Alexander Huth

I was spurred to respond to your paper by the fact that it's being shared widely outside neuroscience. It raised the specter of the dead salmon paper (the number of times I've heard about that damn salmon as a gotcha from non-neuroscientists..), so I wanted to dig in and understand the results.

Finally, you may not consider the 40% number the "headline result" of the paper but it is (1) in the abstract, and (2) the literal headline for news stories about this paper. Broadcasting the idea that most fMRI results are hogwash may not be your intent, but it is definitely the outcome

When we repeatedly show fixating subjects short movie clips, these areas often show negative BOLD correlation between repetitions! But with longer and more engaging stimuli the correlations flip to positive. Real functional nonstationarity is anatomically localized and could contribute strongly here

I agree with you that the spatial clustering is a strong signal, and replicating it makes it stronger. But the localization of discordant voxels to VAN/DMN makes me uneasy. These are the areas where we see the most nonstationarity in BOLD as subjects' attention wanders in and out during experiments

The flip from +ve to -ve correlation btwn OEF and CMRO2 in discordant v concordant voxels (your Fig. 5d) is actually also replicated by this simulation — I think it emerges naturally from selecting based on discordance.

I am less confident in the assumed noise levels, which seem plausible but are not empirically based. Still, the only noise that seems to matter is that in the CBF measurement, and its scale here (stdev=1 ml/100g/min) is not crazy, I think.

Thank you for your responses, Valentin! On this: I tried to match the contrast values to your results — the "true" ΔCBF is based on your reported 7.7% from Table S1; others are from Figure 2b,c. The noise-free simulation shows ΔCBF ranging from -7..7%, well within the -15..30% range in your Fig. 3b.

yes, that is a stronger signal. But all of the underlying datasets are also spatially autocorrelated, which will induce some clustering of the results. I also would not be surprised if noise level in the CBF measurement varied across the cortical surface, which would cause consistent clustering

I totally agree that negative BOLD is weird and probably not indicative of decreased metabolism. That's the part of this recent paper that I found most compelling.

yes absolutely. I simulated uniform "true" activations symmetric about zero, but the real data has more +BOLD than -BOLD.

yes, I think it depends on the distribution of "true" values. I used uniform values symmetric about zero in the simulation, so the % discordant is equal for + and -. Shifting the distribution to have more positive values than negative (realistic) makes more of the negatives discordant.

In short, I think the headline result is wrong, or at least statistically unsupported. The results are consistent with BOLD perfectly tracking CMRO2. (I'm not saying it does, but I don't think the data says it doesn't.)

notebook-sharing/bold-cmro2-cbf-r2.ipynb at master · alexhuth/notebook-sharing Contribute to alexhuth/notebook-sharing development by creating an account on GitHub.

My simulation is here if anyone wants to play with it: github.com/alexhuth/not...

The statistically nasty thing here is that CMRO2 is a vey derived metric, so noise in one part affects all. It turns out you only need noisy CBF measurement to get the discordance effect, because CMRO2 is a product of CBF and some other factors.

So: simulating the data-generating process with no discordance but realistic noise gives exactly the same result as the paper (40% discordance). I think this means the result does not exclude the null hypothesis that BOLD and CMRO2 always covary positively.

But (shocker) with realistic amounts of noise you recover something that looks exactly like their plot. It even has roughly the same level of discordance, about 40% in total!

So I made a little simulation where there is no actual discordance, i.e. the underlying values from which the CBF, CBV, BOLD, and T2* are measured all move in lockstep. With zero noise this gives you zero discordant voxels (nice little diagonal line).

(Also the measures here are not that simple: ΔCMRO2 in particular is an extremely derived metric that includes both BOLD and CBF!)

My gut check: that’s a big noisy blob, and they’re saying all points have to fall in those two triangles? Zero chance with real data. Each measure is noisy, and they’re making hay out the fact that if you stratify on one of the measures, the other is not also perfectly stratified

They go on to claim that ~40% of all voxels are discordant, including ~25% of voxels with +BOLD and ~60% of voxels with -BOLD. The 40% figure is in the abstract.

The key result is Figure 3b, which shows ΔCMRO2 vs ΔCBF vs ΔBOLD for a bunch of voxels. Voxels that fall (1) to the right of ΔCMRO2=0 and below the diagonal or (2) to the left of 0 and above the diagonal are called “discordant” because their ΔBOLD and ΔCMRO2 have different signs

They also include several other measures, including especially CBF (cerebral blood flow), which is used to compute CMRO2.

CMRO2 is compared to the more classic BOLD (blood-oxygen-level-dependent) signal, which is easy to measure but not quantitative and complex in origin (it involves changes in blood oxygen, blood volume, and blood flow).

The paper is based on the quantitative functional metric CMRO2 (cerebral metabolic rate of oxygen), which is a calibrated measure of how much oxygen is consumed in a piece of brain tissue during some period of time.

This paper had a pretty shocking headline result (40% of voxels!), so I dug into it, and I think it is wrong. Essentially: they compare two noisy measures and find that about 40% of voxels have different sign between the two. I think this is just noise!

the world is ~1/f. therefore the brain that perceives and models the world should be ~1/f. 🤷‍♂️

📣 New preprint from the Braga Lab! 📣

The ventral visual stream for reading converges on the transmodal language network

Congrats to Dr. Joe Salvo for this epic set of results

Big Q: What brain systems support the translation of writing to concepts and meaning?

Thread 🧵 ⬇️

Sintel - Wikipedia

Always include some stimuli that are permissively licensed so they can be used as examples! E.g. we have a video stimulus set that's mostly Pixar short films, but also includes a segment from the Blender movie Sintel (en.wikipedia.org/wiki/Sintel), which is licensed CC-BY.

Happy and proud to see @rjantonello.bsky.social’s work awarded by SNL!