Machine learning models for the differential diagnosis of vascular parkinsonism and Parkinson's disease using [(123)I]FP-CIT SPECT

Depósito de Investigación de la Universidad de Sevilla https://idus.us.es/ This version of the article has been accepted for publication, after peer review and is subject to Springer Nature’s AM terms of use, but is not the Version of Record and does not reflect post-acceptance improvements, or any corrections. The Version of Record is available online at: https://doi.org/10.1007/s00259-014- 2882-8 or https://link.springer.com/article/10.1007/s00259-014-2882-8 AUTHOR'S PROOF! UNCORRECTEDPROOF 1 2 3ORIGINAL ARTICLE 4Machine learning models for the differential diagnosis of vascular 5parkinsonism and Parkinson’s disease using [ 123 I]FP-CIT SPECT 7I. Huertas-Fernández &F. J. García-Gómez &D. García-Solís &S. Benítez-Rivero & 8V. A. Marín-Oyaga &S. Jesús &M. T. Cáceres-Redondo &J. A. Lojo & 9J. F. Martín-Rodríguez &F. Carrillo &P. Mir 10 11 12 Received: 30 April 2014 /Accepted: 28 July 2014 13 #Springer-Verlag Berlin Heidelberg 2014 14 Abstract 15 Purpose The study’s objective was to develop diagnostic 16 predictive models using data from two commonly used 17 [ 123 I]FP-CIT SPECT assessment methods: region-of-interest 18 (ROI) analysis and whole-brain voxel-based analysis. 19 Methods We included retrospectively 80 patients with vascu- 20 lar parkinsonism (VP) and 164 patients with Parkinson’s 21 disease (PD) who underwent [ 123 I]FP-CIT SPECT. Nuclear- 22 medicine specialists evaluated the scans and calculated bilat- 23 eral caudate and putamen [ 123 I]FP-CIT uptake and asymmetry 24 indices using BRASS software. Statistical parametric map- 25 ping (SPM) was used to compare the radioligand uptake 26 between the two regionsQ1 at the voxel level. Quantitative data 27 from these two methods, together with potential confounding 28 factors for dopamine transporter availability (sex, age, disease 29 duration and severity), were used to build predictive models 30 following a tenfold cross-validation scheme. The performance 31 of logistic regression (LR), linear discriminant analysis and 32 support vector machine (SVM) algorithms for ROI data, and 33 their penalized versions for SPM data (penalized LR, 34penalized discriminant analysis and penalized SVM), were 35assessed. 36Results Significant differences were found in the ROI analy- 37sis after covariate correction between VP and PD patients in 38[ 123 I]FP-CIT uptake in the more affected side of the putamen 39and the ipsilateral caudate. Age, disease duration and severity 40were also found to be informative in feeding the statistical 41model. SPM localized significant reductions in [ 123 I]FP-CIT 42uptake in PD with respect to VP in two specular clusters 43comprising areas corresponding to the left and right striatum. 44The diagnostic predictive accuracy of the LR model using 45ROI data was 90.3 % and of the SVM model using SPM data 46was 90.4 %. 47Conclusion The predictive models built with ROI data and 48SPM data from [ 123 I]FP-CIT SPECT provide great discrimi- 49nation accuracy between VP and PD. External validation of 50these methods is necessary to confirm their applicability 51across centres. 52Keywords Vascular parkinsonism .Parkinson’sdisease . 53[ 123 I]FP-CIT SPECT .Statistical parametric mapping . 54Predictive models 55Introduction 56Vascular parkinsonism (VP) is a parkinsonian syndrome 57resulting from cerebrovascular lesions and is characterized 58by the presence of gait difficulties, symmetrically lower body 59bradykinesia and postural instability, and the absence of rest- 60ing tremor [1–3]. Although recent neuropathology and epide- 61miological studies have identified hallmarks distinguishing 62VP from idiopathic Parkinson’s disease (PD), overlap in 63symptom presentation is not rare and their differentiation is 64still a clinical challenge, especially in early stages [4–8]. I. Huertas-Fernández :S. Benítez-Rivero :S. Jesús : M. T. Cáceres-Redondo :J. F. Martín-Rodríguez :F. Carrillo : P. Mir (*) Unidad de Trastornos del Movimiento, Servicio de Neurología y Neurofisiología Clínica, Instituto de Biomedicina de Sevilla (IBiS), Hospital Universitario Virgen del Rocío/CSIC/Universidad de Sevilla, Avda. Manuel Siurot s/n., 41013 Seville, Spain e-mail: [email protected] F. J. García-Gómez :D. García-Solís :V. A. Marín-Oyaga : J. A. Lojo Servicio de Medicina Nuclear, UDIM, Hospital Universitario Virgen del Rocío, Seville, Spain D. García-Solís :P. Mir Centro de Investigación Biomédica en Red sobre Enfermedades Neurodegenerativas (CIBERNED), Seville, Spain Eur J Nucl Med Mol Imaging DOI 10.1007/s00259-014-2882-8 JrnlID 259_ArtID 2882_Proof# 1 - 05/08/2014 AUTHOR'S PROOF! UNCORRECTEDPROOF 65 Visualization of the dopamine transporter (DAT) through the 66 use of [ 123 I]FP-CIT SPECT is a commonly used tool that may 67 help in differentiating VP and PD. However, the status of the 68 striatal DAT in VP patients is controversial due to its heteroge- 69 neity, and accuracy in the differential diagnosis is still poor [6, 70 9–14]. This heterogeneity was reflected in a recent study involv- 71 ing a large cohort of patients with VP in which the [ 123 I]FP-CIT 72 SPECT scans in about one-third of the patients were normal, 73 while the scans in the other two-thirds were abnormal, and the 74 imaging pattern in a small percentage of patients overlapped the 75 typical pattern seen in PD [14]. Furthermore, it has been sug- 76 gested that a normal scan in patients with VP may be associated 77 with negative responsiveness to levodopa treatment [14], al- 78 though this association was not seen in another study [9]. 79 The majority of studies including patients with VP have 80 evaluated [ 123 I]FP-CIT SPECT imaging through visual as- 81 sessment according to standardized scales [15]or 82 semiquantification of striatal ligand uptake involving region- 83 of-interest (ROI) analysis. Such methods may be suboptimal 84 mainly because of first a certain degree of subjectivity in 85 visual interpretation and in manual ROI delineation and sec- 86 ond the focus primarily on DAT uptake in the striatum, thus 87 missing the extent of radioligand binding to the DAT, seroto- 88 nin and noradrenergic transporters in other brain regions. In 89 contrast, voxel-based analysis has proven to be a reliable and 90 unbiased tool for the analysis of whole-brain imaging. Statis- 91 tical parametric mapping (SPM) is one of the most popular 92 tools for whole-brain voxel-based analysis and some studies 93 have used it with success in the differentiation of PD from 94 other neurodegenerative diseases [16–19]. However, voxel- 95 based studies including VP patient series are still lacking. 96 We have recently reported the results of a detailed clinical 97 study in a large cohort of VP and PD patients [20]. We 98 presented a newly developed visual scoring system with an 99 accuracy in the differentiation of VP and PD greater than 94 % 100 and a clustering method using ROI data with an accuracy of 101 82 %. The first objective of the present study was to build a 102 diagnostic predictive model using the ROI data from the same 103 dataset with improved performance and applying a more 104 suitable methodology for the problems of classification from 105 the machine learning theory. The second objective was to 106 conduct a whole-brain voxel-based comparison of imaging 107 data between VP and PD patients using SPM and following 108 the same strategy as for the ROI data to build a predictive 109 model with the voxel data. 110 Materials and methods 111 Patients 112 We included a total of 80 patients with VP and a control group 113 of 164 patients with PD seen at our centre from 2006 to 2011. 114This is the subset of patients with [ 123 I]FP-CIT SPECT 115scans available from our previous work, and detailed 116clinical information for the whole population of patients 117was given in the reort of that study [20]. For this study, 118the features sex, age, disease duration and severity mea- 119sured according to the Hoehn & Yahr (H&Y) scale were 120reviewed when carrying out SPECT (Table 1). The 121diagnosis of VP was made according to the diagnostic 122criteria proposed by Zijlmans et al. [5] and the diagno- 123sis of PD was made according to the UK Parkinson’s 124Disease Society Brain Bank clinical diagnostic criteria 125[4]. Patients gave written informed consent for the 126[ 123 I]FP-CIT SPECT scan after a full discussion of 127possible risks and benefits as is the general practice in 128our hospital. This study was approved by the local 129ethics committee and conducted in accordance with the 130Declaration of Helsinki. 131SPECT imaging 132Patients underwent a brain SPECT scan with a dual-head 133rotating gamma camera (Philips Axis) fitted with LEHR fan- 134beam collimators. In order to block the thyroid uptake of free 135radioactive iodide, patients were given potassium perchlorate 136500 mg orally 30 min before intravenous injection of 137185 MBq of [ 123 I]FP-CIT (Ioflupane, Datscan®; GE 138Healthcare). Image acquisition was started between 3 and 1394 h after radioligand injection. A total of 120 projections of 14030 s each over a 360° circular orbit were acquired on a 128× 141128 matrix (zoom 1.5). Reconstruction was performed by 142filtered back-projection using a Butterworth filter without 143attenuation or scatter correction and further reorientation to 144obtain transaxial slices. 145ROI analysis 146An automated semiquantitative analysis was performed to 147evaluate specific-to-nondisplaceable [ 123 I]FP-CIT binding 148potential (BP ND ) using HERMES-BRASS software (version 1493.5). ROIs were constructed around the left and right 150striatum, the striatal subregions caudate and putamen, and 151in the background brain (occipital cortex). The automated 152method in HERMES-BRASS first normalized the patient 153scans to a [ 123 I]FP-CIT mean template and then delineated 154the regions using the standardized 3-D volume-of-interest 155(VOI) maps. Further details of the procedure and creation 156of the [ 123 I]FP-CIT mean template and standardized 3-D 157VOI maps have been provided by Koch et al. [21]. 158[ 123 I]FP-CIT BP ND for the left and right putamen and 159caudate were calculated by normalizing the subregional 160radioactivity counts by the background counts (for each 161striatal subregion: BP ND =[(striatal subregion counts−occip- 162ital counts)/occipital counts]). We defined the more affected Eur J Nucl Med Mol Imaging JrnlID 259_ArtID 2882_Proof# 1 - 05/08/2014 AUTHOR'S PROOF! UNCORRECTEDPROOF 163 side as the hemisphere with the lower putamen BP ND ,and 164 the ROI variables were defined as those from the putamen 165 and caudate ipsilateral to the more affected side (Put_I, 166 Cau_I), and the putamen and caudate contralateral to the 167 more affected side (Put_C, Cau_C). The asymmetry index 168 (AI) was calculated using the following formula [9]: 169 AI=[(contralateral striatum binding−ipsilateral striatum 170 binding)/(contralateral striatum binding+ipsilateral striatum 171 binding)]×2×100. 172 SPM analysis 173 A semiquantitative whole-brain voxel-based analysis was per- 174 formed using SPM8 (Wellcome Department of Cognitive 175 Neurology, London, UK; http://www.fil.ion.ucl.ac.uk/spm/ 176 software/spm8/) running under a Matlab environment 177 (MathWorks, Sherborn, MA). 178 SPECT images were first manually reoriented, setting 179 the anterior commissure as the origin of the coordinates. 180 Each scan was then spatially normalized into the standard 181 stereotactic MNI (Montreal Neurological Institute) space 182 using a [ 123 I]FP-CIT template developed by our group 183 (available atQ2 http://www.nitrc.org/projects/spmtemplate) 184 [22]. Next, spatially normalized images were smoothed 185 using an isotropic 8-mm full-width at half-maximum iso- 186 tropic gaussian kernel (FWHM). For further details about 187 processing underlying the normalization and smoothing 188 steps, the reader is referred to the SPM manual (http:// 189 www.fil.ion.ucl.ac.uk/spm/doc/manual.pdf). For the 190 analysis stage, to account for the interindividual variability 191 of [ 123 I]FP-CIT uptake, the option “proportional scaling” 192 was enabled to intensity-normalize each scan. Also, since 193 DAT densities are known to be low in the occipital lobe 194 and the cerebellum, a general brain mask for those areas 195 was created using an automated anatomical labelling atlas 196 and applied to all images for statistical comparison. A total 197 of 152,673 voxels were analysed. Clusters of a minimum 198 of 16 (twice FWHM of the gaussian filter) 3-D contiguous 199 voxels with a threshold of P FWE <0.05 corrected for multi- 200 ple comparisons based on family-wise error (FWE) were 201 considered to be statistically significant. 202Data analysis 203Statistical analyses were performed using IBM SPSS Statistics 20420.0 software and the free software environment R (http:// 205www.r-project.org/). Descriptive statistics are reported with 206percentages, means and standard deviations and medians and 207interquartile ranges when appropiate. Univariate analyses 208were first performed to compare the demographic and 209clinical features, and the ROI variables between VP and PD 210patients. Sex distribution was compared using the chi-squared 211test. Scale variables (i.e. age, disease duration, H&Y stage and 212ROI variables) were compared using the t-test (parametric) or 213the Mann–Whitney test (nonparametric). To decide whether to 214use parametric or nonparametric tests with the scale variables, 215we assessed the assumptions for a normal distribution using 216the Shapiro-Wilk test and the homogeneity of the variance 217(homoscedasticity) using the Levene test. To assess differ- 218ences in the ROI variables between the VP and PD patients 219taking into account the effects of the demographic and clinical 220features, we further performed multivariate analysis using 221logistic regression (LR) introducing the ROI variables as 222factors and sex, age, disease duration and H&Y stage as 223covariates. Additionally, since these covariates are known to 224influence radioligand uptake [23,24], interaction terms were 225included to check their role as effect modifiers. In the SPM 226analysis, the [ 123 I]FP-CIT uptake was compared between VP 227and PD patients with a two-sample t-test contrast (VP>PD). 228Moreover, due to its clinical relevance and in order to 229clarify the inconsistencies apparent in the literature [9,14], 230subanalyses of VP patients comparing levodopa responders 231and nonresponders were performed using LR for ROI data, 232and again using a two-sample t-test contrast in SPM (VP 233nonresponder>VP responder). 234Predictive models 235Predictive models using the quantitative data from the ROI 236and SPM analyses were built for diagnosis classification. The 237clinical diagnosis, as defined in the section Patients, was 238considered the gold standard in this study and was used as 239the dependent variable in building the models. For the models t1:1Table 1 Demographic and main clinical features of VP and PD patients t1:2VP patients (n=80) PD patients (n=164) Pvalue (intergroup comparison) t1:3ttest Mann–Whitney test Logistic regression t1:4Sex (M/F), n57/23 101/68 t1:5Age (years), mean±SD 75.11±6.70 60.26±10.84 <0.001 <0.001 t1:6Disease duration (years), median (interquartile range) 4 (2, 8) 2 (1, 4) <0.001 <0.01 t1:7H&Y stage, median (interquartile range) 2.5 (2.5, 3) 2 (2, 2.5) <0.001 <0.05 Eur J Nucl Med Mol Imaging JrnlID 259_ArtID 2882_Proof# 1 - 05/08/2014 AUTHOR'S PROOF! UNCORRECTEDPROOF 240 using ROI data, the independent variables were the significant 241 factors from the LR, while for the models using SPM data, the 242 independent variables were the intensity values of the voxels 243 (after normalization and smoothing) contained in the signifi- 244 cant clusters from the SPM contrasts. Significant covariates 245 were also included in the models. 246 Since the number of independent variables for the models 247 using SPM data was large (a few hundred voxels), regularized 248 algorithms were used. These algorithms weight the indepen- 249 dent variables according to their information content, 250 priorizing some and penalizing others through tunable shrink- 251 age functions. We opted for comparing three algorithms rec- 252 ommended elsewhere [25]: penalized LR (PLR), penalized 253 discriminant analysis and penalized support vector machine 254 (SVM). The ROI data were analysed using equivalent 255 methods: LR, linear discriminant analysis (LDA) and SVM. 256 Tuning parameters for the algorithms were chosen based on 257 the package default grid of iterations. 258 The models were assessed using a tenfold cross-validation 259 scheme, which randomly split the dataset into ten parts (K= 260 10), 90 % used for training and the remaining 10 % for testing, 261 for every kth=1, 2, …, 10. This strategy prevented overfitting 262 the model with our dataset, thus allowing model generaliza- 263 tion of data from other centres. The final model and perfor- 264 mance results were obtained from averaging the ten runs, 265 which were given in terms of area under the receiver operating 266 characteristic curve, accuracy, sensitivity and specificity. All 267 calculations were done using the R package “caret”(http:// 268 caret.r-forge.r-project.org/). 269 270 Results 271 Demographic and clinical features 272 Age and disease duration fulfilled normality and homoscedas- 273 ticity assumptions and were compared in the univariate anal- 274 ysis using a t-test. The H&Y stage did not fulfil the assump- 275 tions and was compared using the Mann–Whitney test. There 276were significant differences between VP and PD patients in 277terms of age, disease duration and H&Y stage (P<0.001; see 278Table 1). These associations were consistent in the regression 279analysis. As already described in our previous work [20], our 280VP patients were older, with longer disease duration and 281higher H&Y stage than our PD patients. 282Discrimination between VP and PD patients using ROI 283analysis 284Regional [ 123 I]FP-CIT uptake and AI values of the VP and PD 285patients and intergroup statistics are shown in Table 2.The 286variables Put_I, Cau_I, Put_C and Cau_C uptake values ful- 287filled normality and homoscedasticity assumptions and were 288compared in the univariate analysis using a t-test. AI did not 289fulfil the assumptions and was compared using the Mann– 290Whitney test. Univariate analyses showed significantly lower 291[ 123 I]FP-CIT BP ND values for all four regions along with a 292higher AI in PD patients than in VP patients (P<0.001). 293Regression analysis indicated that these findings were consis- 294tent after covariate correction for the more affected t2:1Table 2 Mean regional putamen and caudate [ 123 I]FP-CIT BP ND in VP and PD patients t2:2VP patients (n=80) PD patients (n=164) Pvalue (intergroup comparison) t2:3ttest Mann–Whitney test Logistic regression t2:4[ 123 I]FP-CIT BP ND , mean±SD Cau_I 1.54±0.54 1.06±0.46 <0.001 <0.001 t2:5Cau_C 1.64±0.49 1.23±0.50 <0.001 t2:6Put_I 1.20±0.52 0.53±0.30 <0.001 <0.001 t2:7Put_C 1.43±0.52 0.78±0.37 <0.001 t2:8Asymmetry index, median (interquartile range) 7.04 (3.00, 18.94) 20.69 (8.82, 39.00) <0.001 3 2 1 00123 Ipsilateral caudate [123I]FP-CIT BPND Ipsilateral putamen[123I]FP-CIT BPND PD VP Fig. 1 Ipsilateral caudate [ 123 I]FP-CIT BP ND as function of ipsilateral putamen [ 123 I]FP-CIT BP ND (red circles VP patients, blue circles PD patients) Eur J Nucl Med Mol Imaging JrnlID 259_ArtID 2882_Proof# 1 - 05/08/2014 AUTHOR'S PROOF! UNCORRECTEDPROOF 295 hemisphere regions (Cau_I, P<0.001;Put_I,P<0.001). None 296 of the interaction terms reached significance. 297 These significant variables, along with the covariates age, 298 disease duration and H&Y stage, were further used to build 299 the predictive models. Figure 1displays the scatter plot of the 300 two input factors, where Cau_I [ 123 I]FP-CIT BP ND is plotted 301 as a function of Put_I [ 123 I]FP-CIT BP ND .MostoftheVP 302 patients grouped separately from most of the PD patients, and 303 the decision boundaries between the two entities could be 304 fitted with linear algorithms. Hence, LR, LDA and SVM fed 305 by the first-order terms of the input factors were adequate 306 approaches. The cross-validation results for the three methods 307 are shown in Table 3. LR demonstrated slightly better dis- 308 crimination accuracy than SVM and LDA (accuracy 0.903, 309 0.899 and 0.898, respectively), and its equation is given by the 310 following formula: logit diagnosisðÞ¼14:55 3:92 Cau Iþ7:29 Put I þ0:18 age þ0:75 H&Y 0:28 DisDur 311312 313 where logit represents the logarithm of p/(1−p)andpis the 314 probability of being a VP patient, and DisDur is disease 315 duration. 316 These results indicate that, despite being a very good 317 model, a small percentage of scans were misdiagnosed. To 318 improve the discrimination accuracy, we established a cut-off 319 of 80 % for the class probability. In other words, we assigned a 320 diagnosis only if the probability of belonging to that class 321 applying the formula was above 80 %. We tested the LR 322 model in the whole dataset and the accuracy was increased 323 to 95 %, although the data from 17 % of the patients were 324under the threshold and their diagnosis remained tagged as 325“doubtful”(Fig. 2). 326Discrimination between VP and PD patients using SPM 327analysis 328Voxel-based analysis of [ 123 I]FP-CIT SPECT scans supported 329the results of the striatal ROI analysis. SPM contrasts revealed 330decreased intensity values in PD patients compared with VP 331patients in two specular clusters (1,113 and 1,320 voxels) that 332comprised areas corresponding to, respectively, the left and 333right striatum (Table 4). 334The predictive models were built using all intensity values 335of voxels contained in the significant clusters as independent 336variables, and the same covariates as in the ROI analysis (age, 337disease duration and H&Y stage). The cross-validation results 338from the penalized methods are summarized in Table 5.SVM 339showed slightly better accuracy in discriminating between VP 340and PD than PLR and LDA (accuracy 0.904, 0.887 and 0.884, 341respectively). 342Comparison between levodopa responders and nonresponders 343Neither LR with ROI data nor SPM analysis revealed an 344association between [ 123 I]FP-CIT uptake and levodopa re- 345sponsiveness in VP patients. 346 347 Discussion 348In this study, we investigated the accuracy of methods for 349distinguishing between VP and PD using [ 123 I]FP-CIT 350SPECT. We developed predictive models using the t3:1Table 3 Average tenfold cross-validation performance results (mean± standard deviation) for the diagnostic predictive models built with ROI data, given as area under the ROC curve (AUC), accuracy, sensitivity and specificity . The methods tested were logistic regression (LR), linear discriminant analysis (LDA) and support vector machine (SVM) t3:2Method AUC Accuracy Sensitivity Specificity Parameters t3:3LR 0.951±0.046 0.903±0.058 0.944±0.062 0.794±0.142 – t3:4LDA 0.940±0.042 0.898±0.065 0.963±0.049 0.775±0.138 – t3:5SVM 0.950±0.045 0.899±0.049 0.947±0.061 0.784±0.145 C=1 Fig. 2 Voxel clusters representing significant decreases in [ 123 I]FP-CIT uptake in PD patients with respect to VP patients. The areas include the putamen and caudate nucleus, and are represented in MNI- normalized MRI scans Eur J Nucl Med Mol Imaging JrnlID 259_ArtID 2882_Proof# 1 - 05/08/2014 AUTHOR'S PROOF! UNCORRECTEDPROOF 351 semiquantitative data from the SPECT evaluations of a large 352 cohort of patients using two widely differing methods: striatal 353 ROI analysis and whole-brain voxel-based analysis. Our pre- 354 vious study [20], as well as a similar multicentre study per- 355 formed in Italy [14], confirmed what previous studies have 356 indicate: VP is a different and distinguishable entity from PD, 357 but clinical manifestations and imaging patterns are heteroge- 358 neous. [ 123 I]FP-CIT SPECT is a widely available tool helping 359 the physician in the diagnosis of VP, and numerous studies 360 have investigated visual assessment and ROI quantification 361 using [ 123 I]FP-CIT SPECT [6,26]. Some authors have found 362 significant differences in the AI in PD patients [9,12], but 363 these studies had small sample sizes and their sensitivity was 364 as low as 50 % [12]. These results have led to questioning the 365 accuracy of [ 123 I]FP-CIT SPECT in the diagnosis of VP, and 366 indeed, a very recent study considered the inclusion of cardiac 367 [ 123 I]MIBG SPECT and the use of the smell identification 368 UPSIT test in the differential diagnosis [13]. 369 In our previous study we used the [ 123 I]FP-CIT BP ND 370 values of the more affected side of the putamen and the 371 ipsilateral caudate and the AI in a clustering method, and 372 achieved an accuracy of 82 %. However, this approach did 373 not exploit all the information available from the patient and 374 contained in the image, nor did it provide a generalizable 375 mathematical formula for use by other groups. In contrast, 376 other studies have successfully applied elegant methods for 377 distinguishing atypical parkinsonisms and other diseases from 378 PD using DAT SPECT imaging [16–19]. Scherfler et al. used 379 ROI analysis and SPM to extract mean voxel cluster values 380 and introduced their parameters into a stepwise discriminant 381 analysis [17]. Some years later, the same group elaborated a 382computer-assisted image algorithm (CAIA) using voxel data 383that outperformed a multinomial regression using ROI data 384[18]. In this study, we investigated images from VP patients 385using these types of approaches. In the ROI analysis, in 386agreement with the findings of previous studies [10,12], we 387found that in PD patients, in comparison with VP patients, the 388striatal DAT availability is markedly reduced and the AI is 389significantly higher. LR revealed that the more affected 390side of the putamen and the ipsilateral caudate, along with 391the covariates age, disease duration and H&Y stage, were 392informative in feeding the predictive model. Cross- 393validation procedures demonstrated that the algorithms 394LR, LDA and SVM were excellent classifiers using these 395variables. In the case of LR, the model achieved a diag- 396nostic accuracy of 90.3 %. Moreover, the results could be 397improved to 95 % accuracy by thresholding the class 398probability and creating a pool of patients with a doubtful 399diagnosis. For these patients, we assumed that the ROI 400analysis of the [ 123 I]FP-CIT SPECT scans was inconclu- 401sive and that it would be necessary to evaluate their 402clinical profile and structural neuroimaging to determine 403a more reliable diagnosis. 404Despite the diagnostic accuracy for the newly developed 405visual scoring system in our previous work that reached above 40694 %, we acknowledge that the application of this system 407requires highly trained nuclear medicine specialists, and that 408the intraobserver and interobserver rates are not perfect. Al- 409though we strongly encourage specialists to learn and apply 410the new visual scoring system, we believe that the application 411of the LR formula could be used more easily to achieve 412diagnostic accuracies above 90 %. t4:1Table 4 Significant findings of the SPM comparison of VP and PD patients (VP>PD) t4:2Cluster localization Cluster size MNI coordinates Tvalue Zvalue p FWE value t4:3xyz t4:4Left striatum 1,113 −26 −10 2 8.31 7.77 <0.001 t4:5−20 14 2 5.59 5.42 <0.001 t4:6−22 −18 4 5.46 5.29 <0.001 t4:7Right striatum 1,320 28 −6 2 8.08 7.59 <0.001 t4:822 −16 0 5.32 5.17 0.001 t4:920 10 12 5.27 5.12 0.001 t5:1Table 5 Average tenfold cross-validation performance results (mean± standard deviation) for the diagnostic predictive models built with SPM data, given as area under the ROC curve (AUC), accuracy, sensitivity and specificity. The methods tested were penalized logistic regression (PLR), penalized discriminant analysis (PDA) and support vector machine (SVM) t5:2Method AUC Accuracy Sensitivity Specificity Parameters t5:3PLR 0.960±0.039 0.887±0.049 0.981±0.034 0.704±0.144 α=0.1, λ=0.1 t5:4PDA 0.878±0.073 0.884±0.069 0.944±0.058 0.769±0.135 λ=3 t5:5SVM 0.954±0.057 0.904±0.059 0.954±0.056 0.801±0.172 C=1 Eur J Nucl Med Mol Imaging JrnlID 259_ArtID 2882_Proof# 1 - 05/08/2014 AUTHOR'S PROOF! UNCORRECTEDPROOF 413 Regarding SPM, the comparison gave significant differ- 414 ences in [ 123 I]FP-CITuptake in two specular clusters of voxels 415 including areas of the striatum. We took advantage of this high 416 level of information by introducing all voxel values together 417 with the covariates into a penalized classification algorithm, 418 and found that SVM was able to achieve 90.4 % accuracy. 419 Furthermore, as for the ROI data, it would also be possible to 420 raise this accuracy by restricting the allocation to high class 421 probabilities. This method demonstrated that the use of whole- 422 brain voxel data is a powerful alternative with two great 423 advantages with respect to the previous method, i.e. no a priori 424 assumptions about the location of the ligand uptake and more 425 importantly, the method is conducted in an unbiased and 426 automated fashion. 427 It is also important to note that our models made use of 428 basic clinical information, namely age, disease duration and 429 H&Y stage. Differences in these factors are not uncommon 430 between VP and PD cohorts. Antonini et al. obtained differ- 431 ences in H&Y stage in a large cohort of patients [14]. Other 432 studies have also shown differences in age [9,10,12,13]and 433 disease duration [13]. These differences give these factors 434 predictive ability to differentiate VP from PD, and they are 435 also potential confounders for determining differences in 436 [ 123 I]FP-CIT uptake. For these reasons, it is necessary to 437 incorporate these variables in the discriminative models using 438 [ 123 I]FP-CIT SPECT that seek to be applicable to the general 439 populations of VP and PD patients. However, previous studies 440 using [ 123 I]FP-CIT SPECT for differentiating VP from PD did 441 not fully take into account this information. Age, disease 442 duration and H&Y stage were not quantitatively included in 443 the studies that used visual assessment [13,14], and other 444 studies that used ROI semiquantification have shown differ- 445 ences in the ROI variables between these groups of patients 446 without accounting for them [9–12]. In this study we observed 447 that these factors, apart from directly influencing radioligand 448 uptake per se, were simple, accessible and very informative 449 for differentiating VP from PD. Hence, we recommend their 450 inclusion in the models. 451 We also sought to determine if there was higher striatal 452 ligand uptake in VP patients with a negative response to 453 levodopa treatment than in positive responders, as found by 454 Antonini et al. [14]. Our results were all negative for this 455 association indicating that the [ 123 I]FP-CIT uptake is not a 456 good predictor of responsiveness to dopamine replacement 457 therapy. 458 Finally, it is interesting to speculate as to why these models 459 did not reach 100 % accuracy. In our opinion, a major limita- 460 tion influencing the accuracy might have been that our gold 461 standard was based on clinical criteria that did not take into 462 account the SPECT findings, and perhaps a few patients were 463 wrongly diagnosed. Some of the patients who were diagnosed 464 as having VP, even though they fulfilled the criteria for VP 465 when included in the study, had a PD-like scan pattern. It is 466possible that some of these patients truly had VP with a 467[ 123 I]FP-CIT SPECT scan pattern indistinguishable from that 468in PD, while others had in reality underlying PD accompanied 469by cerebrovascular damage. In this case, updating our models 470would have resulted in an increase in the accuracy and there- 471fore a boost in the credibility of the SPECT-aided diagnosis. 472Nevertheless, to confirm this hypothesis it would be necessary 473to perform a long-term follow-up to verify how these patients 474evolve clinically, or preferably, an MRI scan or an 475anatomopathological examination in the most misleading 476cases. 477In conclusion, this study provided accuracies above 90 % 478in discriminating between VP and PD using two common 479methods for SPECT scan evaluation: ROI analysis and SPM. 480We provide a mathematical formula for the ROI analysis 481model for evaluation by other groups. We also introduce a 482method for processing voxel-based data: the use of penalized 483algorithms implemented in R packages. This approach pro- 484vides an automated and therefore objective, fast and efficient 485solution that would be very beneficial for decision-making in 486nuclear medicine. Future work will investigate the method 487including more types of parkinsonism and its implementation 488in a distributable application for external evaluation. 489Acknowledgments This work was supported by grants from the 490Ministerio de Economía y Competitividad de España (SAF2007- 49160700), the Instituto de Salud Carlos III (PI10/01674, CP08/00174, 492PI13/01461), the Consejería de Economía, Innovación, Ciencia y 493Empresa de la Junta de Andalucía (CVI-02526, CTS-7685), the 494Consejería de Salud y Bienestar Social de la Junta de Andalucía (PI- 4950377/2007, PI-0741/2010, PI-0437-2012), the Sociedad Andaluza de 496Neurología, the Fundación Alicia Koplowitz, the Fundación Mutua 497Madrileña and the Jaques and Gloria Gossweiler Foundation. 498 499 Conflicts of interest None. 500References 501 502 1. Rektor I, Rektorova I, Kubova D. Vascular parkinsonism - an update. 503J Neurol Sci. 2006;248:185–91. 5042. Fitzgerald PM, Jankovic J. Lower body parkinsonism: evidence for 505vascular etiology. Mov Disord. 1989;4:249–60. 5063. Winikates J, Jankovic J. Clinical correlates of vascular parkinsonism. 507Arch Neurol. 1999;56:98–102. 5084. Hughes AJ, Daniel SE, Kilford L, Lees AJ. The accuracy of clinical 509diagnosis of idiopathic Parkinson’s disease: a clinicopathologic 510study. J Neurol Neurosurg Psychiatry. 1992;55:181–4. 5115. Zijlmans JC, Daniel SE, Hughes AJ, Revesz T, Lees AJ. 512Clinicopathological investigation of vascular parkinsonism, includ- 513ing clinical criteria for diagnosis. Mov Disord. 2004;19:630–40. 5146. Kalra S, Grosset DG, Benamer HT. Differentiating vascular parkin- 515sonism from idiopathic Parkinson’s disease: a systematic review. 516Mov Disord. 2010;25:149–56. 5177. Jellinger KA. Prevalence of cerebrovascular lesions in Parkinson’s 518disease. A post-mortem study. Acta Neuropathol. 2003;105:415–9. Eur J Nucl Med Mol Imaging JrnlID 259_ArtID 2882_Proof# 1 - 05/08/2014 AUTHOR'S PROOF! UNCORRECTEDPROOF 519 8. Zijlmans J, Katzenschlager R, Daniel S, Lees A. The L-dopa re- 520 sponse in vascular parkinsonism. J Neurol Neurosurg Psychiatry. 521 2004;75:545–7. 522 9. Zijlmans J, Evans A, Fontes F, Katzenschlager R, Gacinovic S, Lees 523 AJ, et al. [123I]FP-CIT SPECT study in vascular parkinsonism and 524 Parkinson’s disease. Mov Disord. 2007;22:1278–85. 525 10. Gerschlager W, Bencsits G, Pirker W, Bloem BR, Asenbaum S, 526 Prayer D, et al. [123I]beta-CIT SPECT distinguishes vascular par- 527 kinsonism from Parkinson’s disease. Mov Disord. 2002;17:518–23. 528 11. Lorberboym M, Djaldetti R, Melamed E, Sadeh M, Lampl Y. 123I- 529 FP-CIT SPECT imaging of dopamine transporters in patients with 530 cerebrovascular disease and clinical diagnosis of vascular parkinson- 531 ism. J Nucl Med. 2004;45:1688–93. 532 12. Contrafatto D, Mostile G, Nicoletti A, Dibilio V, Raciti L, Lanzafame 533 S, et al. [123I]FP-CIT-SPECT asymmetry index to differentiate 534 Parkinson’s disease from vascular parkinsonism. Acta Neurol 535 Scand. 2012;126:12–6. 536 13. Navarro-Otano J, Gaig C, Muxi A, Lomeña F, Compta Y, 537 Buongiorno MT, et al. (123)I-MIBG cardiac uptake, smell identifi- 538 cation and (123)I-FP-CIT SPECT in the differential diagnosis be- 539 tween vascular parkinsonism and Parkinson’s disease. Parkinsonism 540 Relat Disord. 2014;20:192–7. 541 14. Antonini A, Vitale C, Barone P, Cilia R, Righini A, Bonuccelli U, et al. 542 The relationship between cerebral vascular disease and parkinsonism: 543 the VADO study. Parkinsonism Relat Disord. 2012;18:775–80. 544 15. Benamer TS, Patterson J, Grosset DG, Booij J, de Bruin K, van 545 Royen E, et al. Accurate differentiation of parkinsonism and essential 546 tremor using visual assessment of [123I]-FP-CIT SPECT imaging: 547 the [123I]-FP-CIT study group. Mov Disord. 2000;15:503–10. 548 16. Colloby SJ, O’Brien JT, Fenwick JD, Firbank MJ, Burn DJ, McKeith 549 IG, et al. The application of statistical parametric mapping to 123-I- 550 FPCIT SPECT in dementia with Lewy bodies, Alzheimer’sdisease 551 and Parkinson’s disease. Neuroimage. 2004;23:956–66. 552 17. Scherfler C, Seppi K, Donnemiller E, Goebel G, Brenneis C, 553 Virgolini I, et al. Voxel-wise analysis of [123I]β-CIT SPECT differ- 554 entiates the Parkinson variant of multiple system atrophy from idio- 555 pathic Parkinson’s disease. Brain. 2005;128:1605–12. 55618. Goebel G, Seppi K, Donnemiller E, Warwitz B, Wenning GK, 557Virgolini I, et al. A novel computer-assisted image analysis of 558[123I]β-CIT SPECT images improves the diagnostic accuracy of 559parkinsonian disorders. Eur J Nucl Med Mol Imaging. 2011;38(4): 560702–10. 56119. Nocker M, Seppi K, Donnemiller E, Virgolini I, Wenning GK, Poewe 562W, et al. Progression of dopamine transporter decline in patients with 563the Parkinson variant of multiple system atrophy: a voxel-based 564analysis of [123I]β-CIT SPECT. Eur J Nucl Med Mol Imaging. 5652012;39:1012–20. 56620. Benitez-Rivero S, Marin-Oyaga VA, Garcia-Solis D, Huertas- 567Fernández I, García-Gómez FJ, Jesús S, et al. Clinical features and 568123I-FP-CIT SPECT imaging in vascular parkinsonism and 569Parkinson’s disease. J Neurol Neurosurg Psychiatry. 2013;84:122–9. 57021. Koch W, Radau P, Hamann C, Tatsch K. Clinical testing of an 571optimized software solution for an automated, observer- 572independent evaluation of dopamine transporter SPECT studies. J 573Nucl Med. 2005;46:1109–18. 57422. García-Gómez FJ, García-Solís D, Luis-Simón FJ, Marín-Oyaga VA, 575Carrillo F, Mir P, et al. Elaboration of the SPM template for the 576standardization of SPECT images with 123I-Ioflupane. Rev Esp 577Med Nucl Imagen Mol. 2013;32:350–6. 57823. Varrone A, Dickson JC, Tossici-Bolt L, Sera T, Asenbaum S, Booij J, 579et al. European multicentre database of healthy controls for [123I]FP- 580CIT SPECT (ENC-DAT): age-related effects, gender differences and 581evaluation of different methods of analysis. Eur J Nucl Med Mol 582Imaging. 2013;40:213–27. 58324. Erro R, Pappatà S, Amboni M, Vicidomini C, Longo K, Santangelo 584G, et al. Anxiety is associated with striatal dopamine transporter 585availability in newly diagnosed untreated Parkinson’s disease pa- 586tients. Parkinsonism Relat Disord. 2012;18:1034–8. 58725. Friedman J, Hastie T, Tibshirani R. The Elements of Statistical 588Learning - Data Mining, Inference, and Prediction, 2nd Edition. 589Springer Series in Statistics. Berlin: Springer; 2009. p. 1–745. 59026. Tatsch K, Poepperl G. Nigrostriatal dopamine terminal imaging with 591dopamine transporter SPECT: an update. J Nucl Med. 2013;54: 5921331–8. 593 Eur J Nucl Med Mol Imaging JrnlID 259_ArtID 2882_Proof# 1 - 05/08/2014