Unsupervised Descriptive Phenotyping of Cardiometabolic Laboratory Patterns from a Routine Five-Analyte Panel: Stability, Transportability and a Rule-Based Laboratory Interpretation Layer

Figures

• 

  Figure 1. Cohort selection flow. Exclusion steps are applied sequentially to the deduplicated pool; the transportability analysis (Section 3.6) applies the same filters within the two sub-cohorts.

• 

  Figure 2. The number of clusters. (a) The value of k at which each of the eight criteria is optimal (blue, a value below four (one, two or three); green, four; orange, five). The grids differ by criterion: gap statistic over k=1–7, internal indices and mixture BIC over k=2–8, prediction strength over k=2–6, the validation cohort’s Calinski–Harabasz index over k=3–5 and the bootstrap at the two candidate values only. (b) Best-match Jaccard recovery of each cluster over 100 bootstrap resamples: dot, mean; cap, 10th percentile; hollow diamond, mean when the whole log–robust-scaler–PCA pipeline is refitted inside every resample. C0–C3, primary model; F0–F4, five-cluster configuration. (c) Agreement between the frozen centres-only model and a model re-derived de novo in the identifier-disjoint validation cohort: adjusted Rand index (solid) and matched agreement (light), for k=4 and k=5.

• 

  Figure 3. The four clusters in the PCA space: (left) PC1 vs. PC2 and (right) PC1 vs. PC3: Black crosses denote centroids. Axis scales are retained because centroid displacements are quoted in these units. Clusters are separated along the glycaemic (PC1), cholesterol (PC2) and triglyceride–HDL-C (PC3) axes with smooth rather than gap-separated boundaries. Cluster labels name single-measurement laboratory patterns, not diagnoses.

• 

  Figure 4. Standardised (z-scored) cluster-median analyte profiles; cell annotations give the original median values (mmol/L). Each cluster is defined by a distinct combination of glycaemic and lipid deviations. Cluster labels name single-measurement laboratory patterns, not diagnoses.

• 

  Figure 5. Distribution of each analyte across the four clusters. Boxes provide the interquartile range with the median; whiskers extend to 1.5 times the interquartile range and observations beyond them are not drawn—a suppression that is largest for glucose in C3. The clusters differ on different analytes, i.e., they encode multivariate patterns rather than a single severity axis. Cluster labels name single-measurement laboratory patterns, not diagnoses.