Read on Twitter
Medical puzzle:
Take Black vs. White patients with similar knee x-rays
Why do Black patients have more pain?
Our paper in @NatureMedicine provides one answer: https://go.nature.com/3i6zAuc
Algorithms see causes of knee pain in Black patients, that human radiologists miss.
Take Black vs. White patients with similar knee x-rays
Why do Black patients have more pain?
Our paper in @NatureMedicine provides one answer: https://go.nature.com/3i6zAuc
Algorithms see causes of knee pain in Black patients, that human radiologists miss.
A patient comes into your office with chronic pain. You send her for an x-ray to look for arthritis.
The report comes back: the knee ālooks okā
But what if the radiologist missed something?
What are radiologists looking for anyway?
The report comes back: the knee ālooks okā
But what if the radiologist missed something?
What are radiologists looking for anyway?
What we know about arthritis on x-rays comes from studies like this one:
Coal miners and office workers in Manchester, ca 1950
(No need to comment on gender/ethnic breakdown of study populationā¦ when itās all the same!)
Coal miners and office workers in Manchester, ca 1950
(No need to comment on gender/ethnic breakdown of study populationā¦ when itās all the same!)
So radiologists might miss causes of painā because they arenāt described in medical knowledge.
Could this be a job for an algorithm?
Maybe, but one problem.
We usually train algorithms *to match human performance*, e.g. this (great) paper
Exactly what we donāt want here!
Could this be a job for an algorithm?
Maybe, but one problem.
We usually train algorithms *to match human performance*, e.g. this (great) paper
Exactly what we donāt want here!
What if there were a different way to train the algorithm?
LĢµeĢµaĢµrĢµnĢµ ĢµfĢµrĢµoĢµmĢµ ĢµtĢµhĢµeĢµ ĢµrĢµaĢµdĢµiĢµoĢµlĢµoĢµgĢµiĢµsĢµtĢµ
Listen to the patient
We train the algorithm to predict the *patientās pain*, not the radiologistās read
LĢµeĢµaĢµrĢµnĢµ ĢµfĢµrĢµoĢµmĢµ ĢµtĢµhĢµeĢµ ĢµrĢµaĢµdĢµiĢµoĢµlĢµoĢµgĢµiĢµsĢµtĢµ
Listen to the patient
We train the algorithm to predict the *patientās pain*, not the radiologistās read
This cuts the unexplained pain gap between Black and White patients by 43%
Notice this isnāt āaffirmative actionā
The algorithm does a better job *finding things in knees that hurt*
Previously unexplained symptoms are just more common in Black patients
Notice this isnāt āaffirmative actionā
The algorithm does a better job *finding things in knees that hurt*
Previously unexplained symptoms are just more common in Black patients
Back to the patient in your office:
You miss the problem in her kneeāso you donāt consider things like knee replacement.
But if the problem is in the knee after all, the patient loses out.
More Black patients lose out for this reason.
You miss the problem in her kneeāso you donāt consider things like knee replacement.
But if the problem is in the knee after all, the patient loses out.
More Black patients lose out for this reason.
Dr Said Ibrahim @WCMPopHealthSci makes that point in a wonderful accompanying commentary https://www.nature.com/articles/s41591-020-01196-3
We like this paper because it shows that algorithms can fight bias:
Here, the unsuspected bias built into medical knowledge,
because of how medical knowledge is built.
So algorithms can helpā¦
Here, the unsuspected bias built into medical knowledge,
because of how medical knowledge is built.
So algorithms can helpā¦
ā¦just as easily as they can do harm, as we learned in our previous work. https://science.sciencemag.org/content/366/6464/447.full
I also like this paper because I got to work with @2plus2make5, who led the research.
Please check out her many other wonderful papers on health + ML + policy, here!
https://cs.stanford.edu/~emmap1/
(for the record, @Cutler_econ @jure @m_sendhil are great too)
Please check out her many other wonderful papers on health + ML + policy, here!
https://cs.stanford.edu/~emmap1/
(for the record, @Cutler_econ @jure @m_sendhil are great too)
