Will disagree with the premise, as a practising physician the answer in unlikely on the records, medical records are notoriously, incomplete, and inaccurate. Training a model on medical records will result in one optimized to maximizing billing and minimizing liability instead of getting to the truth of clinical decision making.
We can also simply have multiple correct answers.
I think there is a massive sunken cost fallacy in healthcare where just because we have collected so much data that this data must inherently be valuable. This is not necessarily true.
While medical records contain a vast amount of information, my clinical intuition is, they did not contain the information needed to replicate the work we do.
I get much more salient and understandable help from chatGPT on health issues, supplement issues, analyzing tests and so on that I have ever received from any other human MD, and I have been a patient at Stanford, UCSF, PAMF and a few others.
At least with chatGPT, when I ask a question, I ALWAYS get an answer that is at least marginally helpful.
OTOH, when I have tried to drill into an issue with an MD, assuming they have the time to give me before rushing to the next appointment that they are late for, their standard reply is often, "don't worry about it. It's not a problem." as they hurry out the door with a "nice to see you " and a wave.
And I have yet to encounter an MD who actually takes the time to read my patient record prior to the consult. 99% of the time I'll get a "so what brings you in today?" and I wonder why/if the scheduler didn't update the initial contact record and why I have to rehash what my problem is?
When I was a corporate worker in the sales arena, I ALWAYS spent 5-10 mins prior to any call refreshing my/our last contact with the prospect/customer to see what issues were outstanding, so we could have an informed discussion from the beginning. MD's refuse to do this, wasting both the patients and their own limited time.
Excellent article!! Very insightful critique of the way AI tools as currently developed/deployed in clinical settings or test beds will ultimately fail patients. Like clinical care algorithms today, they seem to be skewing toward solutions that provide “least common denominator” insights, recommendations or options. Yet, the hype that we’re being sold suggests AI will do the exact opposite. This has been one of my big “fears” for AI for many years.
Will disagree with the premise, as a practising physician the answer in unlikely on the records, medical records are notoriously, incomplete, and inaccurate. Training a model on medical records will result in one optimized to maximizing billing and minimizing liability instead of getting to the truth of clinical decision making.
We can also simply have multiple correct answers.
I think there is a massive sunken cost fallacy in healthcare where just because we have collected so much data that this data must inherently be valuable. This is not necessarily true.
While medical records contain a vast amount of information, my clinical intuition is, they did not contain the information needed to replicate the work we do.
A real problem, oversold
Invisible bottlenecks decide who compounds. The scarce edge is deciding which parts-and-power assumptions still deserve capital.
https://paretoinvestor.substack.com/p/electron-is-the-new-oil
I get much more salient and understandable help from chatGPT on health issues, supplement issues, analyzing tests and so on that I have ever received from any other human MD, and I have been a patient at Stanford, UCSF, PAMF and a few others.
At least with chatGPT, when I ask a question, I ALWAYS get an answer that is at least marginally helpful.
OTOH, when I have tried to drill into an issue with an MD, assuming they have the time to give me before rushing to the next appointment that they are late for, their standard reply is often, "don't worry about it. It's not a problem." as they hurry out the door with a "nice to see you " and a wave.
And I have yet to encounter an MD who actually takes the time to read my patient record prior to the consult. 99% of the time I'll get a "so what brings you in today?" and I wonder why/if the scheduler didn't update the initial contact record and why I have to rehash what my problem is?
When I was a corporate worker in the sales arena, I ALWAYS spent 5-10 mins prior to any call refreshing my/our last contact with the prospect/customer to see what issues were outstanding, so we could have an informed discussion from the beginning. MD's refuse to do this, wasting both the patients and their own limited time.
Excellent article!! Very insightful critique of the way AI tools as currently developed/deployed in clinical settings or test beds will ultimately fail patients. Like clinical care algorithms today, they seem to be skewing toward solutions that provide “least common denominator” insights, recommendations or options. Yet, the hype that we’re being sold suggests AI will do the exact opposite. This has been one of my big “fears” for AI for many years.