Non-linearity matters: a deep learning solution to the generalization of hidden brain patterns across population cohorts

Finding an interpretable and compact representation of complex neuroimage data can be extremely useful for understanding brain behavioral mapping and hence for explaining the biological underpinnings of mental disorders. Hand-crafted representations, as well as linear transformations, may not accurately reflect the significant variability across individuals. Here, we applied a data-driven approach to learn interpretable and generalizable latent representations that link cognition with underlying brain systems; we applied a three-dimensional autoencoder to two large-scale datasets to find an interpretable latent representation of high dimensional task fMRI image data. This representation also accounts for demographic characteristics, achieved by solving a joint optimization problem that simultaneously reconstructs the data and predicts clinical or demographic variables. We then applied normative modeling to the latent variables to define summary statistics ( latent indices) to find a multivariate mapping to non-imaging measures. We trained our model with multi-task fMRI data derived from the Human Connectome Project (HCP) that provides whole-brain coverage across a range of cognitive tasks. Next, in a transfer learning setting, we tested the generalization of our latent space on UK Biobank data as an independent dataset. Our model showed high performance in terms of age and predictions and was capable of capturing complex behavioral characteristics and preserving the individualized variabilities using a highly interpretable latent representation.