Literature Review & Evidence Synthesis

Moderate · Communication & Dissemination track · ~35 min hands-on + readings and quiz

← Back to the Communication & Dissemination track

Speed the mechanical parts of a review while keeping the judgment and the citations yours.

What you’ll be able to do

  • Use AI to draft search strategies and screening criteria
  • Summarize and compare studies with verification
  • Avoid fabricated citations and keep a defensible record

Overview

AI can draft a search strategy, suggest inclusion criteria, and summarize what a set of papers says. It can also invent references that look real, so every citation must be checked against the actual paper.

For a defensible review, keep AI in an assisting role: it speeds screening and summarizing, but you decide what’s included and confirm what each source actually says. This searching-and-summarizing role — AI assists, you decide and verify — is exactly the kind of sanctioned use described on the Grant & Proposal Development and Manuscript & Peer Review pages, and literature you find and verify here often feeds directly into a proposal’s significance section or a manuscript’s introduction and discussion.

Practice activities

Activity 1 · Moderate — Build and refine a search strategy

Time ~18 min · Tools ChatGPT Edu + PubMed

Goal. Draft a reproducible search, then refine it against real results.

Setup. A focused question in PICO form, e.g., Among urban pedestrians (P), does improved street lighting (I) vs. none (C) reduce nighttime crashes (O)?

Steps.

  1. Turn it into a search:

    Turn this PICO question into a PubMed search strategy: key concepts, synonyms and likely MeSH terms, and a Boolean string. Note where I’ll need to adjust. [PICO]

  2. Run it in PubMed and record the result count.

  3. Refine against reality:

    My search returned [N] results — too many / too few. Suggest specific changes to tighten or broaden it.

Expected result. A working search string and a refined version, with the exact final string saved.

Check your work. A search you can’t reproduce isn’t usable — save the final string and the date you ran it.

Common pitfalls. AI may invent MeSH terms — confirm each in PubMed. Don’t trust predicted result counts; run the search.

Stretch (optional). Ask for inclusion/exclusion criteria suited to your question.

Activity 2 · Moderate — Summarize and verify the evidence

Time ~15 min · Tools ChatGPT Edu + open-access papers

Goal. Synthesize a few studies while checking every claim.

Setup. Use these three open-access articles on pedestrian safety and the built environment (or substitute your own topic):

  • PMID 31783684 — DiMaggio et al. (2019). “Changes in Pedestrian and Bicyclist Injury Patterns in New York City After Adoption of Vision Zero.” JAMA Netw Open. PubMed
  • PMID 31175717 — Stoker et al. (2019). “Pedestrian Safety and the Built Environment: A Review of the Risk Factors.” J Plan Lit. PubMed
  • PMID 35238777 — Amram et al. (2022). “Association of the Built Environment with Pedestrian Crashes Along Arterial Roads.” Inj Prev. PubMed

Open each abstract in PubMed before starting — you’ll need the actual text to verify the AI’s summary.

Steps.

  1. Build a comparison:

    Compare these abstracts in a table: design, population, key finding, and main limitation. Quote the exact effect sizes. [abstracts]

  2. Open each paper and confirm the numbers and the framing.

  3. Synthesize cautiously:

    Write a 3-sentence synthesis, flagging where the evidence conflicts or is thin.

Expected result. A comparison table plus a careful synthesis paragraph.

Check your work. Verify every effect size against the source; reject any number the abstract doesn’t actually state.

Common pitfalls. AI smooths over conflicting findings and fabricates citations. If you can’t open it, you can’t cite it.

Stretch (optional). Ask what study design would best fill the gap you found.

Activity 3 · Moderate — Verify citations programmatically with Claude Code

Time ~20 min · Tools Claude Code (personal or PI-purchased license) + terminal

Goal. Catch fabricated or misattributed citations with a script instead of your eyeballs — check every PMID against PubMed’s actual records.

Setup. Claude Code is a consumer/PI-purchased tool, not Drexel-supported: Low Risk (public) data only — these citations are public PubMed records, so they qualify. Open Claude Code in an empty folder with no sensitive files. Use this reference list, which mixes the three real citations from Activity 2 with two planted errors (don’t peek yet at which):

  • PMID 31783684 — DiMaggio et al. (2019). “Changes in Pedestrian and Bicyclist Injury Patterns in New York City After Adoption of Vision Zero.” JAMA Netw Open.
  • PMID 31175717 — Stoker et al. (2019). “Pedestrian Safety and the Built Environment: A Review of the Risk Factors.” J Plan Lit.
  • PMID 35238777 — Chen & Rodriguez (2021). “Sidewalk Connectivity and Pedestrian Crash Risk in Mid-Sized Cities.” Am J Public Health.
  • PMID 99887766 — Larsen et al. (2020). “Urban Greening Interventions and Pedestrian Injury: A Systematic Review.” Accid Anal Prev.

Steps.

  1. Ask Claude Code to build and run the check:

    Here are four references: [paste the list]. Write and run a small script that checks each one against PubMed’s free E-utilities API (esearch/esummary — no API key needed for light use). For each PMID: does it exist, and do the title, first author, and year on record match what the reference claims? Print a PASS/FAIL table with a reason for every FAIL.

  2. Before approving execution, read the script: confirm it only calls eutils.ncbi.nlm.nih.gov and sends nothing beyond the PMIDs and citation text.

  3. Ask for the corrections:

    For each FAIL, show the citation PubMed actually has on record (if any) so I can fix the list.

  4. Compare against Activity 2. Which planted error would you have caught by eye, and which might have slipped through under deadline pressure?

Expected result. Two PASSes; PMID 99887766 flagged as nonexistent; PMID 35238777 flagged as a mismatch (PubMed’s record is Amram et al., 2022, Inj Prev — the claimed title and authors belong to no paper at all).

Check your work. Open the two flagged PMIDs in PubMed yourself and confirm the script’s verdicts. The script checks the AI; you check the script.

Common pitfalls. A PASS means the citation exists and matches — not that it’s relevant, well-conducted, or says what your draft claims it says. Keep requests light (E-utilities allows about 3/second without a key). And Claude Code runs real commands on your machine: review before you approve, every time.

Stretch (optional). Ask Claude Code to generalize the script to read a whole .bib or exported reference file, or to verify DOIs against Crossref the same way.

Check your readiness

Answer these, then check — your score suggests whether to dive in or skim the readings first.

Useful resources