Phenotyping, genotyping, and prediction of abdominal pain in children using machine learning

Abstract

AbstractObjectiveMechanisms of abdominal pain in children are not fully understood due to patient heterogeneity. We aimed to identify abdominal pain phenotypes in children to facilitate the investigation of phenotypic-genotypic associations and to determine risk factors for abdominal pain.DesignThis study included 13,789 children from a large birth cohort. The comorbidities of children and mothers and single nucleotide polymorphisms in children were investigated. Machine learning (ML) was used to identify clusters of patients with homogenous characteristics; subsequently, genome-wide association studies and enrichment analyses were performed. The factors contributing to predictive models were identified using ML.ResultsA total of 1,274 children experienced abdominal pain (9.2 %) (average age: 8.4 ± 1.1 years old, male/female: 615/659), who were classified into 3 clusters: cluster 1 with an allergic predisposition (n = 137), cluster 2 with mother’s comorbidities (n = 676), and cluster 3 with minimal comorbidities (n = 340). Enrichment analysis indicated that genetic pathways related to the intestinal barrier or bile acid biosynthesis were associated with abdominal pain in cluster 1; bile acid biosynthesis was also involved in cluster 2. Predictive models demonstrated modest fidelity with AUC values up to 0.65 in predicting children’s abdominal pain, showing mother’s and children’s comorbidities formed significant risk factors.ConclusionThe risk factors and phenotypes of paediatric abdominal pain are embedded within phenotype-genotype associations, which can be targeted in future studies. In particular, the link between allergy and intestinal barrier may be of mechanistic and therapeutic importance.