For more details, please check out the paper! A huge thanks to my coauthors @alaingoriely.bsky.social @jwvogel.bsky.social @alexapb.bsky.social @saadjbabdi.bsky.social @biofinder.bsky.social (and others not on Bluesky) for helping to get this over the line.

With help from Oskar Hansson, @jwvogel.bsky.social, @alexapb.bsky.social and the @biofinder.bsky.social team, we were also able to replicate this result in an independent dataset with a different tau tracer, the BF2 dataset!

We then fit the model to biomarker groups across the AD continuum in ADNI. Our results show an initial transport-dominated phase (A+T-), followed by a production-dominated phase (A+T+). Suggesting early cortical tau deposition via network diffusion is followed by accelerated local tau production.

We used a series of test/train splits on A+T+ ADNI subjects with at least 4 scans to show how adding data changes prediction uncertainty. We show that too few scans can lead to misleading predictions, but with 3 longitudinal scans, regional predictions are pretty good! (here showing left inf temp.)

Through model comparison on an A+T+ group from ADNI, using in-sample and out-of-sample measures, we show that regional heterogeneities are essential for longitudinal modelling of tau progression and highlight deficiencies in a network diffusion model and models with homogenous production dynamics.

First, we introduce a novel network reaction-diffusion model that uses regionally heterogeneous tau carrying capacities (the maximum amount of tau a region can hold), and show that the model reproduces tau spreading as observed through tau PET.