We retrospectively identified patients who presented with PCC between May 2020 and December 2022, satisfied the WHO definition [1] and had a brain 18F-FDG PET scan to investigate suspected brain impairment (fatigue, anosmia/ageusia, cognitive symptoms and/or dysautonomia) at two University Hospitals [4]. This study was approved by the CEMEN (Nuclear Medicine Research Ethics Committee) with the record number CEMEN 2024-04.
These patients were age- and sex-matched with healthy controls who underwent brain 18F-FDG PET before the COVID-19 outbreak with no neuropsychiatric antecedents and normal neuropsychological tests from the two University Hospital local databases (NCT00484523 and NCT03345290).
Brain 18F-FDG PET was performed according to recent guidelines [8] and as previously described [5].
After normalization of images to the Montreal National Institute (MNI) space and smoothing with an 8 mm Gaussian filter, voxel-based group comparisons were performed with SPM 12 (https://www.fil.ion.ucl.ac.uk/spm/software/spm12/) using age, sex and University Hospital centre as covariates, proportional scaling for intensity normalization and an inclusive grey-matter mask.
First, a voxelwise group comparison was performed with SPM between the whole group of patients and the controls (p-voxel value of 0.005, uncorrected, with cluster volume corrected from the familywise error (FWE)). This voxelwise group comparison was then performed between the subgroups of patients equally dichotomized by the median time from COVID-19 infection: less than or more than 9 months (respectively, -9 M and + 9 M). The identified clusters were reported according to the anatomical automatic labelling (AAL) atlas (https://www.gin.cnrs.fr/en/tools/aal/). Second, identified cluster values from the whole group comparison were extracted via Marsbar (https://marsbar-toolbox.github.io/) and used as seeds for IRCA [9] (p-voxel value of 0.005, uncorrected, with cluster volume correction on the basis of Monte Carlo simulations [10]). Third, SICE [11] analyses were also performed. A series of nodes representing brain regions of interest (ROIs) for connectivity analysis were selected (116 ROIs from automated anatomical labelling (AAL) [12] and 3 ROIs from the brainstem) and then extracted via the Marsbar. Using the skggm package [13], the SICE method was applied to the subject-by-node/ROI matrix to estimate the metabolic connectivity matrices for each group (whole population, -9 M, + 9 M and control groups), as detailed elsewhere [11]. The Stability Approach to Regularization Selection (StARS) was applied for optimal density selection [14]. Metabolic connectivity matrices were represented in a brain template via the Nilearn package (https://nilearn.github.io/stable/). A bootstrapping procedure (N = 100) was implemented to rigorously test differences in the number of connections between two groups via multiple independent t tests. An absolute threshold of the T score of 15 was used to highlight significant differences. Finally, a graph theory analysis was carried out with BRAPH (BRain Analysis using graPH theory (http://www.braph.org/)), a software package used to perform graph theory analysis of the brain connectome, to compare the connectivity of the patient and control brains. The SICE matrices for patients and controls previously defined were used as input matrices. The tested global parameters were degree, strength, radius, diameter, eccentricity, path length, global and local efficiency, clustering nodes, transitivity, modularity, assortativity and small-worldness. Nonparametric permutation tests with 100 permutations were carried out to assess differences between groups.
Continuous and categorical variables were compared with the Mann‒Whitney and chi‒square tests, respectively. A p value ≤ 0.05 was considered significant. Statistical analyses were performed in MATLAB version 9.13.0 (R2022b).