BrainSee

Authorization

FDA created a new Class II device type for five-year Alzheimer’s-dementia progression assessment and imposed special controls and postmarket evaluation.

Submission lineageDEN220066; no predicate submission.

Intended use

Aid assessment of how likely a patient with amnestic mild cognitive impairment is to progress to clinical Alzheimer's disease dementia within five years.

Input

3D T1-weighted brain MRI; MMSE and CDR Sum of Boxes scores; Age; Sex

Output

A 0-100 similarity score relative to converter and non-converter reference populations; Interpretive report

Use boundary

Prescription adjunct only; diagnosis, prognosis and clinical management should not be based solely on the BrainSee result.

AI disclosure and function

Algorithm / classifier

The De Novo summary describes an underlying model that combines brain MRI features, cognitive scores, age and sex to generate a prognostic similarity score.

The reviewed FDA material supports a fixed prognostic model; it does not identify an LLM, foundation model or generative-AI interface.

No foundation-model, LLM, or generative-AI function was identified in the reviewed FDA materials.

Validation

Retrospective pivotal algorithm-performance validation against longitudinal clinical reference diagnoses.

Comparator

Five-year clinical conversion versus non-conversion to Alzheimer's disease dementia at a prespecified score threshold.

Ground truth

Longitudinal neurocognitive diagnoses assigned by trained clinical panels at participating sites.

Representation

Of 198 analyzed patients, 178 (89.9%) were White, 109 were male and 89 female; FDA required broader postmarket evaluation because the cohort was not representative of US practice.

Reported performance

Sensitivity

72.9% (95% CI 64.9%-80.9%)Retrospective prognostic classification performance at the prespecified threshold.

Specificity

96.3% (95% CI 92.1%-100%)Retrospective prognostic classification performance at the prespecified threshold.

Positive predictive value

96.6% (95% CI 92.9%-100%)Observed in a cohort with 118 converters and 80 non-converters; transportability depends on prevalence and population.

Negative predictive value

70.6% (95% CI 62.1%-79.2%)Observed in the retrospective validation cohort and not a guarantee for an individual patient.

Limitations

• The validation was retrospective, not a prospective clinical-utility study.
• 178 of 198 analyzed patients were White, creating uncertainty about performance in underrepresented populations.
• The lower 95% confidence bound for sensitivity was below the predefined >70% effectiveness goal.
• The study did not establish that device use improves treatment, function or patient outcomes.

It does not prove clinical utility, improved patient outcomes, neurosurgical applicability or foundation-model use.

PCCP / Change control

No authorized PCCP was identified for this submission in the reviewed FDA database record or decision material.

Absence of an identified PCCP does not prevent every future device change; it means no authorized plan was established in the reviewed public record.

Postmarket check

No exact-name openFDA device-enforcement result was found in the targeted query; FDA separately required a postmarket performance study in more representative clinical populations.

No exact-name openFDA result found as of 2026-09-08; this is not proof that no recall exists. A required postmarket study is not a recall.

Primary FDA sources

1. FDA AI-enabled device list — Submission DEN220066Confirms inclusion, list display name, company, panel, product code and decision date only.
2. FDA submission database — De Novo DEN220066Confirms grant date, classification and product code.
3. FDA decision summary — De Novo summary pp. 1-11; clinical evidence pp. 4-7; special controls and postmarket requirements pp. 7-11Supports intended use, model inputs and output, retrospective validation, performance, demographic limitation and postmarket study requirement.

Sources checked Sep 8, 2026 · Next scheduled review Oct 8, 2026. “No exact-name result” is a bounded search finding, not proof that no safety event or recall exists.