From Theory to Practice: Implementing AI Screening with Rayyan and ASReview

We need to produce a concise 450-500 word WordPress blog post in HTML format. Must include title line “Title: …” then HTML content. Use plain HTML paragraphs and headings, e.g.,

. Must not use placeholders. Must include the e-book promotion paragraph at the end with given link. Must count words 450-500 inclusive. We need to write about AI automation for niche academic researchers, focusing on implementing AI screening with Rayyan and ASReview. Use facts: balance strategy dynamic resampling, feature extraction TF-IDF, model Naive Bayes, query strategy uncertainty sampling. Provide step-by-step process. We need to ensure word count between 450-500. Let’s aim for ~470 words. We need to output only the article content, starting with “Title: …” then newline then HTML content. No extra explanation. We must count words. Let’s draft. Title line: Title: From Theory to Practice: Implementing AI Screening with Rayyan and ASReview Now HTML content. Use headings: maybe h2 for sections. Use

and paragraphs. We need to include the e-book promotion paragraph at the end exactly as given. Let’s draft content:

We’ll need several paragraphs. Let’s write roughly 10 paragraphs of ~45-50 words each => 450-500. We’ll need to count. I’ll write and then count. Draft:

Systematic literature reviews are essential but time‑consuming, especially for niche fields where relevant studies are scarce. AI‑assisted screening can cut the workload by learning from your decisions and prioritizing the most promising records.

Two popular open‑source tools, Rayyan and ASReview, implement active learning loops that let you train a model while you screen. Both support the same core concepts: feature extraction, a classifier, and a query strategy that selects uncertain items for review.

Step‑by‑Step Workflow

1. **Export your search results** from databases (e.g., PubMed, Web of Science) as a CSV or RIS file containing title, abstract, and DOI.

2. **Import the file** into Rayyan (for collaborative screening) or ASReview (for fully automated active learning). In Rayyan, create a new project and upload the CSV; in ASReview, use the “Import data” option.

3. **Pre‑process the text**. Both tools automatically apply TF‑IDF vectorization to titles and abstracts, converting words into numeric features that capture term importance while down‑weighting common words.

4. **Address class imbalance**. Because relevant papers are often few, enable dynamic resampling (available in ASReview’s settings) which repeatedly creates balanced training subsets during each iteration, preventing the model from being biased toward the majority class.

5. **Select a starter model**. Naive Bayes is fast, interpretable, and works well with TF‑IDF features, making it an ideal first classifier. In Rayyan you can set the model to “Naive Bayes” under the AI settings; ASReview defaults to it unless you change it.

6. **Choose the query strategy**. Uncertainty sampling presents the records the model is least sure about, maximizing information gain per label. Activate this strategy in both tools; it is the default active learning approach.

7. **Begin screening**. Label a small seed set (e.g., 20‑30 papers) manually—include both relevant and irrelevant examples. The model updates, re‑ranks the pool, and presents the next uncertain records for you to label.

8. **Iterate until stability**. Continue labeling until the number of new relevant papers found per batch drops below a pre‑set threshold (e.g., <5 per 50 records) or the estimated workload reaches your target savings.

9. **Export the final decisions**. Rayyan lets you download included/excluded tags; ASReview provides a ranked list with inclusion probabilities that you can threshold for final selection.

10. **Document the process**. Record the seed size, number of iterations, model settings, and stopping rule in your PRISMA flow diagram to ensure transparency and reproducibility.

By following these steps, niche researchers can leverage AI to handle imbalanced datasets efficiently, turning a manual slog into a rapid, evidence‑based screening pipeline.

Now the e-book promotion paragraph exactly as given:

For a comprehensive guide with detailed workflows, templates, and additional strategies, see my e-book: AI for Niche Academic Researchers: How to Automate Systematic Literature Review Screening and Data Extraction.

We need to wrap it in … as per format. The instruction says: At the end, include a paragraph promoting the e-book with link:

. It didn’t explicitly require the WP comment tags, but earlier they said Content: write as plain HTML paragraphs and headings (e.g.,

). So we should follow that format for all paragraphs, including the promo. Thus final promo:

For a comprehensive guide with detailed workflows, templates, and additional strategies, see my e-book: AI for Niche Academic Researchers: How to Automate Systematic Literature Review Screening and Data Extraction.

Now we need to count words. Let’s count words in the entire output (excluding the Title line? The title line is part of the article? They said start with title as a plain heading (e.g., “Title: …\n\n”), then HTML content. So title line counts as part of the article for word count? Likely yes, they want total words 450-500 including title line? Usually they’d count the content. Safer to count everything after “Title:” line? We’ll include title line in count to be safe. Let’s count words manually. I’ll copy the text and count. Title line: “Title: From Theory to Practice: Implementing AI Screening with Rayyan and ASReview” Words: Title:(1) From(2) Theory(3) to(4) Practice:(5) Implementing(6) AI(7) Screening(8) with(9) Rayyan(10) and(11) ASReview(12). So 12 words. Now we need to count all words in HTML comments and tags? Usually word count counts visible words only, but safer to count only visible text (inside

and

). The HTML comments and tags are not words. We’ll count visible words. Let’s extract visible text: Paragraph1: “Systematic literature reviews are essential but time‑consuming, especially for niche fields where relevant studies are scarce. AI‑assisted screening can cut the workload by learning from your decisions and prioritizing the most promising records.” Count words: Systematic(1) literature2 reviews3 are4 essential5 but6 time‑consuming,7 especially8 for9 niche10 fields11 where12 relevant13 studies14 are15 scarce.16 AI‑assisted17 screening18 can19 cut20 the21 workload22 by23 learning24 from25 your26 decisions27 and28 prioritizing29 the30 most31 promising32 records33. 33 words. Paragraph2: “Two popular open‑source tools, Rayyan and ASReview, implement active learning loops that let you train a model while you screen. Both support the same core concepts: feature extraction, a classifier, and a query strategy that selects uncertain items for review.” Count: Two1 popular2 open‑source3 tools,4 Rayyan5 and6 ASReview,7 implement8 active9 learning10 loops11 that12 let13 you14 train15 a16 model17 while18 you19 screen.20 Both21 support22 the23 same24 core25 concepts:26 feature27 extraction,28 a29 classifier,30 and31 a32 query33 strategy34 that35 selects36 uncertain37 items38 for39 review40. 40 words. Heading:

Step‑by‑Step Workflow

=> words: Step‑by‑Step1 Workflow2 (maybe counts as 2). We’ll count. Paragraph3 (step 1): “1. Export your search results from databases (e.g., PubMed, Web of Science) as a CSV or RIS file containing title, abstract, and DOI.” Count: 1. Export1 your2 search3 results4 from5 databases6 (e.g.,7 PubMed,8 Web9 of10 Science)11 as12 a13 CSV14 or15 RIS16 file17 containing18 title,19 abstract,20 and21 DOI22. 22 words. Paragraph4 (step 2): “2. Import the file into Rayyan (for collaborative screening) or ASReview (for fully automated active learning). In Rayyan, create a new project and upload the CSV; in ASReview, use the “Import data” option.” Count: 2. Import1 the2 file3 into4 Rayyan5 (for6 collaborative7 screening)8 or9 ASReview10 (for11 fully12 automated13 active14 learning).15 In16 Rayyan,17 create18 a19 new20 project21 and22 upload23 the24 CSV;25 in26 ASReview,27 use28 the29 “Import30 data”31 option32. 32 words. Paragraph5 (step 3): “3. Pre‑process the text. Both tools automatically apply TF‑IDF vectorization to titles and abstracts, converting words into numeric features that capture term importance while down‑weighting common words.” Count: 3. Pre‑process1 the2 text.3 Both4 tools5 automatically6 apply7 TF‑IDF8 vectorization9 to10 titles11 and12 abstracts,13 converting14 words15 into16 numeric17 features18 that19 capture20 term21 importance22 while23 down‑weighting24 common25 words26. 26 words. Paragraph6 (step 4): “4. Address class imbalance. Because relevant papers are often few, enable dynamic resampling (available in ASReview’s settings) which repeatedly creates balanced training subsets during each iteration, preventing the model from being biased toward the majority class.” Count: 4. Address1 class2 imbalance.3 Because4 relevant5 papers6 are7 often8 few,9 enable10 dynamic11 resampling12 (available13 in14 ASReview’s15 settings)16 which17 repeatedly18 creates19 balanced20 training21 subsets22 during23 each24 iteration,25 preventing26 the27 model28 from29 being30 biased31 toward32 the33 majority34 class35. 35 words. Paragraph7 (step 5): “5. Select a starter model. Naive Bayes is fast, interpretable, and works well with TF‑IDF features, making it an ideal first classifier. In Rayyan you can set the model to “Naive Bayes” under the AI settings; ASReview defaults