Medicine's Machine Learning Problem

> AI algorithms _need_ bias to work. This is the bias-variance trade off: https://en.m.wikipedia.org/wiki/Bias–variance_tradeoff

To be clear, statistical bias is in fact distinct from the colloquial term ‘bias’ most people use - but they can be interpreted similarly if given the proper context (which you did)

I certainly see and empathize where you are coming from.

However, I would like to add that it kind of makes sense that you’d have more white people with scans available. Focusing on the USA for a second (and note that this likely applies elsewhere too, since screening programs are really only in full force in developed countries which, surprise surprise, are predominantly white). Non white patients don’t get screened as much. Non white patients don’t go to the doctor as much. Non white patients are inherently fewer than white patients.

I agree that finding a good way to get anonymized data is going to help in future endeavors, but we do need to keep in context the players involved in getting and using that data.

And of course the ultimate goal, to improve health regardless of race, social class, wealth, etc.

> I tried a great deal of ML approaches, only to find recently that by better accounting for the optical specificities of the problem I was trying to solve

I want to point out, that any serious machine learning researcher is not oblivious to this, despite the deep learning boom suggesting to the contrary. Modern methods have shown that we are capable of building predictors with surprisingly complex representations, that can solve large-scale downstream tasks. i.e. our models are "flexible" enough.

The next challenge is whether they favor the "right kind" of solutions. For instance, Convolutional Neural Networks (CNNs) are architecturally just sparse version of Fully-Connected Neural Networks. Why is it then that CNNs perform far better on images? A key reason is that "inductive biases" afforded by MLP aren't strongly favored towards images. Another instance of this is the covariance functions used is Gaussian Processes - the Squared Exponential Kernel is very flexible and can in principle fit anything possible. Nevertheless, if the problem has specific structures, say periodicity, one better use the Periodic Kernel because it's inductive biases rightly align with the kind of solutions we expect.

> The lesson is : physics beats AI.

As a consequence, the single biggest reason physics would beat a generic AI in the short-term is precisely due to our ability to explicitly provide inductive biases that align with our expectations from the physical system.

We haven't found the secret sauce for every possible system in the universe. I don't think we can, either. But what we can do is devise ways to "control" such inductive biases we can encode in machine learning systems, which align with our expectations of the way the system should behave.

May I suggest, in response to your sentiment that applications of AI to medicine are lacking, is that you are seeing applications replace current medical practices. An AI diagnosis of a medical image seems redundant indeed, however in this situation a patient has seen a doctor out of complaints and has been sent to the radiologist for further investigation. This medical practice is reactionary, and suspicions are already present, so of course the AI isn't doing much useful here.

Alternatively, imagine a proactive medical world, in which preventative screenings are commonplace. Currently, the implementation of routine screenings without any complaints is prohibitively expensive on a large scale. This is because it requires manpower, and manpower is prohibitively expensive and the expense of manhours needs to be justified by a medical practitioner. However, AI can help in this proactive medical world by reducing the number of hours real people are looking through data to detect problems of patients, reducing the cost of routine screenings at large. Again, this wouldn't replace doctors, as you'd still need a specialist to analyze any positive hits, but it differs from your scenario in which the AI diagnosis seems redundant.

So, when preventative medical practices are more prevalent, the mass routine screening procedures will need help from machines to keep it cost effective, and that I believe is where this technology will find its application.

"The first rule of any technology used in a business is that automation applied to an efficient operation will magnify the efficiency. The second is that automation applied to an inefficient operation will magnify the inefficiency." -- Bill Gates

In the lifetime of my adult sons the world stopped being predominantly agrarian and rural and hit a point where more than half of all people on the planet live in cities. It wasn't hugely long ago -- a hundred or two hundred years -- that most people lived in little villages or tribes and didn't travel all the far and knew most people they dealt with.

The local doctor -- or medicine man -- was often one of the older, best educated and wisest locals. He tended to come to your home with a little black bag and in the course of walking through your house to check on you if you were ailing and not able to get out of bed, he saw a great many things about your life without having to ask.

This informed his conclusions about what was wrong and about how to treat it. And it did so in a way that was largely invisible to the recipients of care.

Doctors likely often didn't explain that the house was filthy or the spouse was obviously abusive. Topics like that tend to be socially unacceptable and people don't like being criticized in that way, but if someone smarter and more experienced and better educated and wiser walks through your life and then prescribes something "for your health" and he has a track record of fixing the problem, you do as you are told because you were told it.

And then modern medicine invented a lot of diagnostics and what not and office visits by the patient replaced home visits because we haven't invented a Tricorder that can replace a little black bag and let you bring all that diagnostic power with you.

Human health is no longer treated like the logical outcome of all your life choices and your physician is no longer the wisest person you know giving you good advice that takes into account a great many factors you never talked about with him. People get treated like specimens in a petri dish in a way that implicitly denies the fact that their physical state of health is the sum total of all their life choices.

In tribal cultures, medicine men were typically people who tended to both spiritual and physical health. The two were not viewed as separate from each other.

Medicine has become commercialized in a way that doesn't really serve the interests of the patient and if you try to point that out you are likely to be written off as some paranoid fruitcake and conspiracy theorist.

There are a lot of good things about modern medicine, but there are also a lot of systemic issues and this article is correct to point out that AI tends to magnify those sorts of things.

Last, health is best understood as a moving target in 4D. Data capture does a poor job of approaching it that way and I'm not aware of any programs that are well equipped to do a good job with that.

Human doctors were historically put on call for up to 24 hours at a time as part of their learning process in part so they would see a patient's condition evolve over time while the doctor was still young and healthy enough to endure this grueling process. Having seen it for a time as part of their training, they retained that knowledge when they were older and could recognize a stage of a moving target.

I don't know how much that is still done, but I don't think we really frame AI in that way. I don't know how we would get there from here either. I still haven't managed to learn to code, what with being too busy with my own health issues all these years.