Abstract
Smell loss is a frequent and early manifestation of Parkinson's disease (PD), serving as a sensitive - albeit nonspecific - clinical biomarker(1). The notion that PD causes odour-selective hyposmia has been debated for three decades. Previous studies have used healthy controls as the comparator; this is problematic given the majority presumably display normal olfactory function. Using University of Pennsylvania Smell Identification Test data from the Parkinson's Progression Markers Initiative, we trained eight machine learning models to distinguish 'PD hyposmia' (n = 155) from 'non-PD hyposmia' (n = 155). The best-performing models were evaluated on an independent validation cohort. While specific responses (e.g. mistaking pizza for bubble gum) were impactful across models, at best only 63% of PD cases were correctly identified. Given we used a balanced data set, 50% accuracy would be achieved by random guessing. This suggests that PD-related hyposmia does not exhibit a unique pattern of odour selectivity distinct from general hyposmia.