EvidenceMD DeepSearch · Medical deep research
EvidenceMD DeepSearch writes evidence-based clinical reasoning reports. Ask one clinical or biomedical question, from a diagnostic case to a research or treatment question, and it plans the research from the perspective of the right specialty, searches peer-reviewed literature and clinical guidelines, and writes a cited report of typically 20–25 pages, usually in one to three minutes, with a graded bottom line, an evidence table and the clinical reasoning laid out step by step.
Updated
An evidence-based clinical reasoning report answers one clinical question the way a specialist would: it gathers the trials, reviews and guidelines, ranks them by strength, reasons from them to a conclusion and says how certain that conclusion is. DeepSearch writes one by medical deep research, an AI agent that plans several searches of the literature, reads what it finds and writes a structured, cited report rather than a paragraph. It works only for medicine and health, and only from scholarly and official sources: it ranks the evidence by study design, says how certain each conclusion is and where the evidence runs out, and lays out every report the same way, so you know where to look for the answer, the numbers and the caveats.
Quoted from a real report · Vascular neurology
Asked: In acute ischaemic stroke, how do randomized trials of tenecteplase compare with alteplase for functional outcome and safety, including the extended time window and patients going on to thrombectomy?
Each conclusion carries its own certainty.
Three of the report's ten rows.
| Source | Design | Population (n) | Comparison | Key result |
|---|---|---|---|---|
| Dose-restricted randomized-trial meta-analysis | Meta-analysis of 11 RCTs | Acute ischaemic stroke (7,546) | Tenecteplase 0.25 mg/kg vs alteplase | 90-day mRS 0–1: OR 1.14 (95% CI 1.03–1.25) vs alteplase |
| AcT | Phase 3 randomized non-inferiority trial | Thrombolysis-eligible acute ischaemic stroke within 4.5 hours (1,577) | Tenecteplase 0.25 mg/kg vs alteplase | mRS 0–1 at 90–120 days: 36.9% vs 34.8%; unadjusted risk difference 2.1% (95% CI –2.6 to 6.9), meeting the −5% non-inferiority threshold |
| TIMELESS | Randomized placebo-controlled trial | Perfusion-selected large-vessel-occlusion stroke at 4.5–24 hours; most underwent thrombectomy | Tenecteplase vs placebo | No improvement in 90-day clinical outcomes versus placebo; sICH incidence was similar |
64K reasoning tokens · 54.6% on HealthBench Hard
EvidenceMD's clinical reasoning model scores 54.6% on HealthBench Hard, the 1,000 clinical conversations OpenAI selected because frontier models find them hardest: 8.4 points ahead of GPT-5.4 High. On a complex question it reasons through up to 64,000 tokens and shows that reasoning, with every claim bound to a source.
DeepSearch is the long-form side of the same platform. When a diagnostic, research or clinical question needs a full report rather than an answer, DeepSearch plans the research, appraises the evidence and writes the clinical reasoning out step by step.
HealthBench Hard score
How it works · Five steps you can watch
The research panel shows each step while it runs: the brief, the plan, the searches, what the sources cover and each section as it is written. You can leave the page; the report is saved in its thread.
A clinical question, a comparison or a topic. Attach PDFs or images if you have them, and choose a specialty and one of 30 languages.
Each comes with numbered options or room for your own answer, and every one is answered before the research starts, so the report answers the question you meant.
A research brief, the specialist perspective the report will take, what evidence would settle the question, up to six sub-questions and five searches.
Peer-reviewed literature and guidelines are searched, the sources are checked against every sub-question, and a further search runs for anything still missing.
Every section is written at the same time from the same sources and outline, so citation numbers agree throughout, and the references close the report.
Each one is a Findings section of the finished report, in this order.
The same structure, every report
The answer comes first, with its certainty, and the reasoning follows. Because every report is laid out the same way, you always know where to find the numbers, the guideline positions and the caveats. The Guidelines section appears whenever a guideline is among the sources.
01Research summary
The answer first, with a bottom line and how certain it is.
02Evidence at a glance
Key studies by design, population, comparison and result.
03Findings
Each part of your question worked through in depth.
04Guidelines
What the guideline bodies recommend, and where they differ.
05Certainty and limitations
Where the evidence is strong, indirect or conflicting.
06Clinical reasoning
Step-by-step specialist reasoning with illustrative scenarios.
07What this means in practice
Concrete next steps for your patient or your work.
08Open questions and references
What remains unsettled, with the sources cited.
Appraisal, not summary
Every source is read and classified by what it is, who it studied, what it compared and the results it reports, with their numbers. Then the evidence is ranked before an answer is reached.
Never cited, even when on topic: blogs, news, consumer-health and patient-education pages, and advocacy or marketing sites.
Where sources conflict, the report works out why, by population, dose, endpoint, era or risk of bias, instead of averaging them or picking one.
Effect sizes keep their comparator and interval, for example OR 1.14 (95% CI 1.03–1.25) vs alteplase, or the result in words when the source has no number.
Each conclusion comes with what would overturn it: a different patient feature, a stronger trial or a contrary guideline.
Four real reports · Open any of them in full
One for each kind of question: medical research, clinical research, a complex patient case and a topic report. Each is the PDF the app downloads, without citation markers.
Medical research · Cardiology and nephrology
Asked: What mechanisms explain the cardiovascular and kidney benefits of SGLT2 inhibitors beyond glucose lowering, and how strong is the evidence for each?
Clinical research · Vascular neurology
Asked: In acute ischaemic stroke, how do randomized trials of tenecteplase compare with alteplase for functional outcome and safety, including the extended time window and patients going on to thrombectomy?
Complex patient case · GI oncology
Works through: Do confirmed new or enlarging liver lesions constitute RECIST progression despite a stable primary and falling CEA?
Topic report · Dermatology
Topic: Hidradenitis suppurativa: pathogenesis, staging and the current treatment landscape, including biologics.
For clinicians, researchers and pharmacists
Diagnostic reasoning, medical research, clinical research and clinical topic reports, for any question that deserves more than a paragraph. For a quick cited answer at the point of care, clinical reasoning is faster; to turn a topic into slides, use clinical presentations.
Describe the patient and ask the hard question. The report reasons with the specialty's own diagnostic criteria and staging, states how certain the conclusion is and what would change it, as the oncology case does for RECIST progression with a falling CEA.
Read the example reportAsk how strong the evidence is for a mechanism or a claim. The SGLT2 report grades each cardiovascular and kidney mechanism separately and ends with the open questions, the background a grant or a paper starts from.
Read the example reportCompare treatments trial by trial, with effect sizes, safety signals and guideline positions side by side. The stroke report does this for tenecteplase against alteplase, useful for P&T and formulary review.
Read the example reportName a condition and get pathogenesis, staging, the treatment landscape and the trials by name: the background for teaching, journal club and grand rounds.
Read the example reportWhy a medical deep research agent
ChatGPT deep research, Gemini Deep Research and Perplexity are built to research any topic across the open web. DeepSearch is built for one domain, and its rules come from how specialists appraise evidence. For every tool side by side, read the guide to deep research tools by job.
Built to be checked
The bottom line says how sure each conclusion is, and the certainty section says where the evidence is indirect, imprecise or conflicting.
In the app every claim links to its numbered source, and the references list every source the report used.
Questions outside medicine and health are declined before anything runs, and never counted against your allowance.
A report that fails to complete is not counted. HIPAA-aligned security, encryption in transit and at rest, BAA on eligible plans.
A DeepSearch report is a rapid evidence synthesis, not a systematic review, and EvidenceMD is clinical decision support that surfaces evidence and reasoning in support of a licensed clinician's judgment. Read the bottom line and open the sources behind any conclusion that changes what you do. Security details are in the Trust Center.
DeepSearch is included with annual plans: 3 reports a month on Individual, more on Max. The example reports above are free to read and download.
Reports in 30 languages