Evidence & Methodology

You should know exactly
how this works.

Every recommendation Meddy surfaces is backed by real, verifiable sources. It shows you where it came from.

Output Shown To You

Consider reassessing GLP-1 titration schedule — slower titration associated with improved tolerability in comparable cohorts.

Traced To

Diabetes Care, 2024
ADA Standards of Care, 2025
JAMA Intern Med, 2023
01The Data Sources

Built on sources you already recognize.

Nothing proprietary. Nothing unverifiable.

80M+
Peer-Reviewed Academic Papers

The published literature your specialty already cites — not a proprietary research corpus.

Full
FDA Drug Label Database

Every approved indication, dosing guideline, and safety consideration, sourced directly from the label.

Major
Clinical Practice Guidelines

Established standards of care across specialties — the same guidelines you're already trained on.

Nothing proprietaryNothing unverifiable
02Driver Analysis Methodology

How we rank what's moving a marker.

The stronger the evidence, the higher the rank. You see the methodology, not just the output.

CorrelationEvidence-Ranked
Positive and negative contributors. Ranked. Cited. Adjustable by you.
A1cRanked by strength of evidence
01Body Weight+0.4%Strong Evidence12 sources
02Carbohydrate Intake−0.5%Strong Evidence9 sources
03Sleep Consistency+0.3%Strong Evidence7 sources
04Metformin Dosage Change+0.2%Weak Evidence2 sources
03The RAG Architecture

Why the AI doesn't hallucinate.

It cannot cite a source that doesn't exist. It cannot recommend something unsupported by the literature it draws from.

Step 1 · Query

Patient signal

A1c trend worsening, recent semaglutide start, T2DM + HTN.

DiagnosisTrendRecent event

Step 2 · Retrieval

Search the curated knowledge base

Only verified sources are eligible to be retrieved — nothing outside the indexed corpus exists to the model.

14 relevant sources found

Step 3 · Generation

Constrained to what was retrieved

The model can only generate language grounded in the retrieved set — it has nothing else to draw from.

3 sources cited in output

It cannot cite a source that doesn't exist.

Every citation maps back to a real, retrieved document — there is no step where the model invents a reference.

It cannot recommend the unsupported.

Generation is bounded by what retrieval actually returned — not by what the model remembers.

Every output is traceable. · Every citation is real.

04What the AI Won't Do

The boundaries matter as much as the capabilities.

It surfaces relevant evidence and flags considerations — the physician interprets and decides.

Always.

What the AI Does

Surfaces relevant evidence
Flags considerations worth reviewing
Cites the literature it draws from
RDr. R. AhnInterprets and decides

What the AI Does Not Do

Diagnose
Override clinical judgment
Decide on the physician's behalf
05Continuous Improvement

The knowledge base grows.
The standards don't change.

The methodology for how evidence is weighted and cited remains consistent.

Continuously Updated

New papers
Updated guidelines
Revised drug labels

Stays Consistent

How evidence is weighted
How sources are cited
Where the physician decides
Meddy
06Get Started

Want to go deeper?

Speak with a Meddy physician advisor about the methodology behind your specialty.