Abstract
Recent machine-learning techniques may be useful to identify subtypes with distinct spatial patterns of biomarker abnormality in the various neurodegenerative diseases. Using the Subtype and Stage Inference (SuStaIn) technique, we categorized data-driven subtypes of PD by examining the deep gray matter volume and dopamine availability and compared cardiac denervation, cognition, and motor symptoms between these subtypes. The SuStaIn algorithm revealed two distinctive subtypes, which were well replicated in an external dataset. Subtype 1 was characterized by lower dopamine availability apparent at early inferred stages, severe cardiac denervation, mild cognitive dysfunction in the early stage, and patterns suggesting accelerated motor and cognitive dysfunction associated with later stages. In contrast, subtype 2 showed patterns indicative of earlier brain atrophy, mild cardiac denervation, and severe cognitive dysfunction apparent at early inferred stages, with no significant correlation between motor and cognitive status and SuStaIn stage. These findings suggest that the machine-learning model can identify heterogeneity in PD biomarker profiles, offering insights into potential region and stage-specific patterns of biomarker abnormality and their clinical implications.