iiinova — where your digital self keeps evolving

Reposted by @yapeixie.bsky.social

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

In a meta-analysis of 210 biomedical AI studies that statistically compared models under cross-validation, 97% used invalid statistical tests. Here's our new preprint doi.org/10.64898/202... led by @tianchu.bsky.social @hetuli.bsky.social @shaoshiz.bsky.social @nichols.bsky.social 1/N

412159

Reposted by @yapeixie.bsky.social

Thomas Yeo @bttyeo.bsky.social · 27/03/2026

Do you censor high motion frames in fMRI? In two preprints by @twktan.bsky.social @mandymejia.bsky.social, we find that we may be censoring too much! doi.org/10.64898/202... arxiv.org/html/2603.07... Strict censoring leads to worse personalized TMS targets than no censoring, even with high motion!

25627

Reposted by @yapeixie.bsky.social

Thomas Yeo @bttyeo.bsky.social · 17/07/2025

1/11 Excited to share our @Naturestudy led by @leonooi.bsky.social @csabaorban.bsky.social @shaoshiz.bsky.social AI performance is known to scale with logarithm of sample size (Kaplan 2020), but in many domains, sample size can be # participants or # measurements... doi.org/10.1038/s415...

317182

Reposted by @yapeixie.bsky.social

Thomas Yeo @bttyeo.bsky.social · 11/06/2025

Many cool results from this paper, including a multivariate twist of Simpson's paradox! Stronger salience FC at baseline predicts better cognition at baseline. Would children with larger salience FC increase also enjoy greater cognitive gains over time? Or is the reverse true??

1195

yapeixie.bsky.social @yapeixie.bsky.social · 09/06/2025

Really appreciate it！Great to hear it sparked interest

010

yapeixie.bsky.social @yapeixie.bsky.social · 08/06/2025

Huge thanks to my amazing collaborators @bttyeo.bsky.social @shaoshiz.bsky.social @csabaorban.bsky.social @elvisha.bsky.social @avramholmes.bsky.social @lucinauddin.bsky.social @nichols.bsky.social and others not yet in Bluesky Curious to dive deeper? [Read the full paper: doi.org/10.1101/2025...

doi.org

040

yapeixie.bsky.social @yapeixie.bsky.social · 08/06/2025

12/12 Simply put, development refines brain–cognition links as children enter adolescence, but who you are matters more than how much you change.

120

yapeixie.bsky.social @yapeixie.bsky.social · 08/06/2025

11/12 Together, our findings suggest that during early adolescence, stable individual differences in brain functional network organization shape future cognitive outcomes more strongly than short-term changes.

110

yapeixie.bsky.social @yapeixie.bsky.social · 08/06/2025

10/12 This reveals a multivariate twist on Simpson’s paradox: cross-sectional and longitudinal estimates of brain-cognition relationships can tell very different stories.

120

yapeixie.bsky.social @yapeixie.bsky.social · 08/06/2025

9/12 That task? The Little Man Task — a spatial reasoning measure — where FC change significantly predicted how much cognition changed over time. Focusing on this task, we uncovered both converging and diverging predictive network patterns across models.

120

yapeixie.bsky.social @yapeixie.bsky.social · 08/06/2025

8/12 Surprisingly, across most cognitive tasks, neither baseline FC nor FC change predicted how much cognition changed over time. That is, while we can predict who performs well at a later timepoint, it’s much harder to predict how much they improve over time – except for one task.

120

yapeixie.bsky.social @yapeixie.bsky.social · 08/06/2025

7/12 But even after controlling for reliability (e.g., by shortening the scan duration), baseline FC still outperformed longitudinal FC change in predictive power. These results suggest that stable individual differences in baseline FC carry more predictive weight than short-term changes over time

130

yapeixie.bsky.social @yapeixie.bsky.social · 08/06/2025

6/12 Intriguingly, while FC change had some predictive value, baseline FC was a stronger predictor of future cognition. Why might that be? Could it just be a matter of measurement reliability? Baseline FC: mean ICC = 0.56 FC change: mean ICC = 0.24

120

yapeixie.bsky.social @yapeixie.bsky.social · 08/06/2025

5/12 Building on this, we found that baseline FC better predicted future cognition than current cognition. Models trained on baseline FC to predict baseline cognition generalized better when tested on Year 2 FC and cognition. This suggests that brain–cognition relationships strengthen over time.

120

yapeixie.bsky.social @yapeixie.bsky.social · 08/06/2025

4/12 Both group-level longitudinal change and individual-level stability in functional connectivity (FC) are organized along the sensory–association (S-A) axis. We also observed notable individual differences in FC change, especially in somatomotor, visual, and medial prefrontal cortex

120

yapeixie.bsky.social @yapeixie.bsky.social · 08/06/2025

3/12 Despite clear group-level gains in cognitive performance over two years, individual differences remained highly stable from baseline to Year 2. At the same time, we observed substantial variation in how much cognition changed across individuals.

130

yapeixie.bsky.social @yapeixie.bsky.social · 08/06/2025

2/12 To test this, we analyzed longitudinal fMRI and cognitive data at baseline and Year 2 in ~3,000 individuals (ages 8.9–13.5) from the ABCD Study, spanning the transition from childhood to adolescence.

120

yapeixie.bsky.social @yapeixie.bsky.social · 08/06/2025

1/12 Cross-sectional studies offer only a single snapshot of brain–cognition relationships, missing the dynamic changes that longitudinal designs can reveal. We hypothesize that cross-sectional and longitudinal estimates may diverge, echoing classical Simpson’s paradox. See below：

140