Harnessing Machine Learning to Predict Chronic Gvhd: A Novel Risk Stratification Score

Le résumé fourni par la source

Introduction Chronic graft-versus-host disease (cGVHD) is a significant complication following hematopoietic stem cell transplantation (HSCT), impacting patient outcomes and quality of life . Timely prediction of cGVHD onset and severity is crucial for effective management. Recent advances in machine learning (ML) offer opportunities to develop predictive models that can support clinical decision-making. This study applies ML to predict cGVHD onset and severity, focusing on model performance through confusion matrix analysis and the identification of key predictive features. Methods We analyzed de-identified data from the CIBMTR on post-HSCT patients (2008-2017) to build ML models for predicting cGVHD. Using Optuna, 50 optimization trials were conducted with five-fold cross-validation. During the evaluation phase, test accuracy provided an unbiased measure of the model's performance on unseen data. The confusion matrix offered insights into classification errors and the trade-offs between sensitivity and specificity. Results As shown in Figure 1 , the confusion matrix for predicting cGVHD onset revealed a sensitivity of 73.2%, meaning the model correctly identified patients with symptoms. However, specificity was lower at 44.25%, indicating many patients without cGVHD were misclassified. Figure 2 illustrates the feature importance analysis, where the Random Forest model identified antithymocyte globulin (ATG) as the most influential factor in predicting cGVHD severity, contributing 17.5% to the model's predictive capability, while acute GVHD ranked second, accounting for 12%. Together, these two features contributed nearly 30% to the model's overall performance. The model's overall accuracy was 59%, indicating moderate effectiveness in differentiating affected from unaffected patients. Conclusion The ML model demonstrates promising sensitivity but struggles with specificity. Key features like ATG offer valuable insights into cGVHD prediction, suggesting areas for further research. Expanding the dataset could improve model specificity and accuracy. Once the model's accuracy exceeds 75%, future work will focus on developing a decision support tool to help clinicians and patients make informed decisions about cGVHD risk and severity before transplant, potentially reducing the overuse of preventive medications.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.