Table 1.

Performance of all the models on the 2010 i2b2/VA dataset.

Model	F-1 (%)	Precision (%)	Recall (%)
Hidden semi-Markova	85.23	86.88	83.64
LSTMb by Liu et al [39]a	85.78	—c	—c
LSTM by Wu et al [43]a	85.94	85.33	86.56
BiLSTMd + ELMo by Zhu et al [40]a	86.84 (0.16)	87.44 (0.27)	86.25 (0.26)
BiLSTM + Flair	87.01 (0.18)	87.54 (0.15)	86.49 (0.21)
BiLSTM + ELMo	87.01 (0.24)	87.64 (0.19)	86.40 (0.30)
BiLSTM + ELMo + Flair	87.30 (0.06)	87.78 (0.09)	86.85 (0.07)
BiLSTM + ELMo + Flair + semantic embedding	87.44 (0.07)	88.03 (0.14)	86.91 (0.10)

^aModel is trained using the complete dataset of i2b2 2010, which contains 349 notes in the training set and 477 notes in the test set.

^bLSTM: long short-term memory.

^cNot reported.

^dBiLSTM: bidirectional LSTM.