Establishment Risk Predicted Models Of Childhood Leukemia In Henan Province

Childhood leukemia has become a major disease that endangers the lives of children.It not only poses a threat to children's health,but also brings economic burdens and mental pressure to the society and families.However,childhood leukemia is preventable and treatable.It is easier to remit than adult leukemia after standardized treatment.Therefore,a risk assessment model for screening and monitoring the high-risk children will be helpful to detect,diagnose and treat the children with leukemia earlier.Up to date several risk prediction models for other tumors have been established domestic and abroad,but there is still no one about childhood leukemia.The establishment of the index system is one of the keys of the prediction model.And as to the special population of childhood leukemia,the occurrence of the disease is not only related to the children's harmful environmental exposures,but also with their maternal and parental exposures during the prenatal and the breast-feeding periods;In addition,the risk factors in different regions are also different.ObjectiveThe aim of this study is to screen the appropriate predictive factors to establish the index system of prediction model,and then to establish and validate the risk prediction models of childhood leukemia in Henan Province,in order to provide scientific evidence for identifying high-risk individuals and formulating primary prevention strategies of childhood leukemia.Methods1.According to the case-control study design,cases diagnosed with bone marrow histopathology were recruited from four hospitals of Henan Province during2014-2016,2017-2018 respectively;Meanwhile,controls with non-leukemia were selected from respiratory ward,digestive ward,and Chinese medicine rehabilitation ward in the same hospitals by 1:1 proportion with cases.The unified questionnaire was finished by children's parents or other guardians through face-to-face interviews.The questionnaire included children as well as their parental socio-demographic characteristics,potential risk factors associated with childhood leukemia.2.The case-control data collected during 2014-2016 were used to analyze the risk factors by non-conditional Logistic regression,and the percentage of population attributable risk?PARP?was calculated.Combined with the results of OR?Odds Ratio?value and its 95%confidence interval,PARP values,results from literatures and experts'opinions,the index system of this prediction model was established.3.The total samples collected during 2014-2018 were randomly divided into establishment samples,validation samples according to the proportion of 3:1.Artificial Neural Network?ANN?and Logistic regression were used to establish risk prediction models.Then the prediction efficiency was assessed by the relevant indicators of discrimination and calibration.4.All the statistical analysis was performed with IBM SPSS Statistics 21.0 and R 3.3.3 software.Results1.Sample recruitment:During 2014-2016,407 cases and 407 controls which met the criteria were recruited.During 2016-2018,120 cases and 120 controls were recruited.According to the proportion of 3:1,the subjects were randomly divided into establishment samples of 790 cases and validation samples of 264 cases.2.Estabilshment the index system of prediction model:According to the result of OR and PARP,combined with the results of literatures and experts'opinions,15indicators were finally determined as input variables of the model,including child-related factors(X₁ birth weight,X₂ delivery mode,X₃ household per capita annual income,X₄ exposure to air pollution around the residence,X₅ exposure to extremely low-frequency electromagnetic fields in the living environment,X₆ fever history,X₇ family history of cancer,X₈ exposure to pesticides,X₉ regular ventilation in the living room,X₁₀ regular consumption of unhealthy food history),maternally-related factors(X₁₁ often have physical activity,X₁₂ regular folic acid intake,X₁₃ housing renovation history,X₁₄ exposure to pesticides)and paternally-related factor(X₁₅ dye hair history).3.Establishment and validation the predtction models:In the process of establishing model,when the ANN model gradually included prediction indicators,the results showed that when the number of nodes in the hidden layer was 2,the root-mean-square error of the cross validation was the smallest?0.4127?,and the area under the receiver operating characteristic curve?AUC?value was the largest?0.864?.The best critical point of the model was 0.3365,sensitivity was 89.9%,specificity was 61.9%,the goodness of fit test X²=8.789,P=0.773?P>0.05 was thought that the model fitted the data,the calibration degree of the model was better?;The logistic regression model showed that a total of 11 indicators were included in theequation,thepredictionmodelwasP=1/(1+exp[-(1.151+1.313X₈+1.214X₆+1.073X₁₅+0.953X₄+0.941X₇+0.837X₁₃+0.835X₅+0.722X₁₄+0.632X₁₀-0.683X₃-0.734X₁₂)]),the model's AUC was 0.842,the best critical point of the model was 0.3252,sensitivity was 89.6%,specificity was 55.9%,the goodness of fit testX²=4.856,P=0.361.After validated the prediction models,ANN model's AUC?0.878?,sensitivity?89.3?,specificity?79.6?and Youden's index?0.648?were all superior to Logistic regression model's AUC?0.847?,sensitivity?85.2?,specificity?58.5?and Youden's index?0.478?,in addition,the difference of the AUC was statistical significance?P=0.042?.After the two models were tested for goodness of fit,the calibration of the ANN model?P=0.572?was superior to the Logistic regression model?P=0.359?.ConclusionTwo risk predicted models of childhood leukemia in Henan Province were established using methods of ANN and Logistic regression respectively.The prediction efficiency of the ANN model was superior to the Logistic regression model.