Evidence

The evidence.

What we’ve validated, on which data, with what result — including what didn’t work.

Part 1 · For drug development

What Brainify.AI found in completed clinical trial data.

We used the Foundational Brain Model to identify clinically relevant subgroups and clarify responder and placebo effects. These are retrospective analyses of completed studies, so they support development decisions and prospective study design rather than claiming prospective clinical validation.

+11.1 percentage points

CAN-BIND-1 · response rate, N=181 → 70

The Foundational Brain Model identified a model-positive subgroup. Its observed response rate was 58.6%, compared with 47.5% in the full dataset. Mean MADRS improvement changed from −13.7 to −16.2 points.

NCT01655706 · Ontario Brain Institute · 6 Canadian sites · escitalopram

p = 0.01

EMBARC · from p = 0.14, N=199 → 145

The model stratified participants by predicted placebo-response risk. In the lower-risk subset, the observed sertraline-versus-placebo difference changed from −1.33 (p = 0.14) to −2.63 (p = 0.01). The response-rate gap increased from 11.4 to 16.1 percentage points.

NCT01407094 · NIMH-funded · 4 US sites · MDD, HAM-D at week 8

86% / 79%

Alzheimer's · sensitivity / specificity

>800 subjects with primary diagnosis and EEG; >1,000 with secondary diagnosis; >500 with EEG recorded prior to diagnosis. Brain age delta carries independent predictive power for neurodegenerative condition.

Early result · retrospective, internally validated

All three are retrospective re-analyses of completed studies, not prospective results.

Population Ruler Framework

Phase 1 generalized anxiety disorder (GAD) program: reference distributions for 14 QEEG biomarkers from 859 GAD-diagnosed patients and ~103,000 clinically referred comparator subjects.

“Theta power, alpha asymmetry, connectivity and network features showed negligible population-level separation (d < 0.2). We publish what doesn’t work.”

Model performance

TaskResultStatus
Placebo response prediction~70% accuracy · 69% balanced accuracy across 3 independent datasetsPeer-reviewed, Neuroinformatics 2025
Alzheimer's detectionSensitivity 86% · specificity 79%Early result
Seizure detection~92% accuracyEarly result
Sex prediction~84% accuracyNeuroImage 2024
Brain ageMAE ~5 yearsFront. Aging Neurosci. 2022 / 2024

Peer-reviewed publications

Predicting Placebo Responses Using EEG and Deep Convolutional Neural Networks

Neuroinformatics, Vol 23, art. 32 (2025) · DOI: 10.1007/s12021-025-09725-6

Predicting age from resting-state scalp EEG signals with deep convolutional neural networks

Front. Aging Neurosci. 14:1019869 (2022) · DOI: 10.3389/fnagi.2022.1019869

Prediction of brain sex from EEG using a large-scale heterogeneous dataset

NeuroImage 2024 Jan;285:120495 · DOI: 10.1016/j.neuroimage.2023.120495

Optimization of the Deep Neural Networks for Seizure Detection

ICASSP 2023, Rhodes Island · DOI: 10.1109/ICASSP49357.2023.10094645

Data Leakage Problem in Large Multi-site EEG Datasets

ISBI 2023, IEEE, Colombia

All publications

How to read these results

We test whether the model is learning clinically relevant biology, rather than shortcuts such as the hospital or recording device used to collect the data. The main limitation today is that these trial results come from retrospective re-analysis. Prospective validation is still required, and no Brainify.AI biomarker has been qualified by FDA or EMA.

Technical methodology and limitations

Multi-site EEG can leak recording-site information into a model. We published this failure mode (IEEE ISBI, 2023) and use leakage-aware validation splits so performance is less likely to reflect which hospital collected the signal.

The trial results above are retrospective re-analyses, not prospective results. Prospective validation is in progress; regulatory qualification has not yet been obtained.

Part 2 · For wellness partners

What the brain wellness scores are built on.

Plainer, shorter, no p-values. Written for a medical director deciding whether to put this on their menu.

The reference data

Scores are compared against age- and sex-adjusted reference distributions built from 500,000+ brain scans. This is a clinical-population reference, not a screened healthy-control cohort — a score in the typical range means typical relative to that population.

Precision

Our EEG brain-age model has a mean absolute error of approximately 5 years. Brain age reflects current brain state — including sleep, arousal and time of day — and is not a measure of neurological health or cognitive ability.

R&D and people

Research collaborations with Harvard Medical School, Weill Cornell Medicine and UC Irvine. Scientific advisory board: Pizzagalli (UC Irvine), Liston (Weill Cornell), Fava (MGH), Deligiannidis (Northwell), Freeman (Harvard Medical School), Doherty (Acumen, ex-Sage, ex-AstraZeneca), Kelly (ex-Novartis).

The device

Professional-grade 19-channel array. Medical-grade components — Swiss-made sensors, Japanese cap materials, biocompatible. EMC and electrical-safety tested. Passive recording: no current, no stimulation.

What this is not

For general wellness use only. Not a medical device. Not intended to diagnose, treat, cure, mitigate, or prevent any disease or condition. Does not constitute medical advice.

Not a diagnostic. Not a screening test. It does not detect, rule out, or monitor any medical condition. If a recording cannot be scored or shows an atypical pattern, the report is withheld and your center advises the client to consult a physician.