Multimodal protective and susceptibility clusters in paediatric atopic dermatitis: A machine learning-based, data-driven observational study.

BACKGROUND

Atopic dermatitis (AD) is a chronic inflammatory skin disease that typically develops in early childhood. Differences in AD prevalence and allergy sensitisation patterns have been observed in African populations, including the AmaXhosa population in South Africa, suggesting alternative pathogenic and immune mechanisms underlying AD. Differences in AD prevalence have also been documented between urban and rural communities, making AmaXhosa children, who share a common ethnogenetic background but differ in environmental exposures, a unique population in which to investigate environmental and immune mechanisms underlying AD. To address this, we performed a machine learning (ML)-based multimodal observational study to identify features associated with AD in 217 AmaXhosa children.

METHODS AND FINDINGS: To gain deeper insights into AD pathogenesis, we re-analysed a previously established multimodal dataset comprising environmental, cytokine, antibody, and transcriptomic data from healthy AmaXhosa children and children with AD, aged 12-36 months, living in rural or urban settings. We applied ML to analyse each data modality individually and subsequently integrate them to identify multimodal signatures associated with AD. Specifically, we used the GeneSelectR workflow to identify informative genes, SHAP values to explain the ML outputs, and DIABLO to integrate the datasets and identify protective and susceptibility clusters. In the environmental and antibody datasets, we found that the combined effects of environmental features and higher levels of allergen-specific and total IgE antibodies contributed to the prediction of AD. In the transcriptomic dataset, we identified a subset of 560 genes that discriminated between children with and without AD and used these for the subsequent analyses. In the integrated analysis, we identified three multimodal clusters associated with AD status. One cluster associated with the healthy phenotype comprised environmental features primarily found in the rural setting, which correlated with plasma cytokine levels and the expression of autophagy-related genes. Two additional clusters were characterised by correlations between allergen-specific and total IgE antibodies and the cytokines MCP-4 and TARC, and by a transcriptomic feature signature associated with the AD endotype. Limitations of this study include the exploratory nature of the explainable ML framework and the lack of validation in an independent cohort.

CONCLUSIONS

Complementary ML approaches enabled explainable analysis of a complex multimodal dataset. Integrated analyses identified a multimodal protective cluster comprising environmental, cytokine, and transcriptomic features, together with two AD susceptibility clusters, one dominated by transcriptomic features and the other by correlated cytokine and antibody levels. These findings improve our understanding of the factors associated with AD. Furthermore, the feature-selection and integration workflow provides a framework for analysing complex multimodal datasets in future biomedical research.