Today's Web is social and largely driven by a wide variety of online communities. Many such communities are owned and managed by businesses that draw much value from these communities, in the form of efficient and cheaper customer support, generation of new ideas, fast spreading of information, etc. Understanding how to measure the health of online communities and how to predict its change over time, whether to better or to worse health, is key to developing methods and policies for supporting these communities and managing them more efficiently. In this paper we investigate the prediction of community health based on the social behaviour exhibited by their members. We apply our analysis over 25 SAP online communities, and demonstrate the feasibility of using behaviour analysis to predict change in their health metrics. We show that accuracy of health prediction increases when using community-specific prediction models, rather than using a one-model-fits-all approach.