With the availability of accessible and widely used cloud services, it is natural that large components of healthcare systems migrate to them; for example, patient databases can be stored and processed in the cloud. Such cloud services provide enhanced flexibility and additional gains, such as availability, ease of data share, and so on. This trend poses serious threats regarding the privacy of the patients and the trust that an individual must put into the healthcare system itself. Thus, there is a strong need of privacy preservation, achieved through a variety of different approaches. In this paper, we study the application of a random projection-based approach to patient data as a means to achieve two goals: (1) provably mask the identity of users under some adversarial-attack settings, (2) preserve enough information to allow for aggregate data analysis and application of machine-learning techniques. As far as we know, such approaches have not been applied and tested on medical data. We analyze the trade-off between the loss of accuracy on the outcome of machine-learning algorithms and the resilience against an adversary. We show that random projections proved to be strong against known input/output attacks while offering high quality data, as long as the projected space is smaller than the original space, and as long as the amount of leaked data available to the adversary is limited.