The long-term evolution of epidemic processes depends crucially on the structure of contact networks. As empirical evidence indicates that human populations exhibit strong community organization, we investigate here how such mesoscopic configurations affect the likelihood of epidemic recurrence. Through numerical simulations on real social networks and theoretical arguments using spectral methods, we demonstrate that highly contagious diseases that would have otherwise died out rapidly can persist indefinitely for an optimal range of modularity in contact networks.