Using machine learning for concept extraction on clinical documents from multiple data sources

Manabu Torii; Kavishwar Wagholikar; Hongfang Liu

doi:10.1136/amiajnl-2011-000155

PMC full text:

J Am Med Inform Assoc. 2011 Sep-Oct; 18(5): 580–587.

Published online 2011 Jun 27. doi: 10.1136/amiajnl-2011-000155

Copyright/LicenseRequest permission to reuse

Table 1

Performance (F scores) of machine learning taggers on the 2009 i2b2 Challenge corpus

Authors	Data	Method	F scores
Authors	Data	Method	M	Do	Mo	F	Du	R
Patrick et al³²	145 longest summaries for training and 251 for testing	CRF and SVM	0.884	0.893	0.899	0.897	0.446	0.444
Li et al³³	147 random summaries for training and the same test set as Patrick et al	CRF and AdaBoost with decision stumps	0.802	0.802	0.821	0.813	0.180	0.030
Halgrim et al³⁴	The same data sets as Patrick et al	MaxEnt tagger	0.841	0.898	0.933	0.932	0.515	0.471
Doan et al³⁵	10-fold cross-validation on 268 summaries	SVM tagger	0.812	0.864	0.947	0.889	0.214	0.333
Doan et al³⁵	10-fold cross-validation on 268 summaries	SVM tagger with a rule-based system	0.923	0.927	0.954	0.944	0.496	0.484

Open in a separate window

The systems by Patrick et al³² and Li et al³³ are those that participated in the challenge workshop, while those by Halgrim et al³⁴ and Doan et al³⁵ are not.

Do, dosage; Du, duration; F, frequency; M, medication; Mo, mode; R, reason.