iiinova — where your digital self keeps evolving

Thomas Yeo @bttyeo.bsky.social · 08/07/2026

A video of a patient who responded well to personalized TMS: www.temasekreview.com.sg/community-st...

temasekreview.com.sg

Thomas Yeo @bttyeo.bsky.social · 17/06/2026

Excited that Imaging Neuroscience has gotten its impact factor. Note that the impact factor is artificially low for the first year of any journal (because of the way it is calculated), so we expect the impact factor to improve next year. See more statistics below

0152

Reposted by Thomas Yeo

Imaging Neuroscience @imagingneurosci.bsky.social · 17/06/2026

Our first Impact Factor is 3.0 — an important milestone for a new journal. As with most new journals, the first IF is affected by the smaller publication volume in the launch year. Latest citation data are encouraging, and the journal’s IF is set to rise in 2027.

15820

Thomas Yeo @bttyeo.bsky.social · 15/06/2026

For those at @ohbmofficial.bsky.social @ohbmtrainees.bsky.social come check out our poster on using deep learning to accelerate biophysical model fitting & resulting insights into lifespan changes in E/I ratio! Poster number 2304 Wed, June 17 | 13:45-14:45 Thurs, June 18 | 14:45-15:45

051

Reposted by Thomas Yeo

wakeupwoke.bsky.social @wakeupwoke.bsky.social · 14/06/2026

One interpretation: baseline salience FC reflects the attractor basin characteristics of that network or how reliably it settles into coherent functional states. Low FC = shallow, poorly defined basins. The network can get there, but with more variability and less precision.

121

Reposted by Thomas Yeo

Rosanna Olsen @rosannaolsen.bsky.social · 14/06/2026

Attending #OHBM2026.🇫🇷? You will not want to miss Monday morning's keynote by Nanthia Suthana (chaired by yours truly 😊). I can't wait!!

0114

Thomas Yeo @bttyeo.bsky.social · 14/06/2026

Congrats to @shellakeilholz.bsky.social and thank you to the amazing leadership by our outgoing EIC @fmrib-steve.bsky.social Hope the community can continue to support the journal to bring imaging neuroscience to new heights!

060

Reposted by Thomas Yeo

wakeupwoke.bsky.social @wakeupwoke.bsky.social · 14/06/2026

Thank you for sharing!

111

Thomas Yeo @bttyeo.bsky.social · 14/06/2026

But for once we are not the worst haha

110

Thomas Yeo @bttyeo.bsky.social · 14/06/2026

This slide was particularly popular. Never had so many people cheering one of my slides before. haha

0131

Thomas Yeo @bttyeo.bsky.social · 14/06/2026

It was great fun giving the talk at the neuroimaging statistics workshop. Happy to share the slides here: www.dropbox.com/scl/fi/8riom... Hopefully, the talk was pitched at a level that was understandable to non-statisticians!

dropbox.com

Dropbox

0152

Thomas Yeo @bttyeo.bsky.social · 14/06/2026

For those going to @OHBM @OHBM_Trainees you can check out our poster on spectral normative modeling! Poster Number: 1054 Monday, June 15, 14:45-15:45 Tuesday, June 16, 13:30-14:30 Our preprint has also been massively updated: doi.org/10.1101/2025...

093

Reposted by Thomas Yeo

Sina Mansour L. @sinamansourl.bsky.social · 06/02/2025

1/ Excited to share our latest preprint! 🚀 We introduce Spectral Normative Modeling (SNM)—a novel approach leveraging graph spectral methods to advance brain charting towards personalized precision medicine. 🔗 www.medrxiv.org/content/10.1...

medrxiv.org

Spectral normative modeling of brain structure

Normative modeling in neuroscience aims to characterize interindividual variation in brain phenotypes and thus establish reference ranges, or brain charts, against which individual brains can be compa...

1217

Thomas Yeo @bttyeo.bsky.social · 14/06/2026

If you are going to @ohbmofficial.bsky.social @ohbmtrainees.bsky.social , come check out our poster to estimate & validate individualized brain networks in epilepsy! Poster Number: 0033 Mon, Jun 15 | 13:45-14:45 Tues, Jun 16 | 12:30-13:30 The paper has also been published: doi.org/10.1002/epi....

0101

Thomas Yeo @bttyeo.bsky.social · 14/06/2026

I guess the stable architecture we found must be related to your SES findings? @ndosenbach.bsky.social @smarek0502.bsky.social

021

Reposted by Thomas Yeo

Scott Marek @smarek0502.bsky.social · 11/06/2026

What matters most for childhood brain organization? We analyzed 649 variables. The answer: Socioeconomics (SES); with brain patterns pointing at sleep & stress as drivers. Even brain-IQ associations were better explained by SES. In Science today: www.science.org/doi/10.1126/...

6319169

Thomas Yeo @bttyeo.bsky.social · 14/06/2026

I guess our results of a stable architecture can be explained by your recent SES results? @smarektj.bsky.social @ndosenbach.bsky.social

031

Thomas Yeo @bttyeo.bsky.social · 14/06/2026

For those attending @ohbmofficial.bsky.social @ohbmtrainees.bsky.social come check out our poster on E/I imbalance in pre-dementia individuals and the relationships with blood/CSF biomarkers. Poster 1059 Stand-by time: Monday, June 15 | 13:45-14:45 & Tuesday, June 16 | 12:30-13:30

0167

Thomas Yeo @bttyeo.bsky.social · 13/06/2026

Openly shared implementation of TMS targeting is still rare, so we are pleased to make ours freely available for research use: github.com/ThomasYeoLab... Let us know if you have any issue running it!

github.com

GitHub - ThomasYeoLab/Kong2026_TMSTree: Tree-based MS-HBM TMS targeting algorithm

Tree-based MS-HBM TMS targeting algorithm. Contribute to ThomasYeoLab/Kong2026_TMSTree development by creating an account on GitHub.

0169

Thomas Yeo @bttyeo.bsky.social · 13/06/2026

Our TMS algorithm is now published in @imagingneurosci.bsky.social doi.org/10.1162/IMAG... If you are coming to @ohbmofficial.bsky.social @ohbmtrainees.bsky.social, come check out our poster about our open label trial. Poster number 1152 Monday, Jun 15 | 14:45-15:45 Tuesday, Jun 16 | 13:30-14:30

0103

Thomas Yeo @bttyeo.bsky.social · 13/06/2026

For those attending @ohbmofficial.bsky.social @ohbmtrainees.bsky.social , come check out our poster finding that functional MRI brain foundation models do not outperform ridge regression. Poster number 2269 Presenting on Wednesday, June 17 | 12:45-13:45 and Thursday, June 18 | 13:45-14:45

03111

Thomas Yeo @bttyeo.bsky.social · 13/06/2026

If you are attending @ohbmofficial.bsky.social @ohbmofficial.bsky.social come by and check out our poster on how excessive censoring hurts parcellation and personalized TMS target accuracy. Poster number 2399 Presenting on Wednesday, June 17 (12:45-13:45) and Thursday, June 18 (13:45-14:45)

0166

Reposted by Thomas Yeo

OHBM Australia @ohbm-australia.bsky.social · 09/06/2026

🎉 Registration is now OPEN for the 2026 OHBM Australian Chapter Annual Meeting! 🇦🇺🧠🇦🇺 Get your ticket now! 🎟️🎟️🎟️ Registration: events.humanitix.com/ohbm-austral... 📝 Submit your abstract: docs.google.com/forms/d/e/1F... 🏆 Nominate a colleague (or yourself) for an award: ohbm-aus.github.io/awards/

events.humanitix.com

OHBM Australia Meeting 2026

Uniting the Australian human brain mapping community.

11112

Thomas Yeo @bttyeo.bsky.social · 25/05/2026

For those who are attending @ohbmofficial.bsky.social brainhack, I will be giving a short presentation of this work at the neuroimaging statistics workshop, co-hosted with brainhack: sites.google.com/view/nsw2026

sites.google.com

NeuroStats2026

02314

Thomas Yeo @bttyeo.bsky.social · 22/05/2026

(5c) Bootstrap variant #3: bootstrap-t. Same as bootstrap-ET, but applies t-test to bootstrapped samples. Bootstrapped samples aren't independent → t-test breaks → very high FPR. Bootstrap procedures vary widely in validity — papers should describe theirs clearly. 7/7

111

Thomas Yeo @bttyeo.bsky.social · 22/05/2026

(5b) Bootstrap variant #2: bootstrap-ET. Run CV first → bootstrap the fold-level differences → apply empirical test (ET) of differences (CI from 2.5%/97.5% percentiles). Valid, but very low power — trades sensitivity for robustness. Bootstrapping happens after CV. 6/N

111

Thomas Yeo @bttyeo.bsky.social · 22/05/2026

(5a) Bootstrap variant #1: bootstrap-orig (Raschka 2018). Bootstrap before training — sample with replacement, train on unique samples, test on the rest. Repeat to build a CI. Valid + moderate power. The key: bootstrapping happens before model training. 5/N

121

Thomas Yeo @bttyeo.bsky.social · 22/05/2026

(4/5) A common error — including in my own lab 😅. The corrected t-test is valid on fold-averaged stats. But if we average across folds within each repetition, the between-repetition correlation is much higher than the 1/K the test assumes, and it breaks. 4/N

121

Thomas Yeo @bttyeo.bsky.social · 22/05/2026

(3/5) Sample-level vs fold-averaged statistics. Sample-level retains every test sample's stat (instead of 1 per fold). Within-fold correlations are even stronger than between-fold, so sample-level statistics inflate FPR even more. 3/N

121

Thomas Yeo @bttyeo.bsky.social · 22/05/2026

(2/5) The label-shuffling permutation test is valid for "is prediction better than chance?" — but invalid for model comparison. Why? Shuffling labels creates a no-signal scenario, which is the wrong null for comparing two models. Slides explain more. 3/N

141

Thomas Yeo @bttyeo.bsky.social · 22/05/2026

(1/5) The paired (sign-flip) permutation test fails under cross-validation. The swap procedure assumes fold-averaged statistics are exchangeable under the null — but they're positively correlated, not independent. The result: inflated FPR. Slide explains the mechanics. 2/N

132

Thomas Yeo @bttyeo.bsky.social · 22/05/2026

Here's bonus slides on cross-validation tests, separate from our preprint. Covering: 1. paired (sign-flip) permutation test 2. label-swap permutation test 3. sample-level vs fold-averaged stats 4. a common misapplication of the corrected t-test 5. three bootstrap variants 1/N

13015

Reposted by Thomas Yeo

Konrad Kording @kordinglab.bsky.social · 21/05/2026

Statistically crossvalidation folds are not independent. The number of papers treating them as if they were is off the charts.

0101

Thomas Yeo @bttyeo.bsky.social · 21/05/2026

It's actually surprisingly hard for AI to catch it because the reporting by authors are usually not sufficiently clear. In our meta-analysis, we used Opus 4.1 to help filter some studies, but Tianchu, Hetu & Shaoshi still have to manually go through >1000 studies. (Maybe Opus 4.7 might be better?)

100

Thomas Yeo @bttyeo.bsky.social · 21/05/2026

Neuroscience has the lowest prevalence! Yay us! (But still at 84% 😅) bsky.app/profile/btty...

020

Reposted by Thomas Yeo

Shaoshi Zhang @shaoshiz.bsky.social · 21/05/2026

For years, we've known that running a standard t-test on cross-validation folds violates sample independence. We wanted to see how widespread this issue actually is. The result? 97% of the studies used an invalid statistical test. 🧵👇

052

Reposted by Thomas Yeo

Sina Mansour L. @sinamansourl.bsky.social · 21/05/2026

Can't stress this enough 👇 If you use ML to compare predictive models in your research (neuroscience, genetics, you name it), this paper is a must read! 👀 The majority of work in this space (mine included 🙋) misses critical nuances when reporting comparative stats.

061

Reposted by Thomas Yeo

Tianchu Zeng @tianchu.bsky.social · 21/05/2026

So glad this is finally public. Grateful to my wonderful co-authors for the long journey.

083

Thomas Yeo @bttyeo.bsky.social · 21/05/2026

Our meta-analysis also revealed broader reporting gaps: 54% of studies didn't clearly apply a statistical test — either none was done, or reporting was too vague. We recommend reporting: folds, repetitions, the specific test, and how fold-level statistics are aggregated. 12/N

161

Thomas Yeo @bttyeo.bsky.social · 21/05/2026

So we don't claim 97% of studies contain false positives. Rather, 97% employed statistical tests that don't achieve the stated error control. Prior findings warrant reassessment, not wholesale dismissal. 11/N

1171

Thomas Yeo @bttyeo.bsky.social · 21/05/2026

Important caveat: an invalid test doesn't imply a false positive. Sufficiently large effects may remain significant under more conservative procedures, and any given model comparison may be just one component of a broader study whose conclusions hold. 10/N

191

Thomas Yeo @bttyeo.bsky.social · 21/05/2026

Our study focuses on single-dataset comparisons, but our guidelines also cover the multi-dataset case. The seminal Dietterich (1998) doi.org/10.1162/0899... offers a wonderful discussion of how scientific goals should guide test choice — highly recommended! 9/N

1112

Thomas Yeo @bttyeo.bsky.social · 21/05/2026

We benchmarked SHARP against 12 existing tests, including fold-aware tests like the corrected t-test and 5×2 paired t-test. SHARP showed the best overall balance of false positive control, statistical power, and confidence-interval calibration. 8/N

191