Purpose: To investigate the effectiveness of using a knowledge-based prediction model to predict plan adaptation in head and neck (HN) proton patients during treatment. This model can be used to predict the need for adaptation earlier than the clinical practice, which is exclusively based on the evaluation of quality assurance CT (QACT).
Methods: Data were gathered from 177 HN patients. The correlation between dosimetric/clinical features and plan adaptation was calculated. These features are mean beam dose heterogeneity, mean max dose change in robust evaluation, age, tumor site, and surgery/chemo status. The average WET for each beam on the highest dose level CTV was calculated. The average and third-largest % WET change between the planning CT (pCT) and different QACTs were used as features. Two predictive machine learning models were trained: a fine tree model using WET differences between the pCT and first QACT (141 training/35 testing) and a naïve Bayes model using WET differences between the pCT and second QACT (91 training/23 testing). Both models were trained with 10-fold cross-validation. Confusion matrices and ROC curves were used for evaluation.
Results: All dosimetric/clinical features were significant (p < 0.01). The average % changes in WET for the first and second QACTs were not significant (p > 0.75). The third-largest % WET changes for the first and second QACTs were weakly significant (p < 0.10). The first QACT model and the second QACT model had overall prediction accuracies of 67.4% and 74.5%, sensitivity/specificities of 63%/70% and 63%/80%, and areas under the ROC curve of 0.70 and 0.78.
Conclusion: Both machine learning models showed strong predictive power for plan adaptation. The second QACT model, where WET changes are more extreme, showed more predictive power. These models can be integrated into plan adaptation workflow to catch adaptation earlier and improve patient outcomes.
Not Applicable / None Entered.