Comment by kens | Hacker News Reader

Comment by kens | Hacker News Reader

Parent Full thread

Thanks for the paper link. I read the paper and it's really a semiconductor sensor engineering paper, not a biology paper. The paper seems very preliminary; it shows that CFS and non-CFS sample have widely different impedances under osmotic stress, but the mechanism behind this is entirely hand-waving.

If I'm reading the paper correctly, the samples were peripheral blood mononuclear cells (i.e. lots of different cell types) in their own plasma. It would be nice to have samples that weren't so heterogeneous, since it's unclear what type of cell is causing the effect, or if it is something in the plasma.

It would be nice if the paper gave the actual measured resistances, rather than % deviation from normalized impedance which is an unusual unit. The paper's use of machine learning seems entirely pointless, basically drawing the obvious line between the two sets of points.

The discussion section didn't seem to have anything concrete. It reminded me of a patent application where you try to list everything that might possibly turn out to be vaguely related.

I apologize for the negativity; the paper's result is very exciting so I hope it pans out, but I found the paper itself disappointing.

This is less a criticism of the paper than it is a statement of misaligned expectations.

> but the mechanism behind this is entirely hand-waving.

The objective of this paper was not to specify, or even attempt to specify, the mechanism.

> The paper's use of machine learning seems entirely pointless, basically drawing the obvious line between the two sets of points.

Yeah, that is pretty much the entire objective of SVMs.

> Yeah, that is pretty much the entire objective of SVMs.

Which is hard to do by hand in many dimensions but almost trivial in 2. I haven't read the paper but maybe the OP is complaining about something like that?

Yes, the data is essentially 1-dimensional and looks like "CCC MMMMMMMM". (Seriously, look at Figure 2 graph H in the paper.) It seems entirely gratuitous to use machine learning to build a SVM classifier to separate the C's from the M's, when it's obvious that the C's are on the left and the M's on the right.

Perhaps the use of machine learning helped to acquire funding, or enabled the researchers to acquire skills for future projects.

> samples were peripheral blood mononuclear cells

Yeah, I was expecting to see flow cytometric cell sorting (1) for monocytes on CD14 and CD64 (2), which I would believe to be monocytes.

The "sample preparation" procedure they describe (3) is ... somewhat lacking in specificity. They pelletized the sample three times. I'm not even sure those cells are intact, and, mechanistically, the sample no doubts includes granulocytes, lymphocytes, and monos, so who knows what's going on there.

(1) https://www.labome.com/method/Flow-Cytometry-and-Cell-Sortin...

(2) https://www.ncbi.nlm.nih.gov/pubmed/9844593

(3) https://www.pnas.org/content/pnas/suppl/2019/04/25/190127411...

The use of machine learning needs to be explained. I’m half expecting it to boil down to “we evolved the neural nets until it perfectly separated the two groups” — overfitting “learning”.

I also had that impression at first but it seems that they are just using Support Vector Machines to draw the separating boundary between the two group of points.

If you look at the graphs by eye, there is an extremely obvious divide. SVM just let them find it automatically, but if you gave a human those graphs, they could instantly point out where the line is.

Why does it need to be explained if the test can be proven to be extensible? Asprin worked for nearly 100 years before anyone had any clue how (ditto with many, many more modern drugs).

To me different impedances implies different electrolyte concentrations. That's my EE perspective, at least.

Reply on news.ycombinator.com