Appendix 1. MEDLINE search strategy

1. *Beverages/

2. *Alcohol Drinking/

3. (drink$ or drunk$ or alcohol$ or beverage$1 or beer$1 or lager$1 or wine$1 or cider$1 or alcopop$1 or alco‐pop$1 or spirit or spirits or liquor$1 or liquer$1 or liqueur$1 or whisky or whiskey or whiskies or whiskeys or schnapps or brandy or brandies or gin or gins or rum or rums or tequila$1 or vodka$1 or cocktail$1).ti,ab.

4. *Tobacco/

5. *Smoking/

6. (cigar$ or smok$ or tobacco$ or e‐cig$).ti,ab.

7. *Diet/

8. *Food/

9. *Food Intake/

10. *Food Habits/

11. *Food Preferences/

12. *Eating/

13. *Drinking/

14. *Food Dispensers, Automatic/

15. (nutri$ or calori$ or food$ or eat or eats or eaten or eating or ate or low‐fat or meal$ or dessert$1 or snack$ or drink$ or beverage$1).ti,ab.

16. ((increase$1 or increasing or add$1 or added or adding or addition$1 or additional or introduce$1 or introduction$1 or introducing or extend$ or reduc$ or decrease$1 or decreasing or remov$ or restrict$ or limit$ or proxim$ or distal or distanc$ or position$ or reposition$ or visib$ or accessib$ or close or closer or closest or near or nearer or nearest or adjacent or far or farther or farthest or farthermost or further or furthest or furthermost) adj3 (amount$1 or volume$1 or quantity or quantities or availab$ or range$ or assortment$1 or arrangement$1 or array$ or display$ or choice$1 or option$ or item$1 or effort or convenien$ or salien$ or product$1 or packag$ or portion$ or serving$ or glass or glasses or bottle$ or dish$2 or bowl$1 or plate$1 or box or boxes or boxed or bag or bags or bagged or packet$ or carton$1 or vending$)).ti,ab.

17. ((increase$1 or increasing or add$1 or added or adding or addition$1 or additional or introduce$1 or introduction$1 or introducing or extend$ or reduc$ or decrease$1 or decreasing or remov$ or restrict$ or limit$ or proxim$ or distal or distanc$ or position$ or reposition$ or visib$ or accessib$ or close or closer or closest or near or nearer or nearest or adjacent or far or farther or farthest or farthermost or further or furthest or furthermost) adj3 (food$ or fruit$ or vegetable$ or FV or FFV or F&V or low‐fat or meal$ or dessert$1 or snack$ or drink$ or beverage$1 or alcohol$ or cigar$ or tobacco or e‐cig$)).ti,ab.

18. or/1‐3

19. or/4‐6

20. or/7‐15

21. 16 and 18

22. 16 and 19

23. 16 and 20

24. or/17,21‐23

25. exp animals/ not humans/

26. (rat or rats or mouse or mice or murine or rodent or rodents or hamster or hamsters or pig or pigs or porcine or rabbit or rabbits or animal or animals or dog or dogs or cat or cats or cow or cows or bovine or sheep or ovine or monkey or monkeys).ti,ab.

27. or/25‐26

28. 24 not 27

29. (editorial or case reports or in vitro).pt.

30. 28 not 29

Appendix 2. Details of the semi‐automated screening workflow

The semi‐automated screening workflow will proceed in four phases: i) Initial sample; ii) Active learning; iii) Topic modelling; iv) Active learning (final phase).

Initial sample

First, we will screen a random sample of title‐abstract records to establish an initial estimate of the baseline inclusion rate (Shemilt 2014), in order to both inform prospective monitoring of the performance of the semi‐automated screening workflow, and supply an unbiased initial sample of records for machine learning (see Active learning).

Active learning

Second, we will deploy active learning with the aim of identifying records of potentially eligible studies as rapidly as possible. In this phase, title‐abstract records will be prioritised for manual screening using active learning, whereby the machine iteratively ‘learns’ to distinguish between relevant and irrelevant records in conjunction with manual user input (Miwa 2014). We have previously deployed this method in two large‐scale systematic scoping reviews of interventions to change health behaviour (Shemilt 2013; Hollands 2013a; Shemilt 2014). Active learning will initially be trained using small samples of provisionally included and excluded records drawn from a reference set of 24 records of potentially eligible studies identified by a published scoping review on physical micro‐environment interventions (Hollands 2013a) or in the random sample of citations screened in phase 1 (Initial sample). In order to deploy active learning, a stopping criterion is needed that prespecifies when this phase will be truncated. We have set the stopping criterion in terms of the maximum marginal resource the review team is willing to ‘pay’ in order to identify one additional title‐abstract record of a potentially eligible study. We will prospectively monitor and record screening time‐on‐task and stop the active learning phase of the semi‐automated workflow if the review authors complete 15 hours of duplicate screening (i.e. 30 hours time‐on‐task in total for two review authors) without identifying any further records of potentially eligible studies. At this point, we will also screen a second random sample of records to establish a second estimate of the baseline inclusion rate (Shemilt 2014). In this active learning phase of the workflow, we will also alternate between title‐abstract and full‐text screening stages after each set of 2400 title‐abstract records has been manually screened. This is intended to promote more accurate initial title‐abstract screening decisions, and to enable retrospective modelling of the impact of using full‐text screening decisions in training data for active learning.

Topic modelling

Active learning can be expected to have identified the large majority of title‐abstract records of potentially eligible studies that are present in the full set retrieved by electronic searches before the above stopping criterion for that phase is enacted. However, given that active learning iteratively prioritises further title‐abstract records for screening based on the researchers’ preceding eligibility decisions about records that were also prioritised by active learning (that is, the algorithm progressively finds ‘more of the same’), we will next introduce an entirely different, novel method into the semi‐automated workflow, in order to provide a check and balance on the use of active learning alone. In this third phase of title‐abstract screening, records will be allocated for duplicate manual screening based on topic modelling using Latent Dirichlet Allocation (LDA) (Pedregosa 2011). Topic modelling essentially clusters title‐abstract records according to the combinations of terms they contain and returns a set of 'topic terms' for each cluster (hereafter, a ‘topic’).

Topic modelling will be used to generate 50 topics underlying the full set of title‐abstracts retrieved by electronic searches (or included among the reference set), and concurrently to generate a series of ‘membership scores’ for each unscreened record, by topic. The membership score is based on the computed probability that a record is described by the topic (that is, a higher membership score reflects a higher probability of membership of the topic) and is > 0 for all records in all topics. Each unscreened title‐abstract record will next be allocated to the single topic that corresponds with its highest membership score. Results of a preliminary simulation study, conducted to simulate this phase of the workflow in a screening dataset curated from another Cochrane review (Hollands 2015), indicated that the large majority of generated topics contained no unscreened records of potentially eligible studies (that is, most topics are irrelevant), and also that the review authors were able to discriminate accurately between topics that contained the most and fewest records of potentially eligible studies when blinded to this information. Two review authors will therefore next examine each topic, blinded to the number of records allocated to each, and place the 50 topics in rank order based on their inter‐subjective judgement of the likelihood that each set of terms describes a set of records that includes eligible studies. A second ranking of the 50 topics will also be generated based on the number of potentially eligible title‐abstract records each contains among records already screened up to the end of the active learning phase (that is, a data‐generated ranking). We will then compute a composite ranking by adding together the review authors’ ranking and the data‐generated ranking, once the latter has been multiplied by 0.5. This procedure assigns double weight to the review authors’ judgements in the composite ranking, promoting those topics that the review authors rank higher but contain a relatively low number of potentially eligible title‐abstract records among those already screened (and, conversely, demoting those topics that the authors rank lower but contain a relatively low number of potentially eligible title‐abstract records among those already screened).

At the end of the active learning phase, the ‘remaining screening budget’ (that is, the ‘overall screening budget’ minus the number of records already screened) will be calculated and allocated between topics, by drawing a random sample of unscreened title‐abstract records from each topic (that is, the sum of the sizes of the 50 random samples will equal the remaining screening budget). The sizes of random samples drawn from topics will be scaled to approximate a beta distribution (α=0.3, β=3.0) across rank‐ordered topics (highest to lowest), in order to reflect our prior belief (informed by results of the simulation study) about the likely distribution of any further potentially eligible title‐abstract records across rank‐ordered topics. Sampled records will next be allocated for duplicate manual screening in topic rank order, from highest to lowest ranked. This procedure will ensure that records assigned to a higher ranked topic will be more likely to be allocated for screening, relative to those assigned to a lower ranked topic. We will continue the topic modelling phase of title‐abstract screening until either all records allocated using the above procedure have been screened, or the following early stopping criterion is enacted: based on prospective monitoring of time‐on‐task, we will truncate this phase of the semi‐automated screening workflow if the review authors complete 15 hours of duplicate screening (i.e. 30 hours time‐on‐task in total for two authors) without identifying any further records of potentially eligible studies.

Active learning (final phase)

Because the topic modelling phase may detect additional title‐abstract records that alter any subsequent prioritisation of records by active learning, we will conduct a final phase of screening using the active learning method outlined above. Again, we will truncate this phase of the semi‐automated screening workflow if the review authors complete 15 hours of duplicate screening (i.e. 30 hours time‐on‐task in total for two authors) without identifying any further records of potentially eligible studies. Including this further phase will give additional confidence that, within available resources, all relevant title‐abstract records have been identified.