Astatine isotopes can be produced in liquid lead-bismuth eutectic targets through proton-induced double charge exchange reactions on bismuth or in secondary helium-induced interactions. Models implemented into the most common high-energy transport codes generally have difficulties to correctly estimate their production yields as was shown recently by the ISOLDE Collaboration, which measured release rates from a lead-bismuth target irradiated by 1.4 and 1 GeV protons. In this paper, we first study the capability of the new version of the Liège intranuclear cascade model, INCL4.6, coupled to the deexcitation code ABLA07 to predict the different elementary reactions involved in the production of such isotopes through a detailed comparison of the model with the available experimental data from the literature. Although a few remaining deficiencies are identified, very satisfactory results are found, thanks in particular to improvements brought recently on the treatment of low-energy helium-induced reactions. The implementation of the models into MCNPX allows identifying the respective contributions of the different possible reaction channels in the ISOLDE case. Finally, the full simulation of the ISOLDE experiment is performed, taking into account the likely rather long diffusion time from the target, and compared with the measured diffusion rates for the different astatine isotopes, at the two studied energies, 1.4 and 1 GeV. The shape of the isotopic distribution is perfectly reproduced as well as the absolute release rates, assuming in the calculation a diffusion time between 5 and 10hours. This work finally shows that our model, thanks to the attention paid to the emission of high-energy clusters and to low-energy cluster induced reactions, can be safely used within MCNPX to predict isotopes with a charge larger than that of the target by two units in spallation targets, and, probably, more generally to isotopes created in secondary reactions induced by composite particles.