Structural health monitoring is a promising technology to automatically detect structural changes based on permanently installed sensors. Vibration-based methods that evaluate the global system response to ambient excitation are suited to diagnose changes in boundary conditions, i.e., changes in member prestress or imposed displacements. In this paper, these changes are evaluated based on sensitivity-based statistical tests, which are capable of detecting and localizing parametric structural changes. The main contribution is the analytical calculation of sensitivity vectors for changes in boundary conditions (i.e., changes in prestress or support conditions) based on stress stiffening, and the combination with a numerically efficient algorithm, i.e., Nelson's method. One of the main advantages of the employed damage diagnosis algorithm is that, although it uses physical models for damage detection, it considers the uncertainty in the data-driven features, which enables a reliabilitybased approach to determine the probability of detection. Moreover, the algorithm can be trained and the probability of detecting future damages can be predicted based on data and a model from the undamaged structure, in an unsupervised learning mode, making it particularly relevant for unique structures, where no data from the damaged state is available. For proof of concept, a numerical case study is presented. The study assesses the loss of prestress in a two-span reinforced concrete beam and showcases suitable validation approaches for the sensitivity calculation.