Binary classifier for dementia based on electrode power features,source estimation and source level connectivity

I am looking for a good discussion on the above topic to help me with a specific research project i am working on.I do not have much experience with writing ensemble or meta level classifiers and it will be great to have your inputs.

The dataset consists of absolute and relative power features derived from 19 EEG electrodes across all power bands (alpha, beta, theta, delta, gamma), in addition to, source level power features obtained after an sLORETA analysis for the brain ROIs. There are also additional source level connectivity strength estimates between the different brain ROIs. There are 20k features altogether. We would like to predict whether the patient has dementia or not based on the above feature set.

What types of ML and dimensional reduction algorithms work best for such feature? I think that the source estimation and source connectivity strength features are derived from the eeg scalp level time series data. If that is so, then, all of these features are correlated with one another. Would we need to write separate feature extractors for each type of features (scalp electrode power, source power, source connectivity measures) or run a dimensionality reduction algorithm on the entire feature set ?

@pantazis @Alexandre?