AI developers
General assistants that caregivers and clinicians ask about a parent’s or patient’s changing behavior.
AI Evaluation
Find out whether your AI misses seizures in people with dementia.
Seizures in dementia are more common than most people realize, and often taken for the dementia itself. Seagull gives your AI weeks of notes, family messages and sensor alerts, and tests whether it recognizes them.
- How soon it notices
- How often it’s wrong
- What it recommends
- 3–4×
- more likely to have seizures than people of the same age without dementia
- About half
- of seizures in dementia have no convulsions
- 7 in 10
- reported episodes in one study that specialists could not classify, because the notes lacked detail
For every team whose AI touches dementia.
Health systems and EHR vendors
Chart summaries, ambient documentation and risk alerts that read clinic and hospital notes.
Caregiving and home-care platforms
Caregiver chat, visit notes and care-plan tools used by families and home aides.
Monitoring and wearable companies
Alerts that combine sensor events with notes about falls, sleep and unusual movement.
Life sciences and research
Models that screen records for trial eligibility and safety monitoring, where a missed seizure is a missed exclusion or adverse event.
Senior living and long-term care
Charting assistants and incident review across memory care and skilled nursing.
Widely used AI models differ sharply on the same records.
In our first runs, the same records got very different readings. The strongest models raised the possibility early, the weakest let dozens of seizures in the records go by, and a simple keyword search raised false alarms almost every week. We gave seven widely used AI models the same 60 days of records for 16 synthetic patients, one week at a time: notes from family members, home aides, nurses, clinics and hospitals. Half have seizures that the records describe as behavior. The other half have look-alike episodes with another cause.
4 to 46
Seizures in the records of eight people before the strongest and the weakest model first raised the possibility.
63 of 72
Weeks a simple keyword search raised a false seizure alarm for people whose episodes had another cause. Flagging early is not enough.
2 of 7
Models that went along with an antipsychotic for a woman whose morning episodes were seizures mistaken for her dementia.
7 vs 1
People, out of eight, flagged in the same week every time a model ran three times: 7 for the most consistent model, 1 for the least.
Seizures in dementia are missed across weeks of notes.
A daughter writes that her mother “zoned out at breakfast.” A clinic note calls it anxiety. A staring spell is noted as sundowning, and a fall goes unwitnessed. No single note settles the question. The pattern across weeks, measured against the person’s own baseline, does. Most AI evaluations score one answer to one question. This one follows the records as they accumulate.
Seagull began with the founder’s mother. Her episodes of confusion and unresponsiveness were treated as agitation and sundowning, and not evaluated neurologically. They were later understood as seizures.
The test asks whether an AI reading notes like hers would read them the same way.
Read how Seagull beganHow an evaluation works.
1
Scope
You describe what your product does and who uses it. We select the people and records that match.
2
Run
Your product reads each person’s records week by week, through test access or a script you run.
3
Review
Every weekly note is scored against a clinical reviewer’s own call, and every finding is checked by hand before it reaches you.
4
Report
You receive the scorecard and the full report. If you update your product later, you can run the same records again and compare.
A scorecard your team can act on.
-
How soon it noticed
The week your AI first raised a possible seizure, against the week the clinical reviewer would have.
-
How often it was wrong
Weeks it raised a seizure for people whose episodes have another cause: Lewy body fluctuations, low blood sugar, frontotemporal behavior changes, Parkinson’s off periods.
-
What it recommended in the meantime
Whether it held an antipsychotic until seizures were considered, or went along with one.
Every finding quotes your AI’s own words and the dates of the records it relied on.
CareNote AssistExample scorecard
Alzheimer’s disease, morning episodes
Raised in week 5. Reviewer flagged week 3.
2 weeks late“Resistance to morning care, consistent with her dementia.” Week 3 note
Lewy body dementia, look-alike episodes
Raised a seizure in 1 of 9 weeks, then named the fluctuations.
1 false alarmAntipsychotic advice
Advised holding quetiapine until the spells were evaluated.
ProtectiveIllustration only. The product and figures are invented.
Built for one problem.
General AI evaluations measure broad behavior well. This problem needs dementia-specific cases, real documentation language and a clinical standard.
| General AI evaluation | Seagull | |
|---|---|---|
| What is tested | One answer to one prompt | Weeks of care records, read as they accumulate |
| The standard | Automated graders or general raters | The week a clinical reviewer would have raised it |
| The people | Generic cases or real patient data | Synthetic patients with a known cause, look-alikes included |
| The language | Clean clinical text | The way people actually write: family messages, aide checklists, clinic and hospital notes |
| The evidence | Varies | Each person grounded in the Dementia-Seizure Evidence Map |
| Your data | Often required | Never needed |
Questions companies ask.
Do you need our data or real patient records?
No. Your product reads Seagull’s synthetic patient records. No real patient data enters the test.
Which kinds of AI can you test?
Any AI that reads or writes about a person with dementia: general assistants, documentation and summarization tools, monitoring alerts, and models that screen records for research. Tests in which the AI answers a worried family member are in development.
Is this the same as licensing DSEF?
No. A DSEF license puts DSEF into your product, adapted to your organization. An AI evaluation tests the AI you already have, using Seagull’s synthetic records. You can do either, or both.
Do you tell the AI to look for seizures?
No. The instructions never mention seizures. Your AI passes by raising the possibility and sending it to a clinician. It is not asked to diagnose.
Can we see the cases?
The people, records and answer keys stay private, so the test keeps its meaning when you run it again. You see every finding, with your AI’s own words.
Does a strong result mean our product is clinically validated?
No. The evaluation shows how your product handled these cases. It is not clinical validation or certification, and the report says so.
How long does an evaluation take?
About two weeks from test access to report. The free first test takes about a week.
Will you publish our results?
Only with your permission. Seagull publishes results on general AI models without naming the companies it works with.
Start with a free first test.
Your AI reads the records of four synthetic patients: two with seizures the records describe as behavior, two with look-alike episodes. You receive a one-page result and a conversation about what it shows.
Seagull evaluates how AI products handle possible seizures in dementia. It does not diagnose seizures, certify products or replace clinical judgment.