Predicting the 3-year overall survival in patients with Waldenström Macroglobulinemia using machine learning algorithms

Abstract Waldenström Macroglobulinemia (WM) is a rare, incurable B-cell lymphoma characterized by heterogeneous clinical outcomes. Accurate prognosis prediction is critical for risk-adapted therapeutic strategies. We aimed to develop an interpretable machine learning (ML)-based model to predict the 3-year overall survival (OS) of WM patients. A retrospective cohort of 179 WM patients from 4 Chinese centers was included. Patients were partitioned into training (n = 134, 75%) and test sets (n = 45, 25%) through stratified randomization to maintain population representativeness. Five ML algorithms were systematically evaluated: categorical boosting (CatBoost), extreme gradient boosting (XGBoost), light gradient boosting machine (LGBM), Random Forest, and Logistic Regression. Recursive Feature Elimination with cross-validation (RFECV) is a stepwise feature selection method that retains the most prognostically meaningful indicators, and we used it to screen key features. Model performance was assessed via area under the curve (AUC), sensitivity, and specificity, with Shapley Additive Explanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) tools applied to explain the model’s predictive logic for clinical use. The CatBoost algorithm retained 20 pivotal features through RFECV screening, forming an optimized feature subset. The model demonstrated exceptional discriminative performance with training and test set AUCs of 0.9322 [95% Confidence interval (CI): 0.901–0.963] and 0.8235 (95% CI: 0.761–0.886), respectively. SHAP-based feature importance analysis identified treatment status, International Prognostic Scoring System for WM (IPSSWM) risk stratification, and hepatomegaly as critical determinants of 3-year mortality prediction, demonstrating clinically actionable biological interpretability. This ML framework provides a robust, interpretable tool for 3-year OS prediction in WM, enabling personalized risk stratification and timely clinical interventions.