use the model evaluation metrics to evaluate the performance of the family-based dementia risk prediction model and the hospital-based dementia risk prediction model, respectively, to obtain evaluation results; and adjust the parameters of the family-based dementia risk prediction model and the hospital-based dementia risk prediction model based on the evaluation results.

[0107] In one embodiment, the risk prediction module 550 is further configured to establish a mapping relationship between model type and calling interface for the family-based dementia risk prediction model and the hospital-based dementia risk prediction model; obtain the user's scenario model selection instruction, extract the target model type from the scenario model selection instruction, and find the corresponding calling interface based on the mapping relationship to call the target model; determine the input feature vector, and input the input feature vector into the target model for risk prediction.

[0108] In one embodiment, the risk prediction module 550 is also used to output a risk probability summary over multiple time spans through the target model; adjust the risk probability summary based on different predictive factors, output individual dementia risk prediction results, and generate interpretive visualizations of risk changes.

[0109] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 6 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When executed by the processor, the computer program implements a dementia risk prediction method for stratified home and hospital scenarios. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device casing, or an external keyboard, touchpad, or mouse.

[0110] Those skilled in the art will understand that Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0111] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement steps of a dementia risk prediction method for stratified home and hospital scenarios.

[0112] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program being executed by a processor to implement steps of a dementia risk prediction method for stratified home and hospital scenarios.

[0113] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0114] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0115] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for predicting dementia risk in stratified scenarios of home and hospital settings, characterized in that, The method includes: We acquire forward-looking cohort data and divide the scenarios based on data acquisition costs, availability constraints, and differences in feature distribution under different scenarios. The scenarios are divided into a home-side feature set that matches the home scenario and a hospital-side feature set that matches the hospital scenario. The home-side feature set and the hospital-side feature set are respectively subjected to feature filtering and correlation redundancy removal processing to obtain a home-side core feature set suitable for home scenarios and a hospital-side core feature set suitable for hospital scenarios. For the core feature sets of the home and hospital, a forward inclusion mechanism based on feature importance ranking is used to iteratively train the model during the gradual inclusion of features. The feature inclusion termination condition is determined based on the marginal gain change of the model performance index, thereby determining the input features for the home scenario and the input features for the hospital scenario, respectively. A family-based dementia risk prediction model is trained based on the input features from the family end, and a hospital-based dementia risk prediction model is trained based on the input features from the hospital end, thus completing the construction of a scenario-based collaborative dual-model prediction system for family and hospital-based hierarchical scenarios. Obtain the user's scenario model selection instruction, call the target model from the family-based dementia risk prediction model and the hospital-based dementia risk prediction model to perform risk assessment, and output the individual dementia risk prediction result.

2. The dementia risk prediction method for stratified family and hospital settings according to claim 1, characterized in that, Forward-looking cohort data was acquired, and scenario segmentation was performed based on data acquisition costs, availability constraints, and differences in feature distribution under different scenarios. The scenarios were divided into a home-based feature set matching the home scenario and a hospital-based feature set matching the hospital scenario, including: We acquired prospective cohort data, with follow-up of new-onset all-cause dementia as the outcome measure. Individuals with no baseline history of dementia diagnosis and follow-up information were included to form a database. Extract all-dimensional candidate features related to dementia risk from the database to form an all-dimensional candidate feature pool; Based on the data acquisition cost and availability standards in real-world application scenarios, as well as the differences in feature distribution under different scenarios, the features in the full-dimensional candidate feature pool are divided into a home-side feature set that matches the home scenario and a hospital-side feature set that matches the hospital scenario.

3. The dementia risk prediction method for stratified family and hospital settings according to claim 1, characterized in that, The home-side feature set and the hospital-side feature set are respectively subjected to feature filtering and relevance redundancy removal processing to obtain a home-side core feature set suitable for home scenarios and a hospital-side core feature set suitable for hospital scenarios, including: Variables with a missing proportion exceeding a preset threshold in the family-side feature set and hospital-side feature set are removed respectively. Categorical variables are one-hot encoded and missing values ​​are treated as independent categories. Missing indicator features are generated for continuous variables and normalized. Gradient boosting tree models are trained on the processed family-side feature set and hospital-side feature set respectively. The importance scores of each feature are calculated and sorted in descending order. Based on the features sorted in descending order, candidate feature sets for the family side and hospital side are selected. The correlation coefficients between each feature pair in the candidate feature set for the home end and the candidate feature set for the hospital end are calculated respectively. Based on the correlation coefficients, redundancy removal is performed to obtain the core feature set for the home end suitable for the home scenario and the core feature set for the hospital end suitable for the hospital scenario.

4. The dementia risk prediction method for stratified family and hospital settings according to claim 3, characterized in that, For the aforementioned core feature sets for home and hospital scenarios, a sequential forward inclusion mechanism based on feature importance ranking is used. Model iterative training is performed during the gradual feature inclusion process. The feature inclusion termination condition is determined based on the marginal gain change of the model performance index, thereby determining the input features for the home scenario and the input features for the hospital scenario, respectively. Each feature in the family-side core feature set and the hospital-side core feature set is sorted in descending order according to its importance score to form a family-side feature inclusion order list and a hospital-side feature inclusion order list. Gradient boosting tree model was uniformly selected as the training model, the number of features included was set to 0, and the initial discriminative power AUC value of the model was recorded. According to the family-side feature inclusion order list and the hospital-side feature inclusion order list, individual features are progressively input into the gradient boosting tree model for repeated training operations. Each time a feature is added, the model is retrained and the area under the receiver operating characteristic (AUC) curve is calculated based on cross-validation. When, after at least two consecutive feature additions, the increase in the AUC curve in two adjacent rounds is not greater than a preset threshold, feature inclusion is stopped, and the final family-side input features and hospital-side input features are determined.

5. The dementia risk prediction method for stratified family and hospital settings according to claim 1, characterized in that, A family-based dementia risk prediction model is trained based on the input features from the family end, and a hospital-based dementia risk prediction model is trained based on the input features from the hospital end, including: The family-side training dataset and the hospital-side training dataset are determined based on the aforementioned family-side input features and hospital-side input features, respectively. On the aforementioned family-based training dataset and hospital-based training dataset, a preliminary family-based model and a preliminary hospital-based model were trained using the gradient boosting tree algorithm, respectively. The model parameters were then optimized through cross-validation to obtain the final family-based dementia risk prediction model and hospital-based dementia risk prediction model.

6. The dementia risk prediction method for stratified family and hospital settings according to claim 5, characterized in that, The method further includes: Determine the model evaluation indicators, and use the model evaluation indicators to evaluate the performance of the family-based dementia risk prediction model and the hospital-based dementia risk prediction model, respectively, and obtain the evaluation results; Based on the assessment results, the parameters of the family-based dementia risk prediction model and the hospital-based dementia risk prediction model are adjusted.

7. The dementia risk prediction method for stratified family and hospital settings according to claim 1, characterized in that, Obtain the user's scenario model selection instruction, call the target model from the family-based dementia risk prediction model and the hospital-based dementia risk prediction model to perform risk assessment, and output the individual dementia risk prediction result, including: Establish a mapping relationship between model type and calling interface for the family-based dementia risk prediction model and the hospital-based dementia risk prediction model, respectively; Obtain the user's scene model selection instruction, extract the target model type from the scene model selection instruction, and find the corresponding calling interface based on the mapping relationship to call the target model; Determine the input feature vector and input the input feature vector into the target model for risk prediction.

8. The dementia risk prediction method for stratified family and hospital settings according to claim 7, characterized in that, Output individual dementia risk prediction results, including: The target model outputs a summary of risk probabilities across multiple time spans. The risk probability summary is adjusted based on different predictive factors to output individual dementia risk prediction results and generate explanatory visualizations of risk changes.

9. A dementia risk prediction system for stratified scenarios of home and hospital, characterized in that, The system includes: The data acquisition and segmentation module is used to acquire prospective cohort data and segment the scenarios based on data acquisition costs, availability constraints, and differences in feature distribution under different scenarios, dividing them into a home-side feature set that matches the home scenario and a hospital-side feature set that matches the hospital scenario. The data preprocessing module is used to perform feature filtering and correlation redundancy removal on the home-side feature set and the hospital-side feature set respectively, to obtain a home-side core feature set suitable for home scenarios and a hospital-side core feature set suitable for hospital scenarios. The input feature determination module is used to perform iterative model training during the gradual feature inclusion process for the core feature sets of the home end and the core feature sets of the hospital end, based on the sequential forward inclusion mechanism of feature importance ranking, and to determine the feature inclusion termination condition based on the marginal gain change of the model performance index, thereby determining the input features of the home scene and the input features of the hospital scene respectively. The model training module is used to train a family-side dementia risk prediction model based on the family-side input features and a hospital-side dementia risk prediction model based on the hospital-side input features, thereby completing the construction of a scenario-based collaborative dual-model prediction system for family and hospital-level scenarios. The risk prediction module is used to obtain the user's scenario model selection instruction, call the target model from the family-side dementia risk prediction model and the hospital-side dementia risk prediction model to perform risk assessment, and output the individual dementia risk prediction result.

Images