Association of machine learning-derived measures of body fat distribution with cardiometabolic diseases in >40,000 individuals

Background The clinical implications of BMI-independent variation in fat distribution are not fully understood. Methods We studied MRI imaging data of 40,032 UK Biobank participants. Using previously quantified visceral (VAT), abdominal subcutaneous (ASAT), and gluteofemoral (GFAT) adipose tissue volume in up to 9,041 to train convolutional neural networks (CNNs), we quantified these depots in the remainder of the participants. We derived new metrics for each adipose depot – fully independent of BMI – by quantifying deviation from values predicted by BMI (e.g. VAT adjusted for BMI, VATadjBMI) and determined associations with cardiometabolic diseases. Results CNNs based on two-dimensional projection images enabled near-perfect estimation of VAT, ASAT, and GFAT, with r 2 in a holdout testing dataset (r 2 = 0.978-0.991). Using the newly derived measures of local adiposity – residualized based on BMI – we note marked heterogeneity in associations with cardiometabolic diseases. Taking presence of type 2 diabetes as an example, VATadjBMI was associated with significantly increased risk (odds ratio per standard deviation increase (OR/SD) 1.49; 95%CI: 1.43-1.55), while ASATadjBMI was largely neutral (OR/SD 1.08; 95%CI: 1.03-1.14) and GFATadjBMI conferred protection (OR/SD 0.75; 95%CI: 0.71-0.79). Similar patterns were observed for coronary artery disease. Conclusions Deep learning models trained on a simplified MRI input enable near perfect quantification of VAT, ASAT, and GFAT. For any given BMI, measures of local adiposity have variable and divergent associations with cardiometabolic diseases.