We talked about the potential of using both spatial and temporal information like dynamic FC for prediction instead of averaging over time, and how our model can be used to compare individuals, including patients, to a reference population. Paper: elifesciences.org/articles/95125
I've had a chat with Chris from #TheNakedScientists for the @elife.bsky.social podcast about our new paper "Predicting individual traits from models of brain dynamics accurately and reliably using the Fisher kernel": elifesciences.org/podcast/epis... #neuroskyence
Read the short and sweet version of our new paper about predicting from brain dynamics, featured in elife digests magazine: elifesciences.org/digests/9512... #neuroskyence
The intuition is that our approach uses a useful projection - like shining a light on an object and looking at its shadow
The whole workflow, from fitting the HMM to constructing the kernel and predicting from it, is part of the GLHMM toolbox in Python: github.com/vidaurre/glhmm and the old HMM-MAR toolbox in Matlab: github.com/OHBA-analysi.... Find all code to replicate the paper in github.com/ahrends/Fish....
The HMM-Fisher kernel approach has no issues here, but we found that several other kernels were problematic in this respect. The key here is the projection: We found that, like for static FC, the right projection leads to more accurate and more reliable predictions.
Beyond accuracy, we thought a lot about reliability. If we run the model again, using standard CV and regularisation with new randomised folds, do the results change dramatically? And 2. Are there single cases where a prediction is so terrible that it would be useless in real-world applications?
Illustration of the workflow for the HMM-Fisher kernel approach to predict individual traits from models of brain dynamics.
The HMM-Fisher kernel approach allows leveraging the entire rich description of dynamic functional connectivity and amplitude changes to predict, e.g. an individual’s cognitive test scores or demographics. It is also computationally efficient and flexible to be used with various other models.
Predicting individual traits from models of brain dynamics:
In our new paper with Mark Woolrich & Diego Vidaurre in @elife.bsky.social, we combine a generative model of brain dynamics, here the Hidden Markov Model, with the Fisher kernel to predict individual traits elifesciences.org/articles/95125