Skip to main content

Table 3 Three basic models use different types of pre-trained word embeddings to predict performance

From: Refining electronic medical records representation in manifold subspace

Method Embedding Macro AUC Micro AUC Macro F1 Micro F1 Test loss value Top-10 recall
RNN Random 0.854 0.972 0.204 0.653 0.032 0.772
FastText 0.842 0.973 0.149 0.628 0.032 0.774
Glove 0.861 0.974 0.219 0.656 0.031 0.788
Word2Vec 0.851 0.974 0.165 0.642 0.031 0.783
BERT 0.500 0.908 0.000 0.000 0.061 0.442
ALBERT 0.503 0.915 0.026 0.018 0.054 0.446
BioBERT 0.513 0.923 0.051 0.038 0.052 0.457
BlueBERT 0.533 0.939 0.075 0.043 0.050 0.471
Ours 0.857 0.976 0.182 0.659 0.030 0.793
CNN Random 0.825 0.968 0.214 0.626 0.040 0.753
FastText 0.665 0.921 0.012 0.223 0.053 0.488
Glove 0.842 0.972 0.188 0.622 0.034 0.767
Word2Vec 0.692 0.925 0.021 0.313 0.052 0.492
BERT 0.549 0.906 0.000 0.000 0.059 0.442
ALBERT 0.556 0.914 0.014 0.012 0.053 0.453
BioBERT 0.559 0.921 0.015 0.041 0.047 0.459
BlueBERT 0.567 0.929 0.021 0.047 0.042 0.464
Ours 0.852 0.974 0.217 0.628 0.038 0.779
CAML Random 0.855 0.978 0.257 0.656 0.032 0.806
FastText 0.856 0.980 0.270 0.656 0.031 0.809
Glove 0.867 0.978 0.272 0.647 0.033 0.801
Word2Vec 0.855 0.980 0.274 0.662 0.030 0.813
BERT 0.497 0.908 0.000 0.000 0.058 0.442
ALBERT 0.505 0.916 0.026 0.022 0.054 0.457
BioBERT 0.513 0.924 0.045 0.041 0.048 0.465
BlueBERT 0.534 0.934 0.060 0.076 0.042 0.478
Ours 0.886 0.982 0.270 0.673 0.029 0.823
  1. Bold values denote the best result for each row of data(%)