Deep learning-based cognitive signature of brain FDG PET associated with outcome of rapid eye movement sleep behavior disorder

Objectives: As rapid eye movement (REM) sleep behavior disorder (RBD) is frequently associated with cognitive decline at risk of progression to neurodegenerative disorders, an objective biomarker evaluating cognitive dysfunction is crucial. To this end, we evaluated deep learning (DL)-based cognitive signature of [¹⁸F]fluorodeoxyglucose (FDG) brain positron emission tomography (PET) correlated with clinical features of RBD.

Methods: Baseline FDG PET data acquired from prospectively enrolled patients with RBD and controls were analyzed between Jun 2017 and Dec 2019. A DL-based cognitive signature on FDG PET trained for differentiating Alzheimer’s disease from normal controls using Alzheimer’s Disease Neuroimaging Initiative database was used. The model was transferred to FDG PET data of patients with RBD with mild cognitive impairment (MCI) (RBD-MCI) (n=19) and RBD without MCI (RBD-nonMCI) (n=31). The DL-based cognitive dysfunction scores of RBD-MCI and RBD-nonMCI were compared using the Mann-Whitney test. The accuracy of differentiating RBD-MCI from RBD-nonMCI was evaluated by area under curve (AUC) of receiver operating characteristic (ROC) analysis. The DL-based cognitive dysfunction score at the baseline of RBD patients with and without the CERAD score decline at 2-year follow-up was compared using the Mann-Whitney test. In addition, we produced class activation map to visualize the area most indicative of the cognitive dysfunction of individual RBD patients.

Results: The DL-based cognitive dysfunction score was significantly higher in RBD-MCI than RBD-nonMCI (-1.157 ± 2.126 vs. -0.202 ± 1.526, p=0.018). The AUC of ROC curve for differentiating RBD-MCI from RBD-nonMCI was 0.70 (95% CI 0.56 to 0.82). The DL-based cognitive dysfunction score at the baseline was significantly higher in RBD patients who showed decreased CERAD scores during 2 years than those who did not (0.947 ± 1.989 vs -1.000 ± 1.480, p = 0.031). In addition, brain metabolic features related to the cognitive dysfunction-related regions of individual RBD patients mainly included posterior cortical regions.

Conclusions: The DL-based cognitive signature could be used to objectively evaluate cognitive dysfunction in RBD. We suggest that this approach could be extended to an objective biomarker reflecting neurodegeneration in RBD in terms of the cognitive domain.