Development of an interpretable machine learning model for venous thromboembolism risk prediction in multiple myeloma patients

To address the limited discrimination of existing clinical risk assessment tools for venous thromboembolism (VTE) in patients with multiple myeloma (MM), this study aims to construct an interpretable machine learning-based VTE risk prediction model. This study retrospectively examined data from 517 MM patients, finding a 10.06% VTE incidence. Recursive feature elimination selected final variables from 43 candidates. Four machine learning models—logistic regression, random forest, support vector machine, and CatBoost—were developed and compared using receiver operating characteristic curves, with Shapley Additive Explanations (SHAP) used for model interpretation. Seven key predictive variables were identified: serum total calcium, folate, serum iron, potassium, age, BMI, and D-dimer. The random forest model achieved optimal diagnostic performance with a C-statistic of 0.888 (95% CI 0.804–0.972) in the test set, and sensitivity, specificity, and accuracy of 0.812, 0.900, and 0.891, respectively. SHAP analysis showed that serum total calcium, folate, serum iron, and potassium had greater predictive significance than BMI, D-dimer, and age. This study's model utilizes routine clinical indicators to potentially identify MM patients at high VTE risk. By incorporating SHAP interpretability to highlight metabolism-related and traditional risk factors, it could become a quantitative tool for personalized VTE prevention in MM patients, subject to further validation.