Hacker News Reader

Comment by roenxi | Hacker News Reader

The elephant in the room with doctors is that there are many countries where the supply is constrained by some sort of guild system and only very clever people do well.

If AI can lower the bar to the point where Ned the Nitwit can type in the symptoms, read the screen and get a reasonable diagnosis then 'doctors' are going to be a completely different class of people even if they share the same title.

View on HN

There are a lot of current doctor work which involves. "Any issues with the medication your on, ok, can I see your bloodwork, no red values, here is your prescription."

Which really should be the job of a nurse or physician's assistant. Doctors should be reserved for the cases that requires the ~10 years of training they receive.

Most people probably only need a yearly-visit to the physician's assistant or a nurse for bloodworks. There shouldn't be a need to see an actual doctor on a regular basis (unless you have some chronic ailment that nurses / physician assistants can't handle).

------------

The prescription is probably the only part of that routine that requires a doctor.

Entirely accurate. It's the AMA/doctor lobby (purposely constraining supply to keep wages high) that is preventing nurses and nurse practitioners from filling this role. My prescriptions and bloodwork are already in an EHR/EMR (EPIC) and shared with Apple Health, machine intelligence should be doing as much heavy lifting as possible.

This question might be out of place, but why should I trust ML algorithms for my health? To be a doctor you need at least a decade of education and training, as well as licensing and adherence to a code of ethics. To be a tech bro you just need to get hired. You don't even need a degree. You could come to work stoned and as long as your boss doesn't notice, you're fine and can continue to write the code that my health could rely on.

I'd expect the development of healthcare related machine intelligence/assistance technologies to be heavily regulated, in the same way medical devices are by the FDA.

50% of doctors graduate in the lower half of their class each year. Lots of low hanging fruit with ML if implemented properly (big if, I know).

Okay, how do you implement efficient ML with terrible quality data?

Time and effort? ML already finds lung cancer better than human radiologists, so it’s clear there are datasets that are of sufficient quality. We continue to improve datasets wherever possible.

In a word: no. Machines can find a spot on an image better than a radiologist when it is known that there might be one.

Unfortunately, that is of very little use to the clinician. The big hype is a consequence of a big misunderstanding of the practical constraints of healthcare.

Time and effort don't fix bad data and bad statisticians.

But time and effort does enable better pipelines to improve future data quality. Garbage in, garbage out, don’t collect garbage.

I'm sold. Now try to convince the people who are important to convince and have a good laugh. Reliable data collection is the biggest challenge for modern medicine. A challenge far greater than getting better statistical models.

Email in profile. Point me to decision makers and I’ll make the case on my own dime.

Comparing the bottom 50% of people that graduate from medical school to "low hanging fruit" is hilariously snide.

Like, if there is a problem with their performance, it is a problem that they are graduating. And then look at the group that is admitted. It isn't a low bar.

My point is not meant to be inflammatory, nor derogatory. My understanding is that medicine is very hard as a domain (correct me if I'm wrong! i am not a medical practitioner). I am not proposing these people are stupid, nor that they are incompetent, but that humans have human failings and that is what leaning on technology is for. Medicine is hard enough without help. I am proposing more help, just as autopilot is to aircraft pilots. Commercial air travel is very safe, but still has humans making decisions.

The majority of diagnoses are extremely simple. It's a small portion that are difficult. We don't need a complicated computer system to help with the simple ones, we need more people that are allowed to indicate the care they know is needed.

Ay, there's the rub. Oftentimes, it is discerning the difficult in a multitude of simples, hearing the zebra hoofbeats when the sound is clearly a horse. Miss the exotic and a patient could die, no repeats. It haunts a physician for the rest of their days.

And 50% of nurses, nurse practitioners, and PAs graduate in the bottom half of their class too. Why do you think they are not the ones whose jobs are going to be automated away? If an AI can handle the easy cases, it won't be the doctors, who are able to deal with the complicated cases, who will lose their jobs.

How much of an average doctor's day is really appointments like this where they just check you lab values and ask you if you're doing okay on your medication and send you on your way? A large number of these cases are already handled by nurse practitioners and PAs. Doctors in my experience spend more time seeing the sicker patients with many chronic medical conditions where the more in depth knowledge of pathology and pharmacology that they spend those years of training acquiring inform the questions they ask, the tests the order, and the physical exam they preform.

I don't think the large majority of PAs and NPs really want to step into the same role as physicians. The number of PAs and NPs who were rejected from medical school and then chose to take either of those routes is probably very small compared to the number who chose to become a PA or NP because they wanted that role and not a doctor's.

Doctor time is most of expensive. It makes the most sense to automate away their mundane or lower value work first, vs NPs, RNs, or CNAs (who all do more hands on work compared to an MD, but also have lower costs per unit of time). It is silly to continue to have humans do low value work if it can be automated. Save that time for higher value work.

This is not about replacing people wholesale, but giving them better tools.

Of course the next-gen system(including people and machines) should be proven as trustworthy, before being used.

Nobody here is against that.

Might be some sort of liability aspect to this, where the doctor has to certify in case something goes wrong. Do nurses/PAs have to carry malpractice insurance? Not super clear on this myself.

I think NP's do.

Nurse Practitioners can be sued for malpractice, but despite practicing independently they are held to a significantly lower standard of care.

We know how to replace doctors with nurses. It works well, in many cases.

It's really a political problem.

People don't seem to have enough motivation to politically fight to replace doctors with nurses, and doctors have a lot of power and motivation to keep their jobs.

And what about powerful actors from inside the system ? Hospitals, Insurers, etc ? In general, everyone gets a share of revenue. So why bother reducing that ?

Insurers and hospital would only love to keep inflating bills higher and higher, and high physician comp help to justify that.

you over-estimate the amount of training they receive.

physicians usually hit their stride 5-10 years before they kick the bucket.

The article specifically covers this. Current AI is missing common sense for example. This could be lethal for Ned.

Along with that Ned is probably fooling himself already. It takes a skilled human to ask the kind of non leading questions that might even lead to the truth of what the problem is.

So far Picasso's statement still holds true "Computers are useless, they can only give you answers" (ok "useless" is an unfair accusation. Answers are useful but you need a human to ask the question and judge the answer. Picasso's statement holds true for art though.)

> Current AI is missing common sense for example.

And normal doctors do not? Are they infallible? Does papal infallibility apply to doctors to? I know a bunch of doctors, they aren't all that different from normal humans except in Australia we work them until they are sleep deprived. They will be making potentially lethal mistakes continuously and it isn't being detected. Every so often there is a scandal when an egregiously bad doctor is found.

Consistent bias and repeatable mistakes are a lot safer to work with than random bias and random mistakes.

We have a lot of examples where humans could initially compete with machines until one day they no longer could. Patient analysis has hallmarks of being a solvable problem (pattern recognition, image interpretation, consuming huge amounts of data, etc). If someone is allowed to compete it will end well for everyone. Once that goes 'doctor' is a title that means something different to what it does now.

We don't have the data collection pipelines necessary for machines to compete against humans at the moment. And everyone is focusing on the hype, which doesn't help.

> And normal doctors do not? Are they infallible? Does papal infallibility apply to doctors to?

Also Australian, not a clinician but I work in health.

Where I have seen the best use of computers working hand-in-hand with doctors is around the areas of clinical risk and funding.

Computer systems that prevents a clinician from prescribing a lethal dose, or prevents a known adverse interaction between drugs have already saved lives in Australia. In hospitals which have properly implemented these systems, drug mistakes have dropped dramatically.

On the funding side, I have found that clinicians are very interested in ensuring that the maximum public funds are available for their patients as soon as possible. If a patient fulfills the requirements for a funding stream, the clinician wants to know about that immediately so that they can start providing their patients with those services. This is especially important in community care situations where patients can receive at-home support to maintain adequate cleanliness and hygiene in order to keep them out of hospital (hospital being bad for the patient and bad for the public purse).

One of the best AI uses is prevention. You can have a personal doctor AI at home, monitoring you 24/7.

That AI could diagnose many diseases before there are any visible symptoms.

Given the potential liabilty consequences, this will invariably lead to a lot of false positive. There is a high chance that this ends up increasing costs.

Reply on news.ycombinator.com