Team 7 (CLIF)

ICU Horizon

From bench to bedside.

Five 24-hour predictions for every patient in an intensive care unit, summed into what the unit will need, from a model trained and run inside the hospital.

J.C. Rojas, MD, MS, Kaveri Chhikara, MS, Vaishvik Chaudhari, MS, Dema Therese, MS, Zewei (Whiskey) Liao, MS

36M

parameter model trained from scratch

On both GB10 nodes, with a custom 1,231-token vocabulary.

5

predictions from one pass

A new outcome is a new linear probe, with no retraining.

0.77 to 0.93

AUROC on held-out patients

Across the five 24-hour predictions.

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

The problem

Resource-constrained ICUs have only a few life-support machines. When those run out, patients must be transferred, often hours away. These units have no data science team and their data cannot leave the building, so they learn they are short too late.

What they built

A bed board that nurses, doctors and respiratory therapists read before morning rounds. It shows who is on life support, each patient’s 24-hour chance of needing a ventilator, vasopressor or dialysis, and whether devices and staff are enough. An LLM chat answers questions about the forecast.

Most healthcare AI trains one model per prediction. The team trained one generative event model from scratch and reads its hidden layer once per patient. Linear probes turn that embedding into each prediction.

The five predictions

  • Ventilator start: AUROC 0.78.
  • Vasopressor start: AUROC 0.77.
  • Continuous renal replacement therapy start: AUROC 0.88.
  • Transfer to ward: AUROC 0.80.
  • Death, shown as a unit count only: AUROC 0.93.

How they used the cluster

  • One Fuzzball training workflow on both NVIDIA GB10 nodes, with data preparation and probe fitting around it.
  • A live workflow loads the model, runs inference and scoring, and serves the board and chat.
  • The chat agent reads the forecast and passes it to the LLM as tool results. The LLM never reads patient records.
  • A VAST Data volume holds the records, tokens, model checkpoint and probes.

Research prototype, not a medical device. Trained on CLIF-MIMIC, built from the public de-identified MIMIC-IV database.