Linking abdominal imaging traits to electronic health record phenotypes

Quantitative traits obtained from computed tomography (CT) scans performed in routine clinical practice have the potential to enhance translational research and genomic discovery when linked to electronic health record (EHR) and genomic data. For example, both liver fat and abdominal adipose mass are highly relevant to human disease; non-alcoholic fatty liver disease(NAFLD) is present in 30% of the US adult population, is strongly associated with obesity, and can progress to hepatic inflammation, cirrhosis, and hepatocellular carcinoma. We built a fully automated image curation and organ labeling technique using deep learning to identify liver, spleen, subcutaneous and visceral fat compartments in the abdomen and extract 12 quantitative imaging traits from 161,748 CT scans in 19,624 patients enrolled in the Penn Medicine Biobank (PMBB). The average liver fat, as defined by a difference in attenuation between spleen and liver, was -6.4 {+/-} 9.1 Hounsfield units (HU). In 135 patients who had undergone both liver biopsy and imaging, receiver operating characteristic (ROC) analysis revealed an area under the curve(AUC) of 0.81 for hepatic steatosis. The mean fat volume within the abdominal compartment for subcutaneous fat was 4.9 {+/-} 3.1 L and for visceral fat was 2.9 {+/-} 2.1 L. We performed integrative analyses of liver fat with the phenome extracted from the EHR and found highly significant associations with chronic liver disease/cirrhosis, chronic non-alcoholic liver disease, diabetes mellitus, obesity, hypertension, renal failure, alcoholism, hepatitis C, use of therapeutic adrenal cortical steroids, respiratory failure and pancytopenia. Liver fat was significantly associated with two of the most robust genetic variants associated with NAFLD, namely rs738409 in PNPLA3 and rs58542926 in TM6SF2. Finally, we performed multivariate principle component analysis (PCA) to show the importance of each of the quantitative imaging traits to NAFLD and their interrelationships with the phenome. This work demonstrates the power of automated image quantitative trait analyses applied to routine clinical imaging studies to fuel translational scientific discovery.