Abstract
INTRODUCTION: Accurate clinical diagnosis of neurodegenerative diseases remains challenging, particularly when individuals have mixed pathologies. We implemented the generalizable protein-based neurodegenerative disease artificial intelligence (GPND-AI) classifier using the NUcleic acid-Linked Immuno-Sandwich Assay (NULISA) central nervous system (CNS) panel to classify Alzheimer's disease, Parkinson's disease, frontotemporal dementia, dementia with Lewy bodies, and healthy controls, while disentangling mixed pathologies. METHODS: Proteomic and clinical information from the Charles F. and Joanne Knight Alzheimer's Disease Research Center (Knight-ADRC) and Movement Disorder Clinic were used to train and test the GPND-AI classifier. External validation was performed in a Banner Sun Health Research Institute cohort and additional Knight-ADRC samples with neuropathologically confirmed diagnoses. RESULTS: GPND-AI identified 15 proteins that achieve an area under the curve (AUC) of 0.955 and 92.3% accuracy across five diagnostic categories. In validation cohort, predicted co-pathologies significantly correlated with clinical characteristics. DISCUSSION: GPND-AI identified a 15-protein panel that accurately classifies individuals across the four major neurodegenerative diseases. Validation against neuropathology-confirmed diagnoses supports the utility of proteomics-based approaches for mapping disease-specific and co-existing neurodegenerative processes.