Applying a deep learning-based sequence labeling approach to detect attributes of medical concepts in clinical text

Table 4 The overall performance of different approaches on the i2b2-medication dataset in detecting 5 attributes of given medications: dosage (dos), mode (mod), frequency (fre), duration (dur), reason (rea). best results are shown in boldface

Attribute		DOS	MOD	FRE	DUR	REA
Baseline (Bi-LSTM-CRF + SVM)	Acc.	0.9201	0.9584	0.9353	0.9783	0.9473
	P	0.8794	0.9110	0.8762	0.5945	0.5373
	R	0.9292	0.9597	0.9390	0.6680	0.6704
	F	0.9036	0.9347	0.9065	0.6291	0.5965
Baseline (Bi-LSTM-CRF + Bi-LSTM)	Acc.	0.9250	0.9559	0.9302	0.9680	0.9269
	P	0.9305	0.9372	0.9198	0.6168	0.5984
	R	0.9434	0.9658	0.9399	0.6525	0.5717
	F	0.9369	0.9513	0.9298	0.6341	0.5848
Sequence Labeling	Acc.	0.9573	0.9807	0.9556	0.9802	0.9589
	P	0.9728	0.9773	0.9503	0.7785	0.7409
	R	0.9159	0.9528	0.9078	0.4479	0.4953
	F	0.9435	0.9649	0.9286	0.5686	0.5938
Usyd [14]	P	0.9189	0.9073	0.9142	0.5604	0.6687
	R	0.8678	0.8915	0.8795	0.3709	0.3319
	F	0.8926	0.8994	0.8965	0.4464	0.4436

ISSN: 1472-6947