Hacker News Reader

Comment by salimmadjd | Hacker News Reader

How do we know if this is really predicting vs. identifying undiagnosed diseases?

For each patient we considered only the prediction of novel diseases, discarding the re-diagnosis of a disease

It seems to me you need a longer time series (patient with 3,4,5 years of medical history) to be able to predict a future disease state. And I don't believe I saw that in the article (I admit I read through it quickly).

If it turns out, this is just identifying undiagnosed diseases, I wonder if this research will have any legal ramifications.

View on HN

The paper fails to discuss the fact that the data for a given patient is biased by the preliminary (likely unrecorded) diagnoses of the doctor. The doctors are trained to systematically order certain sets of tests depending on their preliminary diagnoses. If some doctors have a pretty good idea of the final diagnosis, the deep learning classifier is simply learning what doctors do when they suspect a particular disease. It is not learning to predict a disease. A truer test of the deep learning method would also require that the deep learning algorithm orders the tests, or that we randomly select tests for a patient -- both of these possibilities are highly unethical.

I think calling it prediction is a bit of a strong claim. It would probably be more accurate to say that deep patient comes to a diagnosis before the EDIT:s/prediction/physician, rather than necessarily predicting the future disease state. To properly claim prediction of the future disease state you would need to show that the patient does not have the disease at the time of the prediction.

From what I have skimmed they do not do this in the paper, although to be fair this would probably not be practical because it would require more tests to have been done in the past which in many cases were not done, e.g. imaging to confirm absence of a cancer at the time of prediction.

That said, I have worked on thia same problem and also lazily call it prediction. Also there is still immense value in earlier diagnosis in cases where early treatment leads to better outcomes, like in sepsis or cancer.

I rather think the incompleteness and inaccuracy of the data makes having more of it less valuable than it might seem at first glance. Medical history information is a mess (and that's being kind) in general.

Why would there be legal ramifications?

I'd say (at least) three ways:

1) Docs using it to double check themselves before diagnosing patients, thereby, hopefully, limiting their liability.

2) Patients getting second opinions from Deep Patient and suing the previous doc for not catching a disease.

3) Patient getting diagnosed at a very late stage of an illness and patient + family thinking earlier docs should've caught disease. Hire lawyer, lawyer runs through Deep Patient simulation, Deep Patient identifies illness, DP identification used to prosecute earlier docs. I guess kinda like 2.

Does your point 2 exist today with human doctors? Can you sue after getting a second opinion?

And couldn't it also make a wrong diagnosis? Who do you sue in that case?

How would you know if it made a wrong prediction? It would take a lot of sampling and time to show it.

Reply on news.ycombinator.com