Title : Decoding cognitive decline: A multimodal AI approach to distinguishing Alzheimer’s disease, Dementia with Lewy bodies, and PSP Using Tau oligomers
Abstract:
Neurodegenerative diseases affect over 55 million people worldwide, yet accurately distinguishing Alzheimer's disease (AD), dementia with Lewy bodies (DLB), and progressive supranuclear palsy (PSP) remains difficult due to overlapping clinical symptoms and limited disease-specific biomarkers, with misdiagnosis rates approaching 30%. Recent biochemical evidence suggests that distinct tau protein polymorphs may help differentiate these disorders, though the complexity of tau experimental data limits interpretation using traditional analytical methods. In this study, quantitative structural and morphological tau features extracted from previously collected, de-identified datasets were analyzed using machine-learning models. To address limited sample size, the training dataset was augmented with 5,000 synthetic samples, while evaluation was restricted to replicate-aware held-out real data. Multiple models were tested alongside a custom ensemble framework, NeuroFoldNet, which achieved a mean classification accuracy of 79.01% under K-fold cross-validation. Feature-importance analysis revealed that a small subset of tau structural measurements consistently drove accurate disease classification, suggesting that specific morphological signatures of tau polymorphs may hold greater diagnostic relevance. These findings demonstrate the potential of machine learning to interpret complex tau datasets and support biomarker-driven differentiation of related neurodegenerative disorders.
