As a doctor as opposed to an AI researcher, so many of the choices this study makes are baffling to me.
First of all, why just one heartbeat? You never capture just one heartbeat on an ECG anyway, and "Is the next heartbeat identical to the first one?" is such an important source of information, it seems completely irrational to exclude it. At least pick TWO heartbeats. If you're gonna pick one random heartbeat, how do you know you didn't pick an extra systole on accident? (Extra systoles look different, and often less healthy, than "normal" heart beats, as they originate from different regions of the heart.)
Secondly, why heart failure and not a heart attack? One definition of heart failure is "the heart is unable to pump sufficiently to maintain blood flow to meet the body's needs," which can be caused by all sorts of factors, many of them external to the actual function of the heart - do we even know for sure that there are ANY ECG changes definitely tied to heart failure? Why not instead try to detect heart attacks, which cause well-defined and well-researched known ECG changes?
(I realize AIs that claim to be able to detect heart attacks already exist. None of the ones I've personally worked with have ever been usable. The false positive rate is ridiculously high. I suppose maybe some research hospital somewhere has a working one?)
To add to this, looking at figure 4, why is their "average" heartbeat so messed up? That's not what a normal average heartbeat looks like. P is too flat, Q is too big, R is blunted, and there's an extra wave between S and T that's not supposed to be there at all. If their "healthy patient" ECGs were bad enough to produce this mess on average, it's no surprise their AI had no trouble telling the data sets apart.
(For comparison, the "CHF beat" looks a lot more like a healthy heartbeat.)
I think it's a sort of academic machismo. "Look what we can do - isn't it amazing?"
I saw the same thing in Robotics recently. An academic came to give a talk on localisation using computer vision: they cross-referenced shop signs that were seen by a robotic camera with the shop's location on a map to get a rough estimate of where the robot was. My first question was "what is the incremental benefit of this approach when it was combined with GPS?". It turned out that the researchers just hadn't used GPS at all - almost like they considered it to be "cheating".
I feel like many academic disciplines have unwritten 'rules' that you need to follow if you want to be included in the conversation. Not all of those rules are sensible.
First of all, why just one heartbeat? You never capture just one heartbeat on an ECG anyway, and "Is the next heartbeat identical to the first one?" is such an important source of information, it seems completely irrational to exclude it. At least pick TWO heartbeats. If you're gonna pick one random heartbeat, how do you know you didn't pick an extra systole on accident? (Extra systoles look different, and often less healthy, than "normal" heart beats, as they originate from different regions of the heart.)
Secondly, why heart failure and not a heart attack? One definition of heart failure is "the heart is unable to pump sufficiently to maintain blood flow to meet the body's needs," which can be caused by all sorts of factors, many of them external to the actual function of the heart - do we even know for sure that there are ANY ECG changes definitely tied to heart failure? Why not instead try to detect heart attacks, which cause well-defined and well-researched known ECG changes?
(I realize AIs that claim to be able to detect heart attacks already exist. None of the ones I've personally worked with have ever been usable. The false positive rate is ridiculously high. I suppose maybe some research hospital somewhere has a working one?)