An interpretable machine learning model for biomarker identification and diagnostic nomogram development in hypo-productive thrombocytopenia.

Thrombocytopenia is a common hematological disorder with diverse etiologies. This study aimed to identify potential biomarkers and to develop an interpretable diagnostic nomogram for the early recognition of hypo-productive thrombocytopenia. This retrospective study included 185 patients with thrombocytopenia who were admitted to the Department of Hematology at the Affiliated Suzhou Hospital of Nanjing Medical University between 2020 and 2025. Patients were categorized into hypo-productive thrombocytopenia (n = 114) and hyper-destructive thrombocytopenia (n = 71) according to etiology. Feature selection was performed using least absolute shrinkage and selection operator (LASSO) regression, support vector machine recursive feature elimination (SVM-RFE), and Boruta algorithms, followed by univariate and multivariate logistic regression analyses. Six machine learning (ML) models were developed and compared based on residual analysis and receiver operating characteristic (ROC) curve. The optimal model was interpreted using SHapley Additive exPlanations (SHAP) analysis. The diagnostic nomogram was evaluated by the ROC, calibration curves, and decision curve analysis (DCA). Finally, age, platelet-to-lymphocyte ratio (PLR), neutrophil-to-lymphocyte ratio (NLR), and mean platelet volume (MPV) were identified as key predictors. The generalized linear model(GLM)-based nomogram demonstrated strong discriminative ability (area under the curve [AUC] = 0.945, 95% confidence interval [CI] 0.900-0.976) with excellent calibration. DCA demonstrated a higher net benefit for the nomogram than individual predictors. The interpretable machine learning derived nomogram may serve as a practical and non-invasive tool to assist in the early identification of hypo-productive thrombocytopenia.