Unsupervised Machine Learning of Intrapancreatic Fat Deposition Reveals Distinct Clinical Phenotypes: A CT Study

Figures

• 

  Figure 1. Study design and analytical workflow. Flowchart illustrating the study population selection and analytical workflow. A total of 558 participants were included in the retrospective cross-sectional cohort. Unsupervised k-means clustering was performed based on five standardized clinical variables, including age, body mass index (BMI), pancreas–spleen attenuation difference (P–S difference), fasting plasma glucose (FPG), and triglycerides (TG), resulting in the identification of three exploratory clinical phenotypes (obese, FPD-like, and Control phenotypes).

• 

  Figure 2. Determination of optimal cluster number and feature distribution heatmap. (A) The Elbow method showing the total within-cluster sum of squares (WSS) across different numbers of clusters (k). (B) The Silhouette method evaluating cluster separation and cohesion, supporting the selection of a three-cluster solution. (C) Heatmap showing the distribution of the five standardized clustering variables across the identified phenotypes. Each cell represents the Z-score of the corresponding variable, with warmer colors indicating relatively higher values and cooler colors indicating relatively lower values.

• 

  Figure 3. Multidimensional clinical and morphometric signatures of the identified IPFD phenotypes. (A–D) Violin plots with embedded box plots showing the distributions of age (A), BMI (B), P–S difference (C), and FPG (D) among the three clusters. The central line represents the median, the box indicates the interquartile range (IQR), and whiskers extend to values within 1.5 times the IQR. Comparisons among clusters were performed using the Kruskal–Wallis test, followed by Dunn’s test with Bonferroni correction for post hoc pairwise comparisons. (E) Radar chart illustrating the normalized multidimensional profiles of the three phenotypes based on the clustering variables. Values were standardized using Z-scores to facilitate visualization of relative differences among clusters.

• 

  Figure 4. Prevalence of overt metabolic morbidities across the three IPFD phenotypes. Bar charts showing the prevalence (%) of type 2 diabetes mellitus (T2DM), coronary heart disease (CHD), metabolic dysfunction-associated steatotic liver disease (MASLD), and hypertension (HTN) across the three clusters. The prevalence estimates are presented with corresponding 95% confidence intervals (CIs) in parentheses. Differences among groups were evaluated using Pearson’s Chi-square test.

• 

  Figure 5. Circulating metabolic biomarkers across the identified phenotypes. Box plots showing serum concentrations of (A) free fatty acids (FFA), (B) uric acid (UA), and (C) high-density lipoprotein cholesterol (HDL-C) across the three clusters. The central horizontal line represents the median, and boxes indicate the interquartile range (IQR). Comparisons among clusters were performed using the Kruskal–Wallis test, followed by Dunn’s test with Bonferroni correction for post hoc pairwise comparisons. Asterisks indicate statistically significant pairwise differences (*P < 0.05; **P < 0.01; ***P < 0.001).