Hacker News Reader

Comment by haldujai | Hacker News Reader

This is a terrible demonstration of an unnecessarily complicated solution using pictures of signals instead of the actual signal data for something that is very unlikely to change patient management.

The Mayo study is marginally less questionable with uncontrolled confounders and an outcome measure carefully chosen to show a positive result for a tool that is patented by the Mayo Clinic. Also doubtful that this changes patient management and they chose not to look at that in their study despite having access to that information.

View on HN

The images are the actual data. Vectorizing/normalizing/transforming the representation of data by means of moving it into an image space is a completely valid approach in ML and can have many advantages, not the least of which is being able to take advantage of models and research that have been done with image data.

After all, this is the same thing that we do for human doctors reading an ECG. Or at least I've never heard one exclaim "This chart is useless, can I please have the CSV?"

> The images are the actual data. Vectorizing/normalizing/transforming the representation of data by means of moving it into an image space is a completely valid approach in ML and can have many advantages.

I'm aware that this is a valid approach in general.

> not the least of which is being able to take advantage of models and research that have been done with image data.

Essentially all of the research, clinical and computer science, that we would be able to take advantage of is based off of digital ECG recordings and numerical data not digitized ECG images.

This is the first publication to my knowledge using an image based approach and their innovation is partially validating a deep learning model that is as accurate as explainable models and algorithms described as early as the mid 2000s and can already be clinically implemented.

> After all, this is the same thing that we do for human doctors reading an ECG. Or at least I've never heard one exclaim "This chart is useless, can I please have the CSV?"

So because a human is incapable of analyzing digital signals the best approach for ML is to artificially use the same limitation?

I would love to hear a reason for why it may be a better approach to analyze a noisier and less rich digitized image space representation an ECG rather than the raw digital signals.

Because you can take advantage of pretrained CNNs and perform transfer learning, which is significantly more data-efficient than training from scratch, which is what you'd likely have to do with raw digital signals. This paper is not unique in this approach and many papers have obtained SOTA results by processing digital signals as images.

The complexity/dimensionality of the data representation is increased considerably when going from time series to images of said time series. Sure one can then use transfer learning to manage this complexity. But do you have any references for this approach being more data effective overall?

It's highly likely that they analyzed images because they could not reliably access the digital signal traces. ECG machines, including networked "digital" ECGs store digital ECGs only in proprietary formats, and do not make them accesible as raw data.

The Mayo models were built using digital signal traces.

> Essentially all of the research, clinical and computer science, that we would be able to take advantage of is based off of digital ECG recordings and numerical data not digitized ECG images.

This project isn't using any of that. It's what one would call a 'novel' approach. It may upset you that they are not building on the existing body of medical knowledge that you might consider important. They are instead building on the body of knowledge that exists in a different space.

> This is the first publication to my knowledge using an image based approach and their innovation is partially validating a deep learning model that is as accurate as explainable models and algorithms described as early as the mid 2000s and can already be clinically implemented.

Then it's notable they achieved this result given they did not inherit any of the previous research.

> I would love to hear a reason for why it may be a better approach to analyze a noisier and less rich digitized image space representation an ECG rather than the raw digital signals.

Once again, transforming data into an "image" representation does not automatically imply a lossy process or a process that introduces noise. There are ML models which operate on the raw bitstream from a CCD camera just as well as ML models which operate on "frames" of image data. Both approaches are valid ways to "see the world."

I have no opinion on the value of the study as a matter of changing practice, or patient outcomes. I referenced it to illustrate that AI algorithms for diagnosing LV function from ECGs are already in the field, being tested.

Reply on news.ycombinator.com