Data-Driven Abdominal Phenotypes of Type 2 Diabetes in Lean, Overweight, and Obese Cohorts

Lucas W. Remedios, Chloe Cho, Trent M. Schwartz, Dingjie Su, Gaurav Rudravaram, Chenyu Gao, Aravind R. Krishnan, Adam M. Saunders, Michael E. Kim, Shunxing Bao, Alvin C. Powers, Bennett A. Landman, John Virostko

Introduction

Body Mass Index (BMI) has long served as the primary clinical metric for assessing Type 2 Diabetes (T2D) risk. However, BMI is a "coarse" measurement that fails to account for the significant variation in metabolic health within weight classes. Many individuals classified as obese remain metabolically healthy, while lean individuals can develop T2D. This discrepancy suggests that the specific distribution and quality of abdominal fat and muscle—the "abdominal phenotype"—are more critical drivers of disease than total body mass.

Recent advancements in artificial intelligence now allow researchers to move beyond BMI by performing high-throughput, 3D body composition analysis on clinical imaging. This study utilizes an AI-driven pipeline to identify the specific anatomical signatures associated with T2D across lean, overweight, and obese populations.

Automated Abdominal Profiling

To analyze internal structures at scale, the researchers developed a multi-stage pipeline centered on interpretability. The process begins with TotalSegmentator, an AI tool that automatically identifies and segments eight key abdominal structures from CT scans: the pancreas, liver, spleen, kidneys, visceral fat, subcutaneous fat, and skeletal muscle.

Automated 3D segmentation of abdominal organs (pancreas, liver, spleen, kidneys) across sagittal, coronal, and axial planes.

From these segments, 88 distinct features were extracted, including volume, shape, and radiodensity (measured in Hounsfield Units, or HU). In soft tissue, lower HU intensity serves as a reliable proxy for higher fat infiltration (steatosis), allowing the model to assess organ quality alongside size.

Predictive Modeling and Pattern Discovery

The study utilized a dataset of 1,728 de-identified patients from Vanderbilt University Medical Center, stratified into lean, overweight, and obese cohorts. Random Forest classifiers were trained to distinguish T2D patients from controls based solely on these anatomical features and basic demographics. To ensure the model's decisions were explainable, the researchers employed SHAP (SHapley Additive exPlanations) to rank the influence of specific anatomical features.

Axial CT slice demonstrating the segmentation of skeletal muscle and adipose tissue depots used to identify "fatty" muscle signatures.

Key Insights: The Universality of Fatty Muscle

A primary finding of the study is the consistent role of "fatty" skeletal muscle as a predictor of T2D. Across all BMI categories, low radiodensity in the skeletal muscle—indicating fat infiltration—emerged as a top predictor. In both the lean and obese subgroups, it was identified as the single most important anatomical feature for identifying T2D.

Other consistent drivers included:

  • Organ-Specific Signatures: T2D was linked to a smaller, fat-laden pancreas and an enlarged, fatty liver. Interestingly, enlarged right kidneys were also a significant predictor across most groups.
  • The Lean Paradox: Within the lean cohort, the model found that lower BMI was actually associated with T2D. This suggests that in lean individuals, diabetes may be driven by sarcopenia (reduced muscle mass) rather than excess adiposity.
  • Phenotype Clusters: Using UMAP and K-means clustering, the researchers identified distinct anatomical "types" of diabetes. For instance, in the obese group, one T2D-enriched cluster was predominantly female (characterized by a smaller, fatty pancreas), while another was male-enriched (characterized by high visceral fat and larger liver size).

Conclusion and Future Directions

The Random Forest models achieved a mean Area Under the Curve (AUC) of 0.72 to 0.74. While not intended for primary diagnosis, these results demonstrate that AI can successfully extract meaningful biological patterns from existing clinical data. This "opportunistic" use of CT scans—originally taken for unrelated medical reasons—could eventually enable automated, large-scale risk profiling for metabolic diseases, providing a much more nuanced understanding of patient health than BMI alone.

Summary updated

BibTeX
@article{remedios2025data,
  author = {Lucas W. Remedios and Chloe Cho and Trent M. Schwartz and Dingjie Su and Gaurav Rudravaram and Chenyu Gao and Aravind R. Krishnan and Adam M. Saunders and Michael E. Kim and Shunxing Bao and Alvin C. Powers and Bennett A. Landman and John Virostko},
  title = {{Data-Driven Abdominal Phenotypes of Type 2 Diabetes in Lean, Overweight, and Obese Cohorts}},
  journal = {Journal of Medical Imaging},
  year = {2025},
  doi = {10.1117/1.jmi.12.6.064006},
}
Figure