Data-driven identification of a candidate neutrophil-enriched inflammatory-nutritional phenotype in osteoarticular tuberculosis

Osteoarticular tuberculosis (OATB) exhibits substantial host-response heterogeneity that is not captured by anatomical classification alone. This study sought to derive clinically interpretable phenotypes from routinely available data and to determine whether the dominant immune signal was reproducible in independent transcriptomic cohorts. In this retrospective clinical phenotyping study, 254 patients were retained following source-record deduplication, retrospective cohort reconstruction, and exclusion of cases with insufficiently documented diagnostic eligibility. Twenty-four baseline variables spanning systemic inflammation, hematology, electrolyte homeostasis, and nutritional status were winsorized, median-imputed, standardized, and subjected to K-means clustering; candidate solutions from K = 2 to K = 6 were compared using silhouette analysis. Phenotypes were clinically characterized, and LASSO regression and random-forest importance were used for feature exploration. A clinically parsimonious three-variable phenotype-assignment model was evaluated using repeated stratified 10-fold cross-validation, while a nested cross-validation sensitivity analysis repeated preprocessing and feature selection entirely within the training folds. Robustness and uncertainty were examined using age-, anatomical-site-, sex-, temporal-, neutrophil-excluded, and missing-data sensitivity analyses. Neutrophil single-sample gene-set enrichment analysis in GSE83456 and GSE28623 was used solely for independent biological corroboration. The selected exploratory two-cluster solution comprised OATB-ML1 ( n = 175) and OATB-ML2 ( n = 79). OATB-ML2 displayed a coordinated neutrophil-enriched inflammatory–nutritional profile characterized by higher C-reactive protein, leukocyte and platelet counts, neutrophil percentage, monocyte count, and globulin, together with lower hemoglobin, lymphocyte percentage, selected electrolytes, and albumin. Median neutrophil percentage, calculated absolute neutrophil count, and neutrophil-to-lymphocyte ratio were 71.35% versus 60.85%, 5.64 versus 3.82 × 10⁹/L, and 4.33 versus 2.48, respectively. The neutrophil association persisted after neutrophil percentage was excluded from reclustering (69.90% versus 61.65%; P < 0.001; ARI with the primary solution = 0.830). A parsimonious CRP–albumin–neutrophil model provided conditional internal discrimination for phenotype assignment; however, these performance estimates represent internal reproduction of a data-derived label rather than independent diagnostic, prognostic, or treatment-response accuracy. Both transcriptomic cohorts independently showed greater neutrophil enrichment in active disease and overrepresentation of active tuberculosis within neutrophil-high immune states. Routine clinical data revealed an exploratory two-group structure containing a candidate neutrophil-enriched inflammatory–nutritional pattern in OATB. A three-variable CRP–albumin–neutrophil model provided a parsimonious means of internally assigning the data-derived grouping but should not be interpreted as a diagnostic or prognostic model. Independent transcriptomes supported the biological plausibility of the neutrophil axis; prospective multicenter reproduction of the grouping structure and external evaluation of the assignment model are required.