The first phase of each structural monitoring project is the operational evaluation. Its purpose is to define relevant damage mechanisms and environmental conditions, to consider the data acquisition limitations on site, and to justify the investment. Subsequently, relevant measurement quantities and damagesensitive features are selected, but very few systematic approaches exist in the literature on how to select the most appropriate features. The presented paper fills this gap and develops an approach to select damage-sensitive features based on probability of detection (POD) curves. The POD curves are generated based on a novel method for statistical damage detection tests that requires a finite element model and vibration data from the undamaged structure. However, no data is required from the damaged state, making it particularly suited for unique or large and complex engineering structures. The approach explicitly considers the uncertainties in the features due to unknown loads, measurement noise, and short measurement durations. Although global damage-sensitive features are considered, such as modal parameters and subspace-based residuals, the detectability is evaluated for local structural components. The paper includes a proof of concept study on a laboratory structure. The results demonstrate that the developed method successfully finds the feature with the highest damage detectability for a chosen damage scenario, and that the detectability varies depending on the monitored local component.