Machine learning-driven probability of cancer therapy–related cardiac dysfunction in patients with non-Hodgkin lymphoma
D BursacovschiAbstract
Introduction
Cardiovascular complications remain a major cause of morbidity and mortality among patients with malignancies, particularly following exposure to various antitumoral treatments. In patients with non-Hodgkin lymphoma (NHL), early detection of cancer therapy–related cardiac dysfunction (CTRCD) is essential to prevent progression to heart failure and improve clinical outcomes.
Materials and Methods
A prospective cohort study included 127 adult patients with NHL undergoing chemotherapy. Baseline clinical, electrocardiographic, echocardiographic, cardiopulmonary exercise testing, and treatment-related parameters were collected prior to therapy initiation and at 6-month follow-up. CTRCD was defined according to the 2022 ESC cardio-oncology guidelines. A machine learning–based predictive model was developed using baseline features to estimate the probability of CTRCD occurrence. The predictive performance of the model was assessed using confusion matrix analysis. Model interpretability was evaluated using SHapley Additive exPlanations, while permutation feature importance analysis was applied to determine the relative contribution of each variable to the model’s predictive performance.
Results
CTRCD was identified in 14.2% of patients at 6-month follow-up. Model performance evaluated by confusion matrix analysis demonstrated an overall accuracy of 94.2%. ROC analysis showed excellent discriminative ability, with an AUC of 0.95, while precision–sensitivity analysis yielded an AUC of 0.824 (figure 1). SHAP analysis identified baseline SDNNi and mean heart rate from 24-hour Holter ECG as the strongest predictors of CTRCD, followed by cumulative doxorubicin dose and NT-proBNP levels. Additional contributors included prolonged QTc interval, advanced age, indexed left atrial volume, and increased E/e′ ratio. Permutation feature importance confirmed SDNNi as the most influential predictor, whereas treatment regimen and rituximab administration showed lower predictive impact (figure 2).
Conclusion
The machine learning–based multimodal model showed high accuracy in predicting CTRCD risk in patients with NHL supporting its potential utility for early risk stratification and personalized cardiovascular monitoring during antitumor treatment.Sensitivity and accuracy of the CTRCD modelPermutation and SHAP impact on CTRCD model