Center effects significantly limit the generalizability of brain imaging-based biomarker candidates. Although our previously published resting state functional connectivity-based predictive signature for pain sensitivity (the RPN-signature) showed remarkable out-of-center generalizability, it remained unclear which connectivity features are the most generalizable across study centers.
Here, we re-trained the RPN-signature on multi-center data and found that it outperforms the single-center model in all three centers (explained variance: 26-38% vs. 16%-19%). Our results highlight that neurobiological interpretation of feature importance in predictive modelling is constrained both by center-specific artifacts and by certain characteristics (e.g. regularization) of the applied machine learning algorithm.
This abstract and the presentation materials are available to members only; a login is required.