Abdominal MRI-Derived Liver, Visceral Fat, and Pancreatic Phenotypes Improve Prediction of Incident Type 2 Diabetes

Abstract

Aims

To evaluate whether abdominal magnetic resonance imaging (MRI)-derived metabolic phenotypes improve the opportunistic prediction of incident type 2 diabetes (T2D) beyond clinical risk models.

Materials and Methods

This study included 36,833 UK Biobank participants who underwent abdominal MRI and were free of T2D at imaging baseline. The original Cambridge Diabetes Risk Score (CDRS) was evaluated, and glycated haemoglobin (HbA1c) was added to form the clinical CDRS model. MRI phenotypes were selected using a LASSO-based selection-frequency analysis and added to construct the clinical CDRS-MRI model. Model performance was assessed in an internal validation set using Harrell's C-index, continuous net reclassification improvement (NRI), and integrated discrimination improvement (IDI).

Results

During a median follow-up of 7.38 years, 570 participants developed incident T2D. The selection-frequency analysis retained four MRI phenotypes: liver proton density fat fraction, liver volume, visceral fat volume, and pancreas volume. In the internal validation set, C-indices were 0.765 (95% CI, 0.731–0.800) for the original CDRS model, 0.820 (95% CI, 0.786–0.849) for the clinical CDRS model, and 0.849 (95% CI, 0.818–0.877) for the clinical CDRS-MRI model. The clinical CDRS-MRI model improved the C-index by 0.084 versus the original CDRS model and by 0.029 versus the clinical CDRS model, with a continuous NRI of 56.2% (95% CI, 40.6%–70.0%) and an IDI of 0.022 (95% CI, 0.010–0.037).

Conclusions

Abdominal MRI-derived phenotypes improved the prediction of incident T2D beyond established clinical risk models. These findings support secondary use of available abdominal MRI data for metabolic risk stratification, pending external validation.