Evaluating Global and Domain-Specific PCA in Exposome Research: A Diagnostic Framework

Abstract

Introduction

The exposome describes the totality of environmental, behavioural, occupational, social and biological exposures experienced across the life course. This framework is particularly relevant for complex chronic diseases, where multiple exposures may interact dynamically with biological susceptibility. Exposome studies generate heterogeneous high-dimensional datasets, including pollutants, lifestyle, diet, socioeconomic factors, clinical variables and molecular biomarkers. In this context, dimension-reduction techniques, particularly principal component analysis (PCA), are frequently used to summarize correlated exposures, but it is often applied with incomplete diagnostics and limited justification for accepting low explained variance. Thus, a dataset may be statistically suitable for PCA while still being poorly represented by a small, stable and interpretable set of components.

Objectives
We aimed to propose a diagnostic framework to distinguish applicability from dimension-reduction techniques usefulness in exposome research, with particular attention to the role of domain-specific PCA as a domain-informed unsupervised approach.

Methods

The framework was applied to two datasets. The first was EXPOSITION, an in-house cross-sectional cohort of 150 Northern Italian patients with multiple sclerosis, including 68 variables across demographic, clinical, sociodemographic, dietary, lifestyle, NO₂ exposure and miRNA domains collected and managed through BIOMATRIX, a privacy-preserving digital platform for biodiversity management and advanced analysis. The second was iMSMS, an international gut microbiome dataset including 1,152 multiple sclerosis patients and paired household controls, with variables across demographic, clinical, sociodemographic, dietary, microbiome, metagenomic and metabolomic domains. For each dataset, we applied a structured pre-PCA diagnostic assessment to evaluate the statistical admissibility of component extraction. This included missingness evaluation, correlation matrix inspection, near-zero variance detection, outlier assessment, scaling sensitivity, Kaiser–Meyer–Olkin measure of sampling adequacy, Bartlett’s test of sphericity and variable-level adequacy. PCA usefulness was then evaluated through explained variance, number of retained components, loading stability, domain purity, interpretability and dimensionality-reduction efficiency. Global PCA was compared with domain-specific PCA, robust PCA, sparse PCA, factor analysis, FAMD and MFA.

Results
The application to EXPOSITION and iMSMS highlighted that conventional pre-PCA diagnostics are necessary but not sufficient. Even when factorability diagnostics supported PCA application, global PCA could still require several components to explain an adequate proportion of variance, limiting its value as an efficient dimensionality-reduction strategy. This issue was particularly relevant when variables belonged to distinct conceptual domains or represented complex biological profiles. Domain-specific PCA, while remaining unsupervised because it does not use outcome information, incorporated a priori knowledge on exposure domains and provided more interpretable and epidemiologically coherent summaries than global PCA in domain-structured data.

Conclusions
PCA should not be applied as a default dimensionality-reduction tool in exposome research. Researchers should report both pre-PCA diagnostics and post-PCA usefulness criteria, especially explained variance, component stability and interpretability. Domain-specific PCA may represent a practical compromise between purely data-driven reduction and meaningful exposure summarization in heterogeneous multi-domain datasets.