Abstract
The assessment of cognitive deficits is pivotal for diagnosis and management in patients with parkinsonisms. Low levels of correspondence are observed between evaluations assessed with screening cognitive tests in comparison with those assessed with in-depth neuropsychological batteries. A new tool, we named CoMDA (Cognition in Movement Disorders Assessment), was composed by merging Mini-Mental State Examination (MMSE), Montreal Cognitive Assessment (MoCA), and Frontal Assessment Battery (FAB). In total, 500 patients (400 with Parkinson's disease, 41 with vascular parkinsonism, 31 with progressive supranuclear palsy, and 28 with multiple system atrophy) underwent CoMDA (level 1-L1) and in-depth neuropsychological battery (level 2-L2). Machine learning was developed to classify the CoMDA score and obtain an accurate prediction of the cognitive profile along three different classes: normal cognition (NC), mild cognitive impairment (MCI), and impaired cognition (IC). The classification accuracy of CoMDA, assessed by ROC analysis, was compared with MMSE, MoCA, and FAB. The area under the curve (AUC) of CoMDA was significantly higher than that of MMSE, MoCA and FAB (p < 0.0001, p = 0.028 and p = 0.0007, respectively). Among 15 different algorithmic methods, the Quadratic Discriminant Analysis algorithm (CoMDA-ML) showed higher overall-metrics performance levels in predictive performance. Considering L2 as a 3-level continuous feature, CoMDA-ML produces accurate and generalizable classifications: micro-average ROC curve, AUC = 0.81; and AUC = 0.85 for NC, 0.67 for MCI, and 0.83 for IC. CoMDA and COMDA-ML are reliable and time-sparing tools, accurate in classifying cognitive profile in parkinsonisms.This study has been registered on ClinicalTrials.gov (NCT04858893).