Unsupervised Descriptive Phenotyping of Cardiometabolic Laboratory Patterns from a Routine Five-Analyte Panel:...

(1) Background: Reference-interval flags read a lipid-glucose panel one analyte at a time. We asked whether a routine five-analyte panel supports reproducible, transportable, auditable descriptive grouping of laboratory profiles. (2) Methods: Retrospective analysis of 21,832 laboratory-order records (adults; glucose, total cholesterol, triglycerides, HDL-C, LDL-C; collection dates not exported) from purchasers of a screening panel in Kazakhstan: log transform, robust scaling, PCA, K-means; eight cluster-number criteria; bootstrap recovery with and without pipeline refitting; frozen-classifier transfer to an identifier-disjoint cohort (n=11,382) with de novo re-derivation; a fully specified reference-interval rule layer. No clinical endpoints were available. (3) Results: Four descriptive laboratory-pattern clusters (lower glucose and lipids; higher LDL-C/TC; higher triglycerides, lower HDL-C; marked hyperglycaemia) had bootstrap Jaccard 0.965-0.981 with and without pipeline refitting and were re-derived independently in the validation cohort (ARI 0.82); a five-cluster configuration was not (0.33). Prevalence-shift point estimates were within 5 percentage points (post hoc exploratory benchmark; the C1 interval slightly crossed it). Rule-based reconciliation changed the delivered category of 8.2% of records from their K-means label; flags alone reproduced 77% of reports but 30% of the triglyceride-HDL-C pattern. (4) Conclusions: The panel supports a stable descriptive partition; prognostic and clinical validity remain to be shown.