Team 4

Care Plan Check

AI that helps improve patient care plans.

Reviews a clinician’s care plan against the visit, the chart and public local data, with the patient side and the web side running as separate Fuzzball workflows.

Brian Chen, Lynna Yang, MD, Kevin Yin

62%

of barriers found with evidence

Against 47% for the same model given only the visit transcript.

43%

of barriers never said aloud

Found from the chart and neighborhood. A transcript-only model found none.

0 of 114

patient-data leaks under attack

One model holding both chart and web leaked in 17 of 114.

32

physician-reviewed cases

12 development and 20 held-out, with 72 label disputes settled.

Figures are the team's own measurements from the hackathon weekend, as reported in their submission.

The problem

A care plan can be medically right and still fail. Dialysis needs a ride. Insulin needs a fridge. A new drug may cost more than the patient can pay. The clues are spread across the visit, the chart and the neighborhood, and the clinician has minutes to put them together. Lynna Yang, a resident physician on the team who trained in free community clinics, brought the problem from there.

It is hard to automate because the check needs both the patient’s chart and the open web, and clinics must keep patient data private.

What they built

An AI that reviews the care plan by cross-checking the visit, the chart and public local data, then suggests how to improve it. Every flag shows its evidence: chart lines, missed visits, a bus trip, nearby help. The clinician accepts, edits or dismisses each flag, then prints a plain-language handout in the patient’s language.

A web page can hide instructions, so the side that holds the patient never browses and the side that browses holds no patient. The private planner asks typed questions through a signed gateway and gets back listings with evidence links. There is no field for a name.

How they used the cluster

  • Each side runs as its own Fuzzball workflow with its own volume, so the boundary is written in a file an auditor can read.
  • One scheduler ran all seven workflows on two nodes: serving the models, a six-task attack array of 228 runs in 124 minutes, then training and scoring.
  • NVIDIA Nemotron 3.5 Super 120B ran three reviews at once on one node, and Nano 30B was trained in one pass.
  • One VAST Data volume per side, with no patient file on the public one. Rocky Linux runs on every node and CPU container.

Where it goes

  • A clinic view that maps barriers by neighborhood, where only counts leave the patient side. The team mapped 298 synthetic patients.
  • Overnight learning on site from accepted reviews. A prototype on a smaller model went from 12% to 35% of barriers found after one pass.
  • New sources through the same gateway: pharmacy stock and prices, weather, ride booking, speech to text.

Synthetic patients only, with real public data.