Guide on automating your literature review in 2025

Here’s my blueprint for tackling a second literature review in 2025, having learned from past mistakes so you won’t have to! The approach is heavy on tools, high on automation, and emphasises minimising the potential for human error. I value open-science principles, and prefer open-source software. I try to adhere to the four-eyes principle, which says that all work should be done twice, independently of each other. This includes: peer reviewing the search strategy, title and abstract screening, full-text screening and data extraction. I also try limiting copy-pasting actions as much as possible, because every time you copy-paste, you can introduce an error. The most important tools and frameworks I used are: PRISMA , OSF , ASReview , Zotero , and Google Drive . While you may not agree with all of my points, the take-away is that it is important to thoroughly plan your workflow before you start.

Tools & Frameworks

• PRISMA . PRISMA guidelines help you write clear and complete systematic reviews, meta-analyses, and also scoping reviews (see PRISMA-ScR ) by providing a structured reporting framework.
• OSF . The Open Science Framework is a free and open-source platform for preregistering research, project management, and much more. A commonly used alternative is PROSPERO (systematic reviews and meta-analyses only).
• ASReview . ASReview is an open-source software tool that accelerates literature reviewing using machine learning to prioritise and screen the scientific literature efficiently. I saved about 75% of time using ASReview.
• Zotero . A free and open-source reference manager with loads of useful features. Similar in purpose to EndNote. Its PDF retrieval functionality regularly finds 70-80% of PDFs with the click of a button. Shared libraries make collaboration easy, allowing multiple people to annotate and discuss PDFs asynchronously. It also has a good duplicate detection feature.
• Google Drive . Don’t go working in offline Excel spreadsheets that you share over email. There is an alternative!

Research Plan

The first thing you write and discuss with your team, the research plan is a high-level document intended for structuring your thoughts in the preparatory phase.

I would include the following headings:

• Research Question & Review Methodology,
• Background & Motivation,
• Literature Databases of Interest,
• Eligibility Criteria,
• Timeline.

It is important that you search the literature for existing reviews, immediately after you formulated the research question. If a review has already been done, ask yourself whether enough has changed since then to justify doing another one.

The databases I searched were: PubMed, CINAHL, Embase Ovid, Scopus and ACM Digital Library. Decide which one will be your primary database. It is most convenient to choose the database that you’re most comfortable with. You’ll use it to construct the initial search query, which will then be translated into the other databases. Decide whether you’re going to peer review your search strategy (see Literature Search ).

Define inclusion and exclusion criteria based on your research question, but also think of things like publication language, publication year, being peer-reviewed, and publication types you might not want to include (e.g., conference papers).

Decide whether you’re going to preregister your review, and on which platform. One benefit is the fact that you’ll be able to cite it on grant applications or resumés, even when it’s far from being finished. I preregistered my scoping review on OSF.

Construct a timeline for when you’ll want to have the protocol finished and published, when you want to start screening, etc. A review requires a significa nt time investment for all collaborators. Reaching agreement on the timeline will reduce the chance of delays.

Protocol

This is a document that outlines the entire procedure of conducting the review, in as much detail as possible. If you decided to preregister your review, this is the document you’ll upload there. Its structure depends on the review methodology. I wrote a scoping review, and used this protocol as an example. I might be stating the obvious here, but the protocol should be written and published before you actually do the literature search.

The rest of this post was written with two goals in mind:

1. To help you guide the writing of your protocol;
2. To give advice for the actual execution part of the review.

---

Literature Search

The first thing you’ll want to do is perfect the search query for your primary database: the index query. That means seeking agreement among collaborators before you translate the index query to the language of other databases. Only when the index query is perfect, you start the translation process. At this point, you might want to involve your university library for advice on your literature search (UM Library support ). In deciding how broad or narrow you want to search the literature, the number of hits is often an important contributing factor. If you decide to screen titles and abstracts with ASReview , it is helpful to realise that the proportion of time you save increases with the number of hits. In my case, 63% remained unseen with 1339 total hits. With more references, this can increase to 92% ¹ .

Translation of search queries is tedious and error-prone. There are many differences between databases. For example, they use different controlled vocabularies. Examples of controlled vocabularies are MeSH terms in PubMed/MedLine, EMTREE in Embase, and Headings in CINAHL. These do not match one-to-one. There is a translation tool for search queries, but I cannot account for its accuracy because I only found out about it after having translated my queries. I found this guide especially useful for navigating these different databases. For peer reviewing the search strategy, I used the PRESS (Peer Review of Electronic Search Strategies) guidelines (direct link to the checklist ).

Exporting and Deduplicating References

You should now have all your search queries executed and lined up. Assuming you’ve got Zotero downloaded and installed, the next step is exporting your search queries. Generally speaking, the RIS format is a good file format. Here is a guide on exporting references from different databases for use in Zotero.

After you have your different queries imported into Zotero, you’ll want to remove duplicate references. Because there is a lot of overlap between databases, there will be many. Zotero identifies possible duplicates automatically for you, but it remains up to you to decide which one you’ll keep. I used this guide on deduplication in Zotero . Always keep the reference with the abstract, otherwise, keep the reference with most metadata. Having the abstract will make your life much easier during title and abstract screening.

Deduplication reduced the number of references with 60% for me. You should now have a single Zotero collection with all of your references in one place.

The Importance of Calibration

Calibration involves strategic planning of checkpoints where you meet and discuss disagreements between collaborators. Disagreements can occur in any step along the way that is done twice, which is most of the steps. Such disagreements are unavoidable, and discussing them can reveal implicit assumptions about the review’s aim that can be scrutinised en groupe during calibration sessions. The overall goal of calibration is to align the perspectives of collaborators. This must happen in any review at the end of each step. However, ironing out teething problems at the start of a step can save you a lot of time (at the expense of some degree of blinding ).

Some ways in which calibration can benefit your review:

• Reveal unclear phrasing of eligibility criteria;
• Force you to motivate eligibility criteria in light of the review’s aim;
• Align different interpretations of eligibility criteria among collaborators;
• Discuss categorical ambiguity (also known as the boundary problem), e.g., does a particular edge case fall just inside or just outside of an eligibility criterium?

Implementing calibration can be done in many different ways. I will first outline some general principles I would adhere to, and then describe the way I did it.

General principles:

• Calibration should be part of each step where you go over many references. It should be part of: title and abstract screening, full-text screening, and data extraction.
• Spend as little time as possible not being calibrated with collaborators. Calibration should be planned early on in a given step, so enough references remain for processing, for you to benefit from the calibration.
• Calibration should not be done too early in a given step, because you must also have caught enough disagreements for you to discuss during the calibration session.
• Before calibration, screeners should be blinded to each other’s decisions. After screening, disagreements are discussed and consensus is reached for those references. Until the next calibration session, screeners should again be blinded to each other’s decisions.

This is how I implemented calibration in my review:

• During title and abstract screening, I asked all collaborators to look at 15 titles and abstracts and decide on whether they would want to keep them or not, and why. We then discussed disagreements and made a shared final decision. These 15 decisions then served as prior knowledge for ASReview . We calibrated again at the end of title and abstract screening, the specifics of which are different than other calibrations for reasons specific to ASReview ² .
• During full-text screening, both screeners looked at 30 full-texts. Before continuing with the rest of full-text screening, disagreements were discussed and resolved. At the end of full-text screening, another calibration session resolved all disagreements that had arisen since the last session.
• During data extraction, both screeners performed data extraction for 15 references. Disagreements were then resolved. At the end of data extraction, all new disagreements were resolved again.

The exact numbers are arbitrary, and are only shown to serve as an example.

Each calibration session involves detailed discussion about disagreements. Sometimes, it took two sessions to resolve all disagreements. To respect the time your collaborators invest in your review, it is important to prepare for these sessions. Highlight disagreements. Read through disagreements and form an opinion about them. Resolve minor disagreements where you agreed with the other screener’s viewpoint.

Note that calibration sessions may lead to new insights, that may or may not require you to revise previous steps. This is normal. Don’t cut corners, just redo those steps. Quality trumps speed.

Collaboration and Blinding

There is a trade-off between collaboration and blinding. The more you collaborate — by calibrating your interpretation of the eligibility criteria, or by seeing other people’s highlights and comments in PDF files — the less blinded you are to the other screener’s viewpoint. This may hamper your ability to independently assess a reference’s relevance. On the other hand, eligibility criteria can be ambiguous and edge cases are unavoidable. Reaching some consensus on their interpretation before doing the bulk of the work can lead to more consistency and shared understanding. This can save you a lot of time and effort. I believe that this calibration method does not substantially compromise blinding. Screeners are blind to each other’s decision unless they are calibrating, after which any decision on a reference’s relevance is final.

Making Your Life Easier For Full-Text Screening

Before jumping into the full-text screening, there are several things I’d do in my next review at this point. All of them will make your life easier at a later step in the process, automate something for you, and reduce the probability of making mistakes.

Enabling Direct URLs to PDFs

Bear with me. During full-text screening, you’ll be using a Google Drive spreadsheet to write down whether you think a given reference should be included or excluded. To do that, you’ll need easy access to the PDF. The following advice will enable you to click on a link in Google Drive that opens the associated PDF in a shared Zotero Library. It is worth your effort.

Modifying the .ris Export

You’ll need to modify the way Zotero exports references in the .ris format. What this modification will do, is add the Zotero Item Key, a unique identifier used by Zotero, to each reference in your .ris export file. This will enable you to construct a URL that will take you to the PDF during full-text screening.

• Locate the Zotero data directory on your computer. This can be found in different locations depending on your operating system.
• Inside the Zotero data directory, there is a folder named translators. Locate the file names RIS.js. Duplicate it in some safe location in case something goes wrong.
• Download the modified RIS.js file here . I’ve added a single line to the original file: if (item.key) addTag('LB', item.key).
• Delete the original RIS.js file. Paste the downloaded RIS.js inside the translators folder.
• Restart Zotero.
• Verify that it worked by exporting any collection to the .ris file format: right-click on a collection $\to$ Export Collection … $\to$ Choose RIS as Format $\to$ click OK. Save the file.
• Open the file in a text editor. Press cmd + F (ctrl + F on Windows) and search for “LB”. You should now find that each entry inside the .ris file has a line that goes like “LB - 6YLDRSR4”. That code is what we’ll need later on to locate our PDFs.

Setting Up a Group Library

Setting up a group library will enable every collaborator to open the same URL to a PDF. Go to Zotero Groups $\to$ Click Create a New Group $\to$ Choose a name and choose the Private Membership option $\to$ Invite collaborators and ensure they have accepted the invitation. Verify that the group was created correctly by having each collaborator check whether the group has appeared in the left sidebar in their local Zotero set-up. The final step involves moving the collection with all the deduplicated references into the newly created group.

Taking Good Care of the Eligibility Criteria

All collaborators should by now agree on the eligibility criteria. We need to store these in a safe place. If you’ve made a group library, that is probably the best place to store them. So, create a new collection in the group library, and call it Criteria. Open it, and then click on the yellow notepad with small green plus sign in the middle at the top of the interface. Click New Standalone Note. Paste your criteria in there as a numbered list.

Some other tips on the eligibility criteria:

• Rephrase them as exclusion criteria. This is because all you will do during full-text screening is exclude references.
• Add an identifier that doesn’t depend on a numbered list. That is because, during calibration sessions, some criteria may go, others may merge. This changes the numbering of your list. That’s why you need an additional identifier. I added shorthands for my criteria to the Zotero note as identifiers (e.g., diagnosis or `no_peer_review).
• Note that the PRISMA guidelines tell you to log the reasons for excluding references in the full-text screening, but not during the title and abstract screening.

Title and Abstract Screening with ASReview

At this point, it is important to decide whether you’re going to use ASReview. Watch this 2 minute video to get a general idea of how it works. Read their Nature publication for a more detailed overview. I shaved off 63% of all references using ASReview. That means 63% less work. Using ASReview is fundamentally different from traditional literature reviewing. Make sure you and your collaborators are aware of its advantages and disadvantages before proceeding. If you decide against using ASReview, I recommend looking into other tools such as Rayyan or Covidence .

There are some important things to consider before using ASReview. First, you won’t see all references. Instead, an algorithm uses machine learning to sort your references, presenting the ones it considers most relevant to your review at the top. You stop screening after you reach a pre-defined stopping criterion. My stopping criterion was: having screened 50 consecutive irrelevant references. This number may be higher for larger datasets. See this GitHub thread for a discussion on how to stop screening. Second, collaboration is a little difficult to figure out at first, but it worked fine in the end. This blog post outlines ways of integrating ASReview when you have multiple reviewers. This is a list of publications that used ASReview with multiple screeners . In my case, the job of second screener was shared among three collaborators, which increased the complexity considerably. These three collaborators ended up screening in turns for a specified amount of time, until the stopping criterion was met. They would export the project file and pass it on to the next one. Third, resolving disagreements on references between first and second screeners in ASReview is a bit of a hassle. It involves merging two data files (.csv or .xlsx). Look at points 6 and 7 to get an idea of how you could do this.

I am very happy I used ASReview. It’s a very interesting tool, it has an easy-to-learn user interface, and it saves you a lot of time at very little cost. To understand exactly how to use ASReview, please refer to the documentation . Any advice given beyond this point will be supplementary to the ASReview documentation.

An overview of how I would use ASReview the second time:

1. Install and open ASReview Lab. Take a look around.
2. Take a small number of references (I used 15) and decide whether they are to be included or excluded. Have other collaborators review them, too. Then meet for a calibration session, and resolve any disagreements. Write the final decision down somewhere. These will be used as prior knowledge for ASReview.
3. Export your references from Zotero in the .ris file format. This ensures you will be able to get your references back into Zotero without doing any coding. Although this may not be necessary, there appear to be only upsides to doing it this way.
4. Create a new ASReview project, load your .ris file, and add the prior knowledge you agreed upon.
5. Export the ASReview project for use by the second screener(s), choose the Project option. Rename it by appending something like _2nd. This file can be imported by collaborators and enables them to continue where the previous person stopped.
6. Have both the first and second screeners review references until the stopping criterion has been met. At this point you can do two things. You can go over the titles and abstracts with a different active learning algorithm. With more prior knowledge (you have screened many references by now), you can train a more complex algorithm (e.g. BERT) that can find relevant references based on subtler cues. This is what ASReview advises you to do, and is what I would do the second time around. If you decide against this or have already gone through the process multiple times, continue to step 7.
7. Have both screeners export both the Project files and the Dataset file with both labelled and unlabelled records. Use the .csv format or the .xlsx format if you don’t want to re-import your references. Store these files away neatly for future reference. Realise that separate screeners might not have screened all the same articles. That is because the algorithm learns from each subsequent decision, which is likely to differ considerably between screeners. That’s why the two datasets will have references you agree on, references you disagree on, references that only the first screener saw, and references that only the second screener saw.
8. Do a calibration session to discuss references you disagree on, references that only the first screener saw, and references that only the second screener saw. The way I went about that is by writing a Python script, but there are ways to merge these datasets in Excel .
9. Duplicate your dataset, and retain only the references that you agree to include. These will be used for the full-text screening.

Full-Text Screening in Google Drive and Zotero

You should now have a list of references in Excel that are ready for the full-text screening phase. If you decided to make your life easier for full-text screening, now is the time to reap the fruits of your labour. Verify that the Excel file contains a column named label that has the item key that we added to the .ris export.

Setting up URLs to PDFs in Google Drive

This is how I would set-up for full-text screening:

• Upload the Excel file to Google Drive.
• Go to Zotero Groups and open your newly made group library by clicking Group Library. Navigate to any random reference and look at the URL. It should be of this format: https://www.zotero.org/groups/<some_group_identifier>/<name_of_your_group>/items/<item_key_you_added_to_the_RIS_export>/library. We are going to recreate this URL based on the item key that is in the label column.
• Create a new column, call it pdflink. In the first row, add this text:

="https://www.zotero.org/groups/<some_random_identifier>/<name_of_your_group>/collections/<some_collection_identifier/items/" & A1 & "/collection"

• This assumes that the label column is the first column. If it is not, replace A with the correct column’s letter or simply move the label column to the first position.
• To repeat this formula for all the other rows automatically; select the cell, and pull down on the blue dot in the bottom-right corner of the cell.

What this does it create a URL that leads you to the correct PDF (assuming there is one).

Modifying the Google Drive

Before you can start screening, you need to add columns to your spreadsheet that assign references to screeners, allow logging screeners’ decision (include or exclude), allow logging reasons for exclusion and additional comments for discussion. In the spirit of blinding, it may be wise to duplicate the spreadsheet, or at least duplicate the sheet in the same spreadsheet, so that screener’s cannot see the other’s decisions.

These are the columns I would add to the first (spread)sheet:

• 1st_screener - initials of the first screener for these references;
• 1st_included - whether the first screener wants to include the reference (TRUE or FALSE)
• 1st_criterion - the reason the first screener may want to exclude the reference;
• 1st_comments;

And to the second (spread)sheet:

• 2nd_screener - initials of the second screener for these references;
• 2nd_included - whether the second screener wants to include the reference (TRUE or FALSE);
• 2nd_criterion - the reason the second screener may want to exclude the reference;
• 2nd_comments;

With these comments in place, calibration simply involves filtering columns for rows where there is disagreement (in Google Drive: select all cells $\to$ Data $\to$ Create a filter).

Discussing PDF highlighting

As I briefly mentioned in Collaboration and Blinding , seeing other people’s highlights and comments in PDF files is subject to what I just coined as the calibration-blinding trade-off. As such, it may be wise to agree on some kind of highlighting policy. The benefit of seeing other people’s PDF highlights is that information that collaborators found helpful is likely to be helpful to you, too. However, if collaborators write extensive comments on their reasons for including or excluding some reference, you compromise blinding. So, prohibiting comments may be an appropriate middle ground.

Data Extraction

Data extraction requires a data extraction table. Each row will be a paper. Each column is a piece of information that you’ll need. This is highly dependent on the question you’re trying to answer, but some things may come in handy in most cases:

• Metadata: these can be copied from the full-text screening spreadsheet. Make sure to include the label and pdflink columns;
• Descriptives of the sample: sample size, mean age, age SD, percent male, percent female;
• Quality appraisal: generally required for systematic reviews, not for scoping reviews;
• Diagnostic instrument.

It might also be useful to search OSF for data extraction tables other studies used. Realise that data extraction is probably the most labour-intensive of all, and definitely the most labour-intensive step per study. You don’t want to forget any columns, because going over all studies again will be very burdensome. At the same time, you don’t want to be collecting information you won’t need for the review. Sadly, data extraction is still manual labor.