In medical dialogue summarization, summaries must be coherent and must\ncapture all the medically relevant information in the dialogue. However,\nlearning effective models for summarization require large amounts of labeled\ndata which is especially hard to obtain. We present an algorithm to create\nsynthetic training data with an explicit focus on capturing medically relevant\ninformation. We utilize GPT-3 as the backbone of our algorithm and scale 210\nhuman labeled examples to yield results comparable to using 6400 human labeled\nexamples (~30x) leveraging low-shot learning and an ensemble method. In\ndetailed experiments, we show that this approach produces high quality training\ndata that can further be combined with human labeled data to get summaries that\nare strongly preferable to those produced by models trained on human data alone\nboth in terms of medical accuracy and coherency.\n