Construction And Validation Of A School-based Early Screening Model For Children With Attention Deficit Hyperactivity Disorder

Object:A school-based screening model for children with Attention deficit hyperactivity disorder(ADHD)was developed using a machine learning approach.The aim is to use modern information technology to rapidly detect and identify ADHD early,to improve the sensitivity and specificity of the original screening methods,to reduce the consumption of medical resources,and to maximize the efficiency of screening children with ADHD.Methods:1、A total of 11,056 students from 8 primary schools in a region of Xinjiang were randomly selected by the whole group sampling method,and 447 cases in the ADHD group were screened while 1341 cases in the normal control group were matched 1:3 according to the same age and sex.2、Using the questionnaire method,parents of both groups of children completed the parent version of the SNAP-IV scale(Swanson,Nolan,and Pelham IV Rating Scale,SNAP-IV),the Strengths and Difficulties Questionnaire(SDQ),the School Age Behavior Rating Scale of Executive Function(BRIEF),and the Weiss Functional Impairment Scale(WFIRS).The classroom teachers of both groups completed the teacher versions of the SNAP-IV,SDQ,and BRIEF scales.Using independent sample t-tests to assess the differences between teacher and parent assessed symptoms,behavior problems,executive function,and functional deficits of children with ADHD and normal children.3、After feature selection using correlation analysis and screening out factors with significant relationships.The random forest model,the support vector machine model,and the e Xtreme Gradient Boosting(XGBoost)model were built using Python software,and the Bootstrap method was used for internal validation of the model.The model screening performance was judged by comparing the area under the ROC curve,accuracy,sensitivity,specificity,and other indicators.Results:1、A total of 11,056 people were screened in this study,from which 447 positive cases of ADHD were detected,for a screening positivity rate of 4.04%.The positivity rate was greater in boys than in girls,with statistically significant differences(X~2=19.008,P<0.001).The differences in the prevalence of ADHD in each grade were statistically significant(X~2=83.131,P<0.001).2、The results of the independent sample t-test showed that the ADHD group and the normal control group had significant differences in all factors of the SNAP-IV,SDQ,BRIEF,and WFIRS scales in the parent and teacher versions,and only the pro-social behavior scores in the SDQ scale were lower in the ADHD group than in the normal control group;the scores of the remaining factors were higher than in the normal control group(P<0.001).3、The correlation analysis indicated that all the variables were related to the diagnostic variable.There were 14 factors with Pearson correlation coefficients greater than or equal to 0.6(P<0.001):P1(0.74),P2(0.76),P3(0.67),P4(0.79),P7(0.64),P8(0.63),P9(0.64),P23(0.63),P24(0.61),P28(0.61),T1(0.74),T2(0.71),T3(0.64),and T4(0.77).4、The area under the ROC curve for the random forest model is 0.957,with an accuracy of 0.950,a precision of 0.934,a sensitivity of 0.884,a specificity of 0.971,and an F1 of 0.900.The area under the ROC curve for the support vector machine model is 0.945,with an accuracy of 0.948,a precision of 0.908,a sensitivity of 0.884,a specificity of 0.970,and an F1 of 0.896.The area under the ROC curve for the XGBoost model is 0.952,with an accuracy of 0.951,a precision of 0.909,a sensitivity of 0.893,a specificity of 0.979,and an F1 of 0.901.5、According to the XGBoost importance analysis,the relative importance ranking of the top ten is P4>T4>P2>T1>P1>P9>T9>T2>P8>P7.Conclusion:1、In this study population,the ADHD screening rate was 4.04%,slightly lower than the national average.2、Total scores on the SNAP-IV parent and teacher versions of the scale,scores on the inattentive and hyperactive-impulsive dimensions of the SNAP-IV parent and teacher versions,total scores on the teacher version of the SDQ scale for difficulties,and total scores on the parent version of the SDQ scale for difficulties,hyperactivity problems,and peer interaction problems were the main influencing factors of the screening model.3、The XGBoost model has higher specificity,sensitivity,accuracy,and F1 than the random forest model and support vector machine model,and the comparison concluded that the XGBoost model has better efficacy,so it is recommended to be used as a model for early screening of children with ADHD in the future and promoted for application within schools.