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

Table 2 Performance of LSTM-CRFs trained with different word embeddings (trained using i2b2 training set and evaluated using i2b2 validation set)

Model	Embedding	Performance on validation set (i2b2/UTHealth)
		Strict			Relax
		Precision	Recall	F1 score	Precision	Recall	F1 score
LSTM-CRFs	GoogleNews	0.9679	0.9263	0.9466	0.9783	0.9362	0.9567
	CommonCrawl	0.9697	0.9401	0.9547	0.9797	0.9498	0.9646
	MIMIC-word2vec	0.9669	0.9341	0.9502	0.9774	0.9443	0.9606
	MIMIC-fastText	0.9631	0.9380	0.9504	0.9758	0.9504	0.9629
	MADE	0.9662	0.9158	0.9403	0.9782	0.9271	0.9520

ISSN: 1472-6947