[FredPret]

I wonder if you can replace a GP with a decision tree. You could update the tree as new research is done.

If you could collect reliable diagnostic data locally, you could serve this globally and for free.

It would also be a treasure trove of data about how we respond to various treatments.

[rscho]

> I wonder if you can replace a GP with a decision tree.

No, you can't.

> If you could collect reliable diagnostic data

And there's the reason. You can't do that either. There is a reason why GPs go through medical school.

[FredPret]

> No, you can't.

Any sound reason, or are you either a) a defeatist, or b) a GP?

>There is a reason why GPs go through medical school

The input data would be basic things like:

- blood pressure

- weight

- images of the ear canals and throat

- blood, urine, saliva samples, perhaps analyzed in a regional centre

You don't need a ton of training to get the above from a patient and into a computer, and to ship the samples.

[rscho]

> Any sound reason

The job of a GP is actually probably one of the top hardest to automate, because the GP's main (and often only) job is to extract information. And that _does not_ consist in performing plenty of tests, but in speaking to and most importantly listening to the patient.

> You don't need a ton of training to get the above from a patient and into a computer, and to ship the samples.

Great! And you know what good that would do to improve diagnostic accuracy? Zilch. Zero. There's a saying that '90% of diagnoses are done on history'. Now tell me why that would be different for an algorithm given identical information? If there was a simple answer to that, we'd already be running statistical models over patient labs all day long, which we're not.

> are you either a) a defeatist, or b) a GP?

I'm an epidemiologist and also a practicing anesthesiologist, which is why the statistical theories of people who have never set foot in the clinics to see what's the job really about make me want to jump off a bridge.

[FredPret]

When I go to the doctor, this happens:

- Doctor says "Say ahhh" and looks in my throat with the thingy

- Doctor looks in my ear canals with another thingy

- On other occasions, my other vitals are taken, maybe some vials of blood, etc. Again, a student can do this.

I'm asked a few general questions, with some follow-up questions based on my answers.

Then the doctor puts this information - along with my patient history - into the decision tree in their head and comes up with a result. If the doctor is stumped, I'm sent to a specialist.

The above can be automated, plain and simple. It would also be an improvement over my experience of the health system - in Canada. I have never seen my GP pull up a multi-year graph of my blood pressure, weight, or whatever. What I am describing is a system for creating regular data points of the kind currently used in diagnosis. What I fail to understand is how you cannot see that there must necessarily be predictive value in such a database.

Even if only 80% of the job can be automated, public health would improve immensely if the global population can do regular checkups like the above cheaply.

[rscho]

> What I fail to understand is how you cannot see that there must necessarily be predictive value in such a database.

I can see allright. But you cannot see that your hypothetical database is lacking most of the info because your doc actually mostly evaluates you by looking at your general composure and relying on X years of experience and a bit of knowledge shoves that into the really complex decision tree in his head: "Hmmmm... this guy looks mostly fine."

Now, you feed your database to the latest deep-learning shiny thingy that tells you: "this guy has X% chance of having a horrible cancer, but I can't explain why". So you enjoy many months of costly investigation because you don't want to miss something, right? And after the fact, it is discovered that the lack of standardization in measurements caused the algorithm to decide that the light hue in the office was a sign of cancer.

All that to say that yes, someday what you are imagining may well be possible, but we are really very far from having the technology to do that now.

[FredPret]

I would have never thought of this, but I'm pretty sure gait, posture, and voice analyses can reliably be classified as "probably ill" or "probably well".

[vladTheInhaler]

That's not really the hard or useful part part. According to a radiologist and machine learning researcher[1]:

"It turns out that deep learning is a very good match for some of the most time consuming (and therefore costly) parts of medicine: the perceptual tasks.

We also saw that many decisions simply fall out of the perceptual process; once you have identified what you are seeing or hearing, there is no more “thinking” work to do."

[1]: https://lukeoakdenrayner.wordpress.com/2017/05/03/the-end-of...