A study of deep learning methods for de-identification of clinical notes in cross-institute settings

Table 3 Performance of LSTM-CRFs models on UF test set

Model	Training data	Fine Tuning	Performance on UF Test
			Strict			Relax
			Pre	Rec	F1	Pre	Rec	F1
LSTM-CRFs	i2b2	NA	0.8883	0.8274	0.8568	0.9288	0.8651	0.8958
LSTM-CRFs+Lexical	i2b2	NA	0.8767	0.8509	0.8636	0.9314	0.9041	0.9175
LSTM-CRFs+Lexical + Knowledge	i2b2	NA	0.8767	0.8706	0.8736	0.9229	0.9166	0.9197
LSTM-CRFs+Lexical + Knowledge	i2b2	UF	0.9474	0.9109	0.9288	0.9776	0.9400	0.9584
LSTM-CRFs+Lexical + Knowledge	UF	NA	0.9408	0.8992	0.9195	0.9705	0.9277	0.9486
LSTM-CRFs+Lexical + Knowledge	i2b2 + UF	NA	0.9352	0.9163	0.9257	0.9681	0.9484	0.9582

ISSN: 1472-6947