Alzheimer’s and Parkinson’s: cross-learning from MRI reveals common morphological markers

When trained jointly on the ADNI and PPMI datasets, an interpretable transfer model is better at diagnosing both diseases than separate models and highlights their shared anatomical abnormalities.

Alzheimer's disease and Parkinson's disease, the two most common neurodegenerative diseases, have common points in their symptoms as well as in their mechanisms, although we do not know precisely how far they go. Yuji Zhao and Boris Gutman propose a learning framework by explicit, interpretable and adaptive transfer in space to bring out morphological biomarkers, derived from brain imaging, related to the two diseases.

Two open bases united

The model was applied to the two major public neuroimaging databases: the Alzheimer's Disease Neuroimaging Initiative (ADNI) and the Parkinson's Progression Markers Initiative (PPMI). The cross-classifier reveals patterns of anatomical alteration consistent with the literature and with the clinical manifestations of patients.

Above all, the trained version with a shared a priori between the two diseases obtains statistically superior diagnostic performance compared to models trained separately for each pathology, which argues for this type of approach when the datasets of each disease are of limited size.