Deep learning of conversation-based 'filmstrips' for robust Alzheimer's disease detection.

Abstract

Early detection of Alzheimer's disease remains complex and costly despite advancements in neurobiological markers. We propose an innovative approach based on the topological and kinetic analysis of verbal exchanges to distinguish patients from healthy individuals. Without requiring full transcription, we leverage a convolutional network capable of identifying discursive patterns indicative of cognitive impairments. Our experiments, conducted with 80 participants, demonstrate performance levels exceeding 95% in cross-validation, comparable to computational approaches relying on biological markers. This robust and minimally invasive methodology could be easily integrated into clinical protocols, enhancing current diagnostics. It also holds the promise of cost-effectively extending monitoring to other neurodegenerative or psychiatric diseases.