Cadence · eMed Chronic Health Hackathon

Closing the snapshot gap in Parkinson's care.

One goal: help people with Parkinson's keep the life they have now, for as long as possible.

Karina · Daniel · Henry · [name] — everything shown is running.

The snapshot problem

A neurologist sees 2 dots. The disease is the line.

OFF mornings · freezing · a fall · dyskinesia creeping in — growing, and invisible ✕ fall clinic visit 15 min · a good day next visit 6–18 months later

Between the dots, dosing runs open-loop. The gold-standard fix? A paper diary — self-rated every 30 minutes, for 3 days, by someone with a movement disorder.

The Parkinson's journey

15–20 years of disease. Care runs on minutes of it.

Today a missed year a 15-min appointment every 6–12 months — often longer · each one a snapshot, on a good day Cadence 🎙 voice · 🎥 movement — captured passively at home, every day of all 15–20 years year 0 · diagnosis year 20 snapshots: ~25 visits ≈ 6 hours of observation longitudinal: ~7,300 days of voice + visual data

The doctor has only ever seen the dots. Cadence hands them the line — a longitudinal record of voice and movement, so every scarce appointment starts from what actually happened.

Why timing is the whole game

Levodopa fades in ~90 min. Wearing-off is a time-of-day pattern.

Month 1 the one moment a clinic sees Month 6 07:3011:30 15:3019:30 6:0023:00
ON — medication working OFF — symptoms back ON + dyskinesia dose

Six months later the OFF wedges are wider and dyskinesia caps the dose — the window is narrowing, and no snapshot can see it happening.

Cadence

We automated the validated instrument.

The ON/OFF diary, made objective and continuous — inferred passively from movement and voice, 365 days instead of 3. Footage never leaves the home; only features do.

Capturecamera + mic · passive, consent-first
→
ExtractMediaPipe pose 33 pts · voice
→
Featurestremor · gait · sway · speech
→
InterpretLLM → one clinical flag
→
Surfacesdoctor · patient · family

Capture + extract is Daniel's working pipeline (OpenCV, MediaPipe, Gemini voice) — not a mock-up. Trend vs own baseline, not UPDRS scores. Decision support — the clinician always decides.

Live — everything runs

Watch it measure, interpret, reassure.

DEMO 1Capture Real patient footage → tremor, freezing, dyskinesia features live. Voice check-in logs the dose; tremor visibly settles. Open capture → DEMO 2Doctor The diary grid, dose-by-dose response, one titration flag — and pull the appointment forward when the trend demands it. Open dashboard → DEMO 3Patient & family Same data, opposite rendering: one state, one action, no scary charts. Reassurance, not surveillance. Toggle in dashboard →
Where this goes
🤖

The sensor learns to walk.

Home robots are shipping — 1X Neo (1x.tech/neo). Same pipeline, plus hands: it brings the 15:30 dose, steadies a transfer, calls for help when it can see help is needed.

Nearer term: smart pillbox + mirror · multilingual voice · training data across ethnicities.

Cadence

A 90-minute drug should not be tuned on a six-month-old memory.

20 years of disease, 5–7 hours with a doctor. We make the other 175,000 hours count. Thank you.

Cadence · closing the snapshot gap← → · 1 / 7