Comparing neural language models for medical concept representation and patient trajectory prediction

Alban Bornet, Dimitrios Proios, Anthony Yazdani, Fernando Jaume-Santero, Guy Haller, Edward Choi, Douglas Teodoro

Abstract: Highlights•Language models can represent medical concept semantics via data-driven embeddings•FastText aligns with the hierarchy of medical terminologies using subword information•GloVe's global scope proves advantageous in patient trajectory prediction•Word2vec and GloVe outperform FastText in clinical outcome prediction