(145) Critical Review of Dimensionality Reduction Approaches in Epidemiological Research

Background: Dimensionality reduction (DR) techniques are pivotal in transforming high-dimensional covariate spaces into more manageable, lower dimensional representations. Although these techniques are increasingly employed in epidemiology and public health research, a comprehensive review of their current applications and a discussion on their potential future uses in these fields is currently lacking.

Objectives: To review the application of DR methods in epidemiology and highlight potential areas of application.

Methods: We performed a targeted literature review on Scopus to identify DR techniques using “dimensionality reduction AND epidemiology” as a search term within titles, abstracts and keywords of indexed publications. While our main interest was in clinical epidemiology, the search did not impose any restrictions on publication time, language or publication type. We then reviewed titles, abstracts and full-texts of publications to identify the subject area, the specific type of DR technique used and its application.

Results: We found 147 publications of which 128 were research articles, 8 were reviews, and the remaining were a mix of book chapters, editorials and conference proceedings. Genomics/genetic epidemiology accounted for 80 (55.2%) articles. There were 37 (25.5%) application-based articles across medicine and public health, of which 19 (13.1%) were based on COVID-19 transmission models. The remaining 15 articles concerned clinical epidemiology, and had used the following DR techniques: principal components analysis (PCA), multiple correspondence analysis (MCA), partial least squares regression (PLS), autoencoder and LASSO. The main application of DR was to generate latent variables or risk stratifications that explained a proportion of the variance in the data. One paper discussed application of DR techniques to generate disease risk scores and propensity scores (PS). Key findings from the literature review show that PCA, which implements an orthogonal linear transformation of the covariate space, works best on continuous data, whereas MCA focusses on categorical variables and allows for non-linear relationships. PLS is a linear decomposition technique that also takes into account a response variable. Autoencoders are a type of neural network that can be used for DR with linear or non-linear activation functions.

Conclusions: The main application of DR techniques in clinical epidemiology appears to be limited to the generation of latent variables as a means to identify phenotypic clusters. The application of DR methods to PS estimation is limited. Our next steps are to evaluate balancing properties and performance of DR methods when used is PS estimation in a case study.