Disease burden prediction and prevention and control decision-making method and system based on machine learning

Abstract

The invention provides a disease burden prediction and prevention and control decision-making method and system based on machine learning, and relates to the technical field of disease prediction and public health decision-making, and the method comprises the steps: obtaining epidemiological data of tuberculosis and related diseases from an authoritative database, and carrying out the preprocessing of the epidemiological data; respectively training an XGBoost model, an RF (Radio Frequency) model and a Prophet model; training a random forest meta-model by adopting a Stacking fusion strategy, and constructing a hybrid prediction model; calculating RMSE, MAE, MAPE and Rindex evaluation model performance; based on the obtained hybrid prediction model; based on the variable importance analysis result and the prediction result, the influence of the key independent variable on the tuberculosis burden dependent variable is quantified, the effect of the independent variable change on the dependent variable is simulated, and a tuberculosis prevention and control intervention strategy suggestion is generated. By fusing multi-source data and a hybrid modeling technology, tuberculosis prediction precision is remarkably improved, confidence interval quantization and prevention and control strategies are linked for the first time, and data-driven decision support is provided for global tuberculosis prevention and control.

Description

Technical Field

[0001] This invention relates to the field of disease prediction and public health decision-making technology, specifically to a method and system for disease burden prediction and prevention and control decisions based on machine learning. Background Technology

[0002] In the global fight against disease, the burden assessment and future trend prediction of infectious diseases (such as tuberculosis) are crucial for public health management. In recent years, with the evolution of infectious diseases and the increase in NAFLD / NASH (non-alcoholic fatty liver disease / non-alcoholic steatohepatitis) due to high dietary intake and low physical activity, various epidemic and prevalent gastrointestinal diseases, such as HIV / AIDS, COVID-19, influenza, and NAFLD / NASH, have had a complex impact on the transmission dynamics and mortality trends of tuberculosis. Traditional statistical models have significant limitations in modeling multi-factor interactions and non-linear changes.

[0003] Existing research often employs single prediction models (such as ARIMA and Prophet), which struggle to fully uncover dynamic patterns in high-dimensional, multivariate long-term health data, particularly under conditions of sudden outbreaks and structural changes in influencing factors, resulting in poor prediction accuracy. Studies have found that various infectious diseases and nutritional factors have synergistic and antagonistic effects on the burden of tuberculosis, and the disease burden of gastrointestinal NAFLD / NASH is increasing. Due to the complexity of data vectors, single models cannot fully capture the dynamic changes in health data. Therefore, a novel disease prediction framework that integrates multiple machine learning models and the influence of changes in other infectious diseases and gastrointestinal NAFLD / NASH is urgently needed to comprehensively mine and construct vector relationships, thereby improving predictive capabilities and the precision of intervention strategies. This framework is suitable for long-term trend extrapolation and possesses the ability to quantify uncertainty. Summary of the Invention

[0004] To address the technical problems in related technologies, this invention provides a method and system for disease burden prediction and prevention and control decisions based on machine learning.

[0005] To achieve the above objectives, the technical solution adopted by the present invention includes:

[0006] According to a first aspect of the present invention, a method for disease burden prediction and prevention and control decision-making based on machine learning is provided, comprising the following steps:

[0007] Step S1: Obtain epidemiological data on tuberculosis and related diseases for a certain number of years from an authoritative database. This epidemiological data includes 10 independent variables and 10 dependent variables related to the burden of tuberculosis. The obtained data is then preprocessed. The 10 independent variables include: year, number of newly diagnosed HIV infections, number of currently living HIV-infected individuals, number of HIV-related deaths, number of newly confirmed COVID-19 cases, cumulative number of confirmed COVID-19 cases, and number of COVID-19 deaths. The number of deaths, the number of new influenza cases, and non-alcoholic fatty liver disease; the 10 dependent variables related to the burden of tuberculosis include: the number of new and relapsed tuberculosis cases, the total number of tuberculosis cases, the number of new HIV-positive tuberculosis cases, the number of multidrug-resistant / rifampicin-resistant tuberculosis cases, the number of people receiving treatment for multidrug-resistant / rifampicin-resistant tuberculosis, the number of tuberculosis deaths excluding HIV-infected individuals, the number of tuberculosis deaths in the HIV-positive population, the total number of tuberculosis deaths, the number of bacteriologically confirmed pulmonary tuberculosis cases, and the global tuberculosis treatment coverage rate;

[0008] Step S2: Train the XGBoost model, RF model, and Prophet model respectively; use the Stacking fusion strategy, with the prediction results of the three models as input features, train the Random Forest Meta-model, and construct the hybrid prediction model; calculate RMSE, MAE, MAPE, and R² metrics to evaluate model performance;

[0009] Step S3: Based on the obtained hybrid prediction model, input the independent variable data for the target time period to predict the trend of the 10 dependent variables related to the tuberculosis burden, and calculate and output the 95% confidence interval of the predicted values.

[0010] Step S4: Based on the results of the variable importance analysis in Step S2 and the prediction results in Step S3, quantify the impact of key independent variables on the tuberculosis burden dependent variable, simulate the effect of changes in independent variables on the dependent variable, and generate recommendations for tuberculosis prevention and control intervention strategies.

[0011] Optionally, in step S1, the preprocessing specifically includes:

[0012] Missing values ​​were filled using mean imputation.

[0013] Outliers based on Identify and eliminate based on principles;

[0014] All independent and dependent variables were standardized using the Z-score formula:

[0015]

[0016] Where, For the standardized variable values, Raw data, Variable mean, The standard deviation is denoted as .

[0017] Optionally, in step S2, the specific methods for training each base model include:

[0018] For the XGBoost model, the objective function is set to minimize the squared error:

[0019]

[0020] In the formula, This represents the overall loss function (i.e., the objective function). This indicates the number of samples (i.e., the number of data points). This represents the total number of trees (i.e., the number of base learners). Represents the true value. Indicates the predicted value. Indicates the first The regularization term for the tree; the parameters set include: nrounds=100, max_depth=6, eta=0.1, lambda=1; the gradient boosting method is applied to iteratively train the model, updating the predicted values ​​and optimizing the objective function in each round; the output feature importance score map is used to explain the contribution of variables to the prediction;

[0021] For the RF model, the decision tree ntree=500, the minimum number of samples per split node in a single tree nodesize=5, and mtry=3 variables are randomly selected during the construction of each tree; Bootstrap sampling is used to generate a sample subset, and the node splitting of each tree is performed using the mean square error minimization criterion; the prediction results output by each tree are aggregated into the final prediction value according to the vote or the mean.

[0022] For the Prophet model, enable the annual periodicity modeling option; set the change point prior scale to 0.05 to allow for flexible trend changes; introduce covariates as external regression terms to reflect external influencing factors; the basic form of the Prophet model is as follows:

[0023]

[0024] In the formula, In time The predicted value (i.e., the target variable). For trend items, It is a periodic term. For holidays / covariates, This is the noise term.

[0025] Optionally, in step S2, the Stacking fusion strategy specifically includes:

[0026] First layer: Generate OOF predictions for XGBoost, RF, and Prophet models respectively;

[0027] The second layer: Using the OOF predicted values ​​of the three basic models as input features and the true values ​​of the tuberculosis burden dependent variable as output labels, a random forest is trained as a meta-model to obtain the fused predicted values. :

[0028]

[0029] In the formula, These are the predicted values ​​from the XGBoost model. These are the predicted values ​​from the random forest model. These are the predicted values ​​from the Prophet model.

[0030] Optionally, in step S3, the 95% confidence interval is calculated using the error propagation formula:

[0031]

[0032] In the formula, Indicates the predicted value. This represents the variance of the prediction error based on the XGBoost validation set. This represents the variance of the Out-of-Band (OOB) error in the random forest. This represents the variance of the Prophet residuals.

[0033] Optionally, in step S3, a dynamic weight adjustment mechanism is adopted for predictions exceeding a certain number of years: the variance weight of the Prophet model increases by 0.005 each year.

[0034] Optionally, in step S4, the step of quantifying the influence of key independent variables includes:

[0035] Based on XGBoost or RF models, SHAP or feature importance scores were used to identify the top three independent variables with the greatest impact on the tuberculosis burden dependent variable; and the marginal effect of intervention on the tuberculosis burden was quantified by constructing a policy response function model.

[0036] Optionally, the policy response function model is:

[0037]

[0038] In the formula, This indicates the change in indicators of tuberculosis burden. This represents the mixed prediction model function. Indicates the first One independent variable, The model represents the first One independent variable The partial derivatives, Indicates the first The amount of change in each independent variable after intervention.

[0039] Optionally, step S3 further includes visualizing the prediction results:

[0040] Use ggplot2 to plot the trend of tuberculosis burden indicators over a certain period of time;

[0041] Furthermore, in the trend chart, solid lines represent actual values, dashed lines represent predicted values, and gray gradient areas represent 95% confidence intervals; the charts for 10 dependent variables are output in pagination, and dashed lines mark the boundary between the current year and the predicted year.

[0042] According to a second aspect of the present invention, a disease burden prediction and control decision-making system based on machine learning is also provided, for executing the disease burden prediction and control decision-making method based on machine learning as described in any of the technical solutions of the first aspect of the present invention, the system comprising:

[0043] An automated model training module is used to perform data preprocessing, basic model training, and Stacking fusion.

[0044] Dynamic visualization system: used to generate predictive trend charts and display confidence intervals;

[0045] Strategy Simulation Interface: Used to simulate the impact of changes in independent variables on the tuberculosis burden and output strategy recommendations.

[0046] According to a third aspect of the present invention, a computer device is also provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, is capable of implementing the steps of the machine learning-based disease burden prediction and prevention decision-making method described in any of the technical solutions of the first aspect of the present invention.

[0047] According to a fourth aspect of the present invention, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed by a processor, is capable of implementing the steps of the machine learning-based disease burden prediction and prevention and control decision-making method described in any of the technical solutions of the first aspect of the present invention.

[0048] Beneficial effects:

[0049] 1. Through the above technical solution, firstly, this invention, by integrating multiple disease covariates (10 independent variables) and tuberculosis burden indicators (10 dependent variables) from authoritative databases (e.g., WHO, GBD, etc.), can overcome the limitations of traditional techniques that rely solely on a single disease variable (such as HIV-TB). Therefore, it is possible to construct a high-dimensional dataset covering the synergistic effects of multiple infectious diseases such as HIV, COVID-19, influenza, and NAFLD / NASH, providing a data foundation for revealing complex etiological chains.

[0050] Second, this invention employs three heterogeneous models—XGBoost, Random Forest (RF), and Prophet—and constructs a hybrid model through a stacking fusion strategy. This not only effectively addresses the problem of insufficient fitting of single models (such as ARIMA and Prophet) to high-dimensional nonlinear data, but also significantly improves prediction accuracy. Furthermore, it enhances the robustness of the model in complex scenarios such as sudden epidemics (e.g., the coupling effect between COVID-19 and TB).

[0051] Third, this invention breaks through the limitation of traditional models that only support short-term forecasts of 5–10 years, and for the first time achieves policy extrapolation capabilities on a larger scale (e.g., 50 years). At the same time, this invention quantifies the uncertainty of forecasts through confidence intervals (e.g., controlling the error of the total number of tuberculosis deaths within ±8.7%), replacing unreliable "point estimates" and providing a basis for risk boundaries for decision-making.

[0052] Fourth, this invention generates resource allocation and cross-departmental collaboration strategies based on variable importance analysis and prediction results. It can not only effectively identify key driving factors (e.g., HIV survival rate, COVID-19 death rate), avoiding the blind allocation of prevention and control resources, but also directly transform prediction results into actionable intervention recommendations (e.g., increasing HIV treatment coverage), establishing a closed-loop link from data prediction to policy implementation.

[0053] Overall, the above-mentioned technical solution, through a four-step framework of "multi-source data integration → heterogeneous model fusion → long-term uncertainty quantification → dynamic strategy generation," can effectively address the problems of insufficient predictive breadth (this invention can overcome the limitations of a single disease variable and cover the synergistic effects of 10 covariates), limited predictive depth (this invention can extend the predictive span from the short term to a longer period), and lack of decision support (this invention can transform predictive results and confidence intervals into quantitative prevention and control strategies, realizing the connection between scientific research and policy).

[0054] 2. Other beneficial effects or advantages of the present invention will be described in detail in the specific embodiments. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In addition, it should be understood that the proportional relationship of each component in the drawings of this specification does not represent the proportional relationship in the actual material selection and design, but is only a schematic diagram of the structure or position, wherein:

[0056] Figure 1 This is a flowchart illustrating the steps of a machine learning-based disease burden prediction and prevention decision-making method provided in an exemplary embodiment of the present invention.

[0057] Figure 2 This is a flowchart illustrating the steps of a machine learning-based disease burden prediction and prevention decision-making method provided in an exemplary embodiment of the present invention.

[0058] Figures 3 to 12 This is the final output result of a novel hybrid model provided in an exemplary embodiment of the present invention; wherein, Figure 3 This is a plot predicting the number of registered cases receiving treatment for MDR / RR-TB. Figure 4 This is a chart predicting the number of tuberculosis cases in HIV-positive individuals. Figure 5 This is a prediction chart for multidrug-resistant / rifampicin-resistant tuberculosis. Figure 6 This is a predictive chart for bacteriologically confirmed pulmonary tuberculosis. Figure 7 This is a chart predicting the number of tuberculosis deaths among HIV-positive individuals. Figure 8 This is a chart predicting non-HIV-related tuberculosis mortality. Figure 9 This is a chart predicting the number of tuberculosis cases. Figure 10 This is a prediction chart for new and recurrent tuberculosis cases. Figure 11 This is a projection map of global tuberculosis treatment coverage. Figure 12 This is a chart predicting the total number of deaths from tuberculosis.

[0059] Figure 13 This is a schematic diagram comparing the evaluation index results of three individual models and three mixed models for two of the ten dependent variables provided in an exemplary embodiment of the present invention. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0061] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0062] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings.

[0063] like Figures 1 to 13 As shown, according to a first aspect of the present invention, a method for disease burden prediction and prevention and control decision-making based on machine learning is provided, comprising the following steps:

[0064] Step S1: Obtain epidemiological data on tuberculosis and related diseases for a certain number of years from an authoritative database. This epidemiological data includes 10 independent variables and 10 dependent variables related to the burden of tuberculosis. The obtained data is then preprocessed. The 10 independent variables include: year, number of newly diagnosed HIV infections, number of currently living HIV-infected individuals, number of HIV-related deaths, number of newly confirmed COVID-19 cases, cumulative number of confirmed COVID-19 cases, and number of COVID-19 deaths. The number of deaths, the number of new influenza cases, and non-alcoholic fatty liver disease; the 10 dependent variables related to the burden of tuberculosis include: the number of new and relapsed tuberculosis cases, the total number of tuberculosis cases, the number of new HIV-positive tuberculosis cases, the number of multidrug-resistant / rifampicin-resistant tuberculosis cases, the number of people receiving treatment for multidrug-resistant / rifampicin-resistant tuberculosis, the number of tuberculosis deaths excluding HIV-infected individuals, the number of tuberculosis deaths in the HIV-positive population, the total number of tuberculosis deaths, the number of bacteriologically confirmed pulmonary tuberculosis cases, and the global tuberculosis treatment coverage rate;

[0065] Step S2: Train the XGBoost model, RF model, and Prophet model respectively; use the Stacking fusion strategy, with the prediction results of the three models as input features, train the Random Forest Meta-model, and construct the hybrid prediction model; calculate RMSE, MAE, MAPE, and R² metrics to evaluate model performance;

[0066] Step S3: Based on the obtained hybrid prediction model, input the independent variable data for the target time period to predict the trend of the 10 dependent variables related to the tuberculosis burden, and calculate and output the 95% confidence interval of the predicted values.

[0067] Step S4: Based on the results of the variable importance analysis in Step S2 and the prediction results in Step S3, quantify the impact of key independent variables on the tuberculosis burden dependent variable, simulate the effect of changes in independent variables on the dependent variable, and generate recommendations for tuberculosis prevention and control intervention strategies.

[0068] Through the above technical solution, firstly, this invention, by integrating multiple disease covariates (10 independent variables) and tuberculosis burden indicators (10 dependent variables) from authoritative databases (e.g., WHO, GBD, etc.), can overcome the limitations of traditional techniques that rely solely on a single disease variable (such as HIV-TB). This allows for the construction of a high-dimensional dataset covering the synergistic effects of multiple infectious diseases such as HIV, COVID-19, influenza, and NAFLD / NASH, providing a data foundation for revealing complex etiological chains.

[0069] Second, this invention employs three heterogeneous models—XGBoost, Random Forest (RF), and Prophet—and constructs a hybrid model through a stacking fusion strategy. This not only effectively addresses the problem of insufficient fitting of single models (such as ARIMA and Prophet) to high-dimensional nonlinear data, but also significantly improves prediction accuracy. Furthermore, it enhances the robustness of the model in complex scenarios such as sudden epidemics (e.g., the coupling effect between COVID-19 and TB).

[0070] Third, this invention breaks through the limitation of traditional models that only support short-term forecasts of 5–10 years, and for the first time achieves policy extrapolation capabilities on a larger scale (e.g., 50 years). At the same time, this invention quantifies the uncertainty of forecasts through confidence intervals (e.g., controlling the error of the total number of tuberculosis deaths within ±8.7%), replacing unreliable "point estimates" and providing a basis for risk boundaries for decision-making.

[0071] Fourth, this invention generates resource allocation and cross-departmental collaboration strategies based on variable importance analysis and prediction results. It can not only effectively identify key driving factors (e.g., HIV survival rate, COVID-19 death rate), avoiding the blind allocation of prevention and control resources, but also directly transform prediction results into actionable intervention recommendations (e.g., increasing HIV treatment coverage), establishing a closed-loop link from data prediction to policy implementation.

[0072] Overall, the above-mentioned technical solution, through a four-step framework of "multi-source data integration → heterogeneous model fusion → long-term uncertainty quantification → dynamic strategy generation," can effectively address the problems of insufficient predictive breadth (this invention can overcome the limitations of a single disease variable and cover the synergistic effects of 10 covariates), limited predictive depth (this invention can extend the predictive span from the short term to a longer period), and lack of decision support (this invention can transform predictive results and confidence intervals into quantitative prevention and control strategies, realizing the connection between scientific research and policy).

[0073] The technical solution of the present invention will be described in more detail below with reference to a specific embodiment.

[0074] This invention provides a hybrid prediction method based on three machine learning models: XGBoost, Random Forest, and Prophet, for fitting and predicting the impact trends of multiple disease factors, such as HIV and COVID-19, on the global tuberculosis burden. For example, this invention is trained using global epidemiological data from 2000 to 2023 and predicts the changing trends of tuberculosis burden from 2024 to 2050.

[0075] The method of the present invention may include the following steps:

[0076] Step 1: Collection and cleaning of high-dimensional covariate data for multiple diseases.

[0077] First, epidemiological data on tuberculosis and related diseases were collected from 2000 to 2023. Data sources included the World Health Organization (WHO) database and the Global Burden of Disease (GBD) database. Specifically, the data included 10 covariates (such as HIV, COVID-19, influenza, NAFLD, NASH, etc.) and 10 dependent variables (such as TB mortality rate, number of new cases, MDR / RR-TB ratio, etc.).

[0078] To ensure the quality of the modeling data, the following preprocessing operations are performed:

[0079] Missing values ​​were filled using mean imputation.

[0080] Outliers are identified and removed based on the 3σ principle (i.e., values ​​exceeding the mean ± 3 standard deviations).

[0081] All independent and dependent variables are standardized using the Z-score formula:

[0082]

[0083] Where, For the standardized variable values, Raw data, Variable mean Standard deviation;

[0084] The processed data is divided into a training set (2000–2023) and a prediction set (2024–2050) for model training and future trend prediction.

[0085] In this way, the technical limitations of existing research, such as few covariates, poor data quality, and difficulty in long-term prediction, can be overcome through these steps. Furthermore, by standardizing and cleaning the data, the comparability and stability of the model input can be ensured.

[0086] Step 2: Modeling and fitting of tuberculosis mortality trends based on multiple models.

[0087] This invention constructs several mainstream machine learning models (including XGBoost, Random Forest (RF), and the time series model Prophet) to fit and predict indicators such as tuberculosis mortality. The fitting process of each model is as follows:

[0088] 1. XGBoost Model Construction and Training

[0089] The boosting tree model is used for regression modeling, with the following specific settings:

[0090] The objective function is set as minimizing the squared error:

[0091]

[0092] in, This represents the overall loss function (i.e., the objective function). This indicates the number of samples (i.e., the number of data points). This represents the total number of trees (i.e., the number of base learners). Represents the actual value. Indicates the predicted value. Indicates the first The regularization term for each tree;

[0093] Set the core parameters: nrounds=100 (number of boost rounds), max_depth=6 (maximum tree depth), eta=0.1 (learning rate), lambda=1 (L2 regularization coefficient);

[0094] Gradient boosting is used to iteratively train the model, updating the predicted values ​​and optimizing the objective function in each round.

[0095] Output a feature importance score plot to explain the contribution of variables to the prediction.

[0096] In this way, this step can overcome the limitation of a single linear model in predicting high-dimensional covariates, thereby improving the model's ability to fit nonlinear complex relationships.

[0097] 2. Random Forest (RF) Model Construction

[0098] The ensemble decision tree method is used for fitting and prediction, with the following specific settings:

[0099] Set the number of decision trees ntree=500, and the minimum number of samples per split node in a single tree nodesize=5;

[0100] During the construction of each tree, 3 variables, mtry, are randomly selected.

[0101] Bootstrap sampling is used to generate a subset of samples, and each tree is partitioned using the mean squared error (MSE) minimization criterion;

[0102] The predictions output by each tree are aggregated into the final prediction value based on either voting or average.

[0103] In this way, the generalization ability of predictions and the robustness of the model can be improved, and the stability and anti-overfitting ability of the model can be enhanced.

[0104] 3. Prophet Model Construction

[0105] Suitable for time series forecasting tasks, primarily modeling long-term trends and periodic changes, with the following settings:

[0106] Enable the annual seasonality modeling option yearly.seasonality=TRUE;

[0107] Setting the changepoint prior scale to 0.05 allows for flexible changes in the trend;

[0108] Covariates are introduced as external regression terms (add_regressor) to reflect external influencing factors.

[0109] The basic form of the Prophet model is:

[0110]

[0111] in, In time The predicted value (i.e., the target variable). For trend items, It is a periodic term. For holidays / covariates, This is the noise term.

[0112] In this way, this step can overcome the shortcomings of traditional models in simulating policy changes and external shocks (such as COVID-19), thereby enabling effective modeling of long-term tuberculosis trends.

[0113] Step 3: Model fusion and future trend prediction (2024–2050).

[0114] This step constructs a fusion prediction model based on the prediction results of multiple machine learning models (XGBoost, Random Forest, Prophet). It uses a weighted average method and a stacking method for combined predictions and calculates the prediction interval. The fusion process includes:

[0115] a) Construct base models (XGBoost, RF, Prophet) and fit them to data from 2000 to 2023 respectively;

[0116] b) Generate forecasts for 2024–2050 for each base model;

[0117] c) Combine the above prediction results using the weighted average method and the Stacking method to optimize the combined weights;

[0118] d) Calculate the confidence interval (95% confidence level) through error propagation;

[0119] e) Set up a dynamic adjustment mechanism for long-term forecasts (e.g., increase the weight of Prophet year by year after 2030).

[0120] f) Evaluate future prediction results and compare the accuracy performance of different fusion strategies (such as R², MAPE, RMSE, etc.).

[0121] In this way, the following can be achieved: First, overcome the problem of limited predictive power of a single model and improve adaptability in complex scenarios (such as sudden changes in covariates and alternating seasonal trends); Second, provide a quantifiable range of uncertainty to replace traditional point prediction and improve the reliability of scientific decision-making; Third, solve the dilemma of model selection in actual deployment, propose a trade-off between accuracy and resources, and facilitate flexible selection under different conditions.

[0122] In other words, this allows fusion models to be more adaptable and stable than single models when dealing with multi-factor coupled data (such as HIV and COVID-19 comorbidity). Meanwhile, the Stacking method automatically learns the optimal combination strategy by training a meta-model (such as a random forest), achieving the highest prediction accuracy (e.g., R² can reach 0.92). Furthermore, it employs a variance dynamic adjustment strategy for long-term predictions beyond 2030, ensuring more reasonable quantification of uncertainty. In addition, the introduction of confidence intervals enhances the model's guidance and credibility for policymakers.

[0123] Step 4: Formulate prevention and control strategies based on model results.

[0124] This step, based on model output and variable importance analysis, quantifies the role of key influencing factors in tuberculosis mortality trends and formulates intervention strategies. It mainly includes:

[0125] a) Identify core variables using feature importance scores (or SHAP values) from models such as XGBoost and RF;

[0126] b) Constructing the response function model Quantify the marginal impact of intervention on TB burden;

[0127] c) Assess the impact of changes in key variables (such as a 10% decrease in new HIV cases) on the number of TB deaths by simulating them;

[0128] d) Based on historical policies and projected trends, propose targeted intervention recommendations (such as increasing HIV treatment coverage and enhancing respiratory influenza infection control).

[0129] In this way, this step not only clarifies the key variables affecting the burden of TB (such as HIV infection rate, new COVID-19 cases, etc.), supporting precise policy implementation; but also, the policy response function provides a quantitative relationship between "input variables and output results," facilitating the simulation of intervention effectiveness; in addition, the visualization of analysis results (such as variable importance graphs and simulated response curves) enhances the transparency and operability of decision-making.

[0130] Correspondingly, it can solve the problems of unreasonable allocation of prevention and control resources and lagging policy response; at the same time, it provides data-driven decision support tools to support dynamic strategy adjustment; in addition, it makes up for the shortcomings of traditional epidemiology in lacking multivariate collaborative modeling and prognosis simulation.

[0131] 4. Model Selection Principles

[0132] (1) Data characteristics dominate

[0133] Complex coupling relationships: If it is necessary to combine the advantages of heterogeneous models (such as coexistence of HIV, COVID-19, and influenza), the Stacking model should be selected.

[0134] (2) Resource and time constraints

[0135] High-resource scenarios: Choose the Stacking model (accuracy is prioritized, but training time will be increased).

[0136] (3) Strategy requirements

[0137] Long-term planning: Stacking model (high precision).

[0138] Optimal model selection: The Stacking model has the best overall performance (highest accuracy) in complex data scenarios, but the computational cost needs to be weighed.

[0139] Recommended for practical applications: When data fluctuates drastically and resources are sufficient, prioritize the use of the Stacking model to improve prediction reliability.

[0140] Constructing a hybrid model and predicting trends from 2024 to 2050:

[0141] Dynamically optimizing weights improves model adaptability (for example, Prophet's ability to predict sudden outbreaks may be enhanced during a pandemic); Bayesian optimization balances exploration (global search) and exploitation (local optimization), making it more efficient than grid search.

[0142] Stacked integration model:

[0143] a) First layer: Generate OOF predictions for XGBoost, RF, and Prophet;

[0144] b) Second layer: Use random forest as the meta-model to fuse the base model output.

[0145] 4. Confidence interval calculation:

[0146] For the mixture model, an error propagation method is used:

[0147]

[0148]

[0149] Implementation steps:

[0150] Single-model variance estimation:

[0151] XGBoost: Calculates prediction error variance using 10-fold cross-validation; RF: Estimates variance using out-of-bag error; Prophet: Based on residual variance from time series decomposition.

[0152] Covariance handling: It is assumed that the model prediction errors are independent and the covariance term is 0; in actual calculations, the independence assumption is verified by Bootstrap sampling.

[0153] Dynamic adjustment mechanism: For long-term forecasts (>2030), the Prophet variance weight increases by 0.005 per year; for short-term forecasts, XGBoost is the dominant factor (σXGB2 weight × 1.2).

[0154] Technical advantages:

[0155] Probabilistic output: Compared to traditional point prediction, it provides a quantifiable range of uncertainty; Model specificity: It distinguishes the error characteristics of different algorithms (XGBoost is sensitive to outliers, while Prophet excels at trend extrapolation); Interpretability: It clarifies the contribution of each model through variance decomposition (as shown in the figure, RF variance accounts for 28.1%); Improved predictive ability, for example: In the Stacking mixture model, the R2 of the number of tuberculosis deaths among HIV-positive individuals is significantly improved by 28.7% compared to the R2 of the Prophet model alone.

[0156] Step 4, Formulating Prevention and Control Strategies:

[0157] 1. Identify key influencing factors based on variable importance analysis.

[0158] 2. Constructing a policy response function to quantify the effects of intervention:

[0159]

[0160]

[0161] Specific implementation forms may include the following:

[0162] In step one, data preprocessing specifically includes:

[0163] 1. Perform logarithmic transformation on count variables such as "new-onset HIV disease".

[0164] 2. Perform logit transformation on the rate variable "global tuberculosis treatment coverage".

[0165] 3. Linear interpolation was used to handle missing values ​​for individual years.

[0166] In step two, the basic model is built.

[0167] 1. XGBoost model:

[0168] Parameter settings: number of iterations `nrounds=100`, objective function `reg:squarederror`, maximum tree depth `max_depth=5`;

[0169] Training method: The standardized independent variable matrix is ​​used as input and the tuberculosis index is used as output. The hyperparameters are optimized through cross-validation.

[0170] 2. Random Forest Model:

[0171] Parameter settings: Number of trees `ntree=500`, feature sampling ratio `mtry=3`;

[0172] Training method: Multiple decision trees are generated based on Bootstrap sampling, and the model stability is evaluated by out-of-bag error (OOB).

[0173] 3. Prophet model:

[0174] Parameter settings: Enable annual seasonality (`yearly.seasonality=TRUE`), disable weekly / daily seasonality;

[0175] Training method: Convert the year to date format (`as.Date(paste0(YEAR, "0101"))`) and fit the time series trend and holiday effect.

[0176] The model evaluation uses four types of indicators:

[0177] RMSE (Root Mean Square Error): Measure the deviation between the predicted value and the actual value;

[0178] MAE (Mean Absolute Error): It reflects the absolute magnitude of the prediction error;

[0179] MAPE (Mean Absolute Percentage Error): Assess relative error;

[0180] R² (coefficient of determination): The explanatory power of the representation model;

[0181] Where, This represents the true value of the i-th observation; This represents the i-th predicted value; Total sample size; The mean of the true values; Indicates the relative error ratio; Represents the total sum of squares (SST); Represents the sum of squared residuals (SSE); R² is an explanatory measure, with 1 indicating a perfect fit. MAE indicates that the contribution of each sample point to the result is linear, making it more robust. RMSE represents the average deviation of predicted values ​​from true values. MAPE converts absolute error into a percentage.

[0182] The example shows the evaluation index results for three individual models and three mixed models with two dependent variables out of ten. Each of the three mixed models has its advantages; the optimal mixture method should be selected based on the specific results. See reference [link / reference]. Figure 13 .

[0183] In step three, the hybrid Stacking model is constructed:

[0184] Meta-model construction: Using the predicted values ​​of XGBoost, RF, and Prophet as input features, the true values ​​of tuberculosis indicators as output, and training a random forest as the meta-model;

[0185] Training process:

[0186] Required parameters and settings for this formula

[0187] Input features: : Predictions from the XGBoost model; : Predictions from the random forest model; : Predictions from the Prophet model.

[0188] Output labels: True values ​​of tuberculosis indicators (such as 10 target variables including TB new and relapse cases).

[0189] Meta-model parameters (using the randomForest package in R as an example):

[0190] Number of trees (ntree): 500 (determined through cross-validation to balance computational efficiency and stability); Feature sampling ratio (mtry): the square root of the number of input features (e.g., when inputting 3 features, the default mtry=2); Minimum number of samples per node (nodesize): 5 (to prevent overfitting); Use replacement sampling: TRUE (Bootstrap sampling); Importance assessment: Enable importance=TRUE to calculate the SHAP value or variance contribution rate to quantify feature importance.

[0191] For example, the training process can be implemented using the R language:

[0192] stack_train <- data.frame(

[0193] xgb = xgb_pred,

[0194] rf = rf_pred,

[0195] prophet = prophet_pred,

[0196] y = y_train # Real labels )

[0198] stack_model <- randomForest(

[0199] x = stack_train[, -ncol(stack_train)], # Input features (xgb, rf, prophet)

[0200] y = stack_train$y, # Output labels

[0201] ntree = 500, # Number of trees

[0202] mtry = sqrt(ncol(stack_train) - 1), # Feature sampling ratio

[0203] nodesize = 5, # Minimum number of samples for leaf nodes

[0204] importance = TRUE # Enable feature importance assessment )

[0206] Stacking_Pred <- predict(stack_model, newdata = stack_input)

[0207] Key optimization methods:

[0208] Input feature processing: The predicted values ​​of XGBoost, RF, and Prophet are standardized (Z-score) to eliminate dimensional differences; if multicollinearity exists among the predicted values ​​(VIF > 5), principal component analysis (PCA) is used for dimensionality reduction.

[0209] Hyperparameter tuning: Optimize the ntree (200~1000) and mtry (1~3) parameters through grid search with the goal of minimizing the RMSE on the validation set; evaluate the generalization performance of the model using 5-fold cross-validation.

[0210] Prediction visualization implementation methods:

[0211] Input independent variable data for 2024 and 2050 (variables such as HIV and COVID-19 are filled with historical means); generate predicted values ​​for each model and 95% confidence intervals for the mixture model (calculated as ±10% error range of predicted value); use ggplot2 to plot a trend chart from 2000 to 2050; use gray gradient areas to represent the 95% CICI coverage (95% confidence interval) of the three mixture models to assess the reliability of the predicted values; distinguish line types: solid line (actual value), dashed line (predicted value), and dotted line (single model); output predicted charts for 10 dependent variable indicators in paginated format; dashed line dividing 2023 and 2024 (historical data and predicted data).

[0212] Step four involves the strategy suggestion generation process:

[0213] 1. Variable Importance Analysis:

[0214] It can identify the top 3 most important independent variables;

[0215] Extract the feature importance of XGBoost and RF (based on SHAP value or variance contribution rate) to identify key driving factors (such as HIV survival rate and COVID-19 death rate).

[0216] Example conclusions: HIV survival rate was negatively correlated with tuberculosis treatment coverage (SHAP value = 0.32); COVID-19 deaths explained 28% of total TB deaths;

[0217] 2. Simulate the effect of a 10% increase or decrease in key variables;

[0218] 3. Develop more effective medium- and long-term targeted strategy recommendations:

[0219] Priority intervention areas: Increase HIV treatment coverage to reduce the risk of co-infection with tuberculosis;

[0220] Resource allocation planning: Based on forecasts (such as an increase in TB deaths after 2030), deploy medical resources in advance;

[0221] Policy formulation: Promote multi-sectoral collaboration (such as combining influenza surveillance with early tuberculosis screening).

[0222] This invention, through an innovative multi-model fusion architecture, achieves the following improvements in tuberculosis burden prediction compared to traditional methods: improved prediction accuracy (average RMSE reduced by 23.6%); quantified uncertainty (providing a 95% confidence interval); enhanced policy support (quantifiable assessment of the effectiveness of interventions); and improved system stability (interannual volatility of the Stacking model reduced by 41.2%).

[0223] Compared with the prior art, the beneficial effects of the present invention include at least the following aspects:

[0224] In terms of model architecture, existing technologies mostly use a single model to predict the incidence or mortality of tuberculosis, which suffers from weak generalization ability and poor stability. This invention constructs a multi-model collaborative prediction system with three basic models as the core and integrates three different strategies (such as weighted average and stacked fusion), which significantly improves the predictive performance and adaptability of the model and comprehensively optimizes the model evaluation indicators (such as RMSE, MAE, R², etc.).

[0225] In terms of uncertainty quantification, existing methods mostly rely on point estimation, lacking an effective expression of the prediction interval and making it difficult to assess the stability and confidence level of the prediction results. This invention introduces a confidence interval estimation method based on error propagation theory, adding a gray-transparent 95% confidence interval expression to the prediction results, effectively improving the uncertainty expression ability and interpretability of the model prediction.

[0226] Regarding the scope of covariates, existing technologies are mostly limited to bivariate analysis of HIV and TB, failing to reveal more complex disease interactions. This invention introduces 10 covariates, including HIV, COVID-19, and influenza, to construct a multi-disease coupling mechanism, which helps to uncover potential multi-causal risk pathways and expands the breadth and depth of disease prediction.

[0227] In terms of time span, traditional models mainly focus on short-term predictions of the past 5-10 years, which is insufficient to support policy-making in assessing medium- and long-term trends. This invention can cover data from 2000 to 2050, and particularly supports long-term continuous predictions of tuberculosis mortality trends from 2024 to 2050, providing a stable medium- and long-term basis for public health intervention strategies. This is mainly reflected in the following aspects:

[0228] First, the accuracy of multi-model hybrid prediction is improved.

[0229] This invention innovatively employs a hybrid architecture of three heterogeneous models: XGBoost, Random Forest (RF), and Prophet. Through a Stacking fusion strategy, it reduces the average RMSE by 23.6% (validation set data) in tuberculosis burden prediction compared to traditional single-model prediction methods. In particular, for the prediction of the "number of new and recurrent tuberculosis cases," the Stacking fusion model achieves an R² of 0.912, a 14.2% improvement over the best single model.

[0230] Second, modeling the synergistic effects of multiple diseases.

[0231] Breaking through the limitations of existing studies that only analyze HIV-TB co-infection, this invention is the first to integrate 10 covariates, including HIV, COVID-19, influenza, and NAFLD / NASH, to construct a predictive system. Through random forest variable importance analysis, it was found that the marginal effects of new COVID-19 cases (β=0.32±0.04) and HIV survivors (β=0.28±0.03) on the tuberculosis burden exceed traditional understanding, providing new evidence for cross-disease joint prevention and control.

[0232] Third, quantification of long-term forecast confidence intervals.

[0233] This innovative technique employs an error propagation formula to calculate the prediction uncertainty of a hybrid model. Compared to traditional point prediction methods, this technique can provide the probability distribution of predicted values ​​from 2024 to 2050, with the confidence interval width of key indicators such as "total tuberculosis deaths" controlled within ±8.7%.

[0234] Fourth, real-time strategy simulation system.

[0235] Constructing a policy response function based on variable importance analysis:

[0236] The intervention effects can be quantified and evaluated, such as "a 10% reduction in new HIV cases will lead to a (4.2±0.6)% reduction in the number of HIV-positive tuberculosis deaths," supporting precise prevention and control decisions.

[0237] Fifth, a multi-timescale verification system.

[0238] Using a rolling forecast origin, validation on data from 2000-2023 shows:

[0239] Short-term (1-3 years) forecasts are primarily driven by XGBoost (weight 0.52).

[0240] Long-term (10+ years) forecasts are dominated by Prophet (weight 0.61).

[0241] The fusion model achieved a MAPE of <15% at all time scales, significantly outperforming the single model.

[0242] Sixth, open-source and scalable architecture.

[0243] The R language implementation provided by this invention includes:

[0244] Automated model training module (train_predict_model())

[0245] Dynamic visualization system (plot_forecast())

[0246] Policy simulation interface (policy_simulator())

[0247] It can support the seamless integration of newly added variables and models.

[0248] According to a third aspect of this application, a computer device is also provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it is able to implement the steps of the machine learning-based disease burden prediction and prevention decision-making method in any of the technical solutions of the first aspect of this application.

[0249] It is understood that in this embodiment, the memory may include volatile memory, such as random access memory; the memory may also include non-volatile memory, such as read-only memory, flash memory, hard disk or solid-state drive; furthermore, the memory may include combinations of the above types of memory. This application does not specifically limit it in this regard.

[0250] Similarly, the processor can implement or execute the various exemplary logical steps described in conjunction with the disclosure of this application. The processor can be a central processing unit, a general-purpose processor, a digital signal processor, an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logical steps described in conjunction with the disclosure of this application. The processor can also be a combination that implements computational functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc.

[0251] According to a fourth aspect of this application, a computer-readable storage medium is also provided, on which a computer program is stored, characterized in that, when executed by a processor, the computer program is capable of implementing the steps of the machine learning-based disease burden prediction and prevention and control decision-making method in any of the technical solutions of the first aspect of this application.

[0252] In this embodiment, the computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), a register, a hard disk, an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof, or any other form of computer-readable storage medium known in the art. An exemplary storage medium is coupled to a processor, enabling the processor to read information from and write information to the storage medium. Of course, the storage medium may also be a component of the processor. The processor and the storage medium may reside in an application-specific integrated circuit (ASIC). In the embodiments of this application, the computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0253] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for disease burden prediction and prevention decision-making based on machine learning, characterized in that, include: Step S1: Obtain epidemiological data on tuberculosis and related diseases for a certain number of years from an authoritative database. The epidemiological data on tuberculosis and related diseases includes 10 independent variables and 10 dependent variables related to the burden of tuberculosis. The acquired data were preprocessed. The 10 independent variables included: year, number of new HIV infections, number of currently living HIV-infected individuals, number of HIV-related deaths, number of newly confirmed COVID-19 cases, cumulative number of confirmed COVID-19 cases, number of COVID-19 deaths, number of new influenza cases, and non-alcoholic fatty liver disease. The 10 dependent variables related to tuberculosis burden included: number of new and relapsed tuberculosis cases, total number of tuberculosis cases, number of new HIV-positive tuberculosis cases, number of multidrug-resistant / rifampicin-resistant tuberculosis cases, number of people receiving treatment for multidrug-resistant / rifampicin-resistant tuberculosis, number of tuberculosis deaths excluding HIV-infected individuals, number of tuberculosis deaths in the HIV-positive population, total number of tuberculosis deaths, number of bacteriologically confirmed pulmonary tuberculosis cases, and global tuberculosis treatment coverage. Step S2: Train the XGBoost model, RF model, and Prophet model respectively; use the Stacking fusion strategy, with the prediction results of the three models as input features, train the Random Forest Meta-model, and construct the hybrid prediction model; calculate RMSE, MAE, MAPE, and R² metrics to evaluate model performance; Step S3: Based on the obtained hybrid prediction model, input the independent variable data for the target time period to predict the trend of the 10 dependent variables related to the tuberculosis burden, and calculate and output the 95% confidence interval of the predicted values. Step S4: Based on the results of the variable importance analysis in Step S2 and the prediction results in Step S3, quantify the impact of key independent variables on the tuberculosis burden dependent variable, simulate the effect of changes in independent variables on the dependent variable, and generate recommendations for tuberculosis prevention and control intervention strategies.

2. The disease burden prediction and prevention decision-making method based on machine learning according to claim 1, characterized in that, In step S1, the preprocessing specifically includes: Missing values ​​were filled using mean imputation. Outliers based on Identify and eliminate based on principles; All independent and dependent variables were standardized using the Z-score formula: ； In the formula, For the standardized variable values, Raw data, Variable mean The standard deviation is denoted as .

3. The disease burden prediction and prevention decision-making method based on machine learning according to claim 1 or 2, characterized in that, In step S2, the specific methods for training each basic model include: For the XGBoost model, the objective function is set to minimize the squared error: ； In the formula, Represents the overall loss function. Indicates the number of samples. This represents the total number of trees. Represents the actual value. Indicates the predicted value. Indicates the first The regularization term for the tree; the parameters set include: nrounds=100, max_depth=6, eta=0.1, lambda=1; the gradient boosting method is applied to iteratively train the model, updating the predicted values ​​and optimizing the objective function in each round; the output feature importance score map is used to explain the contribution of variables to the prediction; For the RF model, the decision tree ntree=500, the minimum number of samples per split node in a single tree nodesize=5, and mtry=3 variables are randomly selected during the construction of each tree; Bootstrap sampling is used to generate a sample subset, and the node splitting of each tree is performed using the mean square error minimization criterion; the prediction results output by each tree are aggregated into the final prediction value according to the vote or the mean. For the Prophet model, enable the annual periodicity modeling option; set the change point prior scale to 0.05 to allow for flexible trend changes; introduce covariates as external regression terms to reflect external influencing factors; the basic form of the Prophet model is as follows: ； In the formula, In time The predicted value of For trend items, It is a periodic term. For holidays / covariates, This is the noise term.

4. The disease burden prediction and prevention decision-making method based on machine learning according to claim 3, characterized in that, In step S2, the Stacking fusion strategy specifically includes: First layer: Generate OOF predictions for XGBoost, RF, and Prophet models respectively; The second layer: Using the OOF predicted values ​​of the three basic models as input features and the true values ​​of the tuberculosis burden dependent variable as output labels, a random forest is trained as a meta-model to obtain the fused predicted values. : ； In the formula, These are the predicted values ​​from the XGBoost model. These are the predicted values ​​from the random forest model. These are the predicted values ​​from the Prophet model.

5. The disease burden prediction and prevention decision-making method based on machine learning according to claim 4, characterized in that, In step S3, the 95% confidence interval is calculated using the error propagation formula: ； In the formula, Indicates the predicted value. This represents the variance of the prediction error based on the XGBoost validation set. This represents the variance of the Out-of-Band (OOB) error in the random forest. This represents the variance of the Prophet residuals.

6. The disease burden prediction and prevention decision-making method based on machine learning according to claim 5, characterized in that, In step S3, a dynamic weight adjustment mechanism is used for predictions exceeding a certain number of years: the variance weight of the Prophet model increases by 0.005 each year.

7. The disease burden prediction and prevention decision-making method based on machine learning according to claim 1, characterized in that, In step S4, the step of quantifying the influence of key independent variables includes: Based on XGBoost or RF models, SHAP or feature importance scores were used to identify the top three independent variables with the greatest impact on the tuberculosis burden dependent variable; and the marginal effect of intervention on the tuberculosis burden was quantified by constructing a policy response function model.

8. The disease burden prediction and prevention decision-making method based on machine learning according to claim 7, characterized in that, The policy response function model is as follows: ； In the formula, This indicates the change in indicators of tuberculosis burden. This represents the mixed prediction model function. Indicates the first One independent variable, The model represents the first One independent variable The partial derivatives, Indicates the first The amount of change in each independent variable after intervention.

9. The disease burden prediction and prevention decision-making method based on machine learning according to claim 1, characterized in that, Step S3 also includes visualization of the prediction results: Use ggplot2 to plot the trend of tuberculosis burden indicators over a certain period of time; Furthermore, in the trend chart, solid lines represent actual values, dashed lines represent predicted values, and gray gradient areas represent 95% confidence intervals. The paged output shows the prediction plots for 10 dependent variables, with dashed lines marking the boundary between the current year and the prediction year.

10. A disease burden prediction and prevention decision-making system based on machine learning, characterized in that, The system is used to execute the machine learning-based disease burden prediction and prevention decision-making method as described in any one of claims 1-9, the system comprising: An automated model training module is used to perform data preprocessing, basic model training, and Stacking fusion. Dynamic visualization system: used to generate predictive trend charts and display confidence intervals; Strategy Simulation Interface: Used to simulate the impact of changes in independent variables on the tuberculosis burden and output strategy recommendations.

Images