Application Of Convolutional Neural Network Model In Risk Assessment Of Dyslipidemia In Steel Workers

Objectives A risk assessment model of dyslipidemia convolutional neural network for steel workers was established based on deep learning theory.At the same time,the performance of risk assessment model was compared with that of traditional machine learning models for dyslipidemia,so as to provide basis for the early prevention of dyslipidemia in steel workers.Methods By using the method of current situation investigation,the front-line employees of the main departments of Tangshan Steel Group who participated in the occupational health examination during 2017-2018 were selected as the research objects.We collected information on their physical examination results,laboratory biochemical examination indicators,questionnaire surveys,and occupational exposure to harmful factors.According to the test results of blood lipid indexes,the steel workers were divided into dyslipidemia group and non-dyslipidemia group based on whether the steel workers were tested for abnormal blood lipids.Then used single factor analysis,multi-factor analysis,RCS model to screen out the factors affecting steel workers’ dyslipidemia,and determine the final model input variables based on expert opinions.We established a convolutional neural network model according to the 6: 2: 2 ratio of the training set,test set,and verification set,used the sklearn library for data regularization and data segmentation,the Keras library for training the convolutional neural network,and pyplot in matplotlib the sub-library performs visual display of model results.At the same time,we compared the performance of convolutional neural network and logistic regression model,BP neural network model,support vector machine model and random forest model.Results 1 According to the inclusion and exclusion,a total of 6857 steel workers were included in the study.The detection rate of dyslipidemia was 38.68%,including 13.46% of hypercholesterolemia,17.24% of hypertriglyceridemia,16.23% of low HDL-C and14.29% of high LDL-C.Single factor analysis showed that dyslipidemia positively associated with gender,age,marital status,BMI,fruits and vegetables class diet score,score lipid diet,smoking index,drink tea,alcohol intake,physical activity,family history of hyperlipidemia fat disease,high blood pressure,diabetes,kidney dysfunction,abnormal liver function,high uric acid hematic disease,CO and high temperature cumulative exposure measurement(CEM),Shift years.(P<0.05).After adjusting for related confounding factors,there was a dose-response relationship between dyslipidemia and shift years,high temperature CEM.Multi-factor analysis showed that overweight and obesity,smoking index rise,score higher lipid diet,diabetes,high blood pressure,abnormal liver function,high uric acid hematic disease,shift years and high temperature CEM increase are the risk factors of dyslipidemia,fruits and vegetables diet score rise,small amounts of alcohol and physical activity are the protection factors of dyslipidemia.2The accuracy of the convolutional neural network model in the training set,test set and verification set is 92.56%,91.10% and 88.40% respectively;The goodness-of-fit test results were 0.939,0.922 and 0.927,respectively.In the training set,test set and validation set,the sensitivity were 91.41%,92.47% and 88.31%,the specificity were 93.28%,90.16%and 88.46%,the Youden index were 0.85,0.83 and 0.77,the positive likelihood ratio were13.59,9.40 and 7.65,the negative likelihood ratio were 0.09,0.08 and 0.13,and the F1 score were 0.90,0.89 and 0.85,the AUC(95%CI)were 0.923(0.915～ 0.931),0.913(0.897～ 0.928)and 0.884(0.866～ 0.900),the coincidence rates were 92.56%,91.10% and88.40%,the Kappa values were 0.84,0.82 and 0.76,the positive predictive values were89.37%,86.58% and 82.47%,and the negative predictive values were 94.62%,94.58% and92.49%,respectively.3 The results showed that the accuracy(specificity,sensitivity,Youden index,positive likelihood ratio,negative likelihood ratio,F1 score,AUC),reliability(Kappa values,coincidence rate)and benefit evaluation(positive predictive values and negative predictive values)of convolutional neural network model for dyslipidemia in steel workers were better than logistic regression model,BP neural network model,support vector machine model and random forest model in training set,test set and validation set.Conclusions 1 Overweight and obesity,smoking index rise,score higher lipid diet,diabetes,high blood pressure,abnormal liver function,high uric acid hematic disease,shift years and high temperature CEM increase are the risk factors of dyslipidemia,fruits and vegetables diet score rise,small amounts of alcohol and physical activity are the protection factors of dyslipidemia.2 The convolutional neural network risk assessment model of steel workers with dyslipidemia is superior to the traditional machine learning model in terms of authenticity,reliability and benefit evaluation,and can accurately evaluate the risk of steel workers with dyslipidemia.Figure 14;Table 37;Reference 178...