Table 3.

The overall performances of the FT-BERT and the FC-BERT methods on the i2b2 dataset. The Cased and Uncased models are original Google models; BioBERT corresponds to a model trained from PMC and PubMed data; and MIMIC BERT is trained on MIMIC III. P, R, and F represent precision, recall and F1 score, respectively. The boldface represents the best performance for each metric.

Dataset	Baseline				FT-BERT				FC-BERT
		SDP	Seq-CNN*	Seq-GCRN*	Cased	Uncased	Bio	MIMIC	Cased	Uncased	MIMIC
i2b2	P	0.7569	0.748	0.772	0.7489	0.7673	0.7346	0.7624	0.6885	0.6843	0.6850
	R	0.7303	0.736	0.743	0.7606	0.7681	0.7681	0.7734	0.6991	0.7056	0.7521
	F	0.7434	0.742	0.758	0.7547	0.7677	0.751	0.7679	0.6938	0.6948	0.7170

*These two methods were developed and evaluated on the original dataset of the 2010 i2b2 challenge.