To identify tongue-phenotype subtypes in patients with diabetic kidney disease (DKD) using unsupervised clustering of quantified tongue features, and to compare tongue characteristics and laboratory profiles across subtypes.
Approach:
Study Design: Enrolled 331 patients with DKD from two hospitals in Beijing, extracting 48 continuous tongue features and 3 ordinal/categorical variables.
Clustering Methods: Used elbow method, silhouette width, and gap statistic to determine optimal clusters; applied K-means clustering, Ward.D2 hierarchical clustering, and partitioning around medoids (PAM) for analysis.
Statistical Analysis: Compared tongue features and laboratory variables between subtypes with Benjamini-Hochberg correction for multiple testing.
Key Findings:
K-means clustering identified two subtypes: Cluster 1 (n = 108, 32.6%) and Cluster 2 (n = 223, 67.4%).
Cluster 2 exhibited higher brightness, lower saturation, lower coating ratio, thicker coating, and more tooth marks compared to Cluster 1.
Significant differences in coating thickness grade and tooth marks between clusters, but fissures did not differ.
Thirty-nine of 48 continuous tongue features remained significant after correction.
No significant differences in laboratory or composite indices after multiple imputation and BH correction.
Interpretation:
Objective tongue phenotyping identified two reproducible DKD subtypes, indicating tongue-based heterogeneity that does not overlap with conventional laboratory profiles.
Limitations:
Study conducted at two medical centers, which may limit generalizability.
Potential biases in subjective assessments of tongue features.
Conclusion:
Tongue phenotyping may offer value for syndrome differentiation in DKD.
