Abstract
Recent advances with functional connectivity magnetic resonance imaging have demonstrated that at rest the brain exhibits coherent activity within a number of spatially independent maps, normally called 'intrinsic' or 'resting state' networks. These networks support cognition and behaviour, and are altered in neurodegenerative disease. However, there is a longstanding perspective, and ample functional magnetic resonance imaging evidence, demonstrating that intrinsic networks may be fractionated and that cortical elements may participate in multiple intrinsic networks at different times, dynamically changing alliances to adapt to cognitive demands. A method to probe the fine-grained spatiotemporal structure of networks may be more sensitive to subtle network changes that accompany heterogeneous cognitive deficits caused by a neurodegenerative disease such as Parkinson's disease. Here we tested the hypothesis that alterations to the latent (hidden) structure of intrinsic networks may reveal the impact of underlying pathophysiologic processes as assessed with cerebrospinal fluid biomarkers. Using a novel modelling approach that we call 'network kernel analysis', we compared fine-grained network ensembles (network kernels) that include overlapping cortical elements in 24 patients with Parkinson's disease (ages 45-86, 17 male) and normal cognition or mild cognitive impairment (n = 13), and 21 cognitively normal control subjects (ages 41-76, nine male). An omnibus measure of network disruption, calculated from correlations among network kernels, was correlated with cerebrospinal fluid biomarkers of pathophysiological processes in Parkinson's disease: concentrations of α-synuclein and amyloid-β42. Correlations among network kernels more accurately classified Parkinson's disease from controls than other functional neuroimaging measures. Inspection of the spatial maps related to the default mode network and a frontoparietal task control network kernel showed that the right insula, an area implicated in network shifting and associated with cognitive impairment in Parkinson's disease, was more highly correlated with both these networks in Parkinson's disease than in controls. In Parkinson's disease, increased correlation of the insula with the default mode network was related to lower attentional accuracy. We demonstrated that in an omnibus sense, correlations among network kernels describe biological impact of pathophysiological processes (through correlation with cerebrospinal fluid biomarkers) and clinical status (by classification of patient group). At a greater level of detail, we demonstrate aberrant involvement of the insula in the default mode network and the frontal frontoparietal task control network kernel. Network kernel analysis holds promise as a sensitive method for detecting biologically and clinical relevant changes to specific networks that support cognition and are impaired in Parkinson's disease.