The use of mathematical models to make predictions about tumor growth and response to treatment has become increasingly prevalent in the clinical setting. The level of complexity within these models ranges broadly, and the calibration of more complex models requires detailed clinical data. This raises questions about how much data should be collected and when, in order to minimize the total amount of data used and the time until a model can be calibrated accurately. To address these questions, we propose a Bayesian information-theoretic calibration protocol for experimental design, using a gradient-based score function to identify optimal times at which to collect data for informing treatment parameters. We illustrate this framework by calibrating a simple ordinary differential equation model of tumor response to radiotherapy to a set of synthetic data.