A positive result means someone is about seven times more likely to actually have an autism diagnosis. This rivals the best imaging markers, while still being found in about 1 in 200 people in the general population. 4/🧵

We turned the problem on its head. Instead of trying to classify everyone, we built a brain signature that only makes predictions when it’s confident. 3/🧵

Is that #neuroAI???

In long-term follow-up (up to 16 months), we observed no sustained hearing loss. While high-frequency measures remained variable, no clinically significant declines were consistently found across the group. 7/🧵

Across participants, we found no clinical signs of ear trauma immediately following scanning. Changes in detection thresholds were typically <10 dB, even in high-frequency ranges (>10 kHz) where variability was greatest. 5/🧵

MRI scanners are loud. We used MRI-compatible Sensimetrics earphones with foam inserts and custom over-the-ear protection (see picture of an early prototype).

Still, we wanted to know: does this level of scanning impact auditory health?

🧪 The results? no major structural changes, and in some cases, metrics were remarkably consistent (check out these average brain T_1^MP2RAGE for a teaser). 4/🧵

🧠📊 Just out: a deeply sample structural brain and spine dataset from the cneuromod.ca team, now published in Imaging Neuroscience! led by Mathieu Boudreau and Julien Cohen-Adad. 📰 Paper here: lnkd.in/ewXA3dED 1/🧵

Alright, I just have too many notebooks, and I don't know where to find anything anymore. New 2025 resolution - topical notebooks. And here's the first specimen: infrastructure.

psychiatric conditions are known to have an overlap in genetic risk. This can be quantified through a genetic correlation based on large gwas studies. The amount of genetic correlation across pair of conditions strongly associate with the spatial correlation of case-control fmri connectomes 5/N

actually, if you estimate the effect of a rare genetic variants in terms of its loss in average IQ loss, it correlates strongly with the effect size on brain connectivity. 4/N

effect size estimates in case-control comparison. top: rare genetic variants (very top some with huge clinical effects, bottom low clinical effects). In the middle common clinical diagnoses, at the bottom polygenic risk scores 2/N fig 2 cdn.ncbi.nlm.nih.gov/pmc/blobs/08...

living caricature from www.explainxkcd.com/wiki/index.p... but on $ steroids and surrounded by an army of sycophants.

@karimjerbi.bsky.social is on bsky! Karim is a human of many wonders. One of his achievements has been to organize a major neuroAI conference (called MAIN), which has helped shape the Montreal neuroAI community since 2017! Please join me in welcoming him here 🙏

btw I believe this reference (cited in the article above) directly implements what you requested, which is image decoding from latent representations arxiv.org/pdf/1412.003... Activation atlases are better though, and could be done in brains through individual models. See arxiv.org/abs/2411.10872

I've always found Van Essen's style diagrams very hard to parse. I gave it a stab. Please tell me what you think. PS: I removed one of the three branches of the frontal aslant track, which had the exact same terminations. PPS: I messed minimally with layout, but moved things for clarity.