Physicians face increasing pressure to provide effective medical treatment. The complexity of the healthcare system, which is dynamic, multi-disciplinary, flexible, humancentered, and knowledge-intensive, makes it difficult for physicians to handle this pressure. Existing support systems for physicians often only provide basic documentation capabilities. Predictive process monitoring, a new area of research, offers techniques that can predict the future behavior of ongoing medical processes. This paper goes one step further by exploring the use of these predictions to make recommendations to physicians. We compare different neural network types, topologies, and input encodings with relevant benchmark techniques to identify the best options for next-activity recommendations in healthcare, based on accuracy, consistency, training speed, and prediction speed. We use data from three healthcare processes, including a real-life emergency medicine department, to evaluate the performance of these techniques.