Radiation therapy delivers dose over many fractions, and the patient's anatomy changes between them. Adaptive radiotherapy monitors these changes and adjusts the plan, but the decision to replan is typically reactive -- triggered when something looks wrong on an image. Robert Boyd and Wolfgang Tome propose treating dose accumulation as a stochastic time series and applying exponential smoothing to forecast where the cumulative dose is heading. Their framework generates not just point predictions but closed-form prediction intervals, tested on 32 prostate and 19 head and neck cancer patients.
The reframing is fundamental. Standard adaptive radiotherapy treats each fraction as a snapshot: acquire image, compare to plan, decide whether deviation is large enough to trigger replanning. This is memoryless -- each decision ignores the trajectory. Boyd and Tome's approach models dose as a process with momentum: if the cumulative dose has been drifting upward for three fractions, the forecast shows where it will be in five more, with uncertainty bounds. The exponential smoothing weights recent fractions more heavily, adapting to the patient's specific pattern of anatomical change rather than applying a population-level threshold. The difference-based formulation adds a second layer: it estimates not just the forecast under the current plan but the forecast under an adapted plan, enabling a direct comparison of whether replanning improves the expected outcome. This transforms the replanning decision from a threshold trigger to a probabilistic cost-benefit analysis. The dose is no longer a number to be checked but a trajectory to be steered, and the prediction interval tells you how much steering room you have.
(arXiv:2603.01130)