DementiaBench v0 · compiled from dementia home-safety guidance (Alzheimer's Association, NIA, Alzheimer's Society)

The flight simulator for home-health AI.

HealthDojo turns real clinical guidelines into synthetic homes with exact, verified labels, so AI companies can pressure-test their models before a single patient trial. Guidelines in, benchmarks out.

HomeBenchDementiaBench
The report card, at a glance

The problem

To someone with dementia, a dark mat is a hole and a mirror is a stranger.

6 in 10

people living with dementia will wander at least once (Alzheimer's Association).

~31%

of falls are caused by the environment: rugs, missing grab bars, low toilets.

26% / 38%

fewer falls with home modification, 38% in high-risk patients (Cochrane 2023).

Home-health AI is shipping vision models into homes. Nobody can test them: real homes have no ground truth. HealthDojo compiles one from the guideline.
01 · guideline

Guideline

Start from dementia home-safety guidance (Alzheimer's Association checklist, NIA, Alzheimer's Society, HEAP).

02 · rubric

Rubric

Draft hazard checklist, pending clinician review. Every row cites a guideline line.

03 · scenes

Scenes

Label before pixels: seed one hazard, render, verify it's really there. Boxes come free.

04 · report card

Report card

Any home-health model runs the same eval out of the box: hits, misses, false alarms.

DementiaBench leaderboard

Who can see a dangerous home?

Every model sees the same scenes and the same 14-hazard checklist. Click a column to sort.

Score = mean of recall and (1 − false-alarm rate). Recall = seeded hazards found. False alarm = model reports a hazard we verified is not in the room (e.g. "no grab bar" when there are three). Precision = hits ÷ (hits + verified false alarms). Loc IoU = box overlap on hits.
Where every model fails

Hardest hazards · mean recall across models

Recall by hazard type

Recall by room

0% found
100% foundn/a: not tested yet
Object = something dangerous within reach (pills, knives, bleach, a heater against bedding). Access / wandering = something that enables an unsafe activity (car keys by the door, an open front door). Misperception = a normal object that dementia makes dangerous (a dark mat read as a hole, a bold rug read as steps, a mirror read as an intruder).
Scene explorer

Open any room. See what each model saw.

Each tile is a verified scene. The bar shows how many models found its hazard. Click to compare the clean room with the hazard edit and overlay any model's boxes.