Wireless sensing has emerged as a key enabling technology for Integrated Sensing and Communication (ISAC) systems, where radio resources are jointly utilized for environmental sensing and communication. However, under limited radio resources, continuously improving sensing accuracy may lead to inefficient resource utilization once the sensing requirement has been satisfied. To address this issue, this paper proposes a Predictive Marginal Value of Sensing (PM-VoS) based resource allocation framework for dynamic wireless sensing systems. The proposed PM-VoS evaluates the long-term marginal utility of current sensing actions by propagating sensing and non-sensing covariance states over a finite prediction horizon. Based on this metric, a hybrid resource allocation algorithm is developed by combining analytical one-step VoS evaluation with learning-assisted residual predictive value estimation. Simulation results in multi-user multi-target scenarios demonstrate that the proposed scheme reduces sensing position error and improves radio resource utilization efficiency compared with non-predictive VoS-based, heuristic, and random allocation methods.