Traditional safety-critical systems are engineered in a way to be predictable in all operating conditions. They are common in industrial automation and transport applications where uncertainties (e.g., fault occurrence rates) can be modelled and precisely evaluated. Furthermore, they use high-cost hardware components to increase system reliability. On the contrary, future systems are increasingly required to be “smart” (or “intelligent”) that is to adapt to new scenarios, learn and react to unknown situations, possibly using low-cost hardware components. In order to move a step forward to fulfilling those new expectations, in this paper we address run-time stochastic evaluation of quantitative safety targets, like hazard rate, in self-adaptive event detection systems by using Bayesian Networks and their extensions. Self-adaptation allows changing correlation schemes on diverse detectors based on their reputation, which is continuously updated to account for performance degradation as well as modifications in environmental conditions. To that aim, we introduce a specific methodology and show its application to a case-study of vehicle detection with multiple sensors for which a real-world data-set is available from a previous study. Besides providing a proof-of-concept of our approach, the results of this paper pave the way to the introduction of new paradigms in the dynamic safety assessment of smart systems.