PICOS framework for building answerable clinical questions

Most unproductive literature searches fail at the question, not at the database. A question such as "does this drug work in heart failure" has no defined patients, no comparator and no outcome, so any result looks relevant and none of them settles anything.

A structured question does three jobs at once. It defines the concepts you search for, it sets the inclusion criteria you apply to what you find, and it tells a reader of your summary exactly what the evidence does and does not cover.

PICO was introduced in evidence based medicine teaching in the 1990s as a way to make questions answerable. PICOS adds the study design element, which is why it is the version most often used in systematic review protocols and in reporting guidance for reviews.

• The question becomes reproducible, so a colleague can run the same search and reach the same set of studies.
• Inclusion and exclusion decisions follow rules written before you saw the results.
• Gaps become visible, because an empty cell in the framework shows a part of the question you have not yet defined.

Each letter stands for one part of the question. Write each one down in a sentence before you open any database.

P for population or problem

Who are the patients, and in what setting? Define the condition, how it is diagnosed, the age range, disease severity or stage, prior treatment and care setting where these matter. "Adults with heart failure with preserved ejection fraction, hospitalised in the past year" is a population. "Heart failure patients" is a topic.

I for intervention

What is being done to the population? A drug with its dose and route, a device, a procedure, a diagnostic test, a care pathway or a behavioural programme. Say whether you mean a specific product, a whole class or any agent with a given mechanism.

C for comparator

What is the intervention compared with? Placebo, usual care, an active comparator, a different dose, or no intervention. Many questions are weakened by leaving this blank, because the meaning of a result changes completely when the comparator changes.

O for outcomes

Which results matter to the decision? Separate patient important outcomes (mortality, hospitalisation, function, quality of life, serious adverse events) from surrogate markers (blood pressure, HbA1c, tumour response). Name the time point when it matters, for example outcomes at 12 months.

S for study design

Which designs can answer the question? For a question about the effect of a treatment, randomized controlled trials are usually the target. For rare harms, long term safety or prognosis, cohort studies, registries or case control studies may be the right design. Choosing the design is part of the question, not an afterthought.

The core idea stays the same across variants. Each one adjusts the elements to fit a different kind of question.

• PICO is the original four element form, common in clinical teaching and point of care questions.
• PICOT adds T for time frame, the period over which outcomes are measured or the follow up duration. It is common in nursing research.
• PICOS adds S for study design and is common in systematic review protocols and screening criteria.
• PECO replaces intervention with exposure, for questions about environmental, occupational or lifestyle exposures where nobody assigns the exposure on purpose.
• PIRD (population, index test, reference standard, diagnosis) and similar forms suit diagnostic accuracy questions.
• SPIDER (the group of participants studied, phenomenon of interest, design, evaluation, research type) was developed for qualitative and mixed methods research, where comparators and numeric outcomes often do not apply.

Pick the variant that fits the question type, then use it consistently from search to inclusion to reporting. Our guide to levels of evidence explains why the design element matters when you later weigh what you found.

The examples below are generic illustrations of how a question is structured. They do not describe any specific study or its results.

A therapy question

A clinician wants to know whether an SGLT2 inhibitor reduces hospitalisation in heart failure with preserved ejection fraction.

• Population: adults with heart failure and preserved ejection fraction.
• Intervention: any SGLT2 inhibitor at a standard dose.
• Comparator: placebo on top of usual care.
• Outcomes: heart failure hospitalisation and cardiovascular death, at least 12 months of follow up.
• Study design: randomized controlled trials.

A safety question

A medical affairs team needs evidence on a rare adverse event of a biologic class.

• Population: patients with an immune mediated inflammatory disease.
• Intervention: any agent from the biologic class.
• Comparator: conventional systemic therapy or no biologic exposure.
• Outcomes: the specific serious adverse event.
• Study design: randomized trials plus large observational cohorts and registries, because rare events are underpowered in trials.

A feasibility question

A clinical operations team asks which phase 2 and phase 3 trials are recruiting a similar population.

• Population: the target population of the new protocol.
• Intervention: drugs with the same mechanism of action.
• Comparator: not critical for the landscape, but noted when present.
• Outcomes: primary endpoints used by comparable trials.
• Study design: interventional trials, phase 2 and phase 3, status recruiting or not yet recruiting.

The framework becomes useful when every element maps to something you can actually type or select. A practical way to do it is to build one block of terms per element and combine the blocks.

• Within one element, combine synonyms with OR: the generic name, brand names, the drug class, the mechanism, common abbreviations and spelling variants.
• Between elements, combine blocks with AND, usually population AND intervention, and add outcome terms only when the results are too broad.
• In bibliographic databases such as PubMed, use controlled vocabulary (MeSH terms) alongside free text words, because recent records and some journals are not yet fully indexed.
• Use study design filters, such as a validated filter for randomized controlled trials, as the S element instead of adding design words by hand.

Trial registries work differently from bibliographic databases. They store structured fields, so several PICOS elements become filters rather than keywords.

• Population maps to the condition field, plus age group and sex filters, plus eligibility criteria text.
• Intervention maps to the intervention field, where you enter the drug, device or procedure.
• Comparator usually lives inside the arm descriptions, so you search it as text within arms and interventions.
• Outcomes map to the primary and secondary outcome measure fields.
• Study design maps to study type (interventional or observational), phase, allocation and masking.

Our step by step guide on how to search ClinicalTrials.gov shows exactly which fields and filters to use. For broader coverage across countries, the EU clinical trial register and the WHO platform add trials that do not appear in a single registry.

These errors appear again and again in search strategies, protocols and quick literature checks.

• Population too broad. "Patients with cancer" returns thousands of irrelevant records. Name the tumour type, stage and line of therapy.
• Population too narrow. Requiring a very specific subgroup in the search terms can drop trials that report that subgroup only in their results.
• Missing comparator. Without a comparator you cannot tell an effect from a natural course of disease, and you mix placebo controlled and active controlled trials without noticing.
• Outcomes searched as required terms. Outcomes are often reported in full text but not in titles or abstracts, so adding them with AND can hide relevant trials. Use them for screening instead.
• Only one name for the intervention. Code names from early development, generic names and brand names often appear in different records of the same programme.
• Design chosen after reading results. Deciding which designs count after you have seen the findings introduces bias, even when it is unintentional.
• No time frame. A question about long term safety answered with 12 week trials is answered only on paper.

You can ask ClinicalEvidence a question in plain English, and a well formed PICOS question gives the best results. The product reads the population, intervention, comparator and outcome in your sentence and searches registered trials and published studies with semantic and full text search, so synonyms and drug classes are covered without a hand built string.

The study design element maps directly to filters for phase, study type, status and results available. You can also narrow by condition, intervention, date range, sponsor type and country, the same way you would in a registry, but across ClinicalTrials.gov, PubMed, PubMed Central, the EU Clinical Trials Register and the WHO ICTRP at once.

• Every sentence of the answer carries an NCT number or PMID, and each citation is verified against the index before you see it, as described on our page about cited AI answers.
• When no evidence matches your PICOS question, the answer says so plainly instead of filling the gap from model memory.
• The clinical trial table lists every matching trial with phase, population, interventions, endpoints, status and enrolment, so you can check your inclusion criteria row by row.
• For larger reviews, see how the same structure supports an AI literature review.