Use of Machine Learning-Based Combined Nomogram Model to Predict Autism Spectrum Disorder (ASD) Screening Needs of Children: A Retrospective Cohort Study of 193,101 Children in the US

Background: Autism Spectrum Disorder (ASD) has continued to increase in incidence rates in the past two decades. The purpose of the current study was to develop nomograms to assist in the prediction and identification of children with possible ASD diagnoses so that early interventions can be deployed to improve overall long-term functioning. Methods: Two model selection methods, least absolute shrinkage and selection operator (LASSO) and logistic regression, were used to develop the nomograms to predict ASD based on a large cohort from the National Survey of Children’s Health (NSCH), which included 193,101 children across all 50 states between 2016 and 2021, in the US. We used the area under the curve (AUC) to verify the discrimination performance. The sensitivity and specificity were also estimated by the optimal cutoff through Youden's J Index, Concordance Probability, and Minimize Euclidean distance. Results: LASSO and logistic regression models selected common predictors for ASD, including age, sex, whether born premature, BMI, brain injury, cerebral palsy diagnosis, anxiety, depression, having behavior or conduct problems, developmental delay, having speech or language disorders, learning disability, deafness or hearing problem, having difficulty in concentrating or remembering, having difficulty in making or keeping friends, and ever lived with anyone who had alcohol or drug problems. Maternal age at delivery and Attention Deficit Hyperactivity Disorder (ADHD) were predictors for ASD selected by the LASSO model. On the contrary, the logistic method determined asthma for the ASD predictor, which was not selected by the LASSO regression. Both LASSO and logistic methods showed excellent discrimination with 0.95 AUC. The correct classification rate, optimal sensitivity, and optimal specificity were 89.2% to 89.8%, 86.3 to 86.4%, and 89.3 to 89.9%, respectively. Conclusions: Based on two well-validated predictor models, we built the nomograms for the early prediction of ASD children. Our easy-to-use nomograms are not intended to replace the standard diagnostic methods; instead, it was used to help parents, clinicians, and researchers better estimate children’s potential risk of ASD by comparing the results of the two models.