Meta-analysis of the literature on diagnostic accuracy of SPECT in parkinsonian syndromes

Meta-analysis was done according to current methodological recommendations [22‐24]. We searched MEDLINE using the following terms: PD, parkinsonian, MSA, PSP, VP and ET. We searched for MeSH terms and free text words. All in combination with SPECT and clinical trial. No beginning data limit was used. The search was updated until 9 January 2006. Only English-, Dutch- and German language studies were considered, because the investigators were familiar with these languages. The bibliographies of selected articles were screened for potentially suitable references which were then retrieved. We also searched the EMBASE and Cochrane database (Wiley InterScience) using the same search strategy.

Two investigators (AV, WW) screened the full text of potential relevant articles using the inclusion criteria. For this we use a standard form combined with a modified QUADAS score, see table 1 (form available upon request) [25]. In all cases the investigators reached consensus. Studies were selected if the subject of the study was in one of the following three areas:

1. Patients who underwent SPECT because of diagnostic uncertainty.

2. Cross-sectional study of already diagnosed patient categories, in which SPECT was tested as a means to differentiate between various parkinsonian syndromes.

3. Cross-sectional studies with patients with known PD in an early stage (Hoehn & Yahr stage 2 or less) vs. normal healthy controls, in which SPECT was tested as a means to provide an early diagnosis.

The following exclusion criteria were used: 1) whole article not available, 2) language different from English, German or Dutch, 3) studies including only advanced PD patients vs. healthy controls, and studies with other main categories, e.g. dementia, 4) study population with less than 10 patients, 5) if the numbers of true positives, false negatives, true negatives and false positives with a cut-off point of 2 standard deviations (SD) from the mean of the control group were not available or could not be derived the study was excluded.

When the study included more than 85 patients we contacted the corresponding author to ask for the raw data (see below).

We chose this approach, because we expected a substantial cut-off point effect in the included studies. We did not want to be dependent upon the assumption that the diagnostic odds ratio's in our study would not be affected by differences in the individual cut-off points. To reduce heterogeneity we decided to choose one common cut-off point for all studies. We took a cut-off point of two standard deviations (SD) after consulting with nuclear imaging experts in our hospital and the University Hospital of Amsterdam. Both departments use a cut-off point of 2 SD's below healthy controls. So we recalculated all results from all studies using the individual data from tables and figures in the published paper, using this new cut-off point. If recalculation was not possible (when data for individual patients were not traceable from the manuscript), we excluded the study. This exclusion leads to bias, of course. We feel, however, that, as these studies did not adhere to recommended guidelines by not providing the raw scan results to allow the construction of the diagnostic 2 × 2 table, we did not exclude the methodologically best studies [24].

Sensitivity, specificity and the odds ratio was calculated for each study separately, and the pooled odds ratio's (OR) for all studies together. Although we tried to reduce heterogeneity by recalculating study results using one common cut-off point, we still expected a threshold effect, because of differences in patients, SPECT machinery, radiotracers etc. Therefore, according to recommendations by Deekes and Egger we used diagnostic OR's [26].

For studies with zeroes in one or more cells 0.5 was added to all four cells of the 2 × 2 table. Trials with a sensitivity of 100% and a specificity of 0% were not excluded, however the pooled OR's were also calculated without such studies (See # in Figure 4, 5 and 6).

All results were put in software SPSS 11.0 for Windows and later converted to Stata/SE9.

The metan and metareg commands were built in Stata/SE9. Because of the heterogeneity of the selected studies we used a random model to calculate the diagnostic odds ratios. Heterogeneity was calculated with the I ² [27].

The search on Medline (SPECT & clinical trial) gave 1503 hits. When we added all parkinsonian disorders we limited the Medline search to 56 relevant hits.

In the Cochrane database we found 26 hits and in Embase 45 hits, but no additional clinical studies above the ones found in Medline. With cross-reference searching we found an additional 128 relevant trials (See additional file 1). Together with our own retrospective study of 248 patients with unclear parkinsonism who underwent SPECT in the period 2001 to 2006, this resulted in 185 possibly relevant studies [28].

Of these 185 we excluded 153 studies (See additional file 2). Seven were excluded because of the language criteria and 85 articles were excluded as they did not deal with one of our three designated areas of clinical relevancy: most of them were about techniques, dementia or drug efficacy. We excluded an additional 61, because the absolute numbers with a cut-off point of 2 standard deviations below the control group were not available or could not be derived. We mailed the authors of the four studies with more than 85 patients, to ask for missing data [29‐32], and received a response from 1 [30]. We wanted to acquire raw data from relatively large studies that would have a substantial impact on our meta-analysis; there were 4 large studies with more than 85 patients, the rest involved smaller numbers in the 20–35 range. We thus sought to contact the authors of these 4 studies, as we felt that studies with less than 85 subjects would have a very limited impact on the overall scores.

Of the 32 trials left (See additional file 3) 7 dealt with diagnostic uncertainty including a clinical follow-up as surrogate gold standard, 20 studies were cross-sectional including subjects with known parkinsonian disorders, and 6 studies involved patients with early PD. One clinical study fitted as well in the early PD groups as in the study group of known parkinsonian disorders [33]. Of the 7 follow-up studies, 5 were prospective. Of these prospective analyses, 2 included untreated new patients with parkinsonism [34, 35], and 3 included patients with inconclusive parkinsonism or with a questionable effect of anti-parkinsonian medication [28, 36, 37]. Two of the 7 follow-up studies were retrospective: Lokkegaard and colleagues retrospectively investigated 90 consecutive patients referred for Beta-CIT SPECT for various reasons, and a non-treating neurologist obtained the final diagnosis from the clinical records of the patients [38]. Stoffers et al retrospectively analysed the SPECT scans of 72 patients with an initial diagnosis of PD, who were then re-diagnosed after various intervals [39]. The demographic and methodological characteristics of all included studies are visible in Table 1. The absolute numbers of the 2 × 2 tables of all included trials are shown in Table 2.

We recalculated the diagnostic power of SPECT for the following clinical problems: 1. diagnosis of PD in an early phase vs. normalcy; 2 diagnostic differentiation between PD and ET; 3. diagnostic differentiation between PD and vascular parkinsonism; 4. diagnostic differentiation between PD and Atypical Parkinsonian Syndromes (APS) consisting of MSA and PSP.

Our meta-analysis confirms the general opinion that SPECT is relatively accurate to differentiate between patients with PD in an early stage and healthy controls. The difference in sensitivity between trials can not be explained solely by different radiotracer usage. Especially the difference with the two studies using TRODAT is obvious [41, 43]. An explanation may be that the use of a template vs. hand-circling of the striatum leads to lower diagnostic specificities [40, 42, 43]. Other possible explanations for the lower sensitivity scores in the study of van Laere et al are their consecutive inclusion of patients and their clinic being a tertiary referral centre [43].

Schwarz and Asenbaum were the only two authors who used two standard deviations below the normal controls as cut-off [33, 42]. In the other four studies we recalculated the absolute numbers of true positive, false negative etc. by ourselves, [40, 41, 43, 44] which led in all 4 studies to lower numbers for diagnostic accuracy. Apart from this different cut-off point, higher sensitivity figures in several large trials (normal SPECT scans in 5–10% of clinically definitive PD patients) are probably explained by the disease stage of the patients [68‐70].

Asenbaum's, Haapaniemi's and Muller's were the only studies which mentioned blinding of the investigators [33, 40, 44]. It is perhaps not surprising that these authors found lower numbers for diagnostic accuracy than Huang and Schwarz (See fig. 1) [41, 42].

The results of our meta-analysis confirm the general opinion that SPECT with presynaptic tracers is highly accurate to differentiate between patients with PD and ET. Lee et al. scored lowest specificity. A possible explanation is that they included not only patients with ET but as well patients with isolated postural tremor and postural in combination with resting tremor [45].

According to the results of the meta-analysis we conclude that presynaptic SPECT scans can accurately differentiate between patients with PD and VP. The specificity, however, is only moderate in the studies of Lokkegaard and Eerola [38, 47]. VP is a somewhat controversial clinical concept and the differences found in the studies we analysed probably reflect the variability in clinical definition. Whereas Lokkegaard et al., Gerschlager et al., and Eerola et al. [38, 49] used strict inclusion criteria, Booij et al. did not [36, 47]. This point is illustrated in the paper by Loberboym et al, who investigated 20 patients with VP with FP-CIT SPECT. Nine had a normal presynaptic SPECT scan but 11 had decreased striatal tracer binding. All nine patients with normal presynaptic SPECT scan had no reaction on levodopa treatment, however 5 of the 11 with decreased striatal FP-CIT binding did have [71]. Although SPECT with presynaptic tracers scored high accuracy in differentiation between PD and VP, conventional techniques like CT and MRI may still be necessary as additional diagnostic tools.

This meta-analysis confirmed the generally accepted view that presynaptic tracers cannot distinguish between PD and APS. However, we also found that postsynaptic SPECT is not very good at this. A negative postsynaptic SPECT scan does not exclude APS. The positive predictive value of abnormal postsynaptic SPECT for the diagnosis of APS is high, making a reduced postsynaptic radiotracer binding likely to exclude a diagnosis of PD. However, patients with PD in our meta-analysis do show loss of dopamine-receptor binding. Studies which used IBZM scored higher accuracy than studies with IBF or Epidepride [50, 53, 56, 59]. If the trials are excluded which used other tracers than IBZM the mean 95%CI odds ratio only increased to 21 (11–44).

Some studies, excluded for this meta-analysis, found excellent accuracy for the postsynaptic tracer to differentiate between PD and APS [19, 64, 72‐77]. Schulz et al, who investigated 32 MSA patients, found similar results as in our meta-analysis: only a significant loss in 63% of the patients using IBZM-SPECT with two standard deviations under controls (PD patients) as cut-off point [78]. Results of the studies by Berding et al [79], and Bettin et al [80], are in line with these.

Possible explanations for the difference in results may be difference in cut-off points (many authors use only one SD below healthy controls), and the use of special techniques claimed by some to enhance diagnostic accuracy, e.g. asymmetric indices, caudate/putamen atio, presynaptic/postsynaptic ratio of tracer binding, speed of decline in follow-up [30, 36, 76, 81, 82]. One other explanation of the low discriminating value of postsynaptic SPECT imaging is the reversible down regulation of dopamine receptors by dopaminergic drugs [83]. If the drugs are not stopped appropriately, the scan result can be false positive. In our meta-analysis we found 9 studies where potentially interfering medication was not discontinued appropriately or these did even not mention whether this medication was discontinued (and if, for how long) before the scan.