Thread by @AlexJohnLondon

1. I’m a supporter of making more efficient use of information. But something about this piece bothers me. Thread. #ai #MachineLearning #medialAI #AIinMedicine

Machine Learning in Medicine | NEJM nejm.org/doi/full/10.10…

2. “this is the promise of machine learning in medicine: the wisdom contained in the decisions made by nearly all clinicians and the outcomes of billions of patients should inform the care of each patient....”

3. “That is, every diagnosis, management decision, and therapy should be personalized on the basis of all known information about a patient, in real time, incorporating lessons from a collective experience.”

4. I don’t think most computer scientists understand the degree of uncertainty in medicine. It isn’t just that we often don’t know which of our decisions were wise and which were not (relevant to quote 2 above)...

5. It is that we often don’t know what the relevant variables are for the systems we’re dealing with. So some of our data is the equivalent of measuring phlegm, black bile and yellow bile. No amount of deep learning can help you there.

6. And this doesn’t mean that we just have too many variables. It means that, for example, if it turned out to be true that Alzheimer’s disease really is caused by gingivitis, we would not have captured the relevant causes in ANY of our prior data.

7. Now for “personalization.” For starters, let’s distinguish two cases: prediction of likely outcomes conditional on what-we-usually-do-as-reflected-in-our-prior-data and a recommendation of how to intervene in this case. These are not the same.

8. Systems reliable for prediction can use yellow fingers to predict cancer. But you can’t intervene on yellow fingers to treat cancer. Intervention requires knowing when sets of patients in our dara are exchangeable and the variables on which that exchangability is conditional.

9. But conditional exchangability is often exactly what we don’t know and, please correct me if I’m wrong, this isn’t something you can learn from purely associationist methods.

10. So all of this is just to say that the present article feels misleading to me, underselling the difficulties in making systems that will do what we want and overselling the potential benefits.