Skip to main content

Table 2 Characteristics of the reviews and screening predictions for each tool

From: Performance and usability of machine learning for screening in systematic reviews: a comparative evaluation of three tools

CharacteristicAntipsychotics, N records (%)Bronchiolitis, N records (%)Visual Acuity, N records (%)
Screening workloada12,156586111,229
Included by title/abstractb1178 (10)518 (9)224 (2)
Included in the reviewb127 (1)137 (2)1 (< 1)
Includes/excludes in training setAbstrackr, 15/185Abstrackr, 12/188Abstrackrc, 4/296
DistillerSR, 14/186DistillerSR, 14/186DistillerSR, 2/198
RobotAnalyst, 20/180RobotAnalyst, 15/185RobotAnalyst, 3/197
Screened by toold11,956 (98)5661 (97)11,029 (98)
Predicted relevant by Abstrackr2117 (18)656 (12)3639 (33)
Predicted relevant by DistillerSR7 (< 1)83 (1)0 (0)
Predicted relevant by RobotAnalyst3488 (29)1082 (19)3221 (29)
  1. aTotal number of records retrieved via the electronic searches. Each record was screened by two reviewers
  2. bIncluded following the initial screening by two independent reviewers (retrospective)
  3. cAll training sets were 200 records, with the exception of the Visual Acuity review which required a 300-record training set in Abstrackr before predictions were produced
  4. dAfter a 200-record training set