ESTRO 2024 - Abstract Book

S4453

Physics - Machine learning models and clinical applications

ESTRO 2024

XGBoost shows better performances, but the model tends to overfit, while Random Forest is more consistent between train and test results. Indeed, Random Forest accuracy with 0.5 threshold is 85.51% and 83.33% for train and test respectively, while XGBoost results were 86.96% and 77.78%. The default selection thresholds = 0.5 led to the models being very specific, but less sensitive. On the contrary, high sensitivity is crucial for this application: it’s more important to early identify failing plans, than measuring plans which will pass. So, we varied the threshold selection level to increase the models’ sensitivity (Table1):

Random Forest

XGBoost