Machine-learning based exploration of determinants of gray matter volume in the KORA-MRI study

To identify the most important factors that impact brain volume, while accounting for potential collinearity, we used a data-driven machine-learning approach. Gray Matter Volume (GMV) was derived from magnetic resonance imaging (3T, FLAIR) and adjusted for intracranial volume (ICV). 93 potential determinants of GMV from the categories sociodemographics, anthropometric measurements, cardio-metabolic variables, lifestyle factors, medication, sleep, and nutrition were obtained from 293 participants from a population-based cohort from Southern Germany. Elastic net regression was used to identify the most important determinants of ICV-adjusted GMV. The four variables age (selected in each of the 1000 splits), glomerular filtration rate (794 splits), diabetes (323 splits) and diabetes duration (122 splits) were identified to be most relevant predictors of GMV adjusted for intracranial volume. The elastic net model showed better performance compared to a constant linear regression (mean squared error = 1.10 vs. 1.59, p < 0.001). These findings are relevant for preventive and therapeutic considerations and for neuroimaging studies, as they suggest to take information on metabolic status and renal function into account as potential confounders.