Change point based dynamic functional connectivity estimation outperforms sliding window and static estimation for classification of early mild cognitive impairment in resting-state fMRI

The most widely used inputs in classification models of brain disorders such as early mild cognitive impairment (eMCI) or Alzheimer’s disease are estimates of static-based functional connectivity (SFC) and sliding window dynamic functional connectivity (swDFC). Although these methods are convenient for estimation and computational purposes, as it keeps the estimation tractable, they present a simplified version of a highly integrated and dynamic phenomenon. Change point dynamic functional connectivity (cpDFC) methods, which are far less commonly used, offer an alternative to swDFC approaches. In this study, we consider a classification task between controls and patients with eMCI using resting-state functional magnetic resonance imaging (fMRI) data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) studies, ADNI2 and ADNIGO. Our results indicate that the DFC methods are generally superior to the SFC methods when used as inputs into the classification model. Most importantly, we find that the cpDFC methods are generally superior to the widely used swDFC methods. We discuss how cpDFC methods offer many distinct advantages over swDFC methods, namely, the parsimony of network features and ease of interpretability. We validate the robustness and consistency of our results by testing the methods on an additional resting-state fMRI dataset of mild cognitive impairment patients. These findings call into question the validity of numerous fMRI studies that have utilized inferior approaches, such as SFC and swDFC, as inputs to classification models to predict various brain disorders. Finally, we present an ensemble model of the best models, which achieves an accuracy of 91.17% from leave-one-out cross-validation of subjects with eMCI. Our results suggest that the underlying functional networks are dynamic, multiscale, and that different FC methods capture distinct information for classification efficacy.