Development of a machine learning-based predictive nomogram for screening children with juvenile idiopathic arthritis: a pseudo-longitudinal study of 223,195 children in the United States

Abstract

Background: Juvenile idiopathic arthritis (JIA) is a prevalent chronic rheumatological condition in children, with reported prevalence ranging from 12. 8 to 45 per 100,000 and incidence rates from 7.8 to 8.3 per 100,000 person-years. The diagnosis of JIA can be challenging due to its symptoms, such as joint pain and swelling, which can be similar to other conditions (e.g., joint pain can be associated with growth in children and adolescents).

Methods: The National Survey of Children's Health (NSCH) database (2016-2021) of the United States was used in the current study. The NSCH database is funded by the Health Resources and Services Administration and Child Health Bureau and surveyed in all 50 states plus the District of Columbia. A total of 223,195 children aged 0 to 17 were analyzed in this study. A least absolute shrinkage and selection operator (LASSO) logistic regression and stepwise logistic regression were used to select the predictors, which were used to create the nomograms to predict JIA.

Results: A total of 555 (248.7 per 100,000) JIA cases were reported in the NSCH. In the LASSO model, the receiver operating characteristic curve demonstrated excellent discrimination, with an area under the curve (AUC) of 0.9002 in the training set and 0.8639 in the validation set. Of the 16 variables selected by LASSO, 13 overlapped with those from the stepwise model. The regression achieved an AUC of 0.9130 in the training set and 0.8798 in the validation set. Sensitivity, specificity, and accuracy were 79.1%, 90.2%, and 90.2% in the training set, and 69.0%, 90.9%, and 90.8% in the validation set.

Discussion: Using two well-validated predictor models, we developed nomograms for the early prediction of JIA in children based on the NSCH database. The tools are also available for parents and health professionals to utilize these nomograms. Our easy-to-use nomograms are not intended to replace the standard diagnostic methods. Still, they are designed to assist parents, clinicians, and researchers in better-estimating children's potential risk of JIA. We advise individuals utilizing our nomogram model to be mindful of potential pre-existing selection biases that may affect referrals and diagnoses.

Keywords: LASSO; NSCH; chronic rheumatology; juvenile idiopathic arthritis; machine learning; nomogram; pediatric arthritis; pediatric joint inflammation.

Figure 1

Identification of the optimal penalization coefficient λ in the LASSO logistic regression. (A) The LASSO coefficient profiles of the 22 variables. Child's age, sex, race, low birth weight, BMI, having a genetic or inherited condition identified through a blood test, anxiety, asthma, allergy to food, drug, or insect, Type 1 Diabetes, heart condition, household's ability to afford the food you need during the past 12 months, chronic physical pain, difficulty with eating or swallowing in the past 12 months, adequacy of current insurance coverage, and child with a personal doctor or nurse were selected using LASSO binary logistic regression analysis. The LASSO coefficient profiles of the features were plotted. (B) The optimum parameter (lambda) selection in the LASSO model performed 10-fold cross-validation through minimum criteria. The partial likelihood deviance (binomial deviance) curve was presented versus log (lambda). Dotted vertical lines were shown at the optimum values by performing the lambda.min (red) and the lambda.1se (blue).

Figure 2

Nomograms for predicting JIA developed by LASSO logistic and logistic regression analysis. (A) Nomogram developed by LASSO logistic regression to predict JIA; (B) nomogram developed by logistic stepwise regression to predict JIA.

Figure 3

ROC curves illustrating the capability in predicting JIA. (A, B) are the result of LASSO logistic regression. (C, D) are the result of logistic regression.

Figure 4

Calibration plots of the binary fringe plot with 1,000 bootstrapping re-sample of LASSO logistic regression for JIA. (A, B) are the result of LASSO logistic regression. (C, D) are the result of logistic regression. The X-axis showed the predicted probability of JIA. The Y-axis showed the actual probability of JIA. The solid line indicates the performance of the developed nomogram model.

Figure 5

Decision curve analysis for the developed nomogram model. (A, B) are the result of LASSO logistic regression. (C, D) are the result of logistic regression.