Brain connectivity and clinical application

Organized by molecular connectivity group (molecularconnectivity.com) * fMRI has been around for a long time, still no clinical application?

classical biomarkers relies on within-sample comparison or association (patients vs. controls)
- problem: misleading biomarkers -also pertain to group means. individual statements are usually not possible
- changing in the context of machine learning, key conceptual point, not testing for group means, train model on available data (some multivariate features) - does model generalize?
preprocessing, compression, target of ML approach matters for predictive accuracy
compound scores tend to be more reliable than individual scores
how well can you actually predict your targets just by looking at non-brain markers? (SES etc.). Too much emphasis on connectome prediction without considering other phenotypic markers?
human variability is likely low-dimensional:
- confounds are a gradient and perhaps everything is a confound
- no such thing as a deconfouned analysis

investigate training N dependence on out-of-sample prediction of endpoints for standard inter-subject covariance analysis
- at what point does adding more data fail to result in any predictive/explanatory gain?
investigate topographic robustness of derived pattenrs - another metric apart from endpoint prediction -> important for understanding mechanisms
META or MEGA analysis: is it better to use all available data in one shot, or split the data into parcels?
Inter-individual (rather than intra-individual)
analysis framework: scaled subprofile modeling (=SSM), form of PCA regression
- step 1: PCA on fMRI data
- brain-behavior modeling
- construction of corresponding brain pattern
- application of brain behavioral model to held-out data
big picture synthesis and caveats
- increasing N in the training sample beyond N=o(1000) does not result in any predictive gain for “shallow” techniques and group-level pattners (flip side ot marek et al. 2022)
- both large trianing and test samples will show increasing sig. for fixed effect sizes - no real scientific insight
- tension between prediction and explanation (yarkoni paper)

decoding fMRI (dys)connectivity via cross-species fMRI
functional networks - mouse brain - better understanding of mechanisms
- fMRI connectivity difference, 20 different mutations
functional and structural connectivity are closely related
does fMRI hypoconnectivity reflect reduced axonal input?
can you silence activity in one region and then observe hypoconnectivity?
methods: chronic silencing of DMN chronically and actutely (overexpress sodium channel)
- difference: more connectivity when DMN has been chronically silenced (regions that have strong projections from PFC)
less is more: cortical silencing results in fMRI hyperconnectivity
increasing excitability of the mouse PFC with chemogenetics
excitation results in fMRI hypoconnectivity
fMRI connectivity is dominatd by changes in slow/infraslow neural coherence (faster to slow, not fast rhythms)

Deobra Peretti