Automated Differential Diagnosis of Early Parkinsonism Using Metabolic Brain Networks: A Validation Study

Abstract

The differentiation of idiopathic Parkinson disease (IPD) from multiple system atrophy (MSA) and progressive supranuclear palsy (PSP), the most common atypical parkinsonian look-alike syndromes (APS), can be clinically challenging. In these disorders, diagnostic inaccuracy is more frequent early in the clinical course when signs and symptoms are mild. Diagnostic inaccuracy may be particularly relevant in trials of potential disease-modifying agents, which typically involve participants with early clinical manifestations.In an initial study, we developed a probabilistic algorithm to classify subjects with clinical parkinsonism but uncertain diagnosis based on the expression of metabolic covariance patterns for IPD, MSA, and PSP. Classifications based on this algorithm agreed closely with final clinical diagnosis. Nonetheless, blinded prospective validation is required before routine use of the algorithm can be considered. Methods: We used metabolic imaging to study an independent cohort of 129 parkinsonian subjects with uncertain diagnosis; 77 (60%) had symptoms for 2 y or less at the time of imaging. After imaging, subjects were followed by blinded movement disorders specialists for an average of 2.2 y before final diagnosis was made. When the algorithm was applied to the individual scan data, the probabilities of IPD, MSA, and PSP were computed and used to classify each of the subjects. The resulting image-based classifications were then compared with the final clinical diagnosis. Results: IPD subjects were distinguished from APS with 94% specificity and 96% positive predictive value (PPV) using the original 2-level logistic classification algorithm. The algorithm achieved 90% specificity and 85% PPV for MSA and 94% specificity and 94% PPV for PSP. The diagnostic accuracy was similarly high (specificity and PPV . 90%) for parkinsonian subjects with short symptom duration. In addition, 25 subjects were classified as level I indeterminate parkinsonism and 4 more subjects as level II indeterminate APS. Conclusion: Automated pattern-based image classification can improve the diagnostic accuracy in patients with parkinsonism, even at early disease stages.

Metabolic Imaging in Parkinson Disease

“…The diagnostic accuracy was similarly high when the model was applied to a subgroup of patients with a very short symptom duration (,2 y) (41).…”

FDG PET, dopamine transporter SPECT, and olfaction: Combining biomarkers in REM sleep behavior disorder

“…We expect that the PDRP will be especially informative, because in contrast to olfaction,31 the PDRP is a progression marker 11. Moreover, PDRP expression is useful in the differential diagnosis of parkinsonian disorders,32 whereas DAT imaging is not 33…”

18F-FDG PET in Parkinsonism: Differential Diagnosis and Evaluation of Cognitive Impairment

“…Since NeuroSTAT/3D-SSP does not cover the striatum, SPM might be preferable, especially if reads are performed by nonexperts. The current automated method based on disease-specific network expression is limited with regard to cases classified as "indeterminate parkinsonism," which accounted for 14% and 19% of patients in 2 studies (42,73). This issue seems to limit sensitivity for early APSs (73) and may also become more relevant if patients with severe structural abnormalities or alternative diagnosis are included.…”

“…Aside from multiclass relevance vector machine analysis (74), the specificity of the PET diagnoses for the APS subgroups MSA, PSP, and CBD usually exceeded 90% (as requested for a confirmatory test), whereas sensitivity was 77%-96% for MSA, 74%-100% for PSP, and 75%-91% for CBD (35,(40)(41)(42)70,71,73). Such subgroup classifications will be important for patient enrichment in future trials with disease-or pathology-specific treatments.…”

“…2) (72). Three studies used observer-independent automated statistical classifications of patients with parkinsonism that applied logistic regression based on the expression of metabolic covariance patterns (42,73) or a relevance vector machine in combination with bootstrap resampling for nonhierarchic multiclass classification (74).…”