Specifications of Radio Frequency (RF) analog integrated circuits have increased strictly as their applications tend to be more complicated and high test cost demanding. This makes them very expensive due to an increased test time and to the use of sophisticated test equipment. Alternative test measures, extracted by means of Built-In Self Test (BIST) techniques, are useful approaches to replace standard specification-based tests. One way to evaluate the efficiency of the CUT measures at the design stage is by estimating the Test Escapes (T E) and the Yield Loss (Y L) at ppm level. Unfortunately, an important number of Monte Carlo simulations must be run in order to guarantee their accuracy. For certain types of circuits, this requires many months or even years to generate millions of circuits. To overcome this limitation, we present in this paper a new technique where a small number of simulations is sufficient to reach an important precision. This method is based on a classification using machine learning methods, such as SVM and Neural Networks based classifiers to determine pass/fail regions. The proposed approach requires a few number of simulations only to determine the region separating the process parameters generating good and faulty, or pass and fail circuits. Then only this region is needed to estimate the test metrics without running any additional simulation. The proposed methodology is illustrated for the evaluation of a filter BIST technique.