Technology-assisted title and abstract screening for systematic reviews: a retrospective evaluation of the Abstrackr machine learning tool

Table 3 Descriptive characteristics of the title and abstract screening processes in Abstrackr, across three trials

Characteristic	Topic
Characteristic	Antipsychotics N = 12,763 records	Bronchiolitis N = 5893 records	Child Health SRs N = 5243 records	Diabetes N = 47,385 records^a
Screened by human^b
N records	277 (32)	607 (340)	210 (10)	323 (206)
% records	2.2 (0.3)	10.3 (5.8)	4.0 (0.2)	0.7 (0.4)
Accepted by human^c
N records	19 (3)	56 (35)	118 (20)	111 (74)
% records	6.9 (1.1)	9.0 (0.9)	56.1 (6.9)	34.1 (1.6)
Predicted as relevant by Abstrackr^d
N records	4259 (1281)	1163 (123)	4535 (173)	5187 (1430)
% records	34.1 (10.2)	22.0 (0.9)	90.1 (3.6)	11.0 (3.0)

All values are mean (SD) across three trials. Standard deviations for proportions (% records) relate to the range of values observed across trials, and not the mean variance across trials
SR systematic review
^aIncluded some duplicates as three EndNote libraries were combined to create the dataset
^bBefore Abstrackr produced predictions
^cBased on the decisions of two independent human reviewers for each screening project
^dRecords that Abstrackr predicted as relevant for further inspection following title and abstract screening (equivalent to “accepted as relevant”)

ISSN: 2046-4053