| Background:Chronic non-communicable diseases(NCDs)are the leading cause of death globally,and their incidence continues to rise with rapid economic development and the resulting profound changes in lifestyles among the 56 million deaths worldwide in 2012,38 million(68%)were due to non-communicable diseases.More than 40%(16million)of them died prematurely under the age of 70.Of the 38 million deaths from non-communicable diseases,16 million(42%)were premature and avoidable.It is estimated that 80%of the deaths from non-communicable diseases are caused by cancer,cardiovascular diseases,chronic respiratory diseases and diabetes.Among them,cardiovascular disease,as a common chronic disease,is the main cause of premature death in Chinese people.Among the changeable risk factors of cardiovascular and cerebrovascular diseases,the top three are"three highs",hypertension,diabetes and dyslipidemia.Controlling the"hypertension、diabetes、dyslipidemia"plays an important role in the prevention and treatment of cardiovascular diseases.The lack of chronic disease management is a major difficulty in health management services in China at present,which limits the smooth development of chronic disease management to a certain extent.Therefore,it is of great significance for the prevention and treatment of chronic diseases to build a prediction model of common chronic diseases represented by"hypertension、diabetes、dyslipidemia"in the physical examination population and establish risk management and control strategies for different risk groups through the model.The physical examination platform in general hospitals is an important basis and information source for the risk assessment of chronic diseases in healthy population.But at present,the physical examination data items of different scale medical institutions are different,most of them are inclined to health evaluation,and lack of data items of chronic disease risk assessment.More importantly,healthy individuals can not get a clear assessment of chronic disease risk after participating in physical examination.At the same time,many clinical prediction models only focus on model development and lack of external verification,which leads to the inability to effectively use the models in clinical practice.In order to improve the effectiveness of prevention and control of common chronic diseases,the physical examination information and risk information of chronic diseases should be deeply excavated through the physical examination platform of general hospitals,the prediction model of chronic diseases should be constructed,externally validated,and the stratified management strategy of risk groups should be explored,so as to realize the effective association between the physical examination information of general hospitals and healthy individuals after taking physical examination and the management and control of chronic diseases.Objectives:1.To understand the prevalence and trend of three common chronic diseases,hypertension,diabetes and dyslipidemia,in order to provide scientific basis for the research of prediction models and control strategies of these three diseases.2.Based on Harvard Cancer Index,synthetic analysis,machine learning and other statistical modeling methods,this paper studies a statistical model-based chronic disease risk level prediction tool for health check-up population.3.Based on the analysis of PEST model,this study provides a theoretical basis for the prediction model applied to chronic disease health management,and then puts forward strategies and suggestions for the stratified management and control of common chronic diseases in health check-up population.Methods1.This study was based on a longitudinal cohort of healthy physical examinees in a physical examination center of a general hospital in Liaoning Province from April 2016to September 2019,and information was collected through physical examination,laboratory tests and questionnaires.2.Using Harvard Cancer Index,the risk grade index of hypertension,diabetes and dyslipidemia was constructed to establish the clinical prediction model.The prediction performance of the model was evaluated by studying the direct correlation between the risk grade of three diseases and the actual incidence,and the risk grade of chronic diseases was evaluated.3.Using the synthetic analysis method,the risk assessment model of hypertension,diabetes and dyslipidemia was constructed,and the area under the ROC curve was used as the main evaluation index to compare the predictive efficacy of the synthetic analysis method and the traditional Logistic regression model.4.Five machine learning methods were used to construct the risk assessment models of hypertension,diabetes and dyslipidemia,and the dominant models were selected to draw the nomogram to realize the visualization of the models.5.Using the PEST model to analyze the external factors of the chronic disease prediction model,the population stratified management and control strategy is developed for the risk level evaluation advantages of the model.ResultsPart I Results1.The prevalence rates of hypertension,diabetes and hyperlipidemia were 25.52%,27.6%and 36.46%respectively.The prevalence rates of hypertension,diabetes and dyslipidemia were 31.69%,34.35%and 49.46%respectively in male subjects and18.12%,20.34%and 22.50%respectively in female subjects.There were significant differences between the sexes(P<0.001).2.The results showed that the prevalence rates of hypertension and diabetes increased significantly with the increase of age(P<0.001),and reached the peak in the old age;The prevalence of dyslipidemia increased with age(P<0.001),but the highest prevalence was 42.94%in middle age.3.To compare the prevalence of the three diseases in different BMI groups.The prevalence rates of hypertension,diabetes and dyslipidemia were significantly increased with the increase of BMI(P<0.001).The prevalence rates of hypertension,diabetes and dyslipidemia were the highest in the obese population,which were 44.73%,46.86%and59.84%respectively.The prevalence rate of the three diseases in thin subjects was the lowest,which was 4.76%,5.74%and 5.98%respectively,which was far lower than that of the obese people.Part II Results1.Results of the Harvard Cancer Index Prediction Model(1)Age,BMI and family history of hypertension were selected to construct the risk index of hypertension.The results showed that the risk scores of high sodium diet,overweight and obesity,excessive drinking,mental stress,age,family history of hypertension,lack of physical activity,diabetes and dyslipidemia were 5,5,5,10,10,5,0,10 and 5,respectively.The average score of hypertension was 6.11.In order to validate and evaluate the hypertension risk grade index,this study analyzed the number of people with hypertension in different risk grade.The results showed that in the hypertension cohort,the number of people with hypertension was very low(20.48%),lower(28.55%),high(28.19%),higher(22.50%),and very high(0.3%)according to the risk grade of hypertension,and there were 223(20.48%),311(28.55%),307(28.19%),245(22.50%),and 3(0.3%)respectively.The proportion of patients with hypertension was different in different risk levels(?~2=243.56,P<0.001).Chi-square test for trend showed that the incidence of hypertension increased with the increase of risk levels(P<0.001).(2)Age,BMI and hypertension were selected to construct the risk index of diabetes.The results showed that the risk scores of age,overweight and obesity,hypertension,hyperlipidemia,family history of diabetes,lack of physical activity and infrequent consumption of vegetables and fruits were 10,0,5,5,0,0,5,respectively.The average score of diabetes was 1.14.In order to validate and evaluate the constructed risk grade index of diabetes,this study statistically analyzed the number of people with diabetes in different risk grades,and the results showed that in the cohort with diabetes mellitus,493people(34.69%)were judged as low,472(33.22%)as high,and 456(32.09%)as very high by the risk grade of diabetes mellitus.The proportion of patients with different risk levels of diabetes was different(?~2=28.42,P<0.001).Chi-square test for trend showed that the incidence of diabetes increased with the increase of risk levels of diabetes(P<0.001).(3)Using Harvard Cancer Index,the risk index of dyslipidemia was constructed by selecting 9 risk factors,such as age,sex,BMI,physical activity,smoking,drinking,hypertension,diabetes and high fat diet.The results showed that the risk scores of the 9risk factors were 10,-5,5,0,10,10,5,5,10,respectively.The average score of diabetes was 12.25.In order to validate and evaluate the constructed risk grade index of dyslipidemia,this study statistically analyzed the number of people with dyslipidemia in different risk grades,and the results showed that in the dyslipidemia incidence cohort,164 people(21.19%)had lower risk grade,255people(32.95%)had low risk grade,190 people(24.55%)had high risk grade,and 165 people(21.31%)had higher risk grade.The proportion of patients with different risk levels of dyslipidemia was different(?~2=23.22,P<0.001).Chi-square test for trend showed that the incidence of dyslipidemia increased with the increase of risk levels of dyslipidemia(P<0.001).2.Predicted results of the synthetic analysis model(1)The risk prediction model of hypertension was constructed by taking age,high sodium diet,BMI,drinking,mental stress,family history of hypertension,lack of physical activity,diabetes and dyslipidemia as study variables.Logit P=Logit P=-1.95+0.72×Age+0.21×High Sodium Diet+0.23×BMI+0.20×Drinking+0.44×Mental Stress+0.32×Family History of Hypertension-0.04×Physical Activity+0.21×Diabetes+0.04×Dyslipidemia.There was no significant difference(P>0.05)between the performance of the composite analysis model and the traditional Logistic regression model.(2)The risk prediction model of diabetes mellitus was constructed by taking age,BMI,hypertension,dyslipidemia,family history of diabetes mellitus,lack of physical activity and infrequent consumption of fruits and fruits as study variables.Logit P=-1.22+0.43×Age+0.01×BMI+0.11×Hypertension+0.15×Dyslipidemia+0.04×Family History of Diabetes Mellitus-0.012×Physical Activity+0.18×Infrequent Consumption of Fruits and Vegetables.There was no significant difference(P>0.05)between the performance of the composite analysis model and the traditional Logistic regression model.(3)The risk prediction model of dyslipidemia was constructed by taking age,sex,BMI,physical activity,smoking,drinking,hypertension,diabetes and high-fat diet as the study variables.Logit P=-1.13+0.41×age+0.14×sex+0.42×BMI-0.03×physical activity+0.8×smoking+0.97×drinking+0.03×hypertension+0.04×diabetes+0.8×high-fat diet.There was no significant difference(P>0.05)between the performance of the composite analysis model and the traditional Logistic regression model.3.Results of five machine learning prediction models(1)In the construction of hypertension machine learning model,15 characteristic variables including age,pulse,hemoglobin concentration,platelet count,serum uric acid,alanine aminotransferase,alkaline phosphatase,γ-glutamyltransferase,total protein,total bilirubin,aspartate aminotransferase,creatinine,D3,BMI and albumin were selected through variance selection and correlation coefficient method.The predictive performance of the 5 models in the test set was as follows:the AUC result of SVM was0.634;The AUC result for MLP was 0.608;The AUC result for RF was 0.834;The AUC result for XGB was 0.666;The AUC result for Voting was 0.791.By drawing nomograms of 15 characteristic variables,the probability of individual hypertension risk can be calculated.(2)In the construction of diabetes machine learning model,11 characteristic variables,such as age,pulse,hemoglobin concentration,platelet count,uric acid,alanine aminotransferase,alkaline phosphatase,γ-glutamyltransferase,creatinine,SBP and DBP,were selected by variance selection and correlation coefficient method.The predictive performance of the 5 models in the test set was as follows:the AUC result of SVM was0.612;The AUC result for MLP was 0.657;The AUC result for RF was 0.885;The AUC result for XGB was 0.707;The AUC result for Voting was 0.829.The RF model was selected as the dominant model,and the individual probability of diabetes risk was calculated by drawing nomograms of 11 characteristic variables.(3)In the establishment of dyslipidemia machine learning model,age,pulse,hemoglobin concentration,platelet count,serum uric acid,alanine aminotransferase,alkaline phosphatase,γ-glutamyltransferase,aspartate aminotransferase,creatinine,SBP and DBP were selected by variance selection and correlation coefficient method.The predictive performance of the 5 models in the test set was as follows:the AUC result of SVM was 0.513;The AUC result for MLP was 0.550;The AUC result for RF was 0.589;The AUC result for XGB was 0.543;The AUC result for Voting was 0.579.By drawing nomograms of 12 characteristic variables,the probability of individual risk of dyslipidemia can be calculated.(4)The results of external validation of the model show that the evaluation indexes of the dominant model RF of the three diseases in two independent external validation data sets are similar to those in the internal test set,which suggests that the prediction model of hypertension,diabetes and dyslipidemia of the physical examination population in general hospitals constructed by machine learning RF method has good generalization and transplantation.Part III Results1.According to the morbidity probability of individuals obtained by the optimal model RF,the population was divided into three percentiles,and the survival analysis of different risk probability groups was statistically different.The high risk group had the highest morbidity,while the low risk group had the lowest morbidity.2.The PEST model analyzes and summarizes the policy environment,economic environment,social environment and technical environment of applying chronic disease prediction model to chronic disease health management,summarizes the existing problems and opportunities under the background of artificial intelligence and big data,and puts forward corresponding strategic suggestions.3.In the application of the model to the management and control of the risk population in general hospitals,different management and intervention strategies for the high-risk population,medium-risk population and low-risk population of chronic diseases are proposed,and different management and control strategies at different levels have the advantages of optimization of the model.Conclusions1.The prevalence rate of hypertension,diabetes and dyslipidemia increased with age.The prevalence rate of male was higher than that of female,and was positively correlated with BMI index.The prevalence rate of dyslipidemia reached the highest value in middle-aged people,which should be paid attention to.2.According to different statistical methods,several risk prediction models of hypertension,diabetes and dyslipidemia were constructed from different angles.In the process of evaluating chronic diseases,according to the specific situation of research institutes and data,the machine learning stochastic forest algorithm,Harvard Cancer Index method and synthetic analysis method were applied together to improve the risk assessment quality of chronic diseases among healthy people.3.The Harvard Cancer Index method is used to construct the risk grade index of hypertension,diabetes and dyslipidemia to establish the clinical prediction model.This method is suitable for the case of lack of partial information or the need for preliminary classification of the risk grade of chronic diseases.4.The risk prediction model of hypertension,diabetes mellitus and dyslipidemia established by synthetic analysis method is also suitable for the absence of some characteristic factors or the establishment of a more accurate"all-factor"chronic disease risk model based on the existing mature models,which can avoid the waste of previous research data.5.The risk models of hypertension,diabetes and dyslipidemia were constructed by using five machine learning methods.The performance of RF model was the best,which was superior to SVM,MLP,XGBoost and Voting method.The method of machine learning is suitable to be used in the case of many feature variables,and it is easy to choose the model with higher prediction efficiency because there are many algorithms to choose6.The RF model,the dominant model of three chronic diseases,is externally verified by two independent external data sets,It is confirmed that the machine learning model can be applied to general hospitals to predict the risk of common chronic diseases of physical examination population,The nomograms is used to visualize the risk of three diseases under the machine learning advantage model RF model selected in this study,which realizes the rapid assessment of individual risk in hospitals,and thus takes accurate intervention and control measures for chronic diseases.7.The population is grouped into three percentiles according to the prediction probability of individuals obtained by the optimal model RF,The results of survival analysis show that the incidence probability of high,medium and low risk groups is statistically different among the three chronic diseases,which suggests that the machine learning model established in this study can accurately stratify the risk of chronic diseases of healthy people and provide path support for the implementation of accurate management and control.8.The PEST model systematically analyzes the external environment,existing problems,opportunities and strategy suggestions of the application of the chronic disease prediction model in the health management of chronic diseases.It is suggested that the application of the model will play an important role in the health management of the physical examination population in general hospitals in the whole life cycle.9.The prediction model constructed in this study can achieve stratified control for the risk of chronic diseases in physical examination population.At the same time of hierarchical management,it can optimize the model,make use of health examination data resources efficiently,and save medical costs. |
PDF Full Text Request