Alzheimer's and Parkinson's diseases are the two most prevalent neurodegenerative disorders, and they share both symptomatic and pathophysiological features. Yet the precise extent and nature of their overlap remains poorly characterized. In a new preprint, Yuji Zhao and Boris Gutman propose an explicit, interpretable, and spatially adaptive transfer learning framework to elucidate morphometric biomarkers common to both conditions.
The model was applied to two widely used open-source neuroimaging datasets: the Alzheimer's Disease Neuroimaging Initiative (ADNI) and the Parkinson's Progression Markers Initiative (PPMI). The cross-disorder classifier revealed neuroanatomical alteration patterns that align with earlier neuroimaging research and reflect patients' clinical presentations. More importantly, the spatially adaptive classifier trained with a shared prior achieved statistically superior diagnostic performance compared to models trained on each disease individually.
The authors argue this demonstrates the value of cross-disorder learning for modeling limited-size datasets across disease domains. As a single preprint, the findings have not yet been independently replicated, but they suggest that borrowing information across related neurodegenerative conditions could improve diagnostic tools where data are scarce.