3073 - Predicting Patient-Specific Benefits in Pancreatic Cancer Adaptive Radiotherapy: A Machine Learning Approach

I. Shuryak¹, A. W. Lee¹, Y. F. Wang², L. A. Kachnic¹, D. P. Horowitz¹, and A. C. Riegel²; ¹Columbia University, New York, NY, ²Columbia University Irving Medical Center, New York, NY

Purpose/Objective(s): CBCT-based online adaptive radiotherapy continually adjusts the treatment plan to accommodate changes in the patient’s body during radiation therapy, utilizing the daily CBCT images, to improve the accuracy of tumor targeting. However, this technique is more complex and resource intensive compared with non-adaptive planning and may not benefit all patients equally. The goal of this study was to identify which features of scheduled non-adaptive plans, and which patient clinical characteristics, can predict substantial benefits from adaptive radiotherapy. Materials/

Methods: Data from 37 pancreatic cancer patients treated with a simultaneous boost (SIB) dose scheme of 8 Gy x 5 and 5 Gy x 5 in adaptive radiotherapy at our institution were analyzed. The potential benefit for the adaptive plan vs the scheduled plan for each patient was assessed using the following four metrics: differences in maximum “hotspot” doses for the most-irradiated 0.03 cc of stomach or bowel, and differences between PTV2500 or PTV4000 between plans. Machine learning (ML) techniques were employed to identify variables predicting the adaptive-scheduled difference for each metric. The data were split 75:25 into training and testing sets. The Boruta algorithm selected the most important predictors. Several ML algorithms (linear regression, random forest, XGBoost, LightGBM, CatBoost, LinBoost) were applied with repeated cross-validation. The stacking approach integrated these diverse ML models to train a meta-model for optimal predictions on the testing data.
Results: The study found that scheduled plan features, rather than clinical variables, strongly predict adaptive plan metrics. Notably, there were nonlinear relationships between maximum scheduled and adaptive organ doses for both the stomach and bowel. The advantage of the adaptive plan became evident only at high scheduled doses (>600 cGy for both organs). The stacking ML approach adequately predicted the difference between adaptive and scheduled plans, achieving R² 0.725, root mean squared error (RMSE) 61.7 cGy, mean absolute error (MAE) 38.6 cGy on testing data for stomach, and even better performance for bowel: R² 0.898, RMSE 32.2 cGy, MAE 26.1 cGy. The differences in PTV metrics between adaptive and scheduled plans were most pronounced when the scheduled PTV coverages were <90%.
Conclusion: This research advances personalized medicine in radiotherapy planning. Considering patient-specific scheduled plan features and characteristics is crucial for predicting the efficacy of online adaptive radiotherapy for pancreatic cancer. Patients with high organ doses under the non-adaptive scheduled plan and sub-optimal PTV coverages stand to benefit most strongly from adaptive planning.