Latent Similarity Identifies Important Functional Connections for Phenotype Prediction

Abstract

<p>We present a new machine learning algorithm, Latent Similarity, and use it to predict subject (endo)phenotypes from fMRI data. fMRI can be used to predict dysfunctional mental states. In addition, endophenotypes are known to be predictive of disease status, and are of interest in developmental studies. However, fMRI studies often suffer from small cohort size and high feature dimensionality, making reproducible prediction challenging. The innovation of our algorithm is to combine a kernel similarity function with metric learning to increase the robustness of prediction. Our algorithm becomes robust by utilizing the N squared connections between the N subjects in the cohort, rather than the features of the N subjects themselves. We identify important functional connections in the default mode and uncategorized functional networks for predicting age, sex, and intelligence. We also find that only a few connections contain most of the information required for any predictive task. We believe that our algorithm can be beneficial in small sample size, high noise, high dimensionality settings.</p>

Somatomotor-visual resting state functional connectivity increases after 2 years in the UK Biobank longitudinal cohort

“…Preprocessing pipeline for converting 4D fMRI volumes into FC using the Power264 atlas 34 . Reproduced with permission from Orlichenko et al 35 …”

“…1 Preprocessing pipeline for converting 4D fMRI volumes into FC using the Power264 atlas. 34 Reproduced with permission from Orlichenko et al 35 the prediction of the scan order was carried out using the average connectivity among the Power264 networks, in which each network consisted of many individual ROIs. As before, logistic regression with 20 bootstrap repetitions and C ¼ 1, with 2000 subjects in the training set and the rest in the test set, was used for this purpose.…”