Differential diagnosis of parkinsonian syndromes. A comparison of clinical and automated - metabolic brain patterns' based approach

Differentiation among parkinsonian syndromes may be clinically challenging, especially at early disease stages. In this study, we used ${}^{18}$F-FDG-PET brain imaging combined with an automated image classification algorithm to classify parkinsonian patients as Parkinson's disease (PD) or as an atypical parkinsonian syndrome (APS) at the time when the clinical diagnosis was still uncertain. In addition to validating the algorithm, we assessed its utility in a "real-life" clinical setting. One hundred thirty-seven parkinsonian patients with uncertain clinical diagnosis underwent ${}^{18}$F-FDG-PET and were classified using an automated image-based algorithm. For 66 patients in cohort A, the algorithm-based diagnoses were compared with their final clinical diagnoses, which were the gold standard for cohort A and were made 2.2 ± 1.1 years (mean ± SD) later by a movement disorder specialist. Seventy-one patients in cohort B were diagnosed by general neurologists, not strictly following diagnostic criteria, 2.5 ± 1.6 years after imaging. The clinical diagnoses were compared with the algorithm-based ones, which were considered the gold standard for cohort B. Image-based automated classification of cohort A resulted in 86.0% sensitivity, 92.3% specificity, 97.4% positive predictive value (PPV), and 66.7% negative predictive value (NPV) for PD, and 84.6% sensitivity, 97.7% specificity, 91.7% PPV, and 95.5% NPV for APS. In cohort B, general neurologists achieved 94.7% sensitivity, 83.3% specificity, 81.8% PPV, and 95.2% NPV for PD, while 88.2%, 76.9%, 71.4%, and 90.9% for APS. The image-based algorithm had a high specificity and the predictive values in classifying patients before a final clinical diagnosis was reached by a specialist. Our data suggest that it may improve the diagnostic accuracy by 10–15% in PD and 20% in APS when a movement disorder specialist is not easily available.