Predicting missing and noisy links via neighbourhood preserving graph embeddings in a clinical knowlegebase