Refining cardiometabolic risk assessment using MRI-derived pancreas volume and fat content: insights from the NAKO and UK Biobank

BackgroundThe pancreas is essential for metabolic homeostasis. Alterations in morphology and parenchymal integrity may impact proper function but are not routinely used for risk stratification. Here, we propose an AI-pipeline to quantify pancreas volume and fat content from MRI to identify individuals at high-risk for cardiometabolic disease in the general population. MethodsWe quantified pancreas volume (milliliters, mL) and intrapancreatic fat content (defined as fat fraction; FF, %) from MRI of UK Biobank (UKB) and German National Cohort (NAKO) participants using deep learning. We 1) analyzed differences in volume and FF across age and sex, 2) computed percentile-curves and z-scores adjusted for age and sex to identify high-risk volumes/FF, and 3) conducted Cox regression to assess associations between z-score categories (volume: reference, z=-1 to 1; low, z=<-1; high, z>1; FF: low, z<1; moderate, z=0-1; high, z>1) and incident outcomes (diabetes, major adverse cardiovascular events (MACE), all-cause mortality) after adjustment for risk factors. ResultsAmong 63,548 UKB and NAKO-participants (57.7{+/-}12.8 years; BMI: 26.3{+/-}4.4 kg/m2, 46.9% female), automated pancreas analysis revealed a positive association between both volume and FF and age. In 33,099 UKB-participants (median 4.8 years follow-up), z-score categories were associated with incident diabetes (low volume, aHR:1.59, 95%CI[1.20-2.11]; high FF, aHR:1.70, 95%CI[1.31-2.19]), MACE (high volume, aHR: 0.79, 95%CI[0.61-1.01]; high FF, aHR: 1.32, 95%CI[1.01-1.73]), and all-cause mortality (low volume, aHR: 1.48, 95%CI[1.16-1.90]) beyond risk factors. Adding z-score categories to a baseline model including risk factors improved discrimination of future diabetes (volume:0.781 to 0.784, p=0.004; FF:0.781 to 0.787, p<0.001) and mortality (volume:0.781 to 0.787, p<0.001) ConclusionsDeviations from normalized pancreas volume and FF predicted cardiometabolic outcomes beyond known risk factors and alcohol intake. This automated approach identifies high-risk individuals who may benefit from cardiometabolic/endocrinology referral.