Machine learning approach for genome-wide polymorphisms to predict childhood allergic rhinitis in a longitudinal follow-up birth cohort study

Abstract

Background: Allergic rhinitis (AR) is a common chronic inflammatory disorder in children. Genetic characteristics and environmental exposures contribute to the development of allergic diseases. This study aimed to develop a predictive model using genetic single nucleotide polymorphisms (SNPs) for the development of childhood AR from a longitudinal follow-up birth cohort study.Methods: Mother-infant pairs in the birth-cohort study were recruited from night Taiwanese hospitals from 2001 to 2005. Information on children’s health status including AR occurrence was obtained via questionnaire interviews at the age of 2, 5, 10, and 14 years old. SNP Array (Axiom Genome-Wide TWB 2.0) was used for genotype measurements. We applied machine learning method to establish predictive model for childhood AR. Multiple variable regression was used to assess the association of childhood AR with genotypes and conventional risk factors.Results: Overall, 247 mother-infant pairs completed all measurements. The prevalence of ever having physician-diagnosed AR by 14 years of age was 40.1%. By decision tree analysis, 19 SNPs were found to efficiently distinguish children with or without AR. The area under the receiver operating characteristic curve (AUROC) of the predictive model was 0.825. We then used backward stepwise regression to select significant SNPs in the predictive model and 8 SNPs were selected (AURIC = 0.806). After adjusted for potential confounders of child sex and maternal education, 6 SNPs of them (i.e. CD28 rs12693993, EGFR rs117179604, FTO rs71390222, IL1RAPL1 rs1016210, SEMA3A rs1520102, and SOCS3 rs4969169) were still significantly associated with childhood AR. The AUROC was slightly improved to 0.827. Conclusion: This study provides a model for predicting the risk of childhood AR according to hereditary factors. Future functional investigations of the SNPs and verifications with additional larger cohorts are warranted.