What Is Screening in a Systematic Literature Review?

What Is Screening in a Systematic Literature Review?

Most people outside research imagine systematic reviews are about reading papers and having insights.

Anyone who’s ever actually done one knows the real story:
you spend a shocking amount of time pressing include / exclude thousands of times in a row.

That part of the process has a name:
screening.

It looks simple from the outside, but it’s the stage that can make or break your review, and your sanity.

In this post, I’ll walk through:

• What screening actually is
• How the study screening workflow fits into a systematic review
• Where people go wrong (and how to avoid it)
• How AI can help you automate screening in a systematic review without losing control

This isn’t a textbook definition. It’s based on what I’ve learned the hard way, screening everything from small biotech projects to complex symptom-mapping reviews in pharma.

---

My First Reality Check With Screening

My first proper screening job was for a small biotech company.

The dataset wasn’t huge by systematic review standards – around 600 articles.

I thought, “That’s fine, I’ll get through it quickly.”

It was incredibly burdensome.

It wasn’t the reading time that drained me, it was the constant micro-decisions:

• Does this really match the population?
• Is this outcome relevant enough?
• Is this actually about the right disease stage?

By the time I’d finished, I realised:
even a few hundred records can feel heavy when every single one demands a yes/no decision tied to a protocol.

Later, I manually screened 4,600 records for another project. That one was torture. It pushed me right up against the limits of what’s reasonable for a human to manage manually.

That’s when I stopped seeing screening as “just a step” and started seeing it as its own world.

---

So, What Is Screening in a Systematic Review?

In simple terms:

Screening is the process of deciding which studies from your search results actually belong in your review, based on pre-defined inclusion and exclusion criteria.

You start with a big pool of potentially relevant records (often 1,000–10,000+).
You end with a much smaller set of studies that truly match your question.

Typically, screening happens in two main stages:

1. Title and Abstract Screening

At this stage, you’re asking:

• Is this clearly irrelevant?
• Or is it potentially relevant enough to keep for full-text?

You usually don’t have enough detail to be 100% sure, so you err on the side of “include if in doubt”.

2. Full-Text Screening

Here you dig into:

• Population
• Intervention/exposure
• Comparator (if applicable)
• Outcomes
• Study design

…all in the context of your pre-specified criteria.

This is where you give the final judgement: in or out, with a documented reason.

---

The Study Screening Workflow (What It Actually Looks Like)

Whether you’re a PhD student, academic, or working in pharma, a typical study screening workflow looks something like this:

1. Run searches across databases
   PubMed, Embase, Web of Science, Scopus, etc.

2. Export and merge results
   Pull everything into one environment.

3. Remove duplicates
   Surprisingly non-trivial when you have multiple databases and messy metadata.

4. Title and abstract screening
   Quick decisions based on limited information, but repeated hundreds or thousands of times.

5. Retrieve full texts
   Library access, interlibrary loans, PDFs from authors… it’s a job in itself.

6. Full-text screening
   Detailed eligibility check against your inclusion and exclusion criteria.

7. Resolve conflicts
   If two reviewers disagree, you discuss or bring in a third opinion.

8. Document everything
   PRISMA flow diagram, reasons for exclusion at full-text stage, any deviations from protocol.

On paper, this looks clean.
In reality? It’s often messy, slow, and cognitively draining.

---

Where Screening Goes Wrong (The Biggest Mistake)

The single biggest mistake I see teams make is this:

Starting screening without concrete, discussed, and agreed inclusion and exclusion criteria.

Everyone thinks they’re on the same page… until they’re not.

A Real Example: Symptom Mapping in Sjögren’s Syndrome

One project that nearly broke me was a symptom-mapping review in Sjögren’s syndrome.

• Hits: 2,380 records
• Problem: every study design was different
• Different measures
• Different outcomes
• Different ways of even defining symptoms

It was incredibly difficult to compare and contrast results, because the underlying conceptual ground wasn’t stable.

Add slightly vague criteria and you get:

• Endless “maybe” decisions
• Reviewers interpreting criteria differently
• Conflicting screening decisions that have to be reconciled later
• Hours lost revisiting the same PDFs

That kind of chaos doesn’t show up in the final PRISMA diagram, but it absolutely shows up in your stress levels.

---

How to Screen Well: Practical Tips From the Trenches

A few things I now consider non-negotiable.

1. Make Your Criteria Boringly Clear

Don’t rely on “we kind of know what we mean”.

Write your inclusion and exclusion criteria in concrete, operational language:

• “Adults” → “Participants aged 18–65 years”
• “Intervention-based” → “Any structured social skills training programme with at least 4 sessions”
• “Relevant outcomes” → list them explicitly

If someone new joins the project, they should be able to understand the criteria without a one-hour call.

2. Pilot Screen ~100 Studies Together

One of the most useful habits I’ve developed:

Screen the first 100 studies, then compare decisions with your co-author or co-reviewer.

You will discover:

• Criteria that are interpreted differently
• Edge cases that weren’t anticipated
• Phrases that are too vague and need tightening

Fix those early and you prevent hundreds of misaligned decisions later.

3. Treat “Maybe” as a Deliberate Category

“Maybe” isn’t a failure, it’s a pressure valve.

Use it when:

• The abstract doesn’t have enough detail
• You’re not sure about the population
• The outcome sounds adjacent but not clearly in or out

Then revisit those systematically, ideally with a second reviewer.

4. Document Reasons for Exclusion at Full Text

It’s tiring when you just want to move on, but essential for:

• Transparency
• PRISMA reporting
• Sanity when someone asks “Why did we exclude this again?” six months later

Even simple categories like “wrong population”, “wrong outcome”, “wrong study design” help a lot.

---

Why AI Fits Screening So Well

Out of all the tasks in a systematic review, screening is uniquely suited to AI support.

Not because it’s trivial, but because:

• It’s highly repetitive
• It follows structured criteria
• It scales brutally with project size
• Yet still benefits from clear, auditable rules

Humans are great at nuance, context, and exceptions.
We are terrible at doing the same classification task 4,600 times in a row without getting tired and inconsistent.

That’s why I see screening as the clearest candidate for an AI screening tool.

Not to take over the final decision.
To stop wasting expert brainpower on endless binary clicks.

---

How I Validate an AI Screening Tool (Before Trusting It)

I don’t treat AI as magic.
I treat it as a very fast, very literal assistant that needs to be tested.

A simple, practical way to validate an AI screening tool is:

Put in a small, known set of studies first – say 10 – and manually evaluate its answers.

For each record, you check:

• Did it correctly identify clear includes?
• Did it correctly identify obvious excludes?
• What did it do with edge cases?
• Are the justifications aligned with your criteria?

If it behaves reasonably on 10, you can scale up cautiously.

You can then:

• Sample-check AI decisions against human decisions
• Look at where it tends to say “maybe”
• Tighten your criteria or instructions to the model if needed

The goal isn’t blind trust.
The goal is calibrated trust.

---

Automating Screening in a Systematic Review (Safely)

When we talk about automating screening in a systematic review, I don’t mean pressing a button and letting an algorithm decide the fate of your dataset.

What makes more sense is:

• Let AI do the first-pass triage
• Let humans handle edge cases and final decisions

That’s the philosophy behind study-screener.com.

I built it out of frustration with exactly the kind of painful, manual workflows I described earlier. The idea isn’t to be clever for the sake of it; it’s to:

• Take the heavy lifting out of title/abstract and full-text screening
• Keep decisions transparent and auditable
• Always err on the side of caution

When the AI is unsure, it doesn’t pretend to know.
It pushes those studies into a “maybe” bucket instead of quietly excluding something that might be important.

Humans still make the final call.
AI just clears enough of the forest that you can actually see the trees.

---

Where Screening Fits in the Bigger Picture

It’s tempting to treat screening as a hurdle to get past so you can reach the “real” work.

But everything downstream depends on it:

• Your evidence base relies on those early include/exclude decisions
• Your risk of bias judgements depend on the right studies being retained
• Your conclusions are only as good as the pool of studies you start from

Messy screening = messy review.

A well-designed study screening workflow, with clear criteria, good communication between reviewers, and sensible use of automation, is what allows the rest of the review to actually mean something.

---

A Takeaway Thought

So what is screening in a systematic literature review?

It’s the process of turning a chaotic pile of search results into a coherent, justifiable set of studies that truly answer your question.

Done badly, it’s torture:
thousands of vague decisions, inconsistent criteria, and endless re-checking.

Done well, it’s deliberate:
clear rules, aligned reviewers, transparent documentation, and smart use of tools to protect human attention.

For me, the turning point was realising that:

• My first 600-article project shouldn’t have felt that bad
• My 4,600-record project didn’t have to be torture
• And that tools like study-screener.com could absorb a lot of the repetitive strain, while leaving humans firmly in charge

If you’re a PhD student, academic, or part of a pharma team about to begin a review, here’s the short version:

• Make your criteria boringly clear
• Pilot-screen your first 100 studies and compare decisions
• Use “maybe” intentionally
• Document exclusions
• And don’t be afraid to let an AI screening tool handle the repetition , as long as you keep control of the judgement.

AI where it helps.
Humans where it matters.
That’s how screening stops being torture and becomes just another deliberate, manageable part of the workflow.