Accurate prediction of beam verification time for online model checking scheduling

Develop a machine-learning-based predictor that estimates the duration of Uppaal SMC statistical model-checking queries used for beam feasibility verification in robotic radiation therapy, with accuracy high enough to be useful for prioritizing verification tasks during online scheduling, using real patient respiratory motion data.

Background

The paper proposes using online model checking to dynamically select feasible radiation therapy beams based on current respiratory motion, aiming to reduce idle time without compromising treatment quality. To further improve scheduling decisions, the authors explored machine learning to predict verification durations and classify breathing patterns, which could help prioritize which beams to verify within tight time windows.

Their experiments with a multilayer perceptron regressor to predict verification times consistently produced low-end estimates and failed to capture meaningful variability, particularly on real patient data. Because accurate predictions would enable better prioritization of verification queries, the inability to obtain sufficiently accurate predictions leaves the usefulness of such a predictor unresolved.