The prognostic and clinical significance of hemogram parameters: a single-center retrospective analysis

Objectives This study explores the prognostic relevance of hemogram parameters and derived ratios in stratifying hospitalization levels, including intensive care unit (ICU) admission, using machine learning models. Methods Clinical and laboratory data were retrospectively obtained from a tertiary-level teaching hospital encompassing a wide range of care settings. Five classification models – Multinomial Logistic Regression, Random Forest, Gradient Boosting Machine (GBM), Support Vector Machine, and Multi-Layer Perceptron Neural Network (MLP-ANN) – were trained and evaluated. External validation was conducted using an independent dataset from affiliated institutions. SHAP (Shapley Additive Explanations) analysis was applied to enhance model interpretability. Results Among the evaluated models, GBM and MLP-ANN demonstrated consistent performance across treatment categories, particularly in distinguishing outpatient from higher-acuity hospital admissions. SHAP analysis identified eosinophil and basophil counts as key predictors, with eosinophil levels inversely associated with hospitalization intensity. External validation supported the models’ applicability across patient subsets, although multicenter generalizability remains to be established. Conclusions This study demonstrates that machine learning models trained on routine hemogram data can effectively differentiate hospitalization levels, offering a scalable approach to early clinical risk assessment. Eosinophil count, in particular, emerged as a key variable with consistent predictive power across models, suggesting its potential relevance in acute care stratification. These findings support the integration of hematological parameters into real-time triage systems and pave the way for future research to validate their role in diverse clinical contexts.