Metabolic network alterations as a supportive biomarker in dementia with Lewy bodies with preserved dopamine transmission

The framework of the European dementia with Lewy bodies (E-DLB) consortium with conception, design and patient selection has been described previously [22]. Within the imaging arm of this study, all patients with available brain FDG-PET scan and additional DaT-SPECT images were included [18] together with additional datasets acquired at LMU Munich between 02/2018 and 10/2019 resulting in a total of 108 patients with DLB. DLB diagnosis was based on the established criteria [4]. In order to minimalize the risk of misdiagnosis, we only included imaging data of patients with initially both probable and possible DLB diagnosis and with cognitive impairment both in the prodromal and the dementia stage who received confirmation of DLB diagnosis based on clinical follow-up in experienced neurological centres. FDG-PET images of 23 HCs imaged in Munich (n = 9) and Genova (n = 14) and DaT-SPECTs of 37 historical similarly aged HC served as controls [23].

For validation of the PCA expression score at the individual patient level, a cohort of patients with a confirmed diagnosis of either AD, PD or MSA were included. FDG-PET scans of this validation cohort have all been acquired previously at LMU Munich in the same time period [24]. Included subjects were 19 patients with MSA (mean age 63.8), 33 patients with PD (mean age 77.7) and 49 patients with AD (mean age 69.7). All patients with AD had a positive amyloid-PET scan and at least minor perfusion alterations (A+/N+). Diagnoses were made by a team of experienced clinicians based on current diagnosis criteria and confirmed through imaging and laboratory parameters as well as clinical follow-up [25‐28]. Details on healthy controls as well as patients with AD, PD and MSA are provided in Supplemental Table 1.

Patients all gave informed written consent for the diagnostic procedures including radiopharmaceutical applications. Local institutional ethics committees approved the retrospective analyses and transfer of imaging data separately for all centres.

DaT-SPECT and FDG-PET images were acquired and preprocessed as described previously [18]. Details on sites and scanners are provided in Supplemental Table 2. In brief, z-score values against HC were calculated for the DaT-ratio in the bilateral putamen (as defined in Hermes BRASS model 5, Hermes Medical Solutions, Stockholm, Sweden) with the bilateral occipital lobe as reference region. FDG-PET images were spatially normalized and scaled to their global mean for assessment of relative regional glucose metabolism using PMOD (V3.5, PMOD technologies, Basel, Switzerland). This approach delivered robust metabolic connectivity analysis, cross-validated by intensity normalization using the cerebellum and a cluster-based approach [18]. Next, images were smoothed using a Gaussian filter (8mm³) and global mean scaled standardized uptake value ratios (SUVrs) were extracted for 77 predefined cortical and subcortical gray matter VOIs of the Hammers atlas [29]. The whole brain VOI was derived from merging all 77 VOIs. Eight composite regions were defined by summarizing Hammers atlas regions of frontal, parietal, temporal, occipital and insular cortex regions as well as limbic regions, basal ganglia and cerebellum within the PMOD software package.

According to the putaminal DaT-ratio z-scores, patients were categorized into DLB patients with significant dopamine deficiency (≥ 2 standard deviations (SD) below HC, DLB-DaT(+)) and DLB patients without significant dopamine deficiency (< 2SD below HC, DLB-DaT(−)). The average of the putaminal z-scores of both hemispheres served as classifier to categorize subjects in one of both groups. The control cohort was used as implemented in the Hermes software package, using age matched comparison.

In order to evaluate metabolic pattern similarities between DLB with and without significant dopaminergic loss, we conducted a metabolic connectivity analysis as described in our previous work by Huber et al. [18]. By calculation of inter-region correlation coefficients (ICCs, Pearson) for all pairs of global-mean scaled regional FDG-PET regional values (77 × 77 matrix), we performed a group-level metabolic connectivity analysis in each of the three subgroups (HC, Dat(+), DaT(−)), followed by Fisher’s transformation to enhance normal distribution. The difference between the Fisher transformed metabolic connectivity values of DLB-DaT(−) to HC and DLB-DaT(+) to HC of all 77 × 77 VOI pairs was calculated and the resulting indices of regional metabolic connectivity alterations were correlated with each other to analyse the similarities between the metabolic connectivity patterns of the two DLB subgroups. A supplemental voxel-based analysis between both DLB groups and controls was performed as described previously (p < 0.001, uncorrected for multiple comparisons) [18].

For individual subject classification, a PCA was conducted using SPSS 25 statistics (IBM Deutschland GmbH, Ehningen, Germany). Figure 1 illustrates the PCA-based calculation of single subject pattern expression. Following a previously described approach [30], the FDG-PET values of the Hammers VOI grid regions were log transformed and double centred by subtracting the mean of the data per row as well as per column of the subject in order to clear covariance and normalize the data to mean metabolic activity. We extracted components with an Eigenvalue > 1.0 and selected the Varimax rotation. Age and sex were included as covariates. The factor analysis was set to list wise exclusion of cases, suppressing small coefficients with an absolute value below 0.1 and sorting coefficients by size. Twelve principal components (PCs) resulted from the PCA for the training cohort of n = 23 HC and n = 86 DLB-DaT(+) patients. The weighting factors (regression coefficients, β) for each PC were determined by multiple linear regression using the DLB status as outcome variable (Supplemental Table 3). To simulate a clinical scenario, single individual subjects of the study test cohorts consisting of DLB-DaT(−), AD, PD and MSA were added into the PCA to simulate a clinical scenario. Thus, the PCA was performed 123 times each with n = 110 cases (n = 109 training plus one test case). Individual factor values of the test cases were extracted and multiplied with the aforementioned PC weighting factors, followed by subsequent summation to a single expression score per individual subject. To exclude floating of PCA expression scores by inclusion of individual test cases, the expression scores of healthy controls were extracted for each PCA run and subject to a variance analysis. PCA expression scores were then converted into z-scores relative to the HC group to allow for compatibility with future studies and to increase interpretability. A ROC analysis was performed for discrimination of DLB-DaT(−), AD, PD and MSA subjects via the individual PCA expression scores. To confirm presence or absence of systematic differences between scanner types and sites that could bias the PCA expression score, we compared all subgroups divided by centre and scanner type. To this end, we used unpaired t-tests for all PCA expression z-scores and corrected the resulting p-values for multiple testing using false discovery rate (FDR) correction.

Demographics and clinical parameters of the DLB subgroups and comparison between groups are shown in Table 1. The disease duration was defined as the elapsed time between the first symptoms and date of FDG-PET scanning. Among the overall 108 patients with sufficient background and imaging data, 86 subjects showed a significant dopaminergic deficiency, whereas 22 subjects evinced no significant pathologic result in their DaT-SPECT scan (z-score < 2 in DaT-SPECT compared to healthy controls). In the DLB-DaT(+) group, 77% received the diagnosis of probable DLB according to the McKeith criteria. In the DLB-DaT(−) group, the percentage of patients with probable DLB diagnosis was 68%. Rapid eye movement sleep behaviour disorder was less frequent in DLB-DaT(−) when compared to DLB-DaT(+) (17% vs. 44%, p = 0.034).

Compared to HC, DLB-DaT(+) patients showed the expected relative reduction in glucose metabolism in parieto-occipital and frontal cortices (superior, middle, inferior and orbitofrontal), whereas a relative glucose hypermetabolism was observed in motor cortices, the basal ganglia, parts of the limbic system and the cerebellum (Fig. 2A). Precuneus and posterior cingulate cortex showed no significant glucose alterations. DLB-DaT(−) patients expressed similar, overall less pronounced, patterns of relative hypo- and hypermetabolism. Single region global mean scaled FDG-PET SUVr values of all three groups are provided in Supplemental Table 4.

Compared to DLB-DaT(−), the DLB-DaT(+) patients showed a more distinct relative hypometabolism in above-mentioned occipital and basal structures of the brain (Fig. 2B). The insula exhibited a more prominent hypermetabolism for those subjects whose DaT-SPECT showed dopamine deficiency. Effect sizes of single region alterations in DLB vs. controls were strongly correlated between DLB-DaT(−) and DLB-DaT(+) patients (R² = 0.479, p < 0.0001; Fig. 2C).

Metabolic connectivity was analysed in both DLB subgroups (DLB-DaT(+)/DLB-DaT(−)) and compared to the control group (23 subjects) with presumably intact nigrostriatal dopamine innervation (Fig. 3). In line with our recent publication [18], the most prominent increase in metabolic connectivity of DLB patients occurred within the basal ganglia, frontal cortices and limbic system as well as between limbic system and basal ganglia [18]. Although some brain regions indicated a different magnitude of metabolic alteration between DLB-DaT(+) and DLB-DaT(−) patients, the overall regional connectivity changes were strongly associated between DLB-DaT(+) and DLB-DaT(−) patients (R² = 0.597, p < 0.0001; Fig. 4). Single region metabolic connectivity values are provided in Supplemental Table 4.

Given the similarity of metabolic connectivity alterations in DLB-DaT(+) and DLB-DaT(−) group, we hypothesized that a data-driven network analysis could facilitate the identification and discrimination of DLB-DaT(−) patients against other neurodegenerative diseases. In order to implement the findings of this study into clinical scenarios, we conducted a PCA on the basis of regional FDG-PET values of the established cohort. We trained the PCA by DLB-DaT(+) patients and controls (Supplemental Fig. 3A) and derived individual expression scores from individual patients of a test cohort including DLB-DaT(−), AD, MSA and PD. The overall variance of PCA expression scores for controls of the training cohort was low (CoV 16.4%). Significant floating of the PCA by adding additional single subjects was excluded by a robust coefficient of variation (7.4%) in the PCA expression scores of controls. PCA expression z-scores of DLB-DaT(−) patients were higher (6.97 ± 2.29) when compared to PCA expression z-scores of other neurodegenerative diseases (MSA: 0.63 ± 2.10; PD: 1.27 ± 1.40; AD: 2.42 ± 3.58; Fig. 5A). The magnitude of PCA expression scores in DLB-DaT(−) patients was similar compared to DLB-DaT(+) patients (Supplemental Fig. 3B). We then conducted separate ROC analyses for the DLB-DaT(−) scores against the scores of the other conditions. ROC curves indicated excellent discrimination of DLB-DaT(−) patients against the whole cohort of degenerative diseases (AUC: 0.912; Fig. 5B). As expected, discriminatory power was highest for the comparison of DLB-DaT(−) against PD (AUC: 0.995) and MSA (AUC: 0.987), and still at a high level for the comparison of DLB-DaT(−) against AD (0.830). Discrimination of DLB-DaT(−) patients by PCA expression scores was stronger compared to discrimination by regional global mean scaled FDG-PET SUVr values (Supplemental Fig. 4). PCA expression scores of the whole DLB cohort correlated weakly but significant with cognitive screening (MMSE; p = 0.019, R² = 0.073), indicating stronger correlation for the DLB-DaT(+) subjects (p = 0.002, R² = 0.160; Supplemental Fig. 5A). The correlation between DaT availability (putaminal z-score) and PCA expression scores was close to significance for the entire DLB-cohort, driven by a substantial correlation in DLB-DaT(−) patients (all DLB: p = 0.050, R² = 0.036; DLB-DaT(−): p = 0.029, R² = 0.216; Supplemental Fig. 5B). PCA expression scores were not significantly different in presence or absence of clinical core features (all p > 0.200, Supplemental Fig. 5C).