Introduction
Parkinson Disease (PD) is the most common neurodegenerative movement disorder in the general population and accounts for 65% of all parkinsonian syndromes (PS) [
1]. It is characterized by progressive akinesia in association with rest tremor and muscular rigidity as well as an excellent response to levodopa therapy [
2]. Clinical diagnosis is not always straightforward and distinction from atypical parkinsonian syndromes (APS) can be particularly challenging in the early phase. In fact, autopsy-based studies have shown that diagnostic accuracy for PD does not exceed 79.6% even when clinical diagnosis is performed by movement disorders experts, or 82.7% when using stringent clinical criteria for PD [
3]. Misclassification involves other degenerative conditions—i.e., multiple system atrophy (MSA), progressive supranuclear palsy (PSP) and corticobasal degeneration (CBD). Compared to PD, APS are characterized by additional clinical features besides parkinsonism (cerebellar or pyramidal signs, postural instability or apraxia), a moderate or transient response to levodopa and a worse prognosis with a mean life expectancy of 7–10 years from disease onset [
4,
5]. The issue of PD clinical misdiagnosis has virtually not improved over the last 25 years, and various biomarkers, notably structural and metabolic imaging, have been examined as potential aids to clinical diagnosis and, to a lesser extent, to separate the different forms of degenerative PS [
6]. As yet, however, only a few of the proposed biomarkers have proved robust enough to be included in the new MDS criteria for PD—olfactory loss and cardiac sympathetic denervation as supportive criteria, and a normal presynaptic dopaminergic system as exclusion criterion [
2].
Presynaptic dopaminergic pathways integrity can be assessed with various compounds,
123I-FP-CIT (ioflupane) being the most widely used due to its fast kinetics, selectivity to dopamine transporters (DaT), SPECT scan availability, and its binding not being affected by concurrent dopaminergic treatment [
7]. While visual and semiquantitative assessments have shown excellent sensitivity and specificity (resp. 97 and 100%) in differentiating early clinically diagnosed PD subjects (showing a reduced striatal uptake) from non-degenerative cases, similar observations could not be drawn for the distinction of PD and APS [
8‐
11]. In fact, there is a large consensus among experts that DaT SPECT is not able to separate PD from APS, although this assumption has not been extensively studied. In fact, recent publications have shown data partly challenging this consensus. At the group level, PD subjects usually harbor an asymmetric rostrocaudal progression of DaT denervation, with a preservation of caudate nucleus uptake in the early phase, whereas MSA-parkinsonism (MSA-P) and PSP usually show a more global uptake impairment [
12]. In patients with corticobasal syndrome (CBS), a highly asymmetric and moderate uptake alteration has been described [
13], with several cases presenting a virtually normal SPECT [
8,
14]. However, individual detection of subjects with degenerative PS based on molecular imaging remains elusive.
Support vector machine (SVM) is a widely used pattern-recognition method that learns to assign labels to feature vectors [
15] and has been used increasingly for automated classification in a variety of medical applications (for a review, see [
16]). SVMs have been applied to distinguish PD from controls (CTL) or APS based on usually complex and time-consuming processing of various imaging techniques—e.g., DaT SPECT, postsynaptic dopamine PET, diffusion tensor imaging (DTI) or voxel-based morphometry (VBM) MRI imaging with accuracies ranging from 70 to 100% [
17‐
25]. Thanks to multivariate pattern-recognition analyses of
123I-FP-CIT SPECT, we recently observed a differential pattern of uptake alteration between PD and APS; i.e., a more severe caudate nucleus impairment in MSA and PSP, and a relative preservation of putaminal uptake in CBS compared to PD [
26].
We here propose to reappraise the issue of distinguishing the various forms of degenerative PS by means of an SVM classification solely based on semiquantitative 123I-FP-CIT SPECT striatal parameters, the most commonly used output to support diagnosis in clinical practice, obtained from a large, single-center cohort of subjects with idiopathic PD, MSA-P, PSP and CBS in the early stage of disease (< 3 years) and with the same SPECT protocol.
Discussion
We here presented an SVM analysis aiming at separating degenerative parkinsonisms strictly based on their semiquantitative 123I-FP-CIT SPECT uptake values and related combined parameters. This analysis was tested on highly homogeneous data collected from a large, single-center cohort of subjects with well-characterized PS scanned with identical acquisition and processing imaging protocol.
Univariate statistics of semiquantitative evaluation confirm previous findings [
33] of significantly decreased uptake in all forms of PS compared to subjects with non-degenerative conditions, namely lower striatal uptake, as well as higher AI and C/P ratio.
Thanks to previously established local age-dependent reference limits for striatal VOIs uptake, AIs and C/P ratio [
28], we also confirmed that CBS subjects have a relative preservation of presynaptic dopamine transporters as they exhibited higher P uptake in comparison to PD (
p < 0.005) and higher P, C and S uptake compared to MSA-P (
p < 0.0005) and PSP (
p < 0.001). In addition, we found significantly higher S-AI in the CBS group in comparison to PSP and lower C/P ratio than PD, MSA-P and PSP (all
p < 0.005) [
34]. PD subjects had an intermediate degree of striatal impairment with higher uptake ratio in all striatal VOIs compared to MSA-P and PSP. These results are in keeping with previous works [
35].
We here report a striking SVM classification accuracy (92.9%, AUC 0.97) in disentangling PS from CTL, with a major contribution of both uptake and asymmetry parameters. Previous SVM studies have already attempted to separate PD from CTL subjects using
123I-FP-CIT SPECT striatal uptake and length/volume [
36] or complex SPECT and biological data (including serum and CSF) from the Parkinson Progressive Markers Initiative cohort [
19] with Acc 96–97%.
Furthermore, binary classification of each combination of PS allowed Acc 62.9–83.7%, with the best results obtained when separating CBS from each other PS. Considering the discussion above, these findings are not surprising, as it is well recognized that CBS harbors a specific pattern of impairment—i.e., relative striatal uptake preservation and moderate-to-high AI, especially in comparison to MSA-P (Acc in the present study 83.7%) and PSP (Acc 73.9%), which in turn exhibit a more severe and symmetrical impairment of striatal uptake. Conversely, we observed an intermediate level of uptake and AI in PD, which could explain why head-to-head classification accuracy of PD vs other PS is lower (63.5–73.1%).
In previous studies, Haller et al. observed 97% accuracy for the classification of PD subjects vs. a heterogeneous group of other PS, including mainly MSA, but also vascular parkinsonism, dementia with Lewy bodies and psychogenic parkinsonism, using fractional anisotropy DTI. Thanks to MRI atlas-based volumetry, Huppertz et al. found balanced accuracies superior to 80% for the distinction of PD from PSP and MSA, whereas a VBM study of Focke [
17] found balanced accuracies superior to 80% for the distinction of PD from PSP and MSA, whereas a VBM study of Focke et al. showed 87.1% and 71.9% accuracies in separating PD from PSP and MSA, respectively [
22]. Cherubini et al. were able to obtain 100% correct classification in classifying PD and PSP based on white matter atrophy. In addition, postsynaptic D2/D3 [
20] were able to obtain 100% correct classification in classifying PD and PSP based on white matter atrophy. Moreover, postsynaptic D2/D3
18F-desmethoxyfallypride (DMFP) PET SVM has shown 70–75% accuracy for the distinction of PD from MSA and PSP) [
18]. Although these efforts are commendable, they also have inherent limitations—e.g., small sample sizes [
20,
22], the debatable choice of merging several APS groups together [
18], the presence of significant gray matter changes only in patients with long disease duration, when the added value as compared with clinical evaluation is limited [
17] or the need for complex and time-consuming processing [
19] that is not compatible with a clinical routine application.
As the diagnosis of PD and APS can be challenging when solely based on clinical evaluation, several functional imaging ligands have been developed to improve diagnostic accuracy. Using multimodal imaging, i.e., presynaptic dopamine SPECT imaging, postsynaptic raclopride PET or IBZM SPECT, and myocardial
123I-metaiodobenzylguanidine (MIBG) scintigraphy can be useful in distinguishing PD from APS, especially in atypical presentations and early cases. However, this requires performing several scans and exposing the patients to a significant amount of radiation, not to mention major costs for the medical facility. With the present study, we have been able to show accurate classification of PD and APS solely based on striatal semiquantitative DaT SPECT, a molecular imaging technique widely available in clinical practice and the only approved one by both 6FDA and European medical agencies for the distinction of PS from non-degenerative forms of parkinsonism and tremor. The tentative flowchart proposed on Fig.
4, although simplified and subject to many exceptions, may guide clinicians in the differential diagnosis of PS and non-degenerative forms of parkinsonism.
The strengths of our study comprise the large cohort of subjects with PS (including almost 300 subjects with PD and 90 with APS) who were scanned at a single-center level with the same acquisition and processing SPECT protocol. In addition, the semiquantitative parameters used for SVM classification are easily obtainable by any nuclear medicine facility. This could ensure replication of the present results in other centers and potential daily utilization of the highest SPECT predictors in helping to disentangle the various forms of PS. Our study also has limitations. First, diagnosis is based on clinical diagnostic criteria and most cases have not been confirmed by autopsy. However, we are confident that the retrospective nature of the study, the assessment of DaT SPECT by two unrelated visual and semiquantitative methods and a thorough follow-up by neurologists specialized in movement disorders (mean follow-up 4.4 years) should have ensured a low rate of diagnostic misattribution. Second, our CTL group is not based on healthy asymptomatic subjects, but on patients with clinically diagnosed non-degenerative forms of parkinsonism or tremor (mainly ET, DIP and PP). Again, the retrospective design of the study and the long follow-up for most cases allowed us to include only subjects whose condition did not convert into any neurodegenerative process over time and who presented a normal DaT SPECT. In addition, we were not able to obtain a clinical rating scale of motor severity (e.g., MDS-UPDRS-III) for every included subject. Using such a clinical scale as a covariate (instead of disease duration) would possibly have allowed an even better performance of SVM-based discrimination between PD and atypical conditions. Finally, the same SPECT machine and gamma camera have been used for all subjects included in the present study. This certainly helped in obtaining homogeneous uptake values. However, despite 6-month periodic quality controls ensured that scanner sensitivity was globally stable, we cannot fully exclude that time slightly affected the image quality.
In conclusion, our results indicate that semiquantitative striatal 123I-FP-CIT SPECT assessment provides a promising approach to distinguish reasonably well CTL, PD and APS, and that, in combination with SVM, a satisfactory classification can be obtained at the individual level. SVM and other computer-aided classification systems represent a valuable tool to assist the clinician’s daily evaluation of patients with PS.