The problem
What's broken
Across our 14 dialysis clinics in the Mountain West region, roughly 1 in 7 same-day appointments doesn't happen. The patient doesn't show. Sometimes they call. Mostly they don't. We hold the chair. The chair is expensive.
We've tried calling everyone the morning of. We've tried text reminders. We've tried a deposit system. None of it has moved the rate below 14%. We think the signal is in our existing data — we just haven't found it.
What we want
The output
A model that ranks every scheduled appointment by likelihood of no-show, run nightly against the next day's schedule, with a confidence score. We need it to be actionable 12 hours in advance — not 2 hours, not 30 minutes.
- Input: our existing scheduling data (3 years), demographics, prior-visit history, weather, distance-to-clinic.
- Output: ranked list of next-day appointments with no-show probability, top 20 flagged for intervention.
- Beats current baseline (calling everyone): same precision, half the calls, or better precision at the same call volume.
- Runs in under 10 minutes nightly on commodity hardware. No GPU clusters.
- Interpretable enough that our care coordinators trust it — they need to know why before they pick up the phone.
A note from the sponsor
We don't need a perfect model. We need one that's better than guessing, that runs reliably, and that our care team understands. If you over-engineer this, you'll lose.
— Christopher M. · COO, Panoramic Health
Data & access
What you get when you enter
Records2,847,193 appointments · 2023-01 → 2026-03
Patients~ 11,400 anonymized IDs
FormatParquet + PostgreSQL dump · ~ 4.2 GB
PHIDe-identified per HIPAA Safe Harbor. No re-identification attempts permitted under any circumstance.
ComputeBring your own. We'll reimburse up to $400 in compute on selection of winner.
NDARequired to access raw data. Standard sponsor NDA, no IP transfer until the prize is paid.