iiinova — where your digital self keeps evolving

Denis Engemann @dngman.bsky.social · 21/10/2025

Join our team as a 12-month PhD intern @ Roche (Basel)! 🧠 Project: Use causal inference & ML on proteomics to find causal drivers of Alzheimer's. For: 3rd-year PhD candidates #Genomics #MachineLearning #CausalInference #CompBio #Alzheimers See details & apply here ➡️ t.co/zMuVjyMW26

040 4

Denis Engemann @dngman.bsky.social · 21/04/2025

🚨 Preprint Alert 🔔 Does EEG brain-age work as a neurodegeneration biomarker? Brain-age was surprisingly under-predicted in neurological populations. Likely due to diverging age trends in EEG power, breaking model assumptions. Preprint: www.biorxiv.org/content/10.1... #biomarker #EEG #neuroskyence

Figure 4 from our paper, proposing a theoretical model for why EEG and other brain-activity measures might deliver inaccurate or surprising and pradoxical brain-age estimates.

060 6

Denis Engemann @dngman.bsky.social · 01/04/2025

9/ Worried about site effects, controlling for MMSE, spatial averaging, impact of head alignment or comparisons VS healthy controls / cortical slowing? We got you covered by dozens of supplementary analyses: www.thelancet.com/cms/10.1016/...

110 1

Denis Engemann @dngman.bsky.social · 01/04/2025

8/ Combining all inputs provided better cross-validation scores than any modality by itself (a). We then used conditional permutation importance to inspect the full model (b). MRI freesurfer volumetric measure, MEG covariance, MMSE and site contributed independent information.

100 1

Denis Engemann @dngman.bsky.social · 01/04/2025

7/ We thank multiple unknown reviewers for motivating us to do a more comprehensive analysis, at the risk of using ML on small data. We used a well tested stacking approach (e.g. elifesciences.org/articles/54055) to have a simple multimodal model while not pre-selecting any inputs.

100 1

Denis Engemann @dngman.bsky.social · 01/04/2025

6/ We next explored phase interaction. Again, the average metrics failed to deliver a clear answer. But the Riemannian distance MANOVA showed systematic group differences in specific frequencies.

100 1

Denis Engemann @dngman.bsky.social · 01/04/2025

5/ The beauty of the wavelet approach is that different spectral metrics can be compared with low friction. We next explored power envelope correlations. The average metric was noisy. Riemannian-distance MANOVA captured differences in a wide range of frequencies.

100 1

Denis Engemann @dngman.bsky.social · 01/04/2025

4/ We next tried to construct a statistical model of progression risk using logistic regression. On a small dataset you want to do simple analyses. We found that combining MEG power and MRI hippocampal ratio with baseline measures provided complementary insights.

100 1

Denis Engemann @dngman.bsky.social · 01/04/2025

3/ Comparing power spectra, we found that converters had lower brain activity in high frequencies above 9Hz (a-c, e). Leveraging Riemannian analysis of the covariance spectrum using distance MANOVA, in addition, we discovered brain activity differences in lower frequencies.

100 1

Denis Engemann @dngman.bsky.social · 01/04/2025

/2 Even if only representing a relatively small sample, BioFIND comes with MEG and MRI measures. This allowed us to explore what differentiates future converters from stable MCI cases using our previously published wavelets pipeline: roche.github.io/neuro-meeglet/

100 1

Denis Engemann @dngman.bsky.social · 01/04/2025

/1 By now it should not be controversial that amyloid beta aggregates are neurotoxic and induce synaptic pathology. Unfortunately, datasets allowing to look at progression from MCI to AD including measures of brain activity are rare. BioFIND is one of those few datasets.

100 1

Denis Engemann @dngman.bsky.social · 01/04/2025

🩺🧠 ✨Excited about our latest work published in @ebiomedicine.bsky.social! In the BioFIND dataset, we found that brain activity provided unique information about progression to Alzheimer’s Disease (AD). #neuroskyence Paper: www.thelancet.com/journals/ebi...; Thread 🧶 ⤵️

Cover figure presenting main results from the paper. Left: altered spectral power and covariance difference between AD progressers and stable MCI patients. Right: multivariate analysis, highlighting the complementarity of MEG covariance spectra, MRI cortical volumes and baseline control measures.

1215 21

Denis Engemann @dngman.bsky.social · 13/02/2025

/9 some additional thoughts. If you ever wondered about the relationship between machine learning and the complex and regulated development of biomarkers in R&D, don’t miss our discussion section in which we tried to share a few ideas about that topic.

010 1

Denis Engemann @dngman.bsky.social · 13/02/2025

/7 We made GREEN with biomarker exploration in mind – as in drug development. Riemannian & deep-learning methods have been popular for BCIs and there is no reason why GREEN should not work here! On MOABB BNCI2014_001 motor imagery, GREEN was among the high performing models.

100 1

Denis Engemann @dngman.bsky.social · 13/02/2025

/6 If you want to get your hands dirty with interpretation work, GREEN has you covered: A partial forward pass to get power topographies from the pooling layer (diagonal of covariance). For the classical Berger effect, we nicely see the occipital eyes-closed alpha modulation.

100 1

Denis Engemann @dngman.bsky.social · 13/02/2025

/5 Comparing learnable wavelets VS a fixed grid of Morlet wavelets allowed us to explore if spectral information was concentrated or diffusely spread. On e.g. sex classification, fixed wavelets were unbeatable (diffuse). Other tasks showed highly localized oscillatory signatures.

100 1

Denis Engemann @dngman.bsky.social · 13/02/2025

/4 Comparing VS complex models e.g. transformers on the CAU data or TUAB, GREEN brings you the same performance with several orders of magnitude fewer parameters and training time. If needed you can expand GREEN via classical phase synchrony measures e.g. PLV - thanks wavelets!

100 1

Denis Engemann @dngman.bsky.social · 13/02/2025

/3 We threw GREEN G1-G3 at five EEG prediction problems (>5000 EEG recordings) and, well, it depends. The flat Riemannian model (G1) was a very competitive baseline. Adding the FC layer (G2) could lead to improvements, yet, most was gained by adding learnable wavelets (G3).

100 1

Denis Engemann @dngman.bsky.social · 13/02/2025

/2 Riemannian models (G1) are effective EEG decoders as the log map matches lognormal brain scaling (Buzsaki & Mizuseki)? If adding a neural network improves => more complex non-linearity (G2). Or is it about hitting the right frequencies (G3)? GREEN will help you figure it out!

100 1

Denis Engemann @dngman.bsky.social · 13/02/2025

/1 GREEN is made of a few building blocks. But those represent operations well tested in EEG research and neuroscience. This not only makes parameter sparse models but facilitates interpretability by comparing what operations improve prediction performance.

100 1

Denis Engemann @dngman.bsky.social · 13/02/2025

📣 ✨Excited about our latest work in Patterns (Cell Press): GREEN a light-weight #NeuralNetwork with Riemannian Geometry + learnable wavelets for #EEG #biomarkers: doi.org/10.1016/j.pa... Joint work with @jpaillard.bsky.social & Jörg Hipp! Thread 🧵👇🏻👇🏾👇

1175 17