The comfortable lie
Step 1Live-room allocation: 3:00
The blood work is real. What comes back from the lab is already an estimate.
A number came back from the lab.
It was computed correctly, by a validated equation, from real blood work, and nobody in the chain did anything wrong.
It was also different depending on the patient’s race, and everyone downstream treated it as a measurement.
Where this goes wrong is not where you expect
Section titled “Where this goes wrong is not where you expect”The obvious story is that somebody put race into an equation and that was the mistake. It is not the story this workshop tells, because it does not survive contact with how the equation came to exist.
It was fitted to data. It was published, peer-reviewed, validated against measured kidney function, and adopted because it was more accurate than what came before. Clinicians used it because it was the standard of care. In 2021 a national task force recommended removing the coefficient, and that recommendation was also careful, also evidence-based, and also correct.
The comfortable lie
Section titled “The comfortable lie”The lie is not “the equation was biased.”
The lie is “the number is the measurement.”
Estimated glomerular filtration rate is an estimate. It is a model output: creatinine and age and sex go in, a number comes out, and that number stands in for a physical quantity nobody measured. Everything downstream then treats it as though somebody had.
A stage boundary reads it as a fact. A referral threshold reads it as a fact. A prescribing rule reads it as a fact. A model trained on the resulting records reads it as a fact. None of them can tell the difference between a measurement and an estimate, because by the time it reaches them, there is no difference to see.
That is the failure this workshop follows. Not one bad decision, but a modelled value crossing into a system that has no way to represent doubt.
The question
Section titled “The question”Have you ever shipped a model with a field in it you never asked how it was computed?
It is a better question if you answer it concretely. Open the feature list of the last model you put into production and find one column that was itself produced by another model, a heuristic, or an imputation. Most feature lists have several. Most people can name where at most one of them came from.
You are not being asked to feel bad about that. Nobody can trace every field. You are being asked to notice that the tracing stops somewhere, and that you do not currently know where.
What happens next
Section titled “What happens next”The next fifty minutes are one number, followed all the way down: from the lab, through a diagnosis, into a mortality model, and out the other side as who is eligible for which medication.