Fig. 6
From: Master clinical medical knowledge at certificated-doctor-level with deep learning model
Example of medical knowledge obtained with “Free Reading” and “Guided Reading”. We use implicit embeddings to represent and capture medical knowledge from a large medical corpus. In the FR phase, we produce a series of embeddings to depict different kinds of medical knowledge. Take disease “Guillain-Barre syndrome” for example, in a “differential-diagnosis” embedding subspace similar diseases are close to each other, while in b “symptom” embedding subspace medical terms describing symptom of Guillain-Barre syndrome are clustering, and while in c “examination” embedding subspace medical terms related to examination to “Guillain-Barre syndrome” are clustering. In the GR phase, medical terms’ embeddings obtained from FR are fine-tuned by supervised learning models to get more rich and precise medical knowledge based on their context. For example, in d the embedding of “nephritis” is fine-tuned into different variants to describe and represent subtle different meanings in different context
