NeuroTrials.ai
NeuroTrials.ai is a curated database of practice changing neurology trials, built by board certified neurologists at NeuroCare Medical LLC in Florida, with an AI assistant that answers clinical questions from it. More than 1,300 trials are summarized and searchable by specialty (stroke, headache, epilepsy, movement disorders, dementia, neuromuscular disease, neuro oncology and neuroimmunology) through a keyword system its neurologists built, alongside a 670 topic NeuroWiki, journal updates and iOS and Android apps. The site reports 105,800 visitors over the last six months.
The AI side is NeuroEvidence.ai, a beta clinical support tool built by neurologist Ahmed Koriesh, MD, which the trial database feeds. It answers neurology questions with inline citations to trials, guidelines, drug labels and diagnostic criteria, and it runs case modes for stroke codes, status epilepticus, myasthenic crisis, Guillain Barre syndrome, spinal cord emergencies and acute headache. In those modes, eligibility for treatments such as tPA comes from hand coded protocol logic rather than from the language model, and a three layer grounding check is designed to stop the model stating a number that is not in a retrieved source.
Both are free for clinicians, trainees and researchers. The terms describe the service as educational, ask users not to enter patient identifiers, and say it is not offered as a HIPAA business associate.
Capability Axes
An AI Health Index grade measures what a buyer can verify from public sources on the date shown. It is not a rating of how good the product is. A vendor can build an excellent system and grade low on an axis because it publishes nothing an outsider can check. How grades read
The core asset is human work. Board certified neurologists choose, review and summarize the trials, and a keyword system they built drives search, so the database would be useful with no model in it at all.
The model layer is NeuroEvidence.ai, a beta assistant that retrieves from those structured trial, drug, guideline and criteria records and writes cited answers. Its acute case modes deliberately keep treatment eligibility, such as tPA and thrombectomy windows, in hand coded rules rather than the model, so the AI writes and retrieves while the decisions with the highest stakes run on fixed logic.
The clinician reads every answer and makes every decision; nothing reaches a patient directly. The boundaries are stated: the model may only give a number, dose, NNT or hazard ratio that comes from a retrieved source, missing evidence should produce an insufficient data reply, every answer shows its citations inline, and treatment eligibility in the acute modes comes from deterministic rules written by a neurologist.
The company says the AI retrieves and cites and does not invent, a promise checkable against the citations under each answer. What is missing is what happens after an error: a report link invites users to flag inaccuracies, but no process says how a wrong answer or protocol rule is corrected, or how fast.
The architecture is described in working detail. Sources are stored as typed, validated records (more than 20,000 data points across trials, guidelines, drugs and diagnostic criteria), vector search finds candidates, a clinical reranking layer promotes the most useful source, and the three layer grounding check sits around generation. Treatment logic in the acute modes is hand coded, and local medical embeddings are in progress.
The answer settings on NeuroEvidence list Claude Haiku and Claude Sonnet as model choices. No version, update schedule or notice of model change is published.
The model layer is named on the product itself: NeuroEvidence's answer settings offer Claude Haiku and Claude Sonnet, Anthropic's models, so questions and retrieved evidence go to Anthropic to be answered.
Nothing else in the chain is named. No hosting provider, email or analytics service, or subprocessor list is published, and no terms say what Anthropic or any other party may keep from the conversations it receives.
Usage is reported and performance is not. The site counts more than 1,300 curated trials, 670 wiki topics, 105,800 visitors and 321,200 trial views over the last six months, which describes reach rather than accuracy or benefit.
NeuroEvidence lists physician evaluation, adversarial testing and accuracy audits as work in progress, and no result from them is published. No named hospital, department or residency program is shown using either product, and no figure for answer accuracy, citation accuracy or protocol engine agreement with clinicians is given.
The data commitments are categorical: no patient identifiers in the service, no sale of personal information, and saved conversations deletable by the user at any time along with the account. Whether conversations or questions are used to improve the models is not stated, and no retention period applies to data kept while an account stays open.
Unusually, the safety engineering is described. NeuroEvidence runs three checks against made up figures: prompt rules that block unsourced numeric claims, an evidence gate that tightens the prompt to what was actually retrieved, and a validator that scans each answer for invented NNT, hazard ratio, odds ratio and p values and blocks them before the user sees the answer. Missing evidence is meant to produce an explicit insufficient data reply.
The position is plain: the service is not designed to store protected health information and is not offered as a HIPAA business associate. Users agree not to enter names, record numbers, dates of birth or admission dates, and must de identify any case before typing it in, which matters because the stroke and seizure modes ask the clinician to describe the patient.
What stands behind that position is thin. The NeuroEvidence terms carry a short privacy section, and the privacy policy linked from NeuroTrials.ai returns a missing page, so no full privacy document covers what is collected from clinicians and kept from their conversations.
No security certification, outside test or security page is published. Accounts use passwordless sign in by emailed one time code, and the terms bar attempts to interfere with the service's security features.
The data held is clinicians' account details, including occupation, specialty and National Provider Identifier, and their saved conversations, which may describe cases even with identifiers removed.
Both terms documents call the service educational and informational, with no medical advice, diagnosis or treatment and no reliance on AI answers for clinical decisions. NeuroEvidence's own pages describe it differently: a clinical decision support platform for real neurology cases, with stroke code and status epilepticus modes that walk the clinician through the case, and tPA eligibility and thrombectomy windows decided by its protocol engines.
Every output goes to a clinician with its sources shown, and the product is labeled beta. No regulatory position is stated for the case modes, the protocol engines or the cited answers.
Testing is described as ongoing work rather than finished results: continuous testing on real clinical scenarios, probes designed to break the grounding, and accuracy audits, with a promise to ship only when the clinical bar holds. Clinical accuracy is named as the success measure.
No framework, test set or result is published, and nothing reports how answers or protocol logic perform for particular patient groups, such as older adults, pregnant patients or children, where neurology guidelines and drug dosing differ.
Ways to check and challenge an answer exist. Every answer shows its trial, guideline, label or criteria sources inline, a validator blocks invented statistics before they appear, and a report link invites users to flag a misquoted trial, a wrong dose or an outdated guideline.
Nothing stands behind an answer that is wrong. The terms disclaim accuracy, exclude liability for clinical decisions made in reliance on the service, and cap total liability at $100.
Nothing connects to a record system. NeuroEvidence can draft the history of present illness section of a clinical note from a case, which the clinician copies to the clipboard and pastes wherever they chart.
No EHR, single sign on, institutional account or interface standard is described, and the terms ask users to keep patient identifiers out of the service, so it is built to sit beside the record rather than inside it.
Nothing is published about where the service runs or where account data and saved conversations rest. It is offered as a website and as iOS and Android apps, and NeuroEvidence sits inside a wider multispecialty platform called MedEvidence.
No hosting provider, region or separation between users is named.
Everything is free. Clinicians, trainees and researchers create a free account online to save searches and conversations, the NeuroTrials apps are free on iOS and Android, and the weekly evidence digest is free as well, with no sales conversation anywhere in the path.
No institutional or enterprise offer is published, and the terms bar reselling or commercially using the service without written permission, so how a hospital or residency program would license it is not stated.
Scope is neurology and nothing else, set out precisely. The database covers stroke, headache, epilepsy, movement disorders, dementia, neuromuscular disease, neuro oncology and neuroimmunology, and NeuroEvidence adds case protocols for stroke codes, status epilepticus, myasthenic crisis, Guillain Barre syndrome, spinal cord emergencies and acute headache. Answers can be tuned for a neurologist, a resident or a generalist.
Nothing yet validates use in any of those settings. The case modes are aimed at the bedside and at rounds, and the product describes itself as in physician testing, with no named site running it.
Alternatives to NeuroTrials.ai
Other products that do the same kind of work as NeuroTrials.ai, and what each one does.
Clinical AI reference platform from CaryHealth (Cary Rx Incorporated, Washington, DC), launched in August 2024 on the back of the company's GalenAI acquisition and available at askclair.ai, on the web and as iOS and Android apps, with an API for embedding.
Clinical evidence search engine for licensed clinicians, operated by Veracity-Health Inc., a Delaware corporation with a San Francisco address, and free to verified clinicians and trainees with unlimited searches.
Doximity Ask, formerly Doximity GPT, is the clinical AI assistant inside Doximity's professional network, whose members include more than 85 percent of US doctors.
Generative chat interface over the American Society of Clinical Oncology's own library of clinical practice guidelines, announced in May 2025 and developed by ASCO and Google Cloud on the Vertex AI platform with Gemini models.
UpToDate is the clinical reference and decision support product of Wolters Kluwer Health, written by more than 7,600 physician authors and editors and used by more than three million clinicians.
Drug information service delivered as software, from an Atlanta company operating since at least 2014. Cofounders Ashish Advani, a pharmacist and the chief executive, and Tulasee Rao Chintha, chief technology officer.
Pricing
Vendor-published figures are labeled as such. Figures labeled “Estimated” are derived from third-party sources and have not been confirmed by the vendor.
| Entry Price | Pricing Basis | BAA Tier | Implementation | Source |
|---|---|---|---|---|
|
Free
$0 baseline
|
Free to use. Clinicians, trainees and researchers create a free account online to save searches and conversations, and the NeuroTrials apps for iOS and Android are free. | Not offered. The terms say the service is not designed to store protected health information and is not offered as a HIPAA business associate. | None. Accounts are created online with an emailed sign in code. | Vendor Published |
NeuroTrials.ai and its AI assistant, NeuroEvidence.ai, are free for individual clinicians, with a free weekly evidence digest by email. No institutional, residency program or enterprise license is published, and the terms bar reselling or commercial use of the service without written permission.