Cardiologist-Level Arrhythmia Detection Using a Deep Neural Network

> Is ECG output really that easy to read and unambiguous?

Not even close. But nurses generally don’t know how to read EKGs except for the most obvious findings, so when they say “it’s all fine” either they mean they don’t see something very conspicuous (ST elevations, widened QRS, absent p waves), or they’re just giving you the default “stay calm, and wait for the doc to read it” comment.

The only non-docs I’ve ever seen impress me with their EKG reads have been some experienced EMTs, really experienced critical care nurses / mid levels, and some experienced cardiology mid-levels. In short, people that do it every single day and have been doing it for a long while. And even they don’t do it with an instant glance.

EKGs are more complicated than they look. Each lead gives you a different slice through the heart; some of the leads are in the xz axis, and some are in the xy axis, and findings are modified by height, weight, position, metabolic profile, etc. Basic EKG reading takes an hour to learn; being good at it takes forever.

The catch about the above study is that they are doing the easiest possible thing: categorizing arrhythmias. That’s the part you can learn in an hour. Doing it “as well as a cardiologist” just isn’t impressive. Give me a couple of weeks with a bright high schooler and they’ll be doing rhythm classification as well as most doctors.

It’s “precisely which part of the heart is malfunctioning, what part of the vasculature or conduction pathway or whatever does that implicate, and based on the patient’s medical history what underlying diagnosis does that imply? And what is the next best step in medical management?” that cardiologists interpret EKGs for.

Yup. We pretty much just ignore what the machine says.

interesting, does it analyze 12 leads? is your data publicly available? is there any article detailing the DNN you developed? thanks.

I find it interesting they don't cite any framework. They presumably wrote all the DNN code from scratch, which can certainly be done, but seems like a lot of effort unless you plan to commercialize, which clearly they are.

It's deep and large model. Maybe unnecessarily so.

In image processing a single pixel or region almost always belongs to one object. Car can be behind a tree, or tree can be behind a car but they are not transparent and mixed together.

In raw 1d radio, audio, ECG or EEG signals there can be multiple additively overlapping singnals across different frequency ranges. Signal patterns are transparent to other signals. ConvNet on raw singnal must learn to discriminate overlapping pattern combinations.

There are various options, but for long periods single lead devices are often used, can't really walk around for 30 days with a 12 lead holter strapped to you...would make bathing rather difficult.

It's possible to implant small loop recorders under the skin on the chest. These can record for months. Those are mostly useful in cases where symptoms occur infrequently.

I think what I had was called a Body Guardian. It consisted of a bunch of sticky strips (photo here:https://imgur.com/I9cXw1V) that had 4 skin contact pads on the sticky side and 4 snap connectors on the other side. I'd stick this to the center of my chest just above nipple line. There were a pair of small pods that snapped to the 4 contacts. Each was about the 1/2 the size of a deck of cards. They would record continuously for about 16 hours. I'd alternate between them every 12 hours or so, recharging the one I wasn't using in the provided recharger. Also included was a customized Samsung cell phone with it's own recharger and an app that communicated wirelessly with the pods. The cell would download results from the pod & upload it to the company that provided the equipment. I carried that phone with me most of the time. When I connected a pod to the recharger, it would signal the cell phone which would download all results from the pod and upload to the company. My cardiologist got the results from them.

The pods had a button that I could press when I felt like I was having "an event". The cell phone app would popup a selection list for possible events like skipped beat, extra beat, weak pulse, etc. I can't remember the entire list. I wasn't actually required to log these when I felt them -- they were just there to let the monitor S/W know I had noticed something. I'd remove the pod to shower but not to sleep or workout. The sticky strips would stay attached for at least a day unless I sweated a lot, like in a sauna. Strips were not reusable. They provide enough strips for the 30 day trial.

My problem is PVCs (premature ventricular contractions ... skipped beats). Everyone gets these every day. I just get them a lot more frequently. Like 3 beats, skip, 4 beats, skip, 60 beats, skip, 10 beats, skip.

I think the monitoring is pretty continuous. Second hand info, so take this with a grain of salt ..... A friend told me that he knew of someone else on the same or a similar system who has a much worse condition. At one point his monitor cell phone, his personal cell phone, his wife's cell phone and the house phone all rang simultaneously. When they answered, someone on the other end told them to get him to the emergency room immediately. He was having an event and not even aware of it. Turns out it was a good thing they did that.

Thats... not quite what I meant, perhaps I worded that a little bit weird. I meant frequency invariant, not amplitude normalised. As in divide the frequencies by the fundamental frequency (i.e. subtract their logs on the 'scope), not divide the amplitudes by the amplitude on the fundamental.

Any deviation from the profile of the FT irrespective of its base frequency, from the normal sinus rhythm, would easily show. There might be a few problems dealing with cardiac load transitions, and I suppose if you are trying to differentiate the exact class of arrhythmia (as in the paper) then perhaps not, but differentiating normal (sinus), noise, abnormal it should work fine.