Catching patient deterioration 12 hours early

These unplanned transfers are only about 5% of admissions, yet they account for roughly a quarter of the provider's ICU admissions, a fifth of in-hospital deaths, and an eighth of all hospital days. The warning signs usually appear 6 to 24 hours before the transfer, and on a busy ward they go unnoticed.

With decades of electronic records across an integrated care system, the provider had the data to find these patients ahead of time. The question, in the words of the provider's principal statistician, was whether the team could recognize who was likely to crash and do something about it.

The provider built an early-warning system with four parts: risk analysis and a deterioration score, monitoring and warning, dissemination to the care team, and intervention.

The score draws on bed history, lab chemistry, vital signs, comorbidities, and demographics to estimate the probability of a sudden ICU transfer or death. The provider's data scientists engineered features and validated every model with clinicians. As the team's data leader put it, when you are saving lives you cannot be off by even a little, so the team ran a rigorous threshold and characteristics review on every case.

After several proofs of concept on infrastructure, processing speed, and algorithms, the team decided open source was the right way to go and settled on H2O-3 with support. The tooling was fast enough to model the full dataset, with no sampling.