E2665. Hepatic Steatosis in the Diabetic Population as Assessed by Using Machine Learning on CT Scans
  1. Pavan Raghupathy; University of Pennsylvania
  2. Richard Tran; University of Pennsylvania
  3. Matthew MacLean; University of Pennsylvania
  4. James Gee; University of Pennsylvania
  5. Daniel Rader; University of Pennsylvania
  6. Walter Witschey; University of Pennsylvania
  7. Hersh Sagreiya; University of Pennsylvania
The prevalence of diabetes has been increasing globally, mainly comprising type 2 diabetes mellitus (T2DM). While non-alcoholic fatty liver disease (NAFLD), a spectrum of conditions of excess fat buildup in the liver, has a prevalence of 25-35% in the general population, this increases in the diabetic population up to 75%. NAFLD is marked by hepatic steatosis, which can be diagnosed with computed tomography (CT) scans through the spleen-hepatic attenuation difference (SHAD). We set out to quantify the relationship between hepatic adiposity and related attributes to T2DM using machine learning.

Materials and Methods:
This study was coordinated through a medical biobank, an IRB-approved protocol that collects data from outpatient visits. Phecodes used were based on International Classification of Diseases, Ninth Revision, Clinical Modification. With 74,988 noncontrast CT studies in 15,259 patients obtained, artificial intelligence methods were used to quantify liver and spleen volumes and attenuation values to determine SHAD–spleen attenuation minus hepatic attenuation. Patients with multiple studies had median hepatic fat reading selected. Phecodes related to alcohol use disorders and end-stage liver diseases served as exclusionary criteria; and patients missing race, BMI and age information were excluded, resulting in 8,746 patient studies. A generalized linear model (Model 1) determined the effects of imaging and demographic characteristics on disease presence. Variables included SHAD, liver volume (LV), and spleen volume (SV) controlled for sex, age, race and body mass index (BMI) category. BMI categories are defined as: underweight (<18.5), normal (18.5 to <25), overweight (25 to <30), obese (30 to <40), morbidly obese (=40). P-values were adjusted with Bonferroni correction with statistical significance threshold of p < 0.05. Model 1: log(p(disease=1)/p(disease=0)) = SHAD + LV + SV + SEX + AGE + RACE + BMI

Logistic regression demonstrated that age [B = 0.039 (p < 0.001)], sex [B = 0.173 (p < 0.01)], and SHAD [B = 0.036 (p < 0.001)] all independently affect chance of having T2DM. BMI classification with normal baseline also independently affects chance of having T2DM: underweight [B = -5.028 (p < 0.001)], obese [B = 0.920 (p < 0.001)], and morbidly obese [B = 1.32 (p < 0.001)]. Race classification, with a baseline of White, also independently affects chance of having T2DM: Asian [B = 0.826 (p < 0.001)], Black [B = 1.163 (p < 0.001)], Indian [B = 2.311 (p < 0.001), Latino/White [B = 0.731 (p = 0.00187)], Latino/Black [B = 0.987 (p = 0.00234)]. Liver metric volume was also observed to independently affect the chance of having T2DM [B = 5.483 x 10-7 (p < 0.001)].

Age, sex, SHAD, LV, and BMI classification are independently associated with T2DM. SHAD value should be computed in a routine clinical setting to account for a growing T2DM population. This study presents an accurate method for examining hepatic steatosis in the context of associated characteristics in diabetic populations using machine learning to fuel translational scientific discovery.