Skip to content

Machine Learning Missteps

Step 4Live-room allocation: 4:00

Open in Colab — 01_act1.ipynb, opens in a new tab
A clinician in a white coat and blue gloves interacts with a holographic medical display of a human figure.

Removing race looks like one tap. It is three operations.

Let’s get into the machine learning implementation trade offs.

What does “remove race” actually mean?

“We removed race from the model” sounds like one decision. It is three different operations, they disagree about the same patients, and only one of them is what the clinical field actually did.

Reach into the formula, strike the coefficient, ship. No refit, no new data. Because the 2009 equation is multiplicative in the race term, this can be computed exactly — divide it back out. Nephrology has a name for it: eGFRcr(ASR-NB).

It is the cheapest of the three. You can do it to an equation already running in production, by editing one line. It is not the clinical recommendation.

Re-estimate every coefficient on the same data with race removed — eGFRcr(AS), CKD-EPI 2021. This is what NKF-ASN recommended and what the country deployed.

A and B change the equation. C leaves the equation alone and trains a model.

You drop race from the feature matrix. The model card can say the model does not use race. The matrix still has eGFR, and that eGFR was computed with race. The model never sees the word “Black.” It still sees the coefficient.

This page does not score C. Step 5 does.

What the team says What it actually is Clinical name
“we dropped the race column” delete the coefficient eGFRcr(ASR-NB)
“we retrained without it” refit eGFRcr(AS) — 2021
“our model doesn’t use race” Strategy C —
The same patients, both operations. Movement into a diagnosis goes right; movement out goes left.

Now we can see why the direction of the count matters

Section titled “Now we can see why the direction of the count matters”

One. Strategy A moves 4.65% of Black adults into CKD. Strategy B — the operation the field actually adopted — moves 2.41%. Deleting the coefficient is cheaper, looks like removing a variable, and moves 1.93× as many people.

Two. Under the refit, Non-Black adults do not move into CKD (0.00%). They move out (2.01%). Step 3’s assessment cannot show that. It only counts into. ‘Zero Non-Black adults were reclassified’ would be the wrong conclusion. The more precise statement is: zero moved into CKD at that threshold. Others moved out. In the spirit of doing no harm, we ensure to note that when a threshold changes, we count both directions.