Public repository. Nothing beyond its existence is verified here.
No account of what was built, because nothing attests one. What is verified is the description, the stack and any link above. Everything below this line is framing — what a thing like this is for, and what this one is made of. None of it is a finding, and none of it is attested.
What was broken.
Differential diagnosis is usually modelled as a correlation problem: learn which findings co-occur with which diagnoses in the record. That is a different question from which condition would explain them, and the two come apart precisely where the stakes are highest. MIMIC-IV makes the correlative version easy and the causal version possible. This is a research tool, not a diagnostic device, and no clinical claims are made for it.
How it is wired.
Causal effect estimation with DoWhy over MIMIC-IV, behind a FastAPI service with Postgres.