Deep-learning derived detailed measures of body fat distribution and cardiovascular outcomes in non-dialysis dependent chronic kidney disease: The Multi-Ethnic Study of Atherosclerosis.

An obesity paradox, i.e., an association between high BMI and reduced all-cause mortality, has been reported in patients with non-dialysis-dependent chronic kidney disease (CKD). We have recently developed and validated a fully automated deep-learning algorithm that can automatically segment and quantify detailed muscle and fat distributions from any conventional chest CT. Here, we generated these robust biomarkers in a study cohort with non-dialysis-dependent CKD (eGFR 15-90 mL/min/1.73 m2, Stage 2-4 CKD) from the Multi-Ethnic Study of Atherosclerosis. In fully adjusted linear models, greater intermuscular, extramyocellular, and perimuscular adipose tissue indices and a lesser lean pectoralis muscle mass index were associated with higher levels of serum cystatin C, a marker of kidney function with well-established associations to cardiovascular risk. In a stratification analysis, a greater perimuscular adipose tissue index was independently associated with an increased risk of developing MACE (i.e., primary cardiovascular outcome) in those with above-median baseline serum cystatin C (≥ 0.95 mg/L). This association was observed in fully adjusted models (p = 0.02, HR 1.18, CI 1.03-1.35, Model 2) and persisted after further adjustment for the sarcopenia index, calculated as (SerumCreatinineSerumCystatinC * 100) (p = 0.04, HR 1.16, CI 1.01-1.34, Model 3). The above associations were not observed in participants with a below-median level of serum cystatin C (< 0.95 mg/L). Distinct biomarkers of body composition and fat distribution (i.e., perimuscular adipose tissue), reliably obtainable from any conventional chest CT using the fully automated DL algorithm, can predict MACE in participants with non-dialysis-dependent CKD and high cystatin C levels at baseline.