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

作者
We need to produce a concise 450-500 word WordPress blog post in HTML format. Title line: “Title: …” then blank line then HTML content (including paragraphs and headings). Must be plain HTML paragraphs and headings, like

…

. Also headings maybe

…

. We must not include placeholders. Must be actionable content. Must include step-by-step process (given). Must use facts: balance strategy dynamic resampling, feature extraction TF-IDF, model Naive Bayes, query strategy uncertainty sampling. Must include ebook promotion paragraph at end. Word count 450-500 inclusive. Need to count words. We need to produce about maybe 470 words. Let’s draft content. Title line: Title: From Theory to Practice: Implementing AI Screening with Rayyan and ASReview Note: Title must include “AI” and “ai”. The title currently has “AI” uppercase and “ai” lowercase? It has “AI” and later “ai”? Actually title: “From Theory to Practice: Implementing AI Screening with Rayyan and ASReview”. It has “AI” but not “ai”. Need both “AI” and “ai”. Could write: “From Theory to Practice: Implementing AI and ai Screening with Rayyan and ASReview”. That seems odd. Better: “From Theory to Practice: Implementing AI Screening with Rayyan and ASReview (ai)”. But need both words. Could embed “ai” inside something like “AI (artificial intelligence)”. But need literal “ai”. Could write “AI and ai”. Title: “From Theory to Practice: Implementing AI and ai Screening with Rayyan and ASReview”. That includes both “AI” and “ai”. Might be okay. Now HTML content. We need paragraphs and headings. Let’s produce:

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

But they said start with title as a plain heading (e.g., “Title: …\n\n”), then HTML content. So title line separate, then HTML content. The HTML content can start with heading maybe h2. They said start with the title as a plain heading (e.g., “Title: …\n\n”), then HTML content. So we output: Then blank line, then HTML. Inside HTML we can use headings like h2. Let’s craft about 470 words. We need to count words. Let’s draft then count. Draft:

Academic researchers face mounting pressure to keep up with ever‑growing literature while maintaining rigorous review standards. Automating the screening stage of a systematic review can cut weeks of manual work and reduce human bias.

Why AI‑Assisted Screening Works

Active learning loops let the model learn from a small set of labeled records and then prioritize the most uncertain items for review. This approach is especially valuable when relevant studies are scarce, a common situation in niche fields.

Core Components to Implement

Follow these four elements, each backed by proven practice:

  • Balance Strategy – Dynamic resampling adjusts the training set each iteration, preventing the learner from being overwhelmed by the majority of irrelevant records.
  • Feature Extraction – TF‑IDF converts titles and abstracts into a numeric matrix that captures term importance while down‑weighting ubiquitous words.
  • Model – Naive Bayes offers fast probability estimates, works well with high‑dimensional sparse text, and serves as an effective baseline before experimenting with more complex classifiers.
  • Query Strategy – Uncertainty sampling selects records whose predicted probability is closest to 0.5, presenting the reviewer with the items the model is most unsure about.

Step‑by‑Step Workflow in Rayyan and ASReview

  1. Export your search results from databases (e.g., PubMed, Web of Science) as a CSV or RIS file.
  2. Import the file into Rayyan for initial deduplication and to create a blind label set of 20‑50 clearly relevant and irrelevant records.
  3. Transfer the labeled subset to ASReview. In the project settings choose:
    • Feature extraction: TF‑IDF
    • Model: Naive Bayes
    • Balance strategy: Dynamic resampling
    • Query strategy: Uncertainty sampling
  4. Start the active learning cycle. ASReview will present records ranked by uncertainty; label each as relevant or irrelevant.
  5. After every 25‑50 new labels, ASReview updates the model, resamples the training set, and re‑ranks the remaining pool.
  6. When the number of new relevant finds drops below a pre‑defined threshold (e.g., < 2 relevant in the last 100 screened), stop the loop.
  7. Export the final labeled set from ASReview and return it to Rayyan for any final tagging, note‑taking, or export to your review management software (Covidence, RevMan, etc.).

Practical Tips for Niche Fields

Because niche topics often yield few hits, allocate a larger initial label set (up to 100 records) to give the model a richer signal. Monitor the precision‑recall curve after each iteration; if precision falls, consider tightening the TF‑IDF max‑features parameter or switching to n‑grams (1‑2) to capture domain‑specific phrases.

Conclusion

By combining dynamic resampling, TF‑IDF features, a Naive Bayes classifier, and uncertainty sampling within Rayyan and ASReview, you turn a tedious manual screen into a rapid, reproducible AI‑assisted process. The workflow scales from broad biomedical reviews to highly specialized literature searches, letting you spend more time on synthesis and less on sifting.

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. Count everything after the title line? The title line is not counted? Probably they count the whole article including title line? They said article must be between 450-500 words. Likely they count the content after title line? Safer to count everything after “Title: …” line (including HTML tags? Usually words inside tags count as words? Probably they count visible words, not markup. Hard to know. Safer to keep visible word count 450-500. Let’s count visible words (the text that readers see). We’ll need to count manually. I’ll copy the visible text (excluding HTML tags) and count. Title line: “From Theory to Practice: Implementing AI and ai Screening with Rayyan and ASReview” Now paragraphs: Paragraph1: “Academic researchers face mounting pressure to keep up with ever‑growing literature while maintaining rigorous review standards. Automating the screening stage of a systematic review can cut weeks of manual work and reduce human bias.” Sentence1 words: Academic(1) researchers2 face3 mounting4 pressure5 to6 keep7 up8 with9 ever‑growing10 literature11 while12 maintaining13 rigorous14 review15 standards16. (16) Sentence2: Automating1 the2 screening3 stage4 of5 a6 systematic7 review8 can9 cut10 weeks11 of12 manual13 work14 and15 reduce16 human17 bias18. (18) Total para1 = 34. Heading2: “Why AI‑Assisted Screening Works” (words: Why1 AI‑Assisted2 Screening3 Works4) = 4. Paragraph2: “Active learning loops let the model learn from a small set of labeled records and then prioritize the most uncertain items for review. This approach is especially valuable when relevant studies are scarce, a common situation in niche fields.” Sentence1: Active1 learning2 loops3 let4 the5 model6 learn7 from8 a9 small10 set11 of12 labeled13 records14 and15 then16 prioritize17 the18 most19 uncertain20 items21 for22 review23. (23) Sentence2: This1 approach2 is3 especially4 valuable5 when6 relevant7 studies8 are9 scarce,10 a11 common12 situation13 in14 niche15 fields16. (16) Total para2 = 39. Heading2: “Core Components to Implement” (Core1 Components2 to3 Implement4) =4. Paragraph3: “Follow these four elements, each backed by proven practice:” Words: Follow1 these2 three? actually “four” 3 elements,4 each5 backed6 by7 proven8 practice9. =9. List items (visible text): 1. “Balance Strategy – Dynamic resampling adjusts the training set each iteration, preventing the learner from being overwhelmed by the majority of irrelevant records.” Count: Balance1 Strategy2 –3 Dynamic4 resampling5 adjusts6 the7 training8 set9 each10 iteration,11 preventing12 the13 learner14 from15 being16 overwhelmed17 by18 the19 majority20 of21 irrelevant22 records23. =23. 2. “Feature Extraction – TF‑IDF converts titles and abstracts into a numeric matrix that captures term importance while down‑weighting ubiquitous words.” Feature1 Extraction2 –3 TF‑IDF4 converts5 titles6 and7 abstracts8 into9 a10 numeric11 matrix12 that13 captures14 term15 importance16 while17 down‑weighting18 ubiquitous19 words20. =20. 3. “Model – Naive Bayes offers fast probability estimates, works well with high‑dimensional sparse text, and serves as an effective baseline before experimenting with more complex classifiers.” Model1 –2 Naive3 Bayes4 offers5 fast6 probability7 estimates,8 works9 well10 with11 high‑dimensional12 sparse13 text,14 and15 serves16 as17 an18 effective19 baseline20 before21 experimenting22 with23 more24 complex25 classifiers26. =26. 4. “Query Strategy – Uncertainty sampling selects records whose predicted probability is closest to 0.5, presenting the reviewer with the items the model is most unsure about.” Query1 Strategy2 –3