Connectome-based predictive modelling estimates individual cognitive status in Parkinson's disease

Introduction: The progressive nature of Parkinson's disease (PD) affords emphasis on accurate early-stage individual-level assessment of risk and intervention appropriateness. In PD, cognitive impairment (CI) may follow or precede motor symptoms but are generally underdetected. In addition to impeding daily functioning and quality of life, CIs increase the risk for later conversion to dementia, providing a pressing need to develop novel tools to detect and interpret them. Connectome-based predictive modelling (CPM) is an emerging machine-learning approach to individual prediction that holds translational promise due to its noninvasiveness and simple implementation. The aim of this study was to investigate CPM's potential to predict and understand CIs in PD. Methods: Resting-state functional connectivity from 58 patients with PD of varying cognitive status was used to train a CPM-model to predict a global cognitive composite (GCC) score. The model was validated using crossvalidation, permutation testing, and internal stability analyses. The combined predictive strength of two brain connectivity networks, positive and negative, directly and inversely correlated with GCC, respectively, was assessed. Results: The model significantly predicted individual GCC scores, r = 0.63, pperm < .05. Separately, the positive and negative networks were similar in performance, rs >= .58, ps < .05, but varied in anatomical distribution. Conclusions: This study identified a connectome predictive of cognitive scores in PD, with features overlapping with established and emerging evidence on aberrant connectivity in PD-related CIs. Overall, CPM appears promising for clinical translation in this population, but longitudinal studies with out-of-sample validation are needed.