Important disease-free risk prediction method based on multi-dimensional health data analysis

Through multi-dimensional health data analysis, combined with machine learning algorithms and statistical models, a high-precision disease risk prediction model is built, which solves the problem of inaccurate prediction results in the existing technology, and realizes personalized and real-time health management services.

CN120544895APending Publication Date: 2025-08-26广州市疾病预防控制中心(广州市卫生监督所)
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202510657860.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-05-21
Publication Date
2025-08-26

AI Technical Summary

Technical Problem

The existing disease risk prediction methods cannot comprehensively consider a variety of factors, resulting in inaccurate prediction results, inability to prevent the occurrence of diseases in advance, and lack personalization and real-timeness.

Method used

By collecting multi-dimensional health data, using correlation analysis, random forest algorithm, Cox proportional hazard model, ARIMA time series model and recurrent neural network, a high-precision disease risk prediction model is constructed, combined with deep learning models for dynamic evaluation, and personalized health warnings and intervention suggestions are provided.

Benefits of technology

It realizes accurate assessment and dynamic monitoring of user health status, provides personalized health management plans, supports real-time data updates and model optimization, and improves the accuracy of disease prediction and personalized services.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120544895A_ABST
    Figure CN120544895A_ABST
Patent Text Reader

Abstract

The invention provides a non-serious disease risk prediction method based on multi-dimensional health data analysis, and the method can collect the multi-dimensional health data of a user, builds a high-precision serious disease risk prediction model through multi-source data fusion and intelligent analysis in combination with a machine learning algorithm and a statistical model, and improves the risk prediction accuracy. The health state of the user is dynamically evaluated based on the deep learning model, the potential disease risk of the user is predicted by analyzing the relevance between historical data and future health risks, the health trend, risk factors and improvement suggestions are covered, the user is helped to make a scientific health management plan, the limitation of a traditional prediction method is broken through, and the user experience is improved. And real-time data updating and model optimization are supported, and a more accurate and personalized prediction result is provided.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of medical health, and in particular to a method for predicting the risk of no major disease based on multi-dimensional health data analysis. Background Art

[0002] In today's society, numerous factors pose significant threats to human health. From a dietary perspective, the prevalence of fast food culture has led to a prevalence of high-fat, high-salt, and high-sugar foods in the daily diet. Long-term consumption can easily lead to obesity, diabetes, and cardiovascular disease. Furthermore, the fast-paced lifestyle has led to a lack of exercise, a significant increase in sedentary time, and the common practice of staying up late, which not only weakens the body's immune system but also disrupts normal circadian rhythms. This, coupled with high work pressure, a harsh working environment, and complex interpersonal relationships, places a heavy psychological burden on people, which, over time, can easily lead to mental illness and other chronic diseases.

[0003] As these adverse effects become increasingly apparent, people are becoming increasingly aware of the value of health and are paying more and more attention to their own well-being. They are actively seeking ways to maintain their health, hoping to prevent illness before it occurs. Therefore, predictive methods that comprehensively consider multiple factors such as diet, lifestyle, and social environment, and accurately assess an individual's risk of developing major diseases through big data analysis and scientific algorithms, are becoming increasingly important. These predictive methods can help people adjust their lifestyles in advance, take timely preventative measures, reduce their risk of illness, and nip illness in the bud, effectively improving their quality of life and safeguarding their health. Summary of the Invention

[0004] To solve the above problems, the present invention provides a method for predicting the risk of no major diseases based on multi-dimensional health data analysis. It can collect data from multiple dimensions of users for integration, and perform intelligent analysis on the collected data to achieve dynamic assessment of the user's health status and predict the user's potential disease risk.

[0005] To achieve the above object, the technical solution adopted by the present invention is:

[0006] The present invention provides a method for predicting the risk of no major disease based on multi-dimensional health data analysis, comprising the following steps:

[0007] S1: Collect users’ health data from multiple channels and integrate them into a dataset;

[0008] S2: Preprocess the collected health data, extract features using correlation analysis, and then combine the preprocessed and feature-extracted data into a complete feature matrix by column based on a feature splicing fusion method.

[0009] S3: Using the random forest algorithm, combined with the Cox proportional hazards model and the ARIMA time series model, a high-precision major disease risk prediction model is constructed;

[0010] S4: Input the user's real-time multi-dimensional health data into the trained prediction model to predict the probability of the user developing a major disease in the future. Combined with the results of the Cox proportional hazards model and the ARIMA time series model, the model predicts the changing trend of the user's disease risk at different time points in the future and provides personalized health warnings and intervention recommendations for the user.

[0011] S5: Based on the recurrent neural network (RNN) and its variant, the long short-term memory (LSTM) model, it dynamically models health data with time series characteristics, and monitors and dynamically evaluates the user's health status in real time.

[0012] S6: Establish a health status scoring system to score the user's health status based on the model's prediction results and changes in the user's health indicators.

[0013] In step S1, the collected data include physiological indicator data, living habit data, genetic information data, environmental data and medical record data.

[0014] Furthermore, in step S2, feature extraction is performed using a correlation analysis method to screen out features that are highly correlated with major disease risks and remove redundant features. The specific process includes:

[0015] For the cleaned and preliminarily processed multidimensional health data, select the target variable to be analyzed and all possible related feature variables;

[0016] Calculate the correlation coefficient. For numerical features with normal distribution, calculate the Pearson correlation coefficient; for data that do not meet the normal distribution or the variables are ordered categorical variables, calculate the Spearman rank correlation coefficient;

[0017] Determine the threshold value of the correlation coefficient;

[0018] The calculated correlation coefficient is compared with a threshold, and features whose correlation coefficient reaches or exceeds the threshold are retained.

[0019] Furthermore, in step S2, in the feature splicing fusion method, corresponding weights are assigned to data from different sources according to their importance to disease risk prediction, and when splicing features, weighted processing is performed on each part of the features according to the weights.

[0020] In step S3, the process of building a high-precision major disease risk prediction model includes:

[0021] From the feature matrix after data fusion, relevant features are extracted as input to the major disease risk prediction model, the target variable is determined, and the dataset is divided into a training set and a test set;

[0022] Use the random forest algorithm to build multiple decision trees and integrate them;

[0023] Set the specific content of outcome variables and covariates and construct the Cox proportional hazards model;

[0024] Use data with time series characteristics to build an ARIMA time series model;

[0025] The predicted category output by the random forest algorithm, the risk score output by the Cox proportional hazard model, and the predicted value of the disease occurrence time output by the ARIMA time series model are used as new features and input into a new meta-model for fusion to obtain the final high-precision major disease risk prediction model.

[0026] Furthermore, in step S3, the specific process of constructing and integrating multiple decision trees using the random forest algorithm includes:

[0027] Randomly extract multiple subsets from the training set with replacement, and build a decision tree for each subset;

[0028] By calculating the information gain of features, we can select features that have a significant impact on disease risk prediction;

[0029] The generated multiple decision trees are integrated, and the prediction results of each decision tree are combined to obtain the final prediction value.

[0030] Furthermore, in step S3, the construction process of the Cox proportional hazards model includes:

[0031] The outcome variable was set as the occurrence or non-occurrence of the disease, and the covariates were physiological indicators, lifestyle habits, genetic information, environmental factors, and pre-processed data of medical records, and a Cox proportional hazard model was constructed;

[0032] The preprocessed data is divided into a training set and a test set, and the constructed model is trained using the training set data;

[0033] Use the maximum likelihood estimation method to estimate the parameters of the model, and find the parameter value that maximizes the likelihood function through iterative calculation;

[0034] Based on the estimated parameters, a risk score is calculated for each individual.

[0035] Furthermore, in step S3, the construction process of the ARIMA time series model includes:

[0036] Perform stationarity tests on data with time series characteristics;

[0037] According to the graphs of the autocorrelation function and partial autocorrelation function of the stabilized data, the autoregressive order, the difference order and the moving average order of the model parameters are determined;

[0038] Use the least squares method to estimate the parameters of the model, and obtain the estimated values ​​of the model coefficients by fitting the model;

[0039] Use the estimated model to predict future time series data.

[0040] In step S5, based on the recurrent neural network (RNN) and its variant, the long short-term memory (LSTM) model, dynamic modeling of health data with time series characteristics is performed, specifically including:

[0041] Map the pre-processed health data collected from multiple channels to specific intervals, and divide the processed data into training sets, validation sets, and test sets;

[0042] Convert time series data into a format suitable for RNN and LSTM model input;

[0043] Introducing memory units and gating mechanisms into the RNN model to construct the LSTM model;

[0044] Train the constructed model;

[0045] Use the loss function to measure the difference between the model prediction results and the actual labels;

[0046] The real-time collected user health data is input into the trained LSTM model. Based on the input time series data, the model outputs the evaluation result of the user's current health status through the state update of the hidden layer and the calculation of the output layer.

[0047] The future health risk assessment results output by the model are analyzed to evaluate the user's health status.

[0048] The beneficial effects of the present invention are: the present invention provides a method for predicting the risk of no major diseases based on multi-dimensional health data analysis, which can collect health data of users in multiple dimensions, and through multi-source data fusion and intelligent analysis, combined with machine learning algorithms and statistical models, construct a high-precision major disease risk prediction model, and dynamically evaluate the user's health status based on the deep learning model. By analyzing the correlation between historical data and future health risks, the user's potential disease risk is predicted, covering health trends, risk factors and improvement suggestions, helping users to formulate scientific health management plans, breaking through the limitations of traditional prediction methods, supporting real-time data updates and model optimization, and providing more accurate and personalized prediction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a flow chart of the prediction method of the present invention. DETAILED DESCRIPTION

[0050] See also Figure 1 As shown, the present invention is about a method for predicting the risk of no major disease based on multi-dimensional health data analysis, comprising the following steps:

[0051] S1: Collect users’ health data from multiple channels and integrate them into a dataset;

[0052] S2: Preprocess the collected health data, extract features using correlation analysis, and then combine the preprocessed and feature-extracted data into a complete feature matrix by column based on a feature splicing fusion method.

[0053] S3: Use the random forest algorithm, combined with the Cox proportional hazards model and the ARIMA time series model, to build a high-precision prediction model for major disease risks and mortality;

[0054] S4: Input the user's real-time multi-dimensional health data into the trained prediction model to predict the probability of the user developing a major disease in the future. Combined with the results of the Cox proportional hazard model and the ARIMA time series model, the model predicts the user's disease risk and mortality trends at different time points in the future, providing users with personalized health warnings and intervention recommendations.

[0055] S5: Based on the recurrent neural network (RNN) and its variant, the long short-term memory (LSTM) model, it dynamically models health data with time series characteristics, and monitors and dynamically evaluates the user's health status in real time.

[0056] S6: Establish a health status scoring system to score the user's health status based on the model's prediction results and changes in the user's health indicators. For example, health status can be divided into three levels: good, fair, and warning. When the user's health status score reaches the warning level, the system automatically sends an alert to the user and relevant medical institutions, reminding the user to take timely health intervention measures.

[0057] Among them, in step S1, the collected data includes physiological indicator data, living habit data, genetic information data, environmental data and medical record data. Before collecting the user's health data, it is first necessary to collect the user's basic information, such as name, gender, age, place of residence and other information. The multi-dimensional health data collected in the present invention includes five categories. The first category is physiological indicator data, which uses wearable devices (such as smart bracelets, smart watches, etc.) to collect the user's heart rate, blood pressure, blood oxygen saturation, sleep quality and other physiological data in real time. The wearable device sends the data to the user's mobile terminal or cloud server in a wireless transmission manner through the sensor; or set up a special health testing site, and the user regularly goes to the site for more comprehensive physiological indicator testing, including blood sugar, blood lipids, liver and kidney function indicators, etc. The testing instrument directly digitizes the results and uploads them to the data management system.

[0058] The second category is lifestyle data, for which a supporting mobile application can be developed. Users can use this application to independently enter information about their daily diet (such as food types, intake, etc.), exercise frequency and duration, smoking and drinking habits, work and rest schedules, and other lifestyle information. The application uses an intuitive interactive interface to facilitate users to record data quickly and accurately; or through cooperation with third-party sports and fitness platforms, diet record platforms, etc., to obtain the lifestyle data recorded by users on these platforms, and realize automatic synchronization and integration of data.

[0059] The third category is genetic information data. By cooperating with professional genetic testing institutions, users can choose to undergo genetic testing to obtain their own genetic information. The genetic testing institutions will send the test results in a standard data format to the database of this system. For existing genetic databases, on the premise of obtaining user authorization, genetic marker data related to disease risk and death can be extracted through data interface docking.

[0060] The fourth category is environmental factor data, which collects meteorological data in the user's area, including temperature, humidity, air quality index (AQI), ultraviolet intensity, etc. Real-time and historical meteorological data can be obtained through the meteorological department's data interface; combined with geographic information system (GIS) technology, it obtains environmental characteristic data of the user's living and working areas, such as whether there are pollution sources and noise levels in the surrounding area, and uses satellite remote sensing data and ground monitoring station data for comprehensive analysis.

[0061] The fifth category is medical records. With user authorization, the system connects to the hospital's information management system (HIS) to obtain the user's outpatient medical records, inpatient medical records, physical examination reports, and other medical records. For examination and test results obtained by the user at other medical institutions, the user can upload them by scanning or manually entering them into the system.

[0062] S2: Preprocess the collected health data, extract features using the correlation analysis method, and use the feature splicing fusion method to splice the preprocessed and feature-extracted data into a complete feature matrix by column.

[0063] After collecting data, preprocessing is required. This involves cleaning the data to remove duplicate, erroneous, and missing data to improve data quality. For missing values, mean or median filling is used, depending on the data type and characteristics. Standardization and normalization are used to convert data of different scales and units into a unified, standardized form to eliminate the impact of different dimensions and facilitate subsequent data analysis and model training. For example, the Z-score standardization method can be used to convert data into a distribution with a mean of 0 and a standard deviation of 1.

[0064] Extract valuable features from the raw data to facilitate subsequent processing. For example, for physiological indicator data, calculate its statistical features, such as mean, variance, maximum, minimum, etc.; for lifestyle habit data, convert it into categorical features or numerical features, such as categorizing smoking habits into categories such as never smoking, occasional smoking, and regular smoking, and assigning corresponding numerical values. Among them, the basis for judging whether it is valuable includes clinical relevance, mainly from the perspective of medical professionals. If there is a correlation between the feature and major diseases in medical theory and clinical practice, it is judged to be valuable; if there is a significant relationship between the feature and disease risk or death through statistical tests, it is judged to be valuable; whether it is valuable is judged based on how much additional information the feature can bring to disease risk and death prediction.

[0065] In step S2, feature extraction is performed using correlation analysis to screen out features that are highly correlated with major disease risks and mortality, remove redundant features, reduce data dimensions, and improve model training efficiency and prediction accuracy. The specific process includes:

[0066] For the cleaned and preliminarily processed multidimensional health data, select the target variable to be analyzed (such as whether a certain major disease is present, usually represented by 0 / 1) and all possible related characteristic variables (such as various physiological indicators, living habits, genetic information, etc.);

[0067] Calculate the correlation coefficient. For numerical features with a normal distribution, calculate the Pearson correlation coefficient, which can be used to measure the degree of linear correlation between two variables. The value range is [-1, 1]. For example, calculate the Pearson correlation coefficient between weight and diabetes risk. If it is close to 1, it indicates that weight and diabetes risk are strongly positively correlated. For data that does not meet the normal distribution or the variable is an ordered categorical variable, calculate the Spearman rank correlation coefficient. The correlation is calculated based on the rank of the data and the value range is also [-1, 1]. For example, the relationship between exercise frequency (low, medium, high) and cardiovascular disease risk;

[0068] Determine the threshold of the correlation coefficient, for example, set the absolute value of the Pearson correlation coefficient greater than 0.3 to indicate a significant correlation. For the chi-square test, set the p value less than 0.05 to indicate a significant correlation;

[0069] The calculated correlation coefficient is compared with a threshold, and features whose correlation coefficient reaches or exceeds the threshold are retained.

[0070] A fusion method based on feature splicing was used to splice the preprocessed and extracted physiological indicator features, lifestyle characteristics, genetic information characteristics, environmental factor characteristics, and medical record characteristics into a complete feature matrix by column. In this feature splicing fusion method, data from different sources were assigned corresponding weights based on their importance to disease risk and mortality prediction. During feature splicing, each component was weighted according to the weights. For example, through a combination of expert experience and data analysis, the weight of genetic information data was determined to be 0.3, the weight of physiological indicator data was 0.3, the weight of lifestyle data was 0.2, the weight of environmental factor data was 0.1, and the weight of medical record data was 0.1. During feature splicing, each component was weighted according to the weights.

[0071] S3: A high-precision prediction model for major disease risk and mortality is constructed using the random forest algorithm, combined with the Cox proportional hazards model and the ARIMA time series model. The specific process of constructing a high-precision prediction model for major disease risk and mortality includes:

[0072] Extract relevant features from the fused data matrix as inputs to the major disease risk and mortality prediction model, such as physiological indicators, lifestyle habits, genetic information, environmental factors, and medical records. Determine the target variable, such as whether a major disease has occurred and when it occurred. Divide the dataset into training and test sets, for example, in a 7:3 ratio, to evaluate model performance.

[0073] Use the random forest algorithm to build multiple decision trees and integrate them;

[0074] Set the specific content of outcome variables and covariates and construct the Cox proportional hazards model;

[0075] Use data with time series characteristics to build an ARIMA time series model;

[0076] The predicted categories output by the random forest algorithm, the risk scores output by the Cox proportional hazards model, and the predicted disease onset time output by the ARIMA time series model are integrated as new features into a new meta-model to produce a high-precision prediction model for major disease risk and mortality. The meta-model can be a logistic regression model or a support vector machine. By training the meta-model, the information from these three models is comprehensively considered to produce the final high-precision prediction model for major disease risk and mortality.

[0077] Use the test set to evaluate the constructed model. Common evaluation metrics include accuracy, recall, F1 score, mean squared error (MSE), and mean absolute error (MAE). Based on the evaluation results, analyze the model's performance and potential problems. Optimize the model by adjusting model parameters, adding or removing features, or changing the model structure. For example, in a random forest, adjust the number of decision trees or the node splitting threshold; in a Cox proportional hazards model, filter or transform features; and in an ARIMA time series model, adjust the differencing order or model order. Repeat the training, evaluation, and optimization process until the model achieves satisfactory performance.

[0078] In step S3, the specific process of constructing and integrating multiple decision trees using the random forest algorithm includes:

[0079] Multiple subsets are randomly sampled from the training set with replacement, each containing a certain number of samples. For each subset, a decision tree is constructed.

[0080] By calculating the information gain of features, we select features that have a significant impact on disease risk and mortality prediction. Random Forest automatically evaluates the importance of each feature and tends to select more important features for splitting when building a decision tree.

[0081] The resulting multiple decision trees are integrated to combine the predictions from each decision tree to obtain the final prediction value. For example, the predictions from each decision tree can be combined by voting or averaging. For classification problems, the final category can be determined based on the classification results from the majority of decision trees. For regression problems, the average of the predictions from all decision trees can be calculated as the final result.

[0082] In step S3, the construction process of the Cox proportional hazards model includes:

[0083] The outcome variable was set as the occurrence or non-occurrence of the disease, and the covariates were the pre-processed data of physiological indicators, lifestyle habits, genetic information, environmental factors, and medical records, and a Cox proportional hazard model was constructed.

[0084] The preprocessed data is divided into a training set and a test set, and the constructed model is trained using the training set data; for example, the occurrence or non-occurrence of cardiovascular disease is set as the outcome variable, and the preprocessed data of physiological indicators (such as blood pressure, blood sugar, blood lipids), lifestyle habits (exercise frequency, dietary structure), genetic information (APOE gene variation), environmental factors (air quality) and medical records (BMI) are used as covariates, and the constructed model is trained using historical data.

[0085] The maximum likelihood estimation method is used to estimate the parameters of the model. Through iterative calculation, the parameter values ​​that maximize the likelihood function are found. These parameters represent the impact of each feature on disease risk and death.

[0086] Based on the estimated parameters, the risk score of each individual is calculated. The risk score is used to assess the risk probability of an individual developing a disease at a certain point in the future. The formula for calculating the risk score is:

[0087] h(t,X)=h0(t)×exp(β1X1+β2X2+...+β p X p ), where h(t, X) represents the risk function of the individual at time t, h0(t) represents the baseline risk function, and β i represents the regression coefficient of the i-th feature, X i Represents the value of the i-th feature.

[0088] In step S3, the construction process of the ARIMA time series model includes:

[0089] For data with time series characteristics, such as the onset time of disease in medical records and the changes in physiological indicators over time, the stationarity test method can use the ADF test method. If the data is not stationary, it needs to be processed by difference to make it stable.

[0090] According to the graphs of the autocorrelation function (ACF) and partial autocorrelation function (PACF) of the stabilized data, the model parameters p (autoregressive order), d (difference order) and q (moving average order) are determined.

[0091] The least squares method is used to estimate the parameters of the model, and the estimated values ​​of the model coefficients are obtained by fitting the model.

[0092] Use the estimated model to predict future time series data. At the same time, according to the changes in actual data and prediction errors, continuously adjust the parameters and structure of the model to improve the accuracy of the prediction.

[0093] S4: Input the user's real-time multi-dimensional health data into the trained prediction model to predict the probability of the user developing a major disease in the future. Combined with the results of the Cox proportional hazard model and the ARIMA time series model, the disease risk and mortality trend of the user at different time points in the future are predicted to provide the user with personalized health warnings and intervention recommendations.

[0094] For example, real-time collection of users' physiological indicators (for example, user A collects physiological data in real time through a smart bracelet. In the past week, the average heart rate is 75 beats / minute, the average systolic blood pressure is 130 mmHg, the average diastolic blood pressure is 85 mmHg, the fasting blood glucose value is 5.5 mmol / L, and the total cholesterol value is 5.0 mmol / L), lifestyle habits (for example, user A records daily diet, exercise, and rest through a mobile phone application. He exercises three times a week for 30 minutes each time, consumes a lot of high-fat foods in his diet, and sleeps an average of 7 hours a day), genetic information (the user undergoes genetic testing, and the results show the presence of APOE gene mutations related to cardiovascular disease), environmental factors (air quality data in the user's area is collected. The average air quality index AQI in the past week was 80, which is at a good level), and medical record data (user A's physical examination report in the past year showed that the electrocardiogram was normal, but the body mass index (BMI was 26, which is overweight).

[0095] The collected health data is cleaned to remove outliers and missing values. For example, if a blood pressure measurement shows a significant error (such as a systolic pressure as high as 300 mmHg), the data point is deleted. Missing blood sugar values ​​are handled using a mean filling method. All data is standardized to make them comparable. For example, physiological indicators such as blood pressure, blood lipids, and blood sugar are standardized to a range of 0-1 to facilitate subsequent model calculations.

[0096] A Cox proportional hazards model was constructed, using the occurrence or absence of cardiovascular disease as the outcome variable and including physiological indicators (such as blood pressure, blood sugar, and blood lipids), lifestyle habits (exercise frequency and dietary structure), genetic information (APOE gene variants), environmental factors (air quality), and medical records (BMI) as covariates. The model was trained using a large amount of historical data (including multidimensional health data of different individuals and information on the occurrence or absence of cardiovascular disease), and the regression coefficients of each covariate in the model were determined using maximum likelihood estimation. For example, the training found that for every 10 mmHg increase in systolic blood pressure, the risk of cardiovascular disease increased by 1.2 times, while for every additional exercise session per week, the risk of disease decreased by 0.8 times.

[0097] User A's physiological indicators (such as blood pressure and heart rate) over a period of time (e.g., the past year) are selected as time series data. This data is tested for stationarity. If the data is not stationary, it is stabilized by performing differential processing. The parameters p, d, and q of the ARIMA model are determined based on the autocorrelation function (ACF) and partial autocorrelation function (PACF). For example, analysis determines that user A's blood pressure time series data is suitable for the ARIMA(1,1,1) model. The ARIMA(1,1,1) model is trained using the past blood pressure data to obtain parameter estimates. This model can be used to predict the user's blood pressure trend over the next period of time.

[0098] User A's real-time, multi-dimensional health data is input into the trained prediction model. The Cox proportional hazards model calculates the probability of User A developing cardiovascular disease within the next year based on User A's real-time data. Assume the probability calculated by the model is 0.15 (i.e., 15%). The ARIMA time series model predicts the trend of User A's blood pressure over the next period of time (e.g., the next three months). The prediction results show that User A's systolic blood pressure may gradually increase over the next three months, with an estimated increase of 5-10 mmHg. Combining the results of the Cox proportional hazards model and the ARIMA time series model, a comprehensive analysis of User A's disease risk and mortality trends at different future time points is conducted. Because elevated blood pressure increases the risk of cardiovascular disease, based on the rising blood pressure trend predicted by the ARIMA model, it is estimated that the user's risk of cardiovascular disease over the next three months may increase from the current 15% to approximately 20%.

[0099] Health warnings and recommendations are made based on the prediction results. If the probability of user A developing cardiovascular disease in the future exceeds a set threshold (such as 10%), a health warning message will be sent to user A. Since user A has a 15% probability of developing cardiovascular disease within the next year, the system will send a text message reminder to the user: "You are currently at a high risk of cardiovascular disease. Please focus on a healthy lifestyle and have regular physical examinations." Given that user A exercises less and consumes a lot of high-fat foods, it is recommended to increase the number of exercises to 4-5 times a week, extend each exercise time to more than 45 minutes, and reduce the intake of high-fat foods and increase the intake of vegetables, fruits, and whole grains. It is also recommended that user A measure blood pressure, blood sugar, and blood lipids every two weeks and undergo an electrocardiogram once a month to promptly detect changes in health indicators. If user A's blood pressure continues to rise and exceeds the normal range, it is recommended that they consult a doctor to consider whether medication is needed to control blood pressure and reduce the risk of cardiovascular disease.

[0100] S5: Based on the recurrent neural network (RNN) and its variant, the long short-term memory (LSTM) model, dynamic modeling of health data with time series characteristics is performed to monitor and dynamically evaluate the user's health status in real time. The specific process includes:

[0101] Preprocessed health data collected from multiple channels is mapped to specific intervals to improve the model's convergence speed and stability. The processed data is then divided into training, validation, and test sets. In this method, 70% of the data is used as the training set, 15% as the validation set, and 15% as the test set. The training set is used for model training, the validation set is used to adjust model hyperparameters, and the test set is used to evaluate the model's final performance.

[0102] Convert time series data into a format suitable for RNN and LSTM model input. Data is typically organized into sequences, with the features of each time step treated as a vector. The length of the sequence is determined based on the actual data and model requirements.

[0103] The traditional RNN consists of an input layer, a hidden layer, and an output layer. The input layer receives the current time step feature vector of the time series data. The hidden layer contains multiple neurons for processing sequence information and retaining historical memory. The output layer outputs the prediction result based on the state of the hidden layer. At each time step, the state of the hidden layer is updated by the following formula: t =δ(W xh x t +W hh h t-1 +b h ), where h t represents the hidden layer state at the current time step t, x t represents the input vector of the current time step, h t-1 represents the hidden layer state of the previous time step, W xh and W hh represents the weight matrix, b h represents the bias vector and δ represents the activation function.

[0104] LSTM is a variant of RNN that introduces memory units and gating mechanisms in the hidden layer to better handle long-term dependencies. The hidden layer of LSTM contains an input gate i t 、Forget Gate t , output gate o t and memory unit c t , at each time step, the update formula is as follows:

[0105] f t =σ(W xf x t +W hf h t-1 +b f )

[0106] i t =σ(W xi x t +W hi h t-1 +b i )

[0107] o t =σ(W xo x t +W ho h t-1 +b o )

[0108]

[0109] h t =o t ☉tanh(c t )

[0110] Among them, W xf , W xi , W xo , W xc , W hf , W hi , W ho , W hc is the weight matrix, b f , b i , b o , b c is the bias vector, ⊙ represents element-wise multiplication. Through these gating mechanisms, LSTM can selectively remember and forget information, thereby better capturing long-term dependencies in time series.

[0111] Train the constructed model, that is, randomly initialize the weight matrix and bias vector of the RNN or LSTM model, such as Xavier initialization or Kaiming initialization, to ensure that the model can converge effectively. Input the training data into the model in batches for forward and backward propagation. During the forward propagation process, the data is gradually propagated from the input layer through the hidden layer to the output layer to obtain the prediction result. The loss between the prediction result and the true label is calculated according to the loss function. During the backpropagation process, the gradient of the loss to the model parameters is calculated by the chain rule, and the parameters are updated according to the update rule of the optimizer. Repeat this process until the loss function converges or the preset number of training rounds is reached. During the training process, the validation set can be used to monitor the performance of the model and adjust hyperparameters such as the learning rate, the number of hidden layer neurons, etc. to prevent the model from overfitting. Use the test set to evaluate the trained model and calculate relevant evaluation indicators such as mean square error (MSE), mean absolute error (MAE), and coefficient of determination R 2etc., to understand the accuracy and reliability of the model's dynamic assessment of the user's health status, and optimize the model based on the assessment results.

[0112] A loss function is used to measure the difference between the model's predictions and the true labels. Examples include mean squared error (MSE) and mean absolute error (MAE). An optimizer is selected to update the model's parameters to minimize the loss function. Examples include stochastic gradient descent (SGD), Adagrad, Adadelta, RMSProp, and Adam. The Adam optimizer is used in this paper because it performs well in practice and can adaptively adjust the learning rate.

[0113] Real-time user health data is fed into a trained RNN or LSTM model. Based on the input time series data, the model updates the hidden layer's state and performs calculations in the output layer to output an assessment of the user's current health status. This result can be a continuous numerical value representing a quantitative indicator of health status, such as disease risk or mortality score, or a categorized result, such as healthy, sub-healthy, or diseased. As new health data is continuously input, the model continuously updates the hidden layer's state, adjusting its assessment of the user's health status in real time. This ensures timely reflection of changes in the user's health status, providing users with dynamic health monitoring and early warning services.

[0114] Analyze the future health risk assessment results output by the model. Historical data can be grouped according to different characteristics or combinations of characteristics, such as age, gender, lifestyle habits and other factors, and then the distribution of future health risks in different groups can be observed. Statistical analysis methods, such as calculating the mean, standard deviation, correlation coefficient, etc., are used to determine the degree of correlation between each factor in the historical data and future health risks. For example, it is found that there is a significant positive correlation between specific lifestyle habits (such as long-term smoking) and a higher risk of disease and death, or that certain physiological indicators (such as high blood pressure) have a strong correlation with the probability of occurrence of specific diseases.

[0115] In step S6, when establishing a health status scoring system, scoring indicators are first determined from the collected health data. For example, physiological indicators can include blood pressure, blood sugar, and heart rate; lifestyle indicators can include exercise frequency, sleep duration, and dietary structure; genetic information includes disease-related gene variants; environmental factors include air quality and noise levels; and medical records include past medical history and physical examination results. After determining the scoring indicators, a hierarchical analysis is performed through expert experience, literature research, or data-driven methods to determine the specific weights of each indicator. Factors with a high correlation with a particular disease are primarily used as scoring indicators. A linear weighted model can be used to establish the health status scoring system, incorporating the prediction results of the aforementioned disease risk and mortality prediction models into the health status scoring system. The system also focuses on how the user's health indicators change over time. If a user's blood pressure continues to rise over a period of time, or if their blood sugar levels change from a normal range to a pre-diabetic trend, this indicates a deterioration in the user's health status, which will be reflected in the score. This change can also be measured by calculating the rate of change of health indicators. Furthermore, because a user's health status is dynamic, the health status scoring system needs to be regularly updated to provide more accurate recommendations.

[0116] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary engineering technicians in this field should fall within the scope of protection determined by the claims of the present invention.

Claims

1. A method for predicting the risk of no major diseases based on multi-dimensional health data analysis, characterized in that: The following steps are involved: S1: Collect users’ health data from multiple channels and integrate them into a dataset; S2: Preprocess the collected health data, extract features using correlation analysis, and then combine the preprocessed and feature-extracted data into a complete feature matrix by column based on a feature splicing fusion method. S3: Using the random forest algorithm, combined with the Cox proportional hazards model and the ARIMA time series model, a high-precision major disease risk prediction model is constructed; S4: Input the user's real-time multi-dimensional health data into the trained prediction model to predict the probability of the user developing a major disease in the future. Combined with the results of the Cox proportional hazards model and the ARIMA time series model, the model predicts the changing trend of the user's disease risk at different time points in the future and provides personalized health warnings and intervention recommendations for the user. S5: Based on the recurrent neural network (RNN) and its variant, the long short-term memory (LSTM) model, it dynamically models health data with time series characteristics, and monitors and dynamically evaluates the user's health status in real time. S6: Establish a health status scoring system to score the user's health status based on the model's prediction results and changes in the user's health indicators.

2. The method for predicting the risk of no major disease based on multi-dimensional health data analysis according to claim 1, characterized in that: In step S1, the collected data includes physiological indicator data, living habit data, genetic information data, environmental data and medical record data.

3. The method for predicting the risk of no major disease based on multi-dimensional health data analysis according to claim 1, characterized in that: In step S2, feature extraction is performed using correlation analysis to screen out features that are highly correlated with major disease risks and remove redundant features. The specific process includes: For the cleaned and preliminarily processed multidimensional health data, select the target variable to be analyzed and all possible related feature variables; Calculate the correlation coefficient. For numerical features with normal distribution, calculate the Pearson correlation coefficient; for data that do not meet the normal distribution or the variables are ordered categorical variables, calculate the Spearman rank correlation coefficient; Determine the threshold value of the correlation coefficient; The calculated correlation coefficient is compared with a threshold, and features whose correlation coefficient reaches or exceeds the threshold are retained.

4. The method for predicting the risk of no major disease based on multi-dimensional health data analysis according to claim 3, characterized in that: In step S2, in the feature splicing fusion method, data from different sources are assigned corresponding weights according to their importance to disease risk prediction, and when splicing features, each part of the features is weighted according to the weights.

5. The method for predicting the risk of no major disease based on multi-dimensional health data analysis according to claim 1, characterized in that: In step S3, the process of building a high-precision major disease risk prediction model includes: From the feature matrix after data fusion, relevant features are extracted as input to the major disease risk prediction model, the target variable is determined, and the dataset is divided into a training set and a test set; Use the random forest algorithm to build multiple decision trees and integrate them; Set the specific content of outcome variables and covariates and construct the Cox proportional hazards model; Use data with time series characteristics to build an ARIMA time series model; The predicted category output by the random forest algorithm, the risk score output by the Cox proportional hazard model, and the predicted value of the disease occurrence time output by the ARIMA time series model are used as new features and input into a new meta-model for fusion to obtain the final high-precision major disease risk prediction model.

6. The method for predicting the risk of no major disease based on multi-dimensional health data analysis according to claim 5, characterized in that: In step S3, the specific process of constructing and integrating multiple decision trees using the random forest algorithm includes: Randomly extract multiple subsets from the training set with replacement, and build a decision tree for each subset; By calculating the information gain of features, features that have an important impact on disease risk prediction are selected; The generated multiple decision trees are integrated, and the prediction results of each decision tree are combined to obtain the final prediction value.

7. The method for predicting the risk of no major disease based on multi-dimensional health data analysis according to claim 5, characterized in that: In step S3, the construction process of the Cox proportional hazards model includes: The outcome variable was set as the occurrence or non-occurrence of the disease, and the covariates were physiological indicators, lifestyle habits, genetic information, environmental factors, and pre-processed data of medical records, and a Cox proportional hazard model was constructed; The preprocessed data is divided into a training set and a test set, and the constructed model is trained using the training set data; Use the maximum likelihood estimation method to estimate the parameters of the model, and find the parameter value that maximizes the likelihood function through iterative calculation; Based on the estimated parameters, a risk score is calculated for each individual.

8. The method for predicting the risk of no major disease based on multi-dimensional health data analysis according to claim 5, characterized in that: In step S3, the construction process of the ARIMA time series model includes: Perform stationarity tests on data with time series characteristics; According to the graphs of the autocorrelation function and partial autocorrelation function of the stabilized data, the autoregressive order, the difference order and the moving average order of the model parameters are determined; Use the least squares method to estimate the parameters of the model, and obtain the estimated values ​​of the model coefficients by fitting the model; Use the estimated model to predict future time series data.

9. The method for predicting the risk of no major disease based on multi-dimensional health data analysis according to claim 1, characterized in that: In step S5, based on the recurrent neural network (RNN) and its variant, the long short-term memory (LSTM) model, dynamic modeling of health data with time series characteristics is performed, specifically including: Map the pre-processed health data collected from multiple channels to specific intervals, and divide the processed data into training sets, validation sets, and test sets; Convert time series data into a format suitable for RNN and LSTM model input; Introducing memory units and gating mechanisms into the RNN model to construct the LSTM model; Train the constructed model; Use the loss function to measure the difference between the model prediction results and the actual labels; The real-time collected user health data is input into the trained LSTM model. Based on the input time series data, the model outputs the evaluation result of the user's current health status through the state update of the hidden layer and the calculation of the output layer. The future health risk assessment results output by the model are analyzed to evaluate the user's health status.

Citation Information

Cited By

  • Neuroimmune disease risk prediction method based on artificial intelligence

    CN121964158A