The Tell-Tale Heart! Boffins build an AI that can tell your sex using just your heartbeat

Neural networks can predict your biological age and sex just by analysing the pattern of your heart beats, according to new research published on Tuesday. A convolutional neural network trained by a group of researchers from the Mayo Clinic College of Medicine and Science, a private research university based in Minnesota, …

[Conundrum1885]

Fascinating

As a recent (2k) graduate, I had to study 12 lead ECGs.

Of course at the time deep learning wasn't exactly mainstream but there were discussions about using neural networks to sort and store ECG data more efficiently eg for ambulatory monitoring.

I did briefly explore whether a RAM chip could be used but at the time 16MB was far too expen$ive which would have been about the bare minimum. Also problematical was the loss of power so most commercial monitors recorded data to tape resembling answerphone storage.

Incidentally thanks to advances you can now approximate a 12 lead using three segmented electrodes with some degree of repeatability.

Even putting the amplifiers on individual electrodes as a very small chip scale (CSP) is a well known method now and if memory serves NASA pioneered this.

[Muscleguy]

Re: Fascinating

When I did my undergrad Physiology degree we used multi channel Grass pen recorders for the ECG lab. Demonstrating the lab during my PhD it had become an analog-digital converter the son of the head of dept developed (company still going) and a Mac Plus on a custom frame. It emulated a pen recorder or an oscilloscope and there was a dot matrix printer so the students could take a record.

I had a 24hr ECG couple of years ago (nothing to worry about, athletic heart doing perfectly safe funky stuff at rest, cf Norwegian X-country skier study). That was all digital and quite lightweight.

During that PhD one of the lecturers got hold of one of the first portable pulse monitors (finger variety) and went for a run with it. Found no matter how hard he tried to run downhill his HR fell. As a consequence from then on I tend to try and run up long gradual hills and down short sharp ones. Unless I'm doing hill reps of course then it's the steepest hill I can find.

[caffeine addict]

Damn them and their last couple of paragraphs. I was just thinking about how interesting it would be to see how it saw trans and intersex individuals, but I do take their point that it would be a very controversial element for them to investigate.

[Robert Carnegie]

If they do invent a machine that can detect prospective transsexuals, that would be ethically complicated, especially when used by the parents. But for people not sure about themselves, a gadget that just tells you could be a blessing - never mind how or whether it works. Something to put in Jeremy Kyle's next show maybe?

[trevorde]

Occam's Razor

Probably just looking at the file name: Patient-123456-M-37

[Anonymous Coward]

Re: Occam's Razor

Meanwhile, the scientist still don't understand how it's doing it because they still haven't found the filename.

[phuzz]

Re: Occam's Razor

That's because they couldn't read their hand writing!

(Don't worry about the moths, they're normal for a joke that's this old)

[Imhotep]

Re: Occam's Razor

Or input from the one of the 12 patches that is placed on the penis.

[Pascal Monett]

The black box nature of convolutional neural networks

We're going to end up creating an honest-to-goodness AI and we won't be able to explain how.

Just you wait an see.

[Tom 7]

Re: The black box nature of convolutional neural networks

You can set up your convolutional networks to give you a 'view' of what the various layers are getting up to in a variety of forms. However you may not get why it makes the decisions it does even if it does tell you in great detail.

Understanding the reasons it makes the decisions will be the next necessary step in AI and should provide us with an interesting tool-box of things we can use to understand how we work,

[Paul Hovnanian]

Re: The black box nature of convolutional neural networks

"Understanding the reasons it makes the decisions will be the next necessary step in AI"

It's been done. Logging which rules or neural network pathways were triggered to support an 'explain' function in some old AI I worked on a few decades ago.

Anything that is done in digital systems or circuits, absent a random input signal (or the odd cosmic ray flipping a bit) can be reproduced given a copy of the inputs and a snapshot of the system. And you can step through it with a debugger if you so desire.

[Baldrickk]

Re: The black box nature of convolutional neural networks

Yes, but the difficulty is working out why that pathway works for that case - you can see what the pathway eventually is, but what features is it really reacting to? It works (the majority of the time) but what is it actually doing to reach that?

A similar question was raised by a story I heard while at Uni. Someone was using genetic algorithms to evolve FPGA designs. The circuit this eventually produced not only worked perfectly, it was also half the size of conventional designs to perform the same purpose. Only one problem, it wasn't clear how it worked. A lot of study later, and it was determined that there was a fault in the FPGA itself. The circuit had evolved around this fault, incorporating it into the design. It was therefore useless as the circuit would not work when applied to a different FPGA, or indeed, a different location on the same FPGA as used for the generation.

The key point here is the amount of study required to actually understand just what had gone on to produce the result, despite having the circuit design in hand.

[Homeboy]

Re: The black box nature of convolutional neural networks

That's basically what happens inside every human skull.

We can all recognise that a new born child has become an intelligent adult but we have no idea what went on inside the "black box" (aka skull) over the years while the change took place. Not sure replicating the problem inside another "black box" advances things a great deal.

[JulieM]

Re: The black box nature of convolutional neural networks

Well, we could always create another Artificial Intelligence to reverse-engineer the first one .....

[My-Handle]

This sounds like...

...sexing a chick :)

When raising poultry, female chicks are highly valued but males are not and are usually cast aside. The problem is that it's very hard to find out what sex a chick actually is, there's no hard and fast method to quickly sort them. Chicken sexers can do it by eye. They are trained by another chicken sexer who simply tells them whether they are correct or not and, after a while, they "get" the pattern. None of them can actually say what it is about the chick that clued them in to whether they are male or female, but they are usually right.

This example of machine learning looks like it's going down a similar route.

[lglethal]

Re: This sounds like...

Obligatory XKCD...

https://www.xkcd.com/2173/

[MonkeyCee]

Re: This sounds like...

"Chicken sexers can do it by eye. They are trained by another chicken sexer who simply tells them whether they are correct or not and, after a while, they "get" the pattern."

I do it by picking them up, and tossing them in the right bin. You can also do two at a time, one in each hand. To me the feathers feel different, and the chick "sits" in your hand differently. So I can tell, but no way to teach it to someone. I'm pretty accurate by eye, but not 100%. So I don't get to teach someone by poking them with a stick :D

The interesting thing with it is you can try and train a robot to do it, and you can get it very good at it, but for whatever reason getting it to 100% hasn't been achieved. IMHO it's a case of "moar sensors" but considering you can get pretty much any human up to 100%, using a meatbag convoluted neural network is still superior.

Icon for the processor coolant.

[Muscleguy]

Re: This sounds like...

I bet I could learn very quickly. Mind you I'm a trained biologist so I can 'see' lots of stuff others can't. I was in demand by the animal techs when they found out I could sex newborn mice (blind, hairless, hotpink). I could tell you how I do it, but getting you to see it is the hard part.

I have all sorts of anecdotes illustrating the effect both from my own experience and others.

I remember when I was very young getting taken to a chicken farm where they were sexing chicks and I remember how fast they were working and all the male chicks being thrown in a bucket to be killed by CO2*. I cannot have been older than 6 since it was here in Scotland and we emigrated to NZ when I was 6 1/2.

*a recognised Schedule 1 (Home Office Animral Experiment Act) method to dispose of large numbers of mice. Looks terrible, they gasp hugely at the end but they have already collapsed and are not conscious. Like the animal rights pictures of animals under surgery with their eyes open. Humans have to have our eyes taped closed so the eyeball doesn't dry out when under anaesthetic.

When doing longish procedures on mice I would periodically drip saline on their wide open eyes. I had already ascertained they were in a surgical plane of anaesthesia.

[A Non e-mouse]

AI Black Box

This "AI is a black box we don't understand" is a problem that's only going to get worse. Without some basic understanding of what the AI is doing to make its decision, you'll never be sure that it's being sane & sensible and isn't looking at the wrong thing in the training data.

[Anonymous Coward]

Re: AI Black Box

"AI is a black box we don't understand"

That's not really the problem, they understand how the AI works, they just don't know what subtle features it's picking out from the training data to categorise the "unknown" data.

"Without some basic understanding of what the AI is doing to make its decision, you'll never be sure that it's being sane & sensible"

Again... They understand what the AI is doing... But I get what you're trying to say. But then that's the point of an AI. To pick up on subtle differences in inputs and correctly categorise it to make fast, appropriate decisions.

"and isn't looking at the wrong thing in the training data."

There is no "wrong thing" if it's coming from the input data, as the training data it's built it's model on is usually the same source as the "unknown" data, only with the answer it should be using supplied.

[Anonymous Coward]

Re: AI Black Box

"you'll never be sure that it's being sane & sensible and isn't looking at the wrong thing in the training data"

To guard against this when training AI the data set being used is can be split (randomly) into two sets. The first "training set" is used to train the AI and then the result is applied to the second "cross-validation set" and the results checked to see if the trained AI achieves an acceptable accuracy on this data - this is to ensure that the AI hasn't just learnt the correct answers from all the questions it saw but is able to generalize.

[Tom 7]

Re: AI Black Box

If you can spare a couple of hours a week you can pop onto futurelearn and do some free courses on Weka and Moa which can give you an reasonable insight into Data Mining and AI which are quite enlightening. It's an art becoming a science and, as illustrated here, coming up with some really useful results which also makes you ask more questions. What more could the inquiring mind want?

[Muscleguy]

Re: AI Black Box

I'm intrigued, to me a Weka is a chicken sized extant NZ bird prone to make a nuisance of itself as it likes shiny things. I once chased one down a path as it tried to make off with my fishing lure, still attached to the line and rod . . .

A Moa is a family of extinct giant rail birds in New Zealand. It is also the fairly generic Polynesian word for chicken. Which means the first Maori to reach NZ saw a giant rail and reckoned it was a big chicken. They probably said 'would you look at the drumsticks on that!'.

[Alan Hope]

In ECGs in women, overall: the heart rate is faster, left axis deviation is twice as common, ST and T-wave abnormalities are almost twice as common, LBBB (left bundle branch block) is slightly more common, etc etc etc for all other ECG parameters.

So I could probably give you a decent guess, from a standard 12-lead ECG, whether it was male or female. I doubt the AI is doing anything more.

But there is a more reliable and faster way to determine whether someone is biologically male or female.

ps the title is slightly misleading, it's not from your "heartbeat" but from a 12-lead ECG (which is a complex surface electrical representation of your "heartbeat").

[Anonymous Coward]

But there is a more reliable and faster way to determine whether someone is biologically male or female.

Ask them to reverse a car?

[Anonymous Coward]

Seems to me there are usually 2 obvious differences in amount of skin, and composition under the skin that should affect electrical measurements. A better test would be if they are identifying breast size, and missing the gender on overweight men, and skinny women.

[Wellyboot]

sex and age affects people’s ECGs

The wires also get in the way and other patients complain about the noise from behind the curtain.

[Cuddles]

Predict?

"Neural networks can predict your biological age and sex just by analysing the pattern of your heart beats"

Neural networks can determine your age by analysing some data. Prediction means trying to figure out what will happen in the future, not assessing something that has already happened or is already known.

[MiguelC]

"you're heart is probably quite healthy"

Yes love, I may be heart, but you're spouting nonsense.

[Anonymous Coward]

Identifying gender now considered controversial?

“At the moment the study only considers a person’s biological sex. Kapa admitted that gender was a trickier and more controversial topic to study.”

How did it come to this where medical researchers censor themselves for fear of retaliation from the woke/diversity crowd, who assert there's no difference between men and women!

sorry for triggering anyone :]

[Imhotep]

Re: Identifying gender now considered controversial?

Walter, I accept your apology for triggering me. After all, just about everything does.

[Muscleguy]

Re: Identifying gender now considered controversial?

Just don't point out biological reality to a Trans person. Doing that is apparently transphobic now.

Sorry folks, even if we could change the X and Y chromosomes in every cell of your body it wouldn't make you female since you would still be the product of the male developmental program, in utero, postnatally and during puberty. Which is males can last until 20/21. I was married at 20 and found I couldn't fit a dress shirt I'd bought at 18. My shoulders had widened and my neck was bigger.

Not an uncommon effect and why we have things like Colts rugby. My wife had no such problem. Well until after the eldest was born she didn't fit her old pre pregnancy jeans, her hips had widened.

Ineluctable biological facts of life there.

BTW it isn't just my axe handle shoulders, I have barrel chest and running gives me tree trunk legs. Probably helped by cycling 5 hilly miles to secondary school and back for 5 years that make me unlikely to pass for female. No amount of hormones are going to change those. Oh yeah and the male version of the family nose is quite different from the female version.

[Anonymous Coward]

Re: Identifying gender now considered controversial?

"[...] that make me unlikely to pass for female."

Many years ago some of our UK engineers were seconded to an Eastern European country to support a mainframe. They found many cultural differences. The single men said that after a while a woman built like a weight-lifter with a small moustache was quite attractive.

[Anonymous Coward]

72% accuracy...

... whereas a purely random answer would be right 50% of the time.

It would be more interesting if it had a confidence value in its output - or even just "M", "F", "not sure".