Validation of voxel-based computed tomography liver fat quantification for assessing incident type 2 diabetes risk in a health check-up population.

BACKGROUND: Conventional computed tomography (CT)-based liver fat assessment relies mainly on mean attenuation or liver-to-spleen ratios, whereas a novel voxel-based approach enables whole-liver, voxel-level quantification of hepatic fat. However, its value for predicting incident type 2 diabetes (T2D) remains unclear. Therefore, this study aimed to validate whether voxel-based liver fat quantification derived from routine CT can be used to assess the risk of incident T2D in a health check-up population. METHODS: In this retrospective cohort study, individuals who underwent routine noncontrast abdominal CT examinations between January 2013 and December 2023 and were free of diabetes at baseline were included. Liver fat was quantified via an automated voxel-based method incorporating spleen-referenced attenuation criteria on the basis of whole-liver segmentation. The average liver fat fraction was defined as the proportion of liver voxels classified as fat. Incident T2D was identified through longitudinal review of outpatient, inpatient, and health check-up records. Associations between liver fat and incident T2D were evaluated via Cox proportional hazards models adjusted for demographic, anthropometric, and metabolic covariates. Restricted cubic spline (RCS) analysis was used to explore dose-response relationships. Predictive performance was assessed via time-dependent receiver operating characteristic (ROC) analysis at 3 and 5 years, calibration plots, and decision curve analysis (DCA). Incremental predictive value was evaluated by comparing liver fat-based models with hemoglobin A1c (HbA1c)-based and clinical models. RESULTS: Among 356 participants [median age, 57 years, interquartile range (IQR), 53-63 years; 56.5% male], 32 individuals (9.0%) developed incident T2D during a median follow-up of 4.5 years. A higher average liver fat fraction was significantly associated with an increased risk of incident T2D after multivariable adjustment [hazard ratio (HR) per standard deviation (SD), 1.65; 95% confidence interval (CI): 1.05-2.60, P=0.031]. RCS analysis demonstrated a monotonic increase in diabetes risk with increasing liver fat fraction. The liver fat fraction showed moderate discrimination for incident T2D, with time-dependent area under the curve (tAUC) values of 0.71 (95% CI: 0.59-0.84) at 3 years and 0.72 (95% CI: 0.60-0.82) at 5 years. Compared with clinical variables or HbA1c alone, the addition of liver fat improved predictive performance for incident T2D, increasing the C-index from 0.574 (0.453-0.680) to 0.740 (95% CI: 0.663-0.803). CONCLUSIONS: Voxel-based CT quantification of liver fat is independently associated with incident T2D and provides incremental value for diabetes risk assessment in a health check-up population.