Radiotherapy response assessment commonly relies on scalar imaging metrics that may fail to capture spatially structured tumor dynamics. When tumor regions interact, such measures can obscure underlying coordination. We introduce a predictive framework to detect non-independent dynamics from longitudinal imaging data. The approach quantifies predictive improvement from spatial information using the Tumor Coupling Index (TCI), defined as the normalized reduction in prediction error between independent and spatial models. Simulations show that TCI remains near zero under independence, increases with spatial coupling, and collapses under randomization, demonstrating sensitivity and falsifiability. In contrast, conventional scalar metrics are insensitive to such structure. TCI provides a model-agnostic observable of non-independent tumor behavior, offering a principled approach for analyzing longitudinal imaging and potential applications in adaptive radiotherapy.