Table 1 A brief summary of several clinical concept embedding studies. Only the largest database used in the study was listed
From: Time-sensitive clinical concept embeddings learned from large electronic health records
Study | Method | Data source | patient size | Time-sensitive | Evaluation strategies |
|---|---|---|---|---|---|
Med2vec [10] | word2vec | EHR/ claims | < 1 million | No | Similarity based on vocabularies, predictive modeling and human assessment |
Cui2vec [11] | word2vec, glove | claims | 60 million | Only in negative sampling for word2vec | Similarity based on vocabularies and human assessment |
MCE [13] | attention- word2vec | EHR | < 2 million | With an attention layer | Similarity based on vocabularies |
Ours’ | word2vec, PMIa, FastText | EHR | 50 million | Dynamic input windows | Similarity based on vocabularies, and predictive \modeling |