Table 4 Comparison among embeddings for weakly supervised Text-to-UMLS linking from MIMIC-III discharge summaries
From: Ontology-driven and weakly supervised rare disease identification from clinical notes
validation (n=142+/400) | test (n=187+/673) | |||||
|---|---|---|---|---|---|---|
Text to UMLS | P | R | \(F_{1}\) | P | R | \(F_{1}\) |
Word2Vec-100 | 86.6 | 50.0 | 63.4 | 85.1 | 61.0 | 71.0 |
Word2Vec-300 | 85.7 | 59.2 | 70.0 | 80.7 | 69.5 | 74.7 |
Word2Vec-768 | 85.1 | 68.3 | 75.8 | 78.9 | 78.1 | 78.5 |
BERT | 88.1 | 83.8 | 85.9 | 79.5 | 91.4 | 85.1 |
PubMedBERT | 88.7 | 77.5 | 82.7 | 79.6 | 87.7 | 83.5 |
SapBERT | 88.3 | 79.6 | 83.7 | 80.8 | 89.8 | 85.1 |
BlueBERT-base | 90.1 | 89.4 | 89.8 | 80.4 | 92.0 | 85.8 |
+ fine-tuning | 84.6 | 88.7 | 86.6 | 73.5 | 92.0 | 81.7 |
BlueBERT-large | 89.1 | 80.3 | 84.4 | 79.0 | 88.8 | 83.6 |