A novel scoring model for predicting efficacy and guiding individualised treatment in immune thrombocytopaenia.

Abstract

Despite diverse therapeutic options for immune thrombocytopaenia (ITP), drug efficacy and selection challenges persist. This study systematically identified potential indicators in ITP patients and followed up on subsequent treatment. We initially analysed 61 variables and identified 12, 14, and 10 candidates for discriminating responders from non-responders in glucocorticoid (N = 215), thrombopoietin receptor agonists (TPO-RAs) (N = 224), and rituximab (N = 67) treatments, respectively. Patients were randomly assigned to training or testing datasets and employing five machine learning (ML) models, with eXtreme Gradient Boosting (XGBoost) area under the curve (AUC = 0.89), Decision Tree (DT) (AUC = 0.80) and Artificial Neural Network (ANN) (AUC = 0.79) selected. Cross-validated with logistic regression and ML finalised five variables (baseline platelet, IP-10, TNF-α, Treg, B cell) for glucocorticoid, eight variables (baseline platelet, TGF-β1, MCP-1, IL-21, Th1, Treg, MK number, TPO) for TPO-RAs, and three variables (IL-12, Breg, MAIPA-) for rituximab to establish the predictive model. Spearman correlation and receiver operating characteristic curve analysis in validation datasets demonstrated strong correlations between response fractions and scores in all treatments. Scoring thresholds S_Glu ≥ 3 (AUC = 0.911, 95% CI, 0.865-0.956), S_TPO-RAs ≥ 5 (AUC = 0.964, 95% CI 0.934-0.994), and S_Ritu = 3 (AUC = 0.964, 95% CI 0.915-1.000) indicated ineffectiveness in glucocorticoid, TPO-RAs, and rituximab therapy, respectively. Regression analysis and ML established a tentative and preliminary predictive scoring model for advancing individualised treatment.