Hacker News Reader

Comment by newyankee | Hacker News Reader

The reason a clean alternative to radiologists in the form of AI is not available because of the inertia of the medical system. Due to its innate conservative nature, a beta testing in a third country with successful results will only be the pathway for it to be adopted by richer countries with stricter medical systems. I feel AI in medicine is a boon for developing countries if used properly. Especially diagnostics.

View on HN

> Especially diagnostics.

I don't think so. Maybe I'm just too old, but I remember vividly that the same was said about expert systems back in the late 90s and early 2000s.

20 years later and no one is even considering expert systems for automated diagnosis anymore. The problem with current machine learning models is their blackbox character.

You cannot query the system for why a diagnosis was made and verify it's "reasoning". Tests rely on using the systems as oracles instead and in medical diagnosis, a patient's medical history is just as important as the latest lab results.

No amount of ML (in its current form) will be able to manage to interview patients accordingly. It might work as a tool for assisting professionals, but it's nowhere near in a state that warrant's its use for automated diagnosing of patients.

Wtf - many ML models today are either full on white boxes or are directly interpretable in various ways (e.g. LIME algorithm). Even neural networks have good interpretability tools (e.g. captum).

ML is not the black box nightmare that I see it described as on here. You can figure out feature contributions and can quite easily (and accurately) verify it's reasoning. If you really need these kind of models, look into various kinds of tree based ML models like random forests or boosted trees...

> look into various kinds of tree based ML models like random forests or boosted trees...

Those are the expert systems that have fallen out of favour over a decade ago, so thanks, but no thanks.

> many ML models today are either full on white boxes or are directly interpretable in various ways

Sources please, and remember to use relevant sources only, i.e. interpretation of medical image analysis like here [1].

Notably activation maps told researchers precisely nothing about what the neural net was actually basing its conclusions on:

> For other predictions, such as SBP and BMI, the attention masks were non-specific, such as uniform ‘attention’ or highlighting the circular border of the image, suggesting that the signals for those predictions may be distributed more diffusely throughout the image.

So much for "fully white boxes" and "direct interpretability"...

[1] https://storage.googleapis.com/pub-tools-public-publication-...

As I said above, it seems fairly clear if you go to the original article that there is a real problem with the existing systems adapting poorly to different setups - they are trained on one system/hospital and then don't generalise well.

Why should medical "beta testing" happen in a third-world country? Is there some reason the higher risk of an experimental procedure is more acceptable there than (say) Boston or Dallas?

Unfortunately doctors are a lot less prevalent in a lot of developing countries, especially in subsaharan Africa. If an AI can do a third of the things that a doctor can do, that means that many more people can be treated. In the US, it just means that the appointments can be cheaper. So the developing countries have a lot more to gain from things like AI. The US and other developed countries should be doing more to help the situation, including training and paying doctors to work in those countries, but AI can potentially save a lot of lives there in the meantime.

Of course, AI can only save lives if it works reliably, which doesn’t seems to be the case yet, but that hopefully can be overcome.

you think doctors are scarce in subsaharan africa but mri machines and xrays and ultrasounds are plentiful?

It’s must easier and faster to buy a machine than train a doctor

It really isn't. Some countries can produce a trained physician for less than the cost of a new MRI machine. And beyond the capital expense those machines are expensive to operate due to technicians, maintenance, consumable supplies, power, etc.

This is in contradiction to my experience working on medicine in Africa, where there are often very well trained people coping with a water system that doesn't always work.

I know that they aren’t, but that doesn’t mean that AI that doesn’t use those tools is impossible.

AI assisted virtual Doctor > No Doctor

90% reliable doctor in Switzerland > 10% reliable AI > 0% reliable no doctor in Africa

Reply on news.ycombinator.com