Author: Leslie Harrell, Sanjay Maraboyina, Romy Megahed, Maida Ranjbar, Xenia Ray, Pouya Sabouri 👨🔬
Affiliation: Department of Radiation Oncology, University of Arkansas for Medical Sciences (UAMS), University of California San Diego 🌍
Purpose: Real-time adaptive radiation therapy (ART) dynamically modifies patients’ treatment plan during delivery to account for anatomical and physiological variations. Addressing ART planning time variability is critical for improving patient throughput and efficiency. However, workflow efficiency is hindered by variability in treatment durations caused by clinical and technical factors. This study develops and evaluates machine learning algorithms to predict ART session duration by accounting for these factors, aiming to improve scheduling and workflow efficiency for ART.
Methods: 543 ART fractions were analyzed, using clinician experience and target volume as predictive variables and ART Time, defined as the interval from initial CBCT to final plan sign-off, as the response variable. Two machine learning models were developed: i) hybrid model incorporating an exponential decay function with polynomial terms and ii) Generalized Additive Model (GAM), a flexible regression model using smooth spline-based terms. Validation included internal testing with leave-points-out cross-validation (LPO) and external validation on data from an independent clinic. Predictive performance was assessed using the coefficient of determination (R²), mean absolute error (MAE), and root mean squared error (RMSE), with scatter plots and error distributions analyzed for reliability and validity.
Results: Treatment duration ranged from 5-43 minutes, averaging 18 minutes (standard-deviation=6.5) per fraction. On the training set, the Hybrid model achieved R²= 0.53, MAE=3.44 and RMSE= 4.41 minutes, while GAM performed better (R²= 0.66, MAE= 2.84, RMSE= 3.76). In external validation, the Hybrid model had R²= 0.27 while GAM failed to predict the response variable (R²=-0.94). In LPO validation, the Hybrid model showed better performance (R²= 0.79, MAE= 2.18 minutes).
Conclusion: A reliable predictive model can enhance ART workflow efficiency by optimizing treatment scheduling. The models predicted ART Time with errors within minutes, demonstrating clinical potential. The Hybrid model showed greater generalizability, while GAM captured patient variability but failed for external validation.