Method for predicting development of comorbidity of the elderly based on integrated machine learning

Abstract

The application relates to the technical field of medical health and artificial intelligence, in particular to a method for predicting the occurrence and development risk of senile comorbidity based on integrated machine learning, which comprises the steps of constructing a standardized data set, screening key variables, training a base learner, combining prediction results, and dynamically evaluating risks, and the like. By integrating multi-dimensional data features, the application uses algorithms such as random forest and support vector machine to construct a prediction model, and combines a verification set to optimize model parameters, and finally outputs a high-precision comorbidity risk prediction result. The application can realize accurate evaluation of the comorbidity risk of the elderly, and provide a scientific basis for personalized health management.

Description

Technical Field

[0001] This invention belongs to the fields of medical health and artificial intelligence technology, specifically a method for predicting the risk of comorbidity development in the elderly based on integrated machine learning. Background Technology

[0002] Studies have shown that comorbidity in the elderly typically follows a progressive developmental process: "health status → single chronic disease → comorbidity." Existing research has identified various comorbidity risk factors, including individual characteristics and environmental exposures, but limitations remain in predicting the risk of comorbidity development. Traditional statistical methods have limitations when processing multidimensional data, especially when integrating individual characteristics and external environmental factors, making it difficult to fully explore the nonlinear relationships between complex variables, resulting in insufficient predictive accuracy and dynamic adaptability.

[0003] In the field of healthcare management, physicians and public health policymakers need accurate risk prediction tools to assess changes in the health status of the elderly population in order to develop personalized intervention strategies. However, existing predictive models often focus on the analysis of single diseases, lacking a dynamic portrayal of the continuous process from health to a single disease and from a single disease to comorbidity. This limits their effectiveness in real-world scenarios. Furthermore, the inclusion of external variables such as socioeconomic data and environmental meteorological data has not received sufficient attention, further affecting the comprehensiveness and applicability of the models.

[0004] In recent years, the development of machine learning technology has provided new ideas for comorbidity risk prediction. Among them, ensemble algorithms have attracted widespread attention due to their ability to combine multiple base learners. Although existing machine learning-based comorbidity prediction methods have improved prediction performance to some extent, there is still room for improvement. Some methods do not fully consider the application of multi-state models when selecting key variables, which may lead to the introduction of redundant information, increasing model complexity and computational cost. At the same time, the combination of results from different base learners and the selection of meta-models lack systematic optimization, affecting the stability and accuracy of the final prediction results.

[0005] Current comorbidity risk assessments largely rely on traditional statistical indicators, lacking in-depth mining and scientific integration of multidimensional data characteristics. Even when some studies attempt to introduce machine learning methods, they often focus on the application of single algorithms, failing to fully leverage the synergistic advantages of multiple algorithms. This can lead to insufficient adaptability of prediction results when faced with diverse population characteristics and regional differences. Furthermore, existing models pay limited attention to local features during training, potentially failing to fully reflect the detailed performance of key variables, thus affecting the accuracy and reliability of predictions.

[0006] In practical applications, the health status of the elderly is influenced by a variety of factors, such as individual behavioral habits, regional environmental conditions, and socioeconomic development levels. This places higher demands on the comprehensiveness and personalization of comorbidity risk prediction. However, existing prediction models often employ a uniform structure and parameter settings, lacking adaptive adjustment mechanisms for different characteristic groups. This can lead to insufficient matching of prediction results when addressing diverse needs. Furthermore, inadequate integration and optimization of multi-source data during model construction may further affect its practical application value. Summary of the Invention

[0007] The purpose of this invention is to provide a method for predicting the risk of comorbidity development in the elderly based on ensemble machine learning, so as to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, this invention provides a method for predicting the risk of comorbidity development in older adults based on ensemble machine learning, the method comprising:

[0009] A vertical cohort database containing multidimensional data features is constructed, and a standardized dataset is formed by integrating demographic information, behavioral habits, physiological indicators, environmental and meteorological data, and socioeconomic variables; the data types in the standardized dataset include continuous variables and categorical variables;

[0010] The standardized dataset is initially screened to remove samples with a missing value ratio exceeding a preset threshold, and the remaining data is normalized. The normalized data is then divided into a training set, a validation set, and a test set.

[0011] Key comorbidity features are extracted from the training set, and representative disease combination patterns are identified through latent category analysis; these disease combination patterns are used to describe the progressive development process from health to a single chronic disease and then to comorbidity.

[0012] Based on the disease combination pattern, the variables in the training set are divided into five dimensions: basic demographic characteristics, personal behavioral characteristics, physiological indicators, intrinsic ability assessment, and environmental factors. The variables in each dimension are respectively included in the multi-state model from health to single disease and from single disease to comorbidity, and variables with significance levels below the preset statistical threshold are selected as key variables.

[0013] Five algorithms—random forest, support vector machine, neural network, extreme gradient boosting, and logistic regression—were selected as base learners. Key variables were input into each base learner for independent training. After training, the results of different base learners were combined using a stacking method to obtain preliminary prediction results.

[0014] The initial prediction results are optimized and adjusted using the validation set. The meta-model is determined by comparing the AIC and AUC values ​​of different meta-models, and a final ensemble model is formed. The final ensemble model is used to generate comprehensive prediction results for the risk of comorbidity development in the elderly population.

[0015] Collect key variable data on comorbidity outcomes and their occurrence in older adults, and dynamically assess the risk of comorbidity in older adults based on the comprehensive prediction results;

[0016] When the prediction results meet the preset accuracy requirements, the current comorbidity risk prediction model and related parameter configurations are output.

[0017] Preferably, the preliminary screening of the standardized dataset includes:

[0018] Calculate the proportion of missing values ​​for each column of variables in the standardized dataset and remove variables with a proportion of missing values ​​higher than a preset threshold; perform Z-score standardization on the continuous data in the remaining variables and one-hot encoding on the categorical data;

[0019] If the proportion of missing values ​​in a certain column of variables is less than a preset threshold, then interpolation is performed to impute the missing values ​​in that column of variables. Linear interpolation is preferred for imputation. If linear interpolation is not applicable, K-nearest neighbor interpolation is used to complete the data imputation.

[0020] Preferably, the identification of representative disease combination patterns through latent category analysis includes:

[0021] Potential category analysis was conducted on the sixteen diseases (hypertension, stroke, angina pectoris, diabetes, chronic lung disease, asthma, depression, anxiety, obesity, cancer, pain, arthritis, cataracts, chronic digestive diseases, chronic kidney disease, and anemia) covered in the normalized dataset. Based on the chronic disease prevalence of the study subjects, the comorbidity pattern characteristics in the population were systematically analyzed, and the final disease combination pattern was formed by combining clinical interpretability.

[0022] Preferably, the step of inputting key variables into each base learner for independent training includes:

[0023] The selected key variables were divided into training, validation, and test sets in a 7:2:1 ratio; cross-validation was performed on the data in the training set with a cross-validation fold of 5.

[0024] In each cross-validation process, random forest, support vector machine, neural network, extreme gradient boosting and logistic regression are used to fit the training set respectively. The base learner is trained with K-1 folds, and the remaining 1-folds generate meta-features. Finally, the prediction results of all the base learners are merged as the meta-features of the training set, and the prediction accuracy of each base learner is recorded.

[0025] Preferably, the method of combining the results of different base learners using a stacking approach includes:

[0026] The predictions of all the base learners are combined as meta-features of the training set and input into different secondary learners for training to generate prediction results.

[0027] Compare the performance of different secondary learners and select the secondary learner with the best performance (highest AUC, lowest AIC) as the final meta-model of ensemble learning.

[0028] Preferably, the optimization and adjustment of the initial prediction results using the validation set includes:

[0029] Input the data from the validation set into the primary prediction model and calculate the predicted probability distribution output by the model;

[0030] The predicted probability distribution is calibrated by dividing the predicted probability into several intervals using the equidistant binning method, and the actual occurrence rate of each interval is calculated.

[0031] Adjust the model parameters based on the deviation between the actual occurrence rate and the predicted probability until the deviation between the predicted probability distribution and the actual occurrence rate is less than the preset error range.

[0032] Preferably, the key variable data collected on comorbid outcomes and their occurrence in older adults include:

[0033] Extract individual characteristics, behavioral habits, physiological indicators, socioeconomic data, and regional environmental and meteorological data of the elderly from external validation datasets;

[0034] The extracted data is matched with the key variables selected by the multi-state model, and the variables that are successfully matched are retained as input data.

[0035] Perform the same normalization and encoding operations on the input data to ensure that it maintains the same format and scale as the training set data.

[0036] Preferably, the dynamic assessment of the comorbidity risk in the elderly based on the comprehensive prediction results includes:

[0037] The input data is substituted into the final integrated model to calculate the comorbidity risk score for each elderly person;

[0038] Based on the scoring results, a risk distribution curve was plotted to analyze the distribution of the number of elderly people in different scoring intervals.

[0039] Based on the risk distribution curve, elderly people whose scores fall within the high-risk range are highlighted and personalized health management plans are developed.

[0040] Preferably, the output of the current comorbidity risk prediction model and related parameter configuration includes:

[0041] Save the structural and training parameters of the final ensemble model and generate an executable prediction script;

[0042] The prediction script and standardized dataset processing workflow are integrated into a unified application for use by healthcare management organizations.

[0043] The application features data import, model prediction, and result visualization capabilities, allowing users to customize input data and view prediction results in real time.

[0044] Preferably, the potential category analysis of the sixteen diseases covered in the normalized dataset includes:

[0045] Based on the chronic disease prevalence of the study subjects, the characteristics of comorbidity patterns in the population were systematically analyzed.

[0046] Based on the maximum likelihood estimation method, a multi-level latent category model from 2 to 7 classes was constructed. The model parameters were iteratively optimized using the expectation-maximization algorithm (EM) until the log-likelihood function converged. The final disease combination pattern was formed by comprehensively comparing the Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC) and combining them with clinical interpretability.

[0047] Preferably, the calibration process for the predicted probability distribution includes:

[0048] The predicted probability is divided into ten equally spaced intervals: [0, 0.1), [0.1, 0.2), ..., [0.9, 1.0].

[0049] Calculate the actual incidence rate of the samples within each interval, and then calculate the mean square error between the actual incidence rate and the predicted probability.

[0050] If the mean squared error is greater than the preset threshold, the decision boundary of the model is adjusted and the predicted probability distribution is recalculated until the mean squared error is less than the preset threshold.

[0051] This method for predicting the risk of comorbidities in older adults based on ensemble machine learning has several beneficial effects, as detailed below:

[0052] 1. Improve prediction accuracy by integrating multi-dimensional data features, including demographic information, behavioral habits, physiological indicators, environmental and meteorological data, and socioeconomic variables, to more comprehensively capture factors influencing the occurrence and development of comorbidities in the elderly. Five algorithms—random forest, support vector machine, neural network, extreme gradient boosting (XGBoost), and logistic regression—are used as base learners. The results are combined using a stacking method and optimized with a validation set. The resulting comprehensive prediction has high accuracy, higher AUC, and better AIC, enabling a more accurate assessment of the risk of comorbidity development in the elderly.

[0053] 2. Enables dynamic risk assessment, collecting key variable data on comorbidity outcomes and their progression in the elderly, and dynamically assessing the risk of comorbidity in the elderly based on comprehensive prediction results. It can calculate a risk score for the comorbidity development status of each elderly person, plot risk distribution curves, and analyze the distribution of the number of elderly people within different score intervals, facilitating timely identification of high-risk groups and adjustment of intervention strategies.

[0054] 3. Facilitating personalized health management: Identifying representative disease combination patterns and describing the progressive development from health to a single chronic disease and then to comorbidity, providing a basis for targeted interventions. High-risk elderly individuals can be specifically identified, enabling the development of personalized health management plans to meet the diverse health needs of seniors and improve the scientific rigor and effectiveness of health management.

[0055] 4. Enhance the model's practicality and generalizability by integrating the prediction script and standardized dataset processing workflow into a unified application. This application features data import, model prediction, and result visualization capabilities, allowing users to customize input data and view prediction results in real time, making it convenient for healthcare management institutions. During model construction, data standardization and filtering ensure data quality and consistency, improving model stability and generalization ability, thus making it more applicable to different regions and elderly populations.

[0056] 5. Optimize variable selection and model combination. By using a multi-state model to select variables with significance levels below a preset statistical threshold as key variables, redundant information is reduced, and model complexity and computational cost are lowered. A stacking method is used to combine the results of different base learners, and the final ensemble model is determined by comparing the AIC and AUC values ​​of different meta-models, which improves the stability and accuracy of prediction results. Attached Figure Description

[0057] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention.

[0058] Figure 2 This is a schematic diagram illustrating the potential disease category analysis.

[0059] Figure 3 The implementation process of training and combining strategies for base learners.

[0060] Figure 4 A schematic diagram of the optimization and adjustment process for the validation set.

[0061] Figure 5 Example diagram for dynamic assessment and visualization of comorbidity risk. Detailed Implementation

[0062] This invention provides a method for predicting the risk of comorbidity development in older adults based on ensemble machine learning, the specific implementation of which is described in conjunction with the appendix. Figure 1 To be continued Figure 5 A detailed explanation follows. First, in the data preprocessing stage, a longitudinal cohort database containing multidimensional data features needs to be constructed. This database integrates demographic information, behavioral habits, physiological indicators, environmental and meteorological data, and socioeconomic variables to form a standardized dataset. The data types in the standardized dataset include continuous variables and categorical variables. (See appendix...) Figure 1 In this process, this corresponds to step 1 of the data preprocessing module. During this step, the proportion of missing values ​​for each column of variables in the standardized dataset is calculated, and variables with a missing value proportion exceeding a preset threshold are removed. For continuous data among the remaining variables, Z-score standardization is used, while one-hot encoding is used for categorical data. If the proportion of missing values ​​for a certain column of variables is less than a preset threshold, interpolation imputation is performed on that column. Linear interpolation is preferred for imputation; if linear interpolation is not applicable, K-nearest neighbor interpolation is used to complete the data imputation. This series of operations ensures the integrity and consistency of the data, laying the foundation for subsequent analysis.

[0063] Next, the data, after initial screening and normalization, is divided into training, validation, and test sets. This step is also described in the appendix. Figure 1 This is part of step 1. Specifically, the data is divided into three parts in a 7:2:1 ratio: the training set is used for model training, the validation set for optimization and adjustment, and the test set for final evaluation. After the division is completed, the key comorbidity feature extraction stage begins. (Appendix) Figure 2 This paper demonstrates the specific process of latent disease category analysis. Latent category analysis was performed on sixteen diseases covered in the normalized dataset. Based on the chronic disease prevalence of the study subjects, the comorbidity pattern characteristics in the population were systematically analyzed. Using maximum likelihood estimation, a multi-level latent category model ranging from 2 to 7 categories was systematically constructed. By comprehensively comparing the Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC), and combining clinical interpretability, a final disease combination pattern was formed to describe the progressive development process from health to a single chronic disease and then to comorbidity.

[0064] After extracting key comorbidity features, the next step is variable screening. Figure 1 Step 2 illustrates this process. Based on the disease combination pattern, the variables in the training set are divided into five dimensions: basic demographic characteristics, individual behavioral characteristics, physiological indicators, intrinsic capacity assessment, and environmental factors. Variables in each dimension are incorporated into multi-state models ranging from health to single disease and from single disease to comorbidity, respectively. Variables with significance levels below a preset statistical threshold are selected as key variables. This process is implemented using statistical analysis software to ensure that the selected variables have high explanatory and predictive power.

[0065] Subsequently, five algorithms—random forest, support vector machine, neural network, extreme gradient boosting, and logistic regression—were selected as base learners, and key variables were input into each base learner for independent training. (Appendix) Figure 3 This process is described in detail. During each cross-validation step, the training set is fitted using the five base learners mentioned above, and the prediction accuracy of each base learner is recorded. The cross-validation fold number is set to 5 to improve the model's stability and generalization ability. The base learners are trained with K-1 folds, predicting the remaining 1-fold generated meta-features. Finally, the prediction results of all generated base learners are merged as the meta-features of the training set, which are then input into the ensemble model for training. The ensemble model automatically adjusts the weights through a corresponding weight optimization algorithm to obtain the optimal weight combination of the base learners that accurately reflects the real situation.

[0066] For the stacking method, the prediction results of each base learner are used as input features, and different models are introduced as secondary learners. After training the secondary learners, the final prediction result is generated. The performance of different secondary learners is compared, and the secondary learner with the best performance (highest AUC, lowest AIC) is selected as the final meta-model of the ensemble learning.

[0067] After the base learner is trained and the combination strategy is completed, the initial prediction results are optimized and adjusted using the validation set. (Appendix) Figure 4 The specific steps of this process are demonstrated. Data from the validation set is input into the primary prediction model, and the predicted probability distribution output by the model is calculated. The predicted probability distribution is calibrated by dividing the predicted probability into ten intervals using an equidistant binning method: [0, 0.1), [0.1, 0.2), ..., [0.9, 1.0]. The actual occurrence rate of samples within each interval is calculated, and the mean squared error between the actual occurrence rate and the predicted probability is calculated. If the mean squared error exceeds a preset threshold, the model's decision boundary is adjusted, and the predicted probability distribution is recalculated until the mean squared error is less than the preset threshold. This process ensures the accuracy and reliability of the model's prediction results.

[0068] After optimization and adjustments, key variable data on comorbid outcomes and their progression in the elderly were collected. (Attached) Figure 5This paper demonstrates the specific process of dynamic assessment and visualization of comorbidity risk. Individual characteristics, behavioral habits, physiological indicators, socioeconomic data, and regional environmental and meteorological data of the elderly were extracted from an external validation dataset. The extracted data were then matched with key variables selected by a multi-state model, and successfully matched variables were retained as input data. The input data underwent the same normalization and encoding processes to ensure consistency in format and scale with the training set data. Subsequently, the input data was substituted into the final ensemble model to calculate the comorbidity risk score for each elderly person. Risk distribution curves were plotted based on the score results to analyze the distribution of the number of elderly people within different score intervals. Based on the risk distribution curves, elderly people with scores in the high-risk interval were highlighted, and personalized health management plans were developed.

[0069] Finally, the structural and training parameters of the final ensemble model are saved, generating an executable prediction script. The prediction script and the processing flow of the standardized dataset are integrated into a unified application for use by healthcare management institutions. The application features data import, model prediction, and result visualization capabilities, supporting user-defined input data and real-time viewing of prediction results. (Appendix) Figure 5 The section shows the specific form of the risk distribution curve and its application in health management.

[0070] The above implementation methods are combined with the appendix. Figure 1 To be continued Figure 5 The document provides a detailed explanation, covering the entire process from data preprocessing to the output of a comorbidity risk prediction model. Through these steps, accurate prediction and dynamic assessment of the risk of comorbidity development in the elderly can be achieved, providing a scientific basis for personalized health management.

[0071] To enable those skilled in the art to fully understand and implement this invention, the specific implementation principle of this invention will be further explained below in conjunction with a specific application scenario.

[0072] I. Scene Background

[0073] A community has 3,200 residents aged 60 and above. The method described in this invention needs to predict their risk of progressing from "health → single chronic disease → comorbidity" to provide a basis for developing personalized health management plans for the community. The implementation process requires integrating multi-dimensional information such as community health records, physical examination data, and regional environmental data, ultimately outputting a feasible risk assessment result.

[0074] II. Specific Implementation Steps and Principles

[0075] Step 1: Construct a standardized dataset

[0076] Data sources: Integrating basic information of elderly residents in the community (demographic data such as age, gender, and education level), behavioral records (smoking history, weekly exercise frequency, dietary habits, etc.), annual physical examination physiological indicators (blood pressure, fasting blood glucose, BMI, blood lipids, etc.), and environmental meteorological data of the community area (average temperature and PM2.5 concentration over the past 5 years). 2.5 Concentration, etc.) and socioeconomic data (household income, type of medical insurance, etc.).

[0077] Data preprocessing:

[0078] Calculate the proportion of missing values ​​for each variable. For example, the missing value proportion of the variable "number of physical examinations per year" is 35% (the preset threshold is 20%), so this variable is removed; the missing value proportion of "daily salt intake" is 12%, so linear interpolation is used to fill in the missing values ​​(if the data is non-linearly distributed, K-nearest neighbor interpolation is used instead).

[0079] Continuous data is standardized using Z-score, which transforms the data into a distribution with a mean of 0 and a standard deviation of 1. Categorical data is encoded using one-hot encoding, for example, "gender = male" is encoded as [1,0], and "gender = female" is encoded as [0,1].

[0080] The dataset is divided in a 7:2:1 ratio: 2240 people are used as the training set for model training, 640 people are used as the validation set for model optimization, and 320 people are used as the test set for final evaluation.

[0081] Step 2: Identify comorbidity patterns

[0082] Disease scope: Focusing on the 16 diseases mentioned in the patent, statistics on the disease status of elderly people in the community (such as "whether they have high blood pressure" or "whether they have diabetes").

[0083] Model Construction: Based on the maximum likelihood estimation method, construct a 2-7 class latent model, and iteratively optimize the parameters using the expectation-maximization algorithm (until the log-likelihood function converges). For example:

[0084] Comparing the two-class model (AIC=1250, BIC=1280) with the four-class model (AIC=1120, BIC=1160), and considering the clinical significance, the four-class model can clearly distinguish between "cardiovascular and metabolic", "musculoskeletal", "psycho-physical" and "low comorbidity". Ultimately, the four-class model was selected as the comorbidity pattern.

[0085] Comorbidity pattern description: Taking cardiovascular and metabolic diseases as an example, it is a progressive process from "health" → "single chronic disease" (such as hypertension) → "comorbidity" (such as hypertension + diabetes or hypertension + diabetes + angina).

[0086] Step 3: Screening key variables

[0087] Variable Dimensioning: The data is divided into 5 dimensions:

[0088] Basic demographics (age, sex, education level);

[0089] Personal behavior (smoking history, frequency of exercise, amount of alcohol consumed);

[0090] Physiological indicators (systolic blood pressure, fasting blood glucose, BMI, renal function indicators);

[0091] Intrinsic abilities (cognitive score, daily living skills);

[0092] Environmental factors (annual average PM2.5) 2.5 (Average annual temperature).

[0093] Multi-state model screening: Variables across all dimensions are incorporated into the "health → single disease" and "single disease → comorbidity" transformation models. Through multi-state modeling, variables with a significance level of P < 0.05 are screened. For example:

[0094] Age (P=0.001), systolic blood pressure (P=0.003), smoking history (P=0.02), PM2.5 concentration (P=0.03), and BMI (P=0.04) were identified as key variables.

[0095] Step 4: Training and combining base learners

[0096] Base learner selection: Five algorithms were used: random forest, support vector machine, neural network, extreme gradient boosting, and logistic regression.

[0097] Training process:

[0098] Key variables are input into each base learner, and 5-fold cross-validation is used. The training set is divided into 5 parts, 4 parts are used for training and 1 part is used for validation. For example, random forest has an average accuracy of 82% in cross-validation and an extreme gradient boosting accuracy of 85%.

[0099] Generative meta-features: The predictions made by each base learner on the training set are used as new input features.

[0100] Stacking method combination: Input the meta-features into the secondary learner (such as linear regression, gradient boosting tree), compare the performance of each secondary learner, such as linear regression AUC=0.88, AIC=156; gradient boosting tree AUC=0.91, AIC=142, select gradient boosting tree as the final meta-model.

[0101] Step 5: Optimize the model on the validation set

[0102] Probability calibration: The validation set (640 people) was input into the primary model to obtain the predicted probability distribution, which was divided into 10 intervals ([0,0.1),[0.1,0.2),...,[0.9,1.0]) using the equal-interval binning method.

[0103] Error adjustment: For example, in the "predicted risk 20%-30%" range, the actual comorbidity rate is 28%, which is relatively close to the predicted value; while in the "70%-80%" range, the actual incidence rate is 62%, which is relatively close to the predicted value (mean squared error = 0.012 > threshold 0.01). By adjusting the decision boundary of the ensemble model, such as increasing the weight of the high-risk range, the mean squared error is eventually reduced to 0.008 (<0.01).

[0104] Step 6: Dynamic Assessment and Health Management

[0105] Real-time data input: Collect new data from elderly people in the community after 6 months, such as the latest blood pressure, changes in exercise frequency, and environmental PM2.5 levels. 2.5 Update the values, match key variables, and then normalize them to match the training set format.

[0106] Risk score calculation: Substitute the scores into the final ensemble model to obtain the comorbidity risk score (0-1) for each elderly person, and plot the risk distribution curve:

[0107] Low risk (0-0.3): 1800 people;

[0108] Medium risk (0.3-0.7): 1100 people;

[0109] High risk (0.7-1.0): 300 people.

[0110] Personalized solutions: For high-risk groups, such as elderly individuals with hypertension, hyperglycemia, and a high BMI, customized plans will be developed.

[0111] Monitor blood pressure / blood sugar weekly;

[0112] Develop a low-salt, low-fat diet plan;

[0113] Community doctors conduct monthly follow-ups and adjust intervention measures based on environmental factors.

[0114] Step 7: Model Output and Application

[0115] Model Deployment: The parameters and data processing flow of the final integrated model are integrated into a desktop application, which community doctors can use to:

[0116] Import new physical examination data;

[0117] Click the "Predict" button to generate a risk score and visualization report in real time, such as a risk distribution histogram and a ranking of key influencing factors;

[0118] Export the list of high-risk groups and intervention recommendations for direct use in community health management.

[0119] III. Implementation Results

[0120] Using this method, the accuracy (AUC) of community prediction of comorbidity risk in the elderly reached 0.89, which is significantly improved compared with traditional single models (such as logistic regression, AUC=0.76). It can also dynamically track risk changes, providing an operable tool for "precise prevention and hierarchical management".

[0121] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0122] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the risk of comorbidity development in the elderly based on ensemble machine learning, characterized in that, The method includes: A vertical cohort database containing multidimensional data features is constructed, and a standardized dataset is formed by integrating demographic information, behavioral habits, physiological indicators, environmental and meteorological data, and socioeconomic variables; the data types in the standardized dataset include continuous variables and categorical variables; The standardized dataset is initially screened to remove samples with a missing value ratio exceeding a preset threshold, and the remaining data is normalized. The normalized data is then divided into a training set, a validation set, and a test set. Based on the chronic disease prevalence of the research subjects, a latent category analysis model was used to systematically analyze different comorbidity patterns in the population and identify representative disease combination patterns. The disease combination pattern is used to describe the progressive development from health to a single chronic disease to comorbidity; Based on the disease combination pattern, the variables in the normalized dataset are divided into five dimensions: basic demographic characteristics, personal behavioral characteristics, physiological indicators, intrinsic ability assessment, and environmental factors. The variables in each dimension are included in the multi-state model from health to single disease and from single disease to comorbidity, and variables with significance levels below the preset statistical threshold are selected as key variables. Five algorithms—random forest, support vector machine, neural network, extreme gradient boosting, and logistic regression—were selected as base learners. Key variables were input into each base learner for independent training. After training, the results of different base learners were combined using a stacking method to obtain preliminary prediction results. The initial prediction results are optimized and adjusted using the validation set. The meta-model is determined by comparing the AIC and AUC values ​​of different meta-models, and a final ensemble model is formed. The final ensemble model is used to generate comprehensive prediction results for the risk of comorbidity development in the elderly population. Collect key variable data on comorbidity outcomes and their occurrence in older adults, and dynamically assess the risk of comorbidity in older adults based on the comprehensive prediction results; When the prediction results meet the preset accuracy requirements, the current comorbidity risk prediction model and related parameter configurations are output.

2. The method for predicting the risk of comorbidity development in the elderly based on ensemble machine learning according to claim 1, characterized in that, The preliminary screening of the standardized dataset includes: Calculate the proportion of missing values ​​for each column of variables in the standardized dataset and remove variables with a proportion of missing values ​​higher than a preset threshold; perform Z-score standardization on the continuous data in the remaining variables and one-hot encoding on the categorical data; If the proportion of missing values ​​in a certain column of variables is less than a preset threshold, then interpolation is performed to impute the missing values ​​in that column of variables. Linear interpolation is preferred for imputation. If linear interpolation is not applicable, K-nearest neighbor interpolation is used to complete the data imputation.

3. The method for predicting the risk of comorbidity development in the elderly based on ensemble machine learning according to claim 1, characterized in that, The latent category analysis model was used to systematically analyze different comorbidity patterns in the population, identifying representative disease combination patterns including: We conducted an in-depth analysis of the sixteen diseases covered in the normalized dataset, systematically explored the comorbidity patterns in the elderly based on the specific disease conditions of the research subjects, and formed the final disease combination pattern by combining clinical interpretability.

4. The method for predicting the risk of comorbidity development in the elderly based on ensemble machine learning according to claim 1, characterized in that, The step of inputting key variables into each base learner for independent training includes: The selected key variables were divided into training, validation, and test sets in a 7:2:1 ratio; cross-validation was performed on the data in the training set with a cross-validation fold of 5. In each cross-validation process, a base learner is trained using K-1 folds to predict the remaining 1-fold generative features. Random forest, support vector machine, neural network, extreme gradient boosting, and logistic regression are used to fit the training set, and the prediction accuracy of each base learner is recorded.

5. The method for predicting the risk of comorbidity development in the elderly based on ensemble machine learning according to claim 1, characterized in that, The method of combining the results of different base learners using a stacking approach includes: In each cross-validation of the base learner, the base learner is trained with K-1 folds, the remaining 1 fold is predicted, and finally the prediction results of all generated base learners are merged as meta-features of the training set. These meta-features are then input into different secondary learners to train and generate prediction results. The weights are automatically adjusted through the corresponding weight optimization algorithm to obtain the optimal weight combination of the base learners that can accurately reflect the real situation. By comparing the performance of different secondary learners, the secondary learner with the best performance is selected as the final meta-model for ensemble learning.

6. The method for predicting the risk of comorbidity development in the elderly based on ensemble machine learning according to claim 1, characterized in that, The optimization and adjustment of the initial prediction results using the validation set includes: Input the data from the validation set into the primary prediction model and calculate the predicted probability distribution output by the model; The predicted probability distribution was calibrated by dividing it into ten intervals using the equidistant binning method: [0,0.1), [0.1,0.2), ..., [0.9,1.0]. The actual incidence rate of samples within each interval is statistically analyzed, and the mean squared error between the actual incidence rate and the predicted probability is calculated. If the mean squared error is greater than a preset threshold, the decision boundary of the model is adjusted, and the predicted probability distribution is recalculated until the mean squared error is less than the preset threshold.

7. The method for predicting the risk of comorbidity development in the elderly based on ensemble machine learning according to claim 1, characterized in that, The key variable data collected on comorbid outcomes and their occurrence in older adults include: Extract individual characteristics, behavioral habits, physiological indicators, socioeconomic data, and regional environmental and meteorological data of the elderly from external validation datasets; The extracted data is matched with the key variables selected through the multi-state model, and the variables that are successfully matched are retained as input data. Perform the same normalization and encoding operations on the input data to ensure that it maintains the same format and scale as the training set data.

8. The method for predicting the risk of comorbidity development in the elderly based on ensemble machine learning according to claim 1, characterized in that, The dynamic assessment of the risk of comorbidities in older adults based on the comprehensive prediction results includes: The input data is substituted into the final integrated model to calculate the risk score of the comorbidity development status for each elderly person; Based on the scoring results, a risk distribution curve was plotted to analyze the distribution of the number of elderly people in different scoring intervals. Based on the risk distribution curve, elderly people whose scores fall within the high-risk range are highlighted and personalized health management plans are developed.

9. The method for predicting the risk of comorbidity development in the elderly based on ensemble machine learning according to claim 1, characterized in that, The output of the current comorbidity risk prediction model and related parameter configurations includes: Save the structural and training parameters of the final ensemble model and generate an executable prediction script; The prediction script and standardized dataset processing workflow are integrated into a unified application for use by healthcare management organizations. The application features data import, model prediction, and result visualization capabilities, allowing users to customize input data and view prediction results in real time.

10. The method for predicting the risk of comorbidity development in the elderly based on ensemble machine learning according to claim 1, characterized in that, The method employs a latent category analysis model to systematically analyze different comorbidity patterns in a population, identifying representative disease combination patterns, including: Based on the maximum likelihood estimation method, a multi-level latent class model from 2 to 7 classes was constructed. The model parameters were iteratively optimized using the expectation-maximization algorithm until the log-likelihood function converged. The Akaike Information Criterion and Bayesian Information Criterion of each model were compared, and the final disease combination model was formed by combining clinical interpretability.

Images