Introduction
Parkinson’s disease (PD) is a common neurodegenerative disorder, for which only symptomatic therapies are available. Efforts to develop neuroprotective or preventive treatments will benefit from a reliable biomarker. Ideally, such a biomarker can identify PD in its early stages, differentiate between PD and other neurodegenerative parkinsonian disorders, track disease progression, and quantify treatment effects.
In PD, abnormal accumulation of α-synuclein in neurons impairs synaptic signaling, causing disintegration of specific neural networks [
1]. Neuro-imaging with [
18F]-fluorodeoxyglucose positron emission tomography (
18F-FDG PET) can capture synaptic dysfunction in vivo. The radiotracer
18F-FDG provides an index for the cerebral metabolic rate of glucose, which is strongly associated with neuronal activity and synaptic integrity [
2].
Brain regions with altered
18F-FDG uptake in PD have been identified with univariate group comparisons using Statistical Parametric Mapping (SPM) [
3‐
7]. However, because metabolic activity is correlated in functionally interconnected brain regions, analysis of covariance is more suitable to assess whole-brain networks. Multivariate disease-related patterns can be identified with the Scaled Subprofile Model and Principal Component Analysis (SSM PCA). Subsequently, a disease-related pattern can be used to quantify the
18F-FDG PET scans of new subjects [
8‐
10]. In this procedure, an individual’s scan is projected onto the pattern, resulting in a subject score. This is a single numeric value which reflects the degree of pattern expression in that individual’s scan.
The PD-related pattern (PDRP) was initially identified by Eidelberg et al. with SSM PCA in 33 healthy controls and 33 PD patients from the USA [
11]. This PDRP (PDRP
USA) has served as a reference in many consecutive studies [
12]. The PDRP
USA is characterized by relatively increased metabolism of the thalamus, globus pallidus/putamen, cerebellum and pons, and by relative hypometabolism of the occipital, temporal, parietal, and frontal cortices. PDRP
USA subject scores were significantly correlated with motor symptoms and presynaptic dopaminergic deficits in the posterior striatum [
13], increased with disease progression [
14], and were shown to decrease after effective treatment [
15,
16]. PDRP
USA was significantly expressed in patients with idiopathic REM sleep behavior disorder (iRBD), a well-known prodromal disease stage of PD [
17], and could discriminate between healthy controls, PD, and patients with multiple system atrophy (MSA) [
18,
19].
Because of these properties, PDRP
USA is considered a neuro-imaging biomarker for PD [
12]. It is essential that the PDRP is thoroughly validated. In collaboration with Eidelberg et al., PDRPs were identified in independent American, Indian, Chinese, and Slovenian populations [
11,
15,
20,
21]. Independently from these authors, the PDRP was recently derived in an Israeli population [
22]. These PDRPs were highly similar to the PDRP
USA, although minor deviations in PDRP regional topography were observed in several of these studies, which may be caused by differences in demographics or clinical characteristics of the cohorts.
We previously identified a PDRP in a retrospective cohort of PD patients scanned on dopaminergic medication [
23], and subsequently in an independent cohort of prospectively included PD patients who were in the off-state (PDRP
NL) [
24]. We used code written in-house, and obtained similar results compared with other PDRP studies. Recently, we demonstrated significant expression of the PDRP
NL in idiopathic REM sleep behavior disorder (a prodromal stage of PD), PD, and dementia with Lewy bodies [
25]. However, the PDRP
NL has not been validated in a larger cohort, and correlations with PDRP
USA were not explored.
The aim of the current study was to validate the PDRPNL in several independent cohorts. We were able to test the PDRPNL in 19 controls and 20 PD patients from our own clinic in the Netherlands, in 44 healthy controls and 38 “de-novo” PD patients from Italy, and 19 healthy controls and 49 late-stage PD patients from Spain. In addition, we newly identified a PDRP in Italian and Spanish datasets and performed a cross-validation between these populations. We compared the three PDRPs to the reference pattern (PDRPUSA).
Discussion
In this study, we cross-validated the previously published PDRP
NL [
24], and additionally identified a PDRP in an Italian (PDRP
IT) and Spanish (PDRP
SP) sample. The three patterns were akin to PDRP
USA, and also to the PDRP described in other populations [
20,
21]. Topographical similarity to PDRP
USA was confirmed for each of the three PDRPs by a significant correlation of region weights, and a significant correlation in subject scores. Furthermore, PDRP
NL, PDRP
IT, and PDRP
SP were significantly expressed in PD patients compared with controls in both identification and validation cohorts, but were not significantly expressed in MSA-p patients.
The typical PDRP topography is characterized by relative hypermetabolism in the thalamus, putamen/pallidum, pons, cerebellum, and motor cortex. These changes co-vary with relatively decreased metabolism in the prefrontal, parietal, temporal, and occipital cortices [
11,
15,
20,
21,
23,
24]. This topography is thought to reflect changes in cortico-striatopallido-thalamocortical (CSPTC) loops and related pathways in PD [
33,
34]. In these circuits, a dopaminergic deficit leads to abnormal basal ganglia output, mediated by hyperactivity of the subthalamic nucleus (STN) and its efferent projections. Several studies support a direct relationship between altered STN output and the PDRP topography [
16,
35‐
38].
The high degree of similarity in PDRP topography across samples is striking considering differences in demographics, clinical characteristics, scanning methods, and reconstruction algorithms. Especially the PDRPNL was highly similar to the reference pattern (PDRPUSA). These two patterns showed the highest subject score correlation and region weight correlation. Furthermore, the PDRPNL achieved the highest AUC in the other cohorts. Like PDRPUSA, PDRPNL was derived in a cohort of off-state patients with a wide range of disease durations (duration 4.4 ± 3.2 years; range 1.5–11.5 years) and severities.
Deviations from the typical PDRP topography are worth exploring further in relation to clinical characteristics. The PDRP
IT is unique in that it is, to our knowledge, the first time the PDRP has been derived in “de-novo,” treatment-naïve PD patients. It is likely that these very early-stage patients have a less severe nigrostriatal dopaminergic deficit compared with the more advanced PD patients in PDRP
USA, PDRP
NL, and PDRP
SP derivation cohorts. This may be reflected by less severe involvement of the frontal cortex in PDRP
IT, as nigrostriatal denervation is known to be positively correlated with hypometabolism in the frontal cortex [
13,
39].
The PDRP
SP was derived in PD patients who were scanned while being on dopaminergic medication. Levodopa is known to decrease metabolism in the cerebellar vermis, putamen/pallidum, and sensorimotor cortex. Levodopa therapy can reduce PDRP expression, but does not completely correct the underlying network abnormalities [
16]. It can be assumed that the effect of dopaminergic therapy on PDRP expression is modest in comparison with the effect of disease progression [
40]. Indeed, the typical PDRP topography could still be identified in these
on-state patients. However, the PDRP
SP did not correlate as well to the other patterns, both in terms of subject scores and region weights. It is not clear whether this is related to the advanced disease stage or the effect of treatment. The PDRP
SP was characterized by more stable involvement of the occipital cortex, possibly related to the presence of mild cognitive impairment and visual hallucinations, which often occur in advanced PD [
41].
Following from the above, it can be concluded that the typical PDRP topography is highly reproducible. Similar topographies have also been identified in studies comparing
18F-FDG-PET scans of healthy controls and PD patients with SPM [
3‐
7]. Such analyses can be supportive in the visual assessment of an
18F-FDG-PET scan in clinical practice. Several studies have evaluated the diagnostic value of observer-dependent visual reads supported by SPM-based comparisons to healthy controls [
3,
4,
42‐
44]. A recent meta-analysis (PD versus “atypical” parkinsonism) estimated a pooled sensitivity of 91.4% and specificity of 94.7% for this semi-quantitative approach [
45].
The merit of SSM PCA over mass-univariate approaches lies in its ability to objectively quantify
18F-FDG PET scans of patients using the pre-defined patterns. Pattern expression scores were shown useful in differential diagnosis, tracking disease progression, and estimating treatment effects [
46]. Although in the current study PDRP
z-scores were significantly higher in PD patients compared with healthy controls, there was a considerable overlap in PDRP
z-scores between patients and controls in almost every cohort. This overlap is not unique to the current data, and is also apparent in other studies [
12].
Some healthy controls appear to express the PDRP. Since we found significant correlations between PDRP
z-scores and age in healthy controls, it could be suggested that ageing and PD share certain metabolic features. Metabolic decreases have been reported in the parietal cortex in normal aging [
47,
48]. This may cause some overlap with the PDRP. However, the correlation with age in our study was not consistent across all datasets and patterns (Table
5). Furthermore, expression of an age-related spatial covariance pattern was shown to be independent from PDRP expression [
49,
50]. Alternatively, a high PDRP
z-score in a healthy subject could signal a prodromal stage of neurodegeneration. For instance, subjects with idiopathic REM sleep behavior disorder (a prodromal stage of PD) were shown to express the PDRP years before onset of clinical parkinsonism [
17,
25].
Low PDRP
z-scores in PD patients could indicate inaccurate clinical diagnosis. A recent meta-analysis of clinicopathologic studies suggests that the clinical diagnosis of PD by an expert, after an adequate follow-up, has a sensitivity of 81.3% and a specificity of 83.5% [
51]. Thus, even under ideal circumstances, the diagnosis is inaccurate in a number of patients.
Overlap in pattern expression scores is not only apparent between controls and PD patients, but also between patients with different parkinsonian disorders. For instance, the PDRP may also be expressed in patients with progressive supranuclear palsy (PSP) [
52]. This means that the expression score for a single disease-related pattern is inadequate to differentiate between multiple disorders. However, this does not hamper application in differential diagnosis. Previous studies have shown that an algorithm combining multiple disease-related patterns (including the PDRP) with logistic regression could accurately distinguish between parkinsonian disorders. With this method, Tang et al
. achieved accurate categorization of patients (
n = 167) with an uncertain diagnosis 3–4 years before a final clinical diagnosis was made by an expert clinician masked to the imaging findings [
18]. Highly similar results were obtained in an independent sample (
n = 129) [
19].
In this study, we compared data from different centers. It is well-known that variations in PET scanners and image reconstruction algorithms influence disease-related pattern scores [
53‐
55] (supplementary Fig
1). In support of this, we recently identified clear center-specific features in the current data using machine-learning algorithms [
56]. Therefore, PDRP subject scores cannot be compared readily between subjects from different centers. In all PDRP studies, this is solved with a
z-transformation using the mean and standard deviation of a small healthy control group. This potentially introduces a bias, depending on which controls are selected. However, this issue is not relevant for
within subject studies. Therefore, PDRP subject scores may be especially useful in tracking disease progression [
14], or treatment effects [
16,
35‐
38].
This study is methodologically different from previous PDRP studies. The PDRPs identified in this study were formed by a combination of principal components (PCs). These combinations were determined based on a forward stepwise logistic regression model [
30]. There are different methods to decide which components are included in the disease-related pattern [
10]. Previous studies have always identified the PDRP as PC1 in isolation [
11,
20,
21]. The process of component selection is not always described in detail. Automatically choosing PC1 as the disease-related pattern, and disregarding consecutive, smaller PCs, increases the risk information loss. On the other hand, a pattern that combines multiple PCs may give a better fit of the initial sample, but may be limited in its relevance or generality across new datasets. Therefore, we re-evaluated the data and included only PC1 for PDRP
IT, PDRP
NL, and PDRP
SP. Indeed, the PC1 patterns correlated better to the reference pattern (PDRP
USA). However, the patterns that included multiple PCs yielded higher diagnostic accuracy . Apart from component selection, several other decisions and cutoffs may influence pattern identification [
10]. More advanced machine-learning algorithms may be of use in determining optimal patterns without the use of arbitrary thresholds and associated loss of potentially useful information [
55‐
58].
There is increasing interest to apply the PDRP in clinical practice and in therapeutic trials [
12]. However, rigorous validation by independent research groups is necessary before widespread application. The current study has contributed to the finding that the PDRP is a universal feature of PD, and it is striking that such similar patterns could be identified in a limited number of
18F-FDG PET scans from three populations, despite overt clinical and methodological heterogeneity. However, our results also show considerable overlap in PDRP subject scores between control and PD groups. Further study is needed to overcome this issue, perhaps by addressing potential center-specific effects in the data or by employing more advanced machine-learning algorithms.
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.