A treatment looks promising in a recent review. Before recommending it, you need to know whether the participants resembled your patient, whether the benefit was meaningful, and whether important harms were missed. A search that stops at the abstract leaves those questions unanswered.
Clinical evidence search connects a clinical question with research that can inform a decision. A practical workflow is to define the decision, choose the right evidence, search efficiently, appraise the findings, verify any artificial intelligence (AI) summary, and apply the results to the patient and setting. This approach is useful across medicine, nursing, pharmacy, dentistry, and rehabilitation.
Define the decision before choosing keywords
Begin with the action the evidence might change: choosing a treatment, ordering a test, estimating prognosis, investigating harm or discussing prevention. Include the care setting and the period over which outcomes matter.
For treatment questions, use PICO: population, intervention, comparator and outcome. The National Library of Medicine recommends this framework to turn a broad concern into a focused question.[1]
Hypothetical search question: In adults with knee osteoarthritis receiving outpatient care, how does supervised exercise compare with a home exercise programme for pain, walking ability and adverse events over six months?
This gives the search a clear comparison. “Exercise for arthritis” does not establish which patients, programme or outcome you mean. Keep the question neutral: ask how options compare rather than asking for evidence that your preferred option works.
Adapt the structure for other questions. A diagnostic search should name the presenting population, test and reference standard. A prognosis search should define when follow-up begins and which outcome is being predicted. These details help avoid applying evidence from a different clinical situation.
Match the evidence to the clinical question
For an established management question, start with a current, relevant guideline or a trustworthy evidence summary, then examine its supporting systematic reviews. Read primary studies when the synthesis is outdated, conflicting or does not answer your question. Study design should fit the decision.[2,3]
| Question | Useful evidence | Key appraisal issue |
|---|---|---|
| Treatment or prevention | Systematic reviews of randomised trials; relevant trials | Benefits, harms and a fair comparator |
| Diagnostic accuracy | Studies comparing a test with an appropriate reference standard | Patient selection, test threshold and clinical setting |
| Prognosis | Cohorts with a defined starting point | Follow-up, missing outcomes and baseline risk |
| Medication harm | Trials, comparative observational studies and regulatory safety information | Confounding, exposure and duration |
For questions about patient experience or barriers to care, qualitative research may be more informative than an intervention trial.[3]
A meta-analysis combines results statistically; that label alone does not establish high certainty. Bias, inconsistent findings, indirect evidence, imprecision and publication bias can weaken a body of evidence.[4] Case reports may raise a safety question, but cannot by themselves establish how often an event occurs.
Search broadly enough to find the answer
Translate the question into a few core concepts. PICO structures your thinking, but including every element in a database query can exclude useful papers. Comparators and outcomes may be missing from titles, abstracts or indexing.[5]
For the hypothetical osteoarthritis question, an illustrative PubMed starting search is:
(knee osteoarthritis) AND (exercise OR physiotherapy OR physical therapy)
OR groups alternatives; AND connects concepts. Inspect PubMed’s Search Details to see how terms were translated. Add Medical Subject Headings (MeSH) alongside text words when refining the strategy. Use the Systematic Review filter to locate syntheses, or Clinical Queries to focus on therapy, diagnosis, prognosis, aetiology or prediction questions. Narrow queries favour specificity; broad queries favour sensitivity.[5,6]
Check publication dates and the last search date of a review. A newly published review may have an older evidence cutoff. Search for important studies published after that cutoff.[5]
If results are sparse, remove restrictive terms and review filters. Age and other indexing-dependent filters can miss relevant records that lack those terms.[6] A quick clinical search is not a systematic review: comprehensive reviews need documented methods, multiple sources and information specialist input.[5]
Return to the exercise example when refining results. If the search retrieves mostly postoperative rehabilitation, add terms describing the intended outpatient population. If it finds very little, remove the supervision requirement initially and examine programme delivery in the full text. Keep a record of the final search terms so the question can be revisited.
Appraise the result before changing practice
Read the methods and results behind the conclusion. Ask whether allocation was concealed, follow-up was adequate, and the comparator represented a reasonable alternative. For diagnostic studies, inspect the reference standard and recruitment setting. For observational findings, consider whether confounding could explain the association.[3]
Separate the size of an effect from confidence in that effect. GRADE, the Grading of Recommendations Assessment, Development and Evaluation approach, rates certainty for each outcome as high, moderate, low or very low. An intervention can have clearer evidence for benefit than for harm.[4]
Look at absolute effects and confidence intervals, not just a relative reduction or a statistically significant result. Hypothetical example: reducing an event from 8% to 6% over one year gives a 25% relative reduction, but an absolute reduction of 2 percentage points. The corresponding number needed to treat is 50 over one year. These invented numbers illustrate interpretation, not a treatment recommendation.[3]
Check whether outcomes matter to patients: function, symptoms, quality of life or serious complications. An improved laboratory marker does not automatically establish better health. Wide confidence intervals may leave clinically important benefit or harm unresolved.[4]
When studies disagree, inspect differences in participants, interventions and outcome measurement before interpreting the pooled average. Unexplained inconsistency can reduce certainty. A large collection of studies does not resolve that problem simply by producing a more precise combined estimate.[4]
Use AI to navigate evidence and verify the important claims
AI can help organise a question and provide a starting summary. ZoeMD supports natural-language medical questions and responses with citations and direct source links. These features provide a route to the evidence; they do not establish that each answer is complete or applicable.
The World Health Organization warns that generative AI can produce inaccurate, biased or incomplete health information. It also highlights automation bias: accepting an output too readily and overlooking errors.[7]
Before relying on a consequential claim, open the cited source. Confirm the population, intervention, outcome and relevant passage. Check for updates, corrections or retractions. A genuine paper may still be cited for a conclusion it does not support. If only the abstract is available, keep conclusions within what it reports.
For example, an AI answer to the exercise question might describe better walking ability even though the cited study measured pain alone. It might also describe a three-month finding as lasting six months. Checking the exact endpoint and time point catches these mismatches; confirming that the paper exists would not.
A reusable evidence search prompt is:
“For [population and setting], compare [intervention] with [alternative] for [important outcomes and time horizon]. Identify current relevant guidelines and systematic reviews, then appropriate primary studies where needed. Provide direct source links, publication dates and review search dates when available. Distinguish study findings from guideline recommendations. Report absolute effects and evidence certainty where available, key harms, exclusions and disagreements. State what remains uncertain and do not invent missing information.”
Treat this prompt as a search aid. Verify the response, and use only the patient information permitted by your organisation’s approved data-handling workflow. The related guide to Evidence Search for Family Physicians offers examples organised by clinical question type.
Apply the evidence and record the reasoning
Apply the evidence and record the reasoning
Applicability requires a separate judgement. Compare the study population, intervention, comparator, outcomes and setting with your own question. Age, comorbidity, baseline risk, treatment intensity and access to monitoring can change the relevance of a finding.[8]
Baseline risk deserves particular attention. Extending the hypothetical numerical example, the same 25% relative reduction applied to a one-year risk of 2% would lower risk to 1.5%. That is a 0.5 percentage point absolute reduction and a number needed to treat of 200. This calculation assumes the relative effect transfers to the lower-risk group; that assumption requires clinical justification.[8]
For medication decisions, verify the exact product, dose, route, interactions and organ-function considerations in current authoritative prescribing information for your jurisdiction. For rehabilitation, examine programme intensity, supervision and adherence. For inpatient care, consider acute illness, competing risks and what follow-up is feasible after discharge.
When recommendations conflict, compare their populations, evidence cutoff dates, outcomes and assumptions before deciding which applies. A recommendation developed for a specialist service may need different implementation where resources or expertise differ.[8]
Make disagreements specific: one source may favour an option for a selected high-risk population, while another addresses routine care. If the populations match, examine whether newer studies or different benefit–harm priorities explain the difference. Record the unresolved issue instead of presenting a false consensus.
Document the question, search date, key sources, important limitations and the reason for the decision. Include the patient’s priorities and any follow-up needed. If uncertainty could materially change care, seek relevant specialist, pharmacist or information specialist input. In urgent care, evidence searching should support timely assessment and approved escalation pathways.
In the hypothetical exercise scenario, the final discussion should address the programme studied, the outcomes it measured, and whether attendance and home practice are feasible. Ask which outcome matters most to the patient, such as walking to local shops or climbing stairs. Agree how progress will be reviewed, rather than ending with the general statement that exercise is beneficial.
For your next clinical evidence search, choose one unresolved decision and follow it from question to source to patient relevance. To explore an AI-supported approach, visit ZoeMD. Keep verification and professional judgement central to the process.
References
1. National Library of Medicine. Using PICO to frame clinical questions. In: Using PubMed in evidence-based practice [Internet]. Bethesda (MD): National Library of Medicine; [cited 2026 Oct 14]. Available from: https://www.nlm.nih.gov/oet/ed/pubmed/pubmed_in_ebp/02-100.html
2. OCEBM Levels of Evidence Working Group. The Oxford 2011 Levels of Evidence [Internet]. Oxford: Oxford Centre for Evidence-Based Medicine; 2011 [cited 2026 Oct 14]. Available from: https://www.cebm.net/wp-content/uploads/2014/06/CEBM-Levels-of-Evidence-2.1.pdf
3. American Academy of Family Physicians. Evidence-based Medicine Toolkit [Internet]. American Academy of Family Physicians; [cited 2026 Oct 14]. Available from: https://www.aafp.org/afp/authors/ebm-toolkit
4. Schünemann HJ, Higgins JPT, Vist GE, Glasziou P, Akl EA, Skoetz N, et al. Chapter 14: Completing “Summary of findings” tables and grading the certainty of the evidence. In: Cochrane Handbook for Systematic Reviews of Interventions. Version 6.5 [Internet]. Cochrane; 2024 [updated 2023 Aug; cited 2026 Oct 14]. Available from: https://www.cochrane.org/authors/handbooks-and-manuals/handbook/current/chapter-14
5. Lefebvre C, Glanville J, Briscoe S, Featherstone R, Littlewood A, Metzendorf MI, et al. Chapter 4: Searching for and selecting studies. In: Cochrane Handbook for Systematic Reviews of Interventions. Version 6.5.1 [Internet]. Cochrane; 2025 [updated 2025 Mar; cited 2026 Oct 14]. Available from: https://www.cochrane.org/authors/handbooks-and-manuals/handbook/current/chapter-04
6. National Library of Medicine. PubMed user guide [Internet]. Bethesda (MD): National Library of Medicine; [cited 2026 Oct 14]. Available from: https://pubmed.ncbi.nlm.nih.gov/help/
7. World Health Organization. WHO releases AI ethics and governance guidance for large multi-modal models [Internet]. Geneva: World Health Organization; 2024 Jan 18 [cited 2026 Oct 14]. Available from: https://www.who.int/news/item/18-01-2024-who-releases-ai-ethics-and-governance-guidance-for-large-multi-modal-models
8. Atkins D, Chang S, Gartlehner G, Buckley DI, Whitlock EP, Berliner E, et al. Assessing the applicability of studies when comparing medical interventions [Internet]. Rockville (MD): Agency for Healthcare Research and Quality; 2011 Jan [cited 2026 Oct 14]. Report No.: 11-EHC019-EF. Available from: https://effectivehealthcare.ahrq.gov/sites/default/files/methods_guide_atkins.pdf
