Penalised Mixed Effects Modelling of Longitudinal Metabolomics Data

Introduction: The Earlybird study is a cohort study of children which investigated the development and determinants of insulin resistance in childhood (1). Data were collected at 12 annual visits and longitudinal serum NMR-spectroscopy metabolomics data were available for 130 participants. Analysis of metabolomics data is complicated by the large number of variables, causing overfitting and instability in standard mixed effects regression. One-at-a-time approaches to modelling metabolite variables fail to account for between-metabolite relationships (2). Missing data is a common issue in longitudinal studies.

Methods: A multilevel least absolute shrinkage and selection operator (LASSO) model was used to perform variable selection (3). Penalisation allows the simultaneous analysis of a large number of predictors that would not be feasible with standard regression. This method has rarely been applied to longitudinal metabolomics data. Random forest multiple imputation was used to estimate missing values. Models included adjustments for age, sex and BMI SD score.

Results: Multilevel LASSO regression was applied across 20 multiply imputed datasets, with insulin resistance as the outcome. Sixteen of the metabolite variables were selected in ≥75% of the multilevel LASSO regression models across the imputed datasets and retained for further analysis. Six of the selected metabolomics features showed statistically significant negative pooled associations with insulin resistance.

Conclusions: A multilevel LASSO modelling approach is applicable to longitudinal metabolomics data, and a reproducible analysis workflow has been established. The approach handles high-dimensional sets of predictors where standard regression is not feasible, manages missing data via multiple imputation, and models individual trajectories over time via random effects. Next, performance of the modelling procedure will be formally evaluated in a simulation study. Preliminary results from the simulation study will also be presented if available.

References:

Voss, L.D. (2003) Pediatr. Endocrinol. Metab., 16(9), 1211–1224.
Harrell, F. (2024) Biostatistics for Biomedical Research.
Tibshirani, R. (1996) R. Stat. Soc. B, 58(1), 267–288.