AI can predict heart attacks more accurately than doctors | Hacker News Reader

AI can predict heart attacks more accurately than doctors

engadget.com

92 pointsdenzil_correa20 comments

As others pointed out, it's an overreach to call this artificial intelligence.

What the authors showed is that by training standard machine learning algorithms (random forests, logistic regression, gradient boosting, and a shallow neural net) on readily-available signals from the medical record (e.g., prior diagnoses), you can increase the c-statistic from 0.728 to 0.764. These machine learning techniques are well-suited to the data set and the empirical evaluation is strong, so this work should really stand on its own without trying to brand it AI.

There is some very high quality work on artificial intelligence in medicine being done today. Google Brain published a validation study of a convolutional net to diagnose diabetic retinopathy in December, and Stanford published similar work but applied to skin cancer. NIPS 2016 had an excellent Machine Learning for Health workshop with several works-in-progress: http://www.nipsml4hc.ws/

What's the difference between AI and 'standard machine learning algorithms'? Seems like they are used interchangeably today. Especially if you count neural nets as a standard algorithm too - which, let's be frank - it is in today's world.

Machine Learning is just a tool belt. AI is the idea of intelligent machines being able to solve problems (and recognizing them) on their own. Therefore of course AI can utilize the tool belt that is Machine Learning.

>AI is the idea of intelligent machines being able to solve problems (and recognizing them) on their own.

From the perspective of a language descriptivist instead of prescriptivist, that strict definition of AI is already long gone. That ship has sailed.

Even wikipedia says that colloquial uses of "artificial intelligence" stands for machine learning that mimics intelligence. It's understandable that The NYT and all mainstream non-compsci publications will use "artificial intelligence" that way. One of the researchers (Stephen Weng) cited in the article also uses the term "artificial intelligence" in his LinkedIn page and in the PDF research article itself.

If and when "solving/recognizing on its own" truly becomes reality, I suspect the world will adapt and call it "True-AI" or "Lifelike-AI" or "Strong-AI". We'll find another term to distinguish it from the watered down "artificial intelligence".

The timeless problem of AI, anything that starts out AI ends up being rebranded as not AI once it's understood. It's the no true Scotsman fallacy. Machine learning is AI, it's just not human level AI.

I'd disagree, there is more here than the no true Scotsman fallacy. It is equivalent to saying that any magic trick ceases to be magic once exposed as a trick. We don't have a good definition of I in AI therefore such arguments go on forever, yet somehow we know that the "I as in human being "I" is different than the "I" as in "AI" statistical smoke and mirrors. My opinion is that a good definition could be established that would clear such BS once and for all, some attempt here: http://blog.piekniewski.info/2017/04/13/ai-confuses-intellig...

You're conflating human level intelligence with artificial intelligence for one, AI doesn't have to be human level to be AI. Secondly "statistical smoke and mirrors" is the Chinese room fallacy; it doesn't matter what it is under the covers, if it behaves intelligently then it's intelligent regardless of whether or not it can be boiled down to some math. Your link covers that near the beginning when he discusses duck typing, but he comes to the wrong conclusion. The article should have ended there, if it behaves intelligently, then it's intelligent, that simply is the truth and looking at the implementation is cheating because we don't know brains aren't doing the same thing.

What's really going on is you are branding "human intelligence" special because we don't understand its implementation and labeling everything else not intelligent because we do, for all we know the human mind itself could be nothing more than statistical smoke and mirrors. The only problem here is human ego.

A car that can drive me somewhere on its own simply by being given a destination, is AI, not matter how it's implemented as long as it's the computer doing the driving and is operating locally by actually having sensors that see the road. It doesn't have to be able to ponder its own existence to be AI.

Neural nets were an attempt to model how the brain works; they are by definition AI regardless of whether they boil down to some maths. Everything a computer does boils down eventually to some maths, that is not an escape hatch to claim something isn't AI.

Machine learning is AI. It is not AGI, but it most certainly is AI.

You are dogmatically defending Turing test which I think is the primary source of this confusion. Turing test says: if it fools humans into thinking its intelligent it is intelligent. That is fair. But once some other humans understand the inner workings of some simple "AI" mechanism it no longer fools humans, since they now know what adversarial questions to ask to uncover it. Therefore it consequently fails the Turing test and we have the AI effect. This test is just a bad idea and it impairs research (for a number of reasons stated in the post which you prematurely dismiss).

The coffee criterion for AGI (https://en.wikipedia.org/wiki/Artificial_general_intelligenc...) is much better, since it requires ability to creatively interact with unpredictable reality as a test for intelligence. It avoids all the philosophical bullshit and all the smoke and mirrors, since you cannot fool physics. Somehow the so called "AI researchers" avoid robotics like fire, sine there stuff actually needs to work (not just statistically) and outrageous BS claims cannot be made.

And yes, ultimately the human brain may be smoke and mirrors. But frankly, quite sophisticated smoke and mirrors, not anywhere close to the crap that is being put forward right now.

No, I'm defending the Chinese room thought experiment. It's cheating to look at the implementation and then claim it isn't AI; you can't look at the human implementation which could very well also be based on simple math we simply haven't figured out yet. It's only fair to judge by inputs and outputs. And you are confusing AI with AGI; something does not have to have human level intelligence to be AI. The Turing test is about AGI, not AI.

Useful and relevant world changing AI will happen long before AGI which could very well be a pipe dream. A car that drives far better than humans is useful AI and yet would fall short of AGI, a robot that can clean my house is useful AI but could fall way short of AGI, there are vast world changing things to be done by AI long before AGI ever becomes a reality and that we understand how something works DOES NOT disqualify it from being AI, even if it boils down to little more than some statistical inference.

Saying it isn't AI because you understand how it works is like saying submarines can't swim; it doesn't have to work like nature to be valid nor does it have to be like human intelligence to be intelligent and any intelligence we build is by definition artificial intelligence. Machine learning that can diagnose better than a doctor... is AI, not matter how well you understand it's just math, it's still AI. Those who conflate AI with consciousness are the ones in err. AI does not and has not ever meant artificial self aware consciousness, while such a thing would be AI, it would be the pinnacle of AI, AGI.

> Machine learning is AI

I disagree. Machine Learning is just super-charged linear regression.

If you build a ML system that autonomously chooses it factors, and automatically adjusts to model drift (by either adjusting existing coefficients (easy) or adding/removing factors (hard), then ML drifts into AI.

> I disagree. Machine Learning is just super-charged linear regression.

Machine learning is neural nets, which started out as an approach to AI and which yes boil down to linear regression, but that's about as useful as saying brains are just super super charged linear regression. And that's the point, AI is a label that keeps getting cast off of things that started out as AI but once understood people decided they no longer were, you are committing the no true Scotsman fallacy. If it started off as AI, it's AI, that doesn't change because you understand the math underneath that it boils down to.

I think you might be talking about what other people call Artificial General Intelligence (AGI)?

https://en.wikipedia.org/wiki/Artificial_general_intelligenc...

What a silly headline. Here is a brief description of what was done, from the original article (helpfully linked by `henrypray):

> Four machine-learning algorithms (random forest, logistic regression, gradient boosting machines, neural networks) were compared to an established algorithm (American College of Cardiology guidelines) to predict first cardiovascular event over 10-years

So, routine learning algorithms were applied to a task and performed better than the simplified (for simple computation) results of a prior learning algorithm.

Based on my experience with several doctors and other medical professionals here in the US, it appears as if they always diagnose problems with a simplified test or some kind of decision tree like the standard test they did here (the "established algorithm"). Doctors I have seen only seem to get the diagnosis right if it is a problem most people have.

The most blatent example, though not the only one, is when I spent months and thousands of dollars (with the help of my insurance company) chasing a problem in my SI joint. Four different physical therapists kept telling me I had referred pain from my spine. This is based on the fact that I felt the pain if I raised my leg when I was lying down. That seems like a pretty poor test. Any other doctor or physical therapist that actually looked at my SI joint said I had an inflammed SI joint problem. The problem turned out to be my hip muscles would get tight when I slept on my side.

I would guess the main problem is not the doctors themselves but the administration that doesn't trust the doctors to use their brains. If that is true, it is not silly that the AI result beat the simple test. Hopefully either they will start using better tests of trust doctors a little more to think on their own.

Academic article PDF: http://journals.plos.org/plosone/article/file?id=10.1371/jou...

Thanks for this url. I am studying the paper now.

How do I interpret Missing BMI as a risk factor. In the ML: Neural Networks run BMI Missing was considered a top 10 risk factor. The body text said this about BMI Missing: "This study suggests that missing values, in particular, for routine biometric variables such as BMI, are independent predictors of CVD."

I'm having a hard time wrapping my brain around this concept.

Thanks

This could be a signal that people who don't visit the doctor often are more likely to develop CVD. It makes sense—if caught early, risk factors like high BMI, high blood pressure, and high LDL cholesterol can be treated and thereby prevent heart attacks and other cardiovascular events.

Or it's selection bias from people who don't normally go to the doctor mostly going when there's a problem.

The people who don't go to the doctor because they feel fine aren't included in the study.

In our ICU data we found that missing fields also predicted worse outcomes. My theory is that in sicker patients nurses are less worried about collecting all the data. But who knows.

Reply on news.ycombinator.com