Unsupervised Global CBC/RUO/CPD Phenotyping Identifies Haematological Clusters Enriched for Thrombocytopenia Severity and Mechanisms.

Abstract

Introduction: Thrombocytopenia has heterogeneous mechanisms, including peripheral destruction, consumption, and reduced marrow production. This study developed an unsupervised global complete blood count/research-use-only/cell population data (CBC/RUO/CPD) phenotyping model and evaluated post hoc whether thrombocytopenia severity and clinically adjudicated mechanisms were enriched within the resulting clusters.

Methods: Retrospective haematology analyser data from 254 289 unique specimens were analysed after cleaning 537 279 extracted rows. Exploratory missingness filtering retained 123 features and 173 851 complete-case specimens. The final model used a refined 41-feature CBC_EXTENDED complete-case dataset containing 241 591 specimens. Features were standardised, reduced using principal component analysis to 20 components explaining 92.94% of cumulative variance, and clustered using k-means. Thrombocytopenia thresholds and electronic health record (EHR) adjudicated primary mechanism labels were evaluated post hoc.

Results: The final global k-means model (k = 4) generated four broad haematological clusters comprising 14%-37% of specimens. Internal separation was modest but interpretable, with sampled silhouette 0.10, Davies-Bouldin 2.21, and Calinski-Harabasz 6536.92 in the fixed 60 000 specimen model comparison analysis. Cluster 3 showed progressive post hoc enrichment with thrombocytopenia severity, accounting for approximately 46%, 60%, and 71% of specimens with platelet counts < 150, < 100, and < 50 × 10⁹/L, respectively. In 225 clinically annotated thrombocytopenic specimens, k-means k = 4 showed limited but non-random concordance with adjudicated mechanisms (purity 0.59; adjusted Rand index 0.14; normalised mutual information 0.18).

Conclusion: Unsupervised global CBC/RUO/CPD clustering identified broad laboratory phenotypes with post hoc enrichment for thrombocytopenia severity and adjudicated mechanisms. These findings support underutilised analyser-derived parameters for laboratory phenotyping and hypothesis generation, not standalone diagnosis.