Critical Care & Sepsis
Sepsis kills on a clock, and most of the interesting decisions — when to start antibiotics, how to weigh their benefit against the kidney damage they can cause — have to be made from a handful of noisy vital signs, hours before the outcome is known. The datasets in this category are built to make that clock visible: each patient trajectory comes from a real inflammatory-response model, and the moment treatment stops working is located explicitly rather than guessed at.
One dataset locates that exact tipping point by bisection search over the treatment-start time, then checks its behaviour against how the real, deployed Epic Sepsis Model performs. Another scores six-trajectory patient cohorts against the actual published NEWS2 early-warning algorithm — the same one used at the bedside. A third layers antibiotic-induced kidney injury on top of the sepsis trajectory using a proper semi-competing-risks structure, so the model knows the difference between "this drug is nephrotoxic" and "this patient died before the kidney damage had time to show up."
What you'll typically find here
- Treatment-timing cohorts — counterfactual treatment-start scenarios per patient, so you can study not just what happened but what would have happened at a different hour.
- Early-warning score datasets — physiology scored against a real, published clinical algorithm, not an invented severity index.
- Competing-risk cohorts — datasets where two clinical processes interact honestly (a treatment's benefit and its harm), with the correct statistical structure between them, not a simplified single outcome.
Every check against a published number is reported as it actually came out — including the cases where the model only partly reproduces the literature. That gap is stated, not smoothed over.