About NeuroEvidence.ai

A clinical decision-support platform being built by a neurologist — designed to manage real neurology cases and answer knowledge questions, with grounded, evidence-based reasoning.

In active development Beta — physician testing

Important: NeuroEvidence is not a finished product. It is in active development and the content, protocols, and recommendations should always be tested against current standard of care and institutional protocols. This is a support tool, not a substitute for clinical judgment or the physician–patient relationship. See the full disclaimer.

What we're trying to solve

Clinicians manage neurology cases by moving between UpToDate, Googling trial names, checking Micromedex for drug dosing, and flipping through guideline PDFs — a fragmented workflow that's slow at the bedside and prone to gaps.

General-purpose LLMs like ChatGPT fill the gaps but carry a very high risk of hallucination — these models can't say "I don't know," so they fabricate numbers, invent trial results, and miscite guidelines. Their output is not 100% grounded in evidence, and in a clinical context that's dangerous.

NeuroEvidence takes a different approach: every clinical claim is anchored to a specific structured object — a trial, a drug label, a guideline recommendation, a diagnostic criterion. The AI retrieves and cites; it does not invent.

Built for two uses

Case management

Step-by-step clinical scenarios — stroke code, status epilepticus, MG crisis, GBS, spinal cord emergencies, acute headache — with stateful protocol engines and deterministic treatment logic.

Knowledge questions

Ask anything — trial summaries, drug dosing, diagnostic criteria, cross-trial comparisons, guideline updates — and get cited, evidence-grounded answers with inline trial / guideline references.

What makes this different

Target: 100% grounded output. The platform is designed so the AI cannot state a number, dose, NNT, or HR unless it comes from a retrieved structured source. Missing evidence triggers explicit "insufficient data" responses — not a guess.
Treatment decisions run on hard-coded medical logic. tPA eligibility, status epilepticus escalation, thrombectomy windows — these are protocol engines written by a neurologist, not LLM outputs.
Every answer cites sources inline. Trials, guidelines, drug labels, and diagnostic criteria appear as reference pills under every response.
Built by a neurologist — clinical accuracy is the success metric, not engagement or completion rate.

How we're building it

Each step has a checkmark when delivered. We ship incrementally and every layer targets clinical accuracy first.

Comprehensive neurology database Done

A hand-curated foundation of every practice-changing trial, drug, guideline, and diagnostic criterion — so retrieval has the right evidence to draw from.
Structured evidence objects Done

Every source is stored as typed, validated data — not prose — so the AI retrieves exact numbers instead of paraphrasing, eliminating misquotes.
Semantic retrieval + clinical reranking Done

Vector search surfaces relevant evidence; a reranking layer promotes the most clinically useful sources so the answer always has a canonical citation.
Three-layer grounding defense Done

Defense-in-depth against AI hallucination: prompt-level guardrails block unsourced numeric claims, a pre-generation evidence gate tightens the prompt based on what was actually retrieved, and a post-generation validator scans the AI's output for fabricated NNT, HR, OR, and p-values and blocks them before the user sees them.
Protocol engine expansion In progress

Hand-coded medical logic for high-stakes cases — adding engines beyond the current six so more of the acute neurology workflow runs on deterministic rules, not AI.
Speed — sub-second bedside response In progress

Local medical embeddings and answer caching to drop round-trip latency so the platform is fast enough to use during rounds.
Physician evaluation + adversarial testing In progress

Continuous testing on real clinical scenarios, grounding-adversarial probes, and accuracy audits. We ship only when the clinical bar holds.

Current scope

What's in the platform today. Numbers update as we add content.

20,000+

Structured data points

Across every trial, guideline, drug, criterion, and condition

1,100+

Neurology trials

Practice-changing, from acute stroke to ALS

Hundreds

Drugs — full profile

MOA, PK, dosing, adjustments, pearls

All

Neurology guidelines

All neurology-related guidelines

All

Diagnostic criteria

Every available diagnostic criterion

Case protocols

Live engines — more in progress

Who built it

NeuroEvidence is built by Ahmed Koriesh, MD — a board-certified neurologist — as part of a broader effort to make high-accuracy clinical decision support practical at the bedside.

Sister projects

NeuroTrials.ai Clinical trial wiki — the curated trial database that feeds NeuroEvidence.

HighYieldNeuro.com Board review — flashcards, study notes, and quiz mode for neurology trainees.

MedEvidence The multi-specialty evidence platform NeuroEvidence sits inside.

Found an inaccuracy?

Clinical accuracy is the whole point. If you've spotted something wrong — a trial misquote, a bad dose, a dated guideline reference — we want to know.

Report it