Abstract
<title>Abstract</title> <p>Background Ischemic heart disease (IHD) is the leading cause of global mortality. Whether routine, low-cost laboratory parameters can reveal clinically meaningful heterogeneity within an already-diagnosed IHD population remains unexplored. Methods In this cross-sectional study, 306 confirmed IHD patients from hospitals in Lahore, Pakistan, underwent measurement of routine hematological, hepatic, renal, and lipid parameters. An unsupervised machine learning framework was used to identify patient phenotypes directly from laboratory profiles. Data were split into training (70%) and test (30%) subsets; five clustering methods; k-means, Gaussian mixture models, fuzzy c-means, hierarchical clustering, and DBSCAN, were compared on UMAP-reduced features, and the best-generalizing configuration was selected. Random Forest, Decision Tree, and fuzzy rule-based classifiers were trained to identify discriminating biomarkers and translate clusters into clinician-readable rules. Results K-means (k = 3) gave the most stable partition (silhouette = 0.361 training, 0.370 test), yielding three phenotypes: a small (12.7%) atherogenic-hepatic-inflammatory cluster (elevated VLDL, triglycerides, transaminases, neutrophils); a larger (26.5%) pure dyslipidemia cluster; and a majority (60.8%) lower-risk cluster with the lowest atherogenic lipid burden. VLDL, total bilirubin, triglycerides, AST, ALT, and HDL were the principal discriminating features (Random Forest test accuracy 78.3%, AUC 0.926). Decision Tree and fuzzy classifiers translated these into interpretable rules (e.g., "IF VLDL is High AND AST is Low"), revealing an accuracy–interpretability trade-off. Conclusions Even without imaging, angiography, or omics data, routine laboratory parameters reveal reproducible phenotypic heterogeneity in IHD, centered on a lipid-metabolic and hepatic-inflammatory axis. This interpretable, low-cost framework may support scalable risk stratification in resource-limited settings, pending external and prognostic validation.</p>