Author: Mark Ashamalla, Renee Farrell, Jinkoo Kim, Kartik Mani, Xin Qian, Samuel Ryu, Yizhou Zhao 👨🔬
Affiliation: Stony Brook Medicine, Stony Brook University Hospital 🌍
Purpose: Adaptive planning is increasingly used in head and neck radiation therapy due to factors like tumor response or changes in patient anatomy. However, methods such as resimulation or offline replanning can lead to treatment interruptions, which are linked to poorer outcomes. This study aims to develop a predictive model using clinical data and radiation therapy plan parameters to identify cases likely to require adaptive planning or resimulation, thereby minimizing treatment breaks.
Methods: Data from 356 patients (468 treatment plans) undergoing head and neck radiation therapy were analyzed. Treatment interruption was determined by remaining fractions in a treatment associated with a resimulation/replanning. Two models, Gaussian Naive Bayes (GNB) and Random Forest (RF), were trained. GNB was chosen for its simplicity, efficiency, and probabilistic outputs, while RF was selected for its ability to capture non-linear relationships, robustness to overfitting, and feature importance analysis. Nine features were inputted into the model. Data was split into 70% training, 15% validation and 15% testing. Adaptive Synthetic Sampling addressed class imbalance, and sensitivity was prioritized to reduce false negatives. Model performance was evaluated using accuracy, AUC, and precision.
Results: Of the 468 plans, 40 required replanning; After oversampling, the dataset included 517 plans with 119 positive labels. Features included metastatic status, inpatient status, obesity/diabetes history, treatment beams, plan type, monitor units, bolus use, and field size. The GNB model achieved moderate accuracy (53%) and AUC (0.68), with high recall (0.77). The RF model performed better, with balanced accuracy (80%) and AUC (0.83), demonstrating strong predictive capability.
Conclusion: Machine learning shows promise in predicting the need for adaptive planning and resimulation in head and neck radiation therapy. The RF model outperformed GNB as benchmark, highlighting its potential for clinical application. Future research will focus on feature engineering and incorporating additional variables to enhance predictive accuracy.