The classic signal shows up: slow theta rhythms rise and alpha falls. Age and sex alone only reach about 59%, so the EEG is doing the work. But telling Alzheimer's from FTD is much harder: our best attempt (* a person-level model plus a second, eyes-open recording) reaches 63.3%, only a little above guessing. More people, not more features, is probably the fix. That's the open problem, and where I'd love help.
Model selection uses nested cross-validation, so the numbers aren't tuned on test people. Honest caveats: a small single-site sample, and it's a research benchmark, not a diagnostic tool.
How it works
- Experts (doctors, scientists) share insights through a short form, no Git needed, or join as scientific advisors.
- Builders pick those ideas up as issues and implement them on open data.
- Results go back to experts for review (
needs-expert-review) before anything is presented.
Every change is reviewed, CI blocks committed data files and wording that sounds like a diagnosis, and everyone, expert or developer, is credited by name.
Where you can help
- Finding more open dementia EEG datasets (more people is the main bottleneck)
- Pretrained EEG foundation models for Alzheimer's vs FTD
- Testing on a second, independent EEG dataset
- Leading a new module: tau spread simulation, drug repurposing with knowledge graphs, or speech markers
Website: https://lordave1.github.io/openengram/
Code: https://github.com/LorDave1/openengram
Cite (Zenodo DOI): https://doi.org/10.5281/zenodo.24281637
If you work in neurology, neuroscience or ML, I'd be grateful for your critique. That's exactly what the project needs.
Top comments (0)