What a size-only threshold for holding cancer treatment loses in a spatial tumour model

Abstract Gallagher et al. derived a size threshold for holding adaptive therapy between clinic visits in non-spatial ordinary differential equation (ODE) models. I tested it in one spatial agent-based model (ABM) against a tolerance of 0.05 for premature progression (progression during a treatment hold). At the derived threshold, premature progression occurred in 0.976, 0.974 and 0.480 of runs at intervals of 7, 14 and 28 days, and at most 0.014 at 42 to 84 days. The ABM’s safe threshold differed from the ODE’s own by − 4.40% to − 8.95% at 7 to 56 days, and recalibration to the ABM’s drug-free regrowth did not remove this shortfall. Most of it (84.2% to 94.1% at 7 to 28 days) arose in the step from a well-mixed stochastic model to the ABM; that step changes several features at once, so no mechanism is identified. With error-free size readings, measurement history reduced the test log-loss of predicting premature progression at equal size by 48.6%; with 10% measurement error, history was only 2.0% better than exact size. Under a common nominal risk constraint, a hold rule using history and image geometry changed time on drug by − 1.4 percentage points and restricted mean time to progression by + 4.0 days over 750 days (+ 0.95 over 365). The gain rested on two of four intervals; non-inferiority sat on the margin at 7 days and was not shown at 14. On new runs its risk was not shown to meet the tolerance at any interval (upper Wilson bounds above 0.05).