Fig. 2: Spatiotemporal encoding procedure and classification performance.
From: Deep learning of conversation-based ‘filmstrips’ for robust Alzheimer’s disease detection
a Schematic depiction of four speech acts in the \((x,y)\) matrix. b At each temporal increment ε, the cell (highlighted in yellow) corresponding to the topological position at that moment is selected, generating a sequence of “snapshots”. c The snapshots are then concatenated to form a “filmstrip” (horizontal axis) that simultaneously encodes topological structure \((x,y)\) and kinetic progression ε. d Results (percentage accuracy) comparing the “Experimental” condition (real dataset: AD vs. HC) and the “Control” condition (artificially mixed groups). e Boxplots illustrating the model’s sensitivity (blue) and specificity (orange) for the same comparison. The higher scores for the experimental condition confirm the algorithm’s ability to effectively distinguish AD patients.
