Skip to main content
. 2021 Apr 21;6:88. [Version 1] doi: 10.12688/wellcomeopenres.16718.1

Table 3. Summary table with performance metrics reported as median (95% CI) and F 1 interquartile variance (IQV) after 200 bootstrap iterations.

The performance metrics are compared across pipelines using different fields from PubMed entries.

Pipeline Precision (%) Recall (%) F 1 (%) F 1 IQV
Title 65.3 (59.0,72.7) 65.8 (55.0,72.8) 65.0 (59.5,71.0) 11.5
Abstract 77.0 (69.8,82.6) 79.8 (73.4,86.1) 78.2 (73.6,82.6) 9.0
Authors * 76.4 (69.4,82.6) 80.4 (72.8,86.1) 78.2 (73.3,82.6) 9.3
Journal * 76.4 (70.2,82.2) 79.8 (72.8,85.4) 78.0 (73.6,82.0) 8.4
Publication Type * 78.0 (71.6,84.3) 81.6 (74.7,87.4) 79.6 (75.5,84.2) 8.7
Keywords * 76.6 (70.2,83.0) 80.4 (72.8,85.5) 78.2 (73.8,82.2) 8.4
MeSH terms * 79.2 (72.1,85.2) 79.8 (72.8,86.1) 79.5 (74.3,83.3) 9.0
Chemicals * 76.0 (69.5,81.9) 80.4 (73.4,86.1) 77.8 (73.0,82.0) 9.0
Affiliations * 76.6 (69.8,82.1) 80.4 (72.8,86.7) 78.3 (73.2,81.9) 8.7
All fields 80.1 (73.0,86.1) 81.6 (74.1,87.4) 80.5 (75.7,84.9) 9.2
Optimal Fields ** 80.1 (73.9,86.0) 82.3 (74.1,88.6) 80.6 (75.8,85.2) 9.4

*Tokens from the title and abstract were also included when encoding this field.

**The optimal fields were the title, abstract, MeSH terms and publication type.