How to Read a Scientific Paper — A Researcher's Guide
Reviewed
How to Read a Scientific Paper — The Three-Pass Method for Biomedical Research
Learning how to read a scientific paper efficiently is a skill most researchers acquire by accident. Biomedical papers are not written to be read linearly — they are structured documents with redundancy built in, and a trained reader extracts the core in 10–15 minutes before deciding whether to read in full. Here is the three-pass method, adapted for biomedical research, plus where AI summarisers help and where they fail.
Pass 1 — the 10-minute biomedical paper skim
The first pass answers one question: should you read this paper at all? Spend 10 minutes, no more. Read in this order: (1) title and abstract, (2) introduction's last paragraph — which usually states the hypothesis and summarises the findings, (3) figure legends for every main figure, (4) conclusion or discussion's first paragraph.
You are looking for three things: the research question, the main claim, and the type of evidence (clinical trial, cohort study, mechanistic in-vitro, animal model). Do not read the methods. Do not read the results prose. Do not read the discussion's nuance.
At the end of 10 minutes write two sentences in your notes — what the paper claims and how they tested it. If you cannot write those sentences, the paper is either badly written or not relevant to your work. This pass is where AI summarisers earn their place: a two-paragraph AI summary is a reasonable substitute for the first-pass skim on papers you are triaging, not papers you are citing.
Pass 1 output — two-sentence note for a biomedical paper
Claim: Faecal microbiota transplantation reduced gastrointestinal symptoms in children with ASD at 12 weeks vs placebo (n = 48, RCT). Method: Double-blind, sham-controlled trial with validated GI symptom scores; secondary outcomes included microbiome diversity (16S rRNA sequencing). — If those two sentences capture the core, pass 1 is done.
Pass 2 — the 45-minute biomedical paper read
If pass 1 clears the bar, commit 45 minutes for pass 2. Read in a different order from the paper's structure: (1) introduction, fully, to understand what question the authors are answering and what prior work they build on; (2) all figures and tables, in order, with their legends — do this before reading the results prose; (3) the results section, using the figures you have already examined as scaffolding; (4) the discussion. Skim the methods for technique names but do not read them in detail yet.
By the end of pass 2 you should be able to state, in your own words, the experimental design, the primary result, the authors' interpretation, and at least one limitation they acknowledged. If you are taking notes for a literature review, this is the pass where you populate extraction fields: sample size, population, intervention, comparator, outcome, effect size, limitations.
Figure-first reading — the single biggest upgrade for biomedical researchers
Papers are written prose-first, but trained readers read figures-first. Figures carry the actual evidence; the results prose is the authors' interpretation. For each panel ask: what is plotted, what are the axes, what is the sample size, what is the direction and magnitude of the effect? When you then read the results prose you will catch over-statements — "significantly reduced" when the effect is 8% with n = 3, "robust" when error bars span the control mean, "dose-dependent" when only two doses were tested.
Pass 3 — the deep read for biomedical methods and controls
Pass 3 is for load-bearing papers — papers you will cite as evidence for a specific claim, or whose method you plan to replicate. Budget 2–4 hours. Now you read the methods in full, well enough that you could (in principle) reproduce the experiment.
Check specifics: antibody catalogue numbers, cell line sources and passage numbers, statistical tests used and whether they fit the data distribution, sample size justification, blinding and randomisation, exclusion criteria. For clinical trials, check CONSORT compliance — the CONSORT 2010 statement provides a 25-item checklist for reporting parallel-group randomised trials (PMID: 20332509).
Check the supplementary materials — load-bearing controls and negative results usually live there. Read the discussion critically: are the conclusions actually supported by the figures, or are they over-reaching? If the paper cites a prior result as foundational, pull that reference too.
This is also where you check for retractions and post-publication concerns. Bozzo et al. (2017) found that 76% of cancer-research retractions occurred in the most recent decade, yet 29% of retracted articles remained available online in their original form without any retraction watermark (PMID: 29451549). Check PubMed for retraction notices, Retraction Watch for the database of retractions, and PubPeer for critical comments.
What to skim and what to scrutinise in biomedical papers
| Section / Element | Skim | Scrutinise |
|---|---|---|
| Introduction (first half) | Context-setting prose; rehash of broad field | — |
| Introduction (last paragraph) | — | Hypothesis statement; gap the authors claim to fill |
| Sample size / n values | — | n per group in every figure; biological vs technical replicates |
| Statistical tests | — | Whether the test fits the data; multiple-comparison correction |
| Error bars | — | SD vs SEM vs 95% CI — these are not interchangeable |
| Primary outcome | — | Prespecified or defined post hoc? |
| Control conditions | — | Appropriate controls in every figure panel |
| Discussion (speculative paragraphs) | Later paragraphs with future-direction speculation | — |
| Supplementary material | — | Controls, negative results, full Western blots, raw data |
| Funding / COI disclosures | — | Does not invalidate a paper but contextualises it |
| CONSORT / PRISMA diagrams | — | For clinical trials and systematic reviews respectively |
Bariani et al. (2015) reviewed 140 phase III oncology trials and found that only 27.9% provided all parameters required for proper sample-size calculation, despite nearly 80% reporting a target sample size (PMID: 24401665). Incomplete statistical reporting is common even in top-tier journals; a careful reader checks the numbers, not just the narrative.
Error bars — a common source of misinterpretation in biomedical figures
SD (standard deviation) describes the spread of the data. SEM (standard error of the mean) describes the precision of the mean estimate — SEM is always smaller than SD (SEM = SD / √n), which makes results look tighter. 95% CI gives a range likely to contain the true population parameter. If a paper reports SEM on a figure with n = 3, the error bars are artificially small. Always check which statistic is displayed; papers that do not specify are a red flag.
When to read biomedical full text vs when to rely on the abstract
A good abstract captures the question, method, result, and conclusion in 250 words. For most screening decisions the abstract is sufficient. Read the full text when: (1) you plan to cite the paper as evidence for a specific claim — the abstract may omit the exact effect size or qualification, (2) you plan to use the paper's method, (3) the abstract's claim seems surprising relative to the field, (4) you are conducting a systematic review or meta-analysis, or (5) the paper is load-bearing for your own hypothesis.
Relying on abstracts alone is how error propagates through the biomedical literature: a nuanced result ("effect observed in subgroup A only, with caveats") becomes "effect observed" in the citing paper's introduction, then "established" three citation hops later. Do not be the citation that loses the caveat.
For triage-heavy workflows, AI research-paper summarisers can expand abstract content usefully — but they can also smooth over caveats. Safrai & Orwig (2024) found that while ChatGPT-4 produced a relevant biomedical review, 16% of its generated references were completely fabricated and 48% contained errors in authorship, journal, or date (PMID: 38619763). Use AI for triage; verify any load-bearing claim in the full text.
Common mistakes when reading biomedical research papers
| Mistake | Why it matters | Fix |
|---|---|---|
| Reading linearly, start to finish | Papers are structured documents, not novels; linear reading buries the key result | Abstract → figures → results → intro → discussion → methods |
| Trusting the abstract for a citation | Abstracts flatten nuance; effect qualifications and subgroup restrictions disappear | If you are going to cite it, read the paper |
| Skipping the supplementary material | In biomedical papers, supplementary often contains controls and negative results | Always open supplementary for pass-3 papers |
| Not checking sample size or n | A three-panel figure with n = 3 per group is a pilot, not evidence | Check n per group in every figure; distinguish biological from technical replicates |
| Ignoring retractions and corrections | 29% of retracted cancer papers remain online without retraction marks | Check PubMed retraction notices, Retraction Watch, PubPeer |
| Letting AI replace reading | AI summaries are triage tools; 16% of ChatGPT-4 biomedical references are fabricated | Use AI for pass 1 triage; read figures and methods yourself for any cited paper |
Tools for reading biomedical papers — who should use what
BioSkepsisBiomedical researchers needing grounded paper summaries
Paste a DOI and get a grounded summary of the question, method, and primary result, with every claim linked to the source paragraph — so you can jump straight to the figures you need to verify. For methodology comparison across related papers, the mechanistic-links table surfaces how different studies converge or diverge on the same biomedical claim. Accelerates passes 1 and 2 without replacing pass 3.
SciteResearchers auditing citation context in biomedical literature
Shows how every citation in a paper has been used — supporting, contrasting, or merely mentioning. Useful when evaluating whether a highly-cited biomedical paper is being supported by follow-up work or is being systematically contradicted. Complements the pass-3 deep read by revealing the paper's reception over time.
ZoteroTeams managing structured reading notes across many biomedical papers
Free reference manager with PDF annotation. Use it to keep structured pass-2 notes (question, design, primary result, limitation) linked to each paper. Export directly from PubMed. Tag papers by pass level — pass 1 only, pass 2 done, pass 3 done — to track reading progress across a review.
PubPeer + Retraction WatchResearchers verifying biomedical paper integrity before citing
PubPeer hosts post-publication peer review with flagged concerns on specific papers. Retraction Watch maintains a searchable database of retractions and expressions of concern. Check both before citing a biomedical paper for the first time — especially papers with surprising claims or high citation counts.
Frequently asked questions
How long should it take to read a biomedical research paper?
It depends on your goal. Pass 1 (deciding if it is worth reading) takes 10 minutes. Pass 2 (understanding the paper) takes about 45 minutes. Pass 3 (deep read for papers you plan to cite or replicate) takes 2–4 hours. Most papers you encounter only need pass 1; only load-bearing papers need pass 3.
Should I read the methods section first or last in a biomedical paper?
Last, unless the paper's method is the reason you are reading it. In the three-pass approach, skim methods for technique names in pass 2, then read methods in full detail only in pass 3 — when you need to assess whether the experimental design actually supports the paper's claims. Reading methods first often buries you in protocol detail before you know what the paper is claiming.
Is it OK to use AI to summarise biomedical papers I am reading?
For triage (pass 1), yes — an AI summary is a reasonable substitute for a first-pass skim on papers you are deciding whether to read. For papers you plan to cite, no. AI summarisers can smooth over caveats, and general-purpose LLMs have a documented citation-fabrication problem: one study found 16% of ChatGPT-4 biomedical references were completely fabricated (PMID: 38619763). Use AI for triage, never for citation-level claims.
What do I do if I don't understand a biomedical paper?
Start with the figures — they carry the actual evidence and are often more interpretable than dense prose. If the field is unfamiliar, read two or three papers the authors cite as foundational. BioSkepsis can identify the five papers a given study most depends on and explain them in context. If a specific methods section is opaque, search for a methods paper or protocol that describes the technique in isolation.
How do I know if a biomedical paper is good quality?
Check sample sizes and statistical power (n = 3 per group is a pilot, not evidence), whether appropriate statistical tests were used, whether controls are adequate, whether the primary outcome was prespecified or defined post hoc, and whether error bars represent SD, SEM, or 95% CI. For clinical trials, check CONSORT compliance (PMID: 20332509). Check PubMed for retraction notices and PubPeer for flagged concerns. For systematic reviews, check the PRISMA diagram and heterogeneity statistics.
Why should I read biomedical figures before the results text?
Figures carry the actual data; the results prose is the authors' interpretation of that data. Reading figures first lets you form your own conclusions before being influenced by the narrative. You will catch overstatements more readily — "significantly reduced" when the effect is 8% with n = 3, "robust" when error bars span the control mean, "dose-dependent" when only two doses were tested. Figure-first reading is the single biggest upgrade most researchers can make to how they read biomedical papers.
Speed up pass 1 and pass 2 on biomedical papers
BioSkepsis gives you grounded, citation-linked summaries across 40M+ biomedical papers. Upload a DOI and jump straight to the figures and claims that matter. Free tier — 100 papers per session, no credit card.
Start freeSources & further reading
- Schulz KF, Altman DG, Moher D; CONSORT Group. CONSORT 2010 statement: updated guidelines for reporting parallel group randomised trials. BMJ. 2010;340:c332. PMID: 20332509. doi:10.1136/bmj.c332
- Bariani GM, de Celis Ferrari ACR, Precivale M, et al. Sample size calculation in oncology trials: quality of reporting and implications for clinical cancer research. Am J Clin Oncol. 2015;38(6):570–574. PMID: 24401665. doi:10.1097/01.coc.0000436085.23342.2d
- Bozzo A, Bali K, Evaniew N, Ghert M. Retractions in cancer research: a systematic survey. Res Integr Peer Rev. 2017;2:5. PMID: 29451549. doi:10.1186/s41073-017-0031-1
- Safrai M, Orwig KE. Utilizing artificial intelligence in academic writing: an in-depth evaluation of a scientific review on fertility preservation written by ChatGPT-4. J Assist Reprod Genet. 2024;41(7):1871–1880. PMID: 38619763. doi:10.1007/s10815-024-03089-7