Alzheimer’s disease classification using dynamic geometric pattern map: A novel image representation of EEG signals

Abstract

Electroencephalography (EEG) offers a non-invasive and low-cost basis for computer-aided dementia detection, but existing systems face limitations: data-dependent feature representations, deep models that process channels independently and discard inter-channel spatial topology, and optimistic accuracy from segment-level k-fold cross-validation (CV) that may collapse under subject-independent validation. This study introduces the Dynamic Geometric Pattern Map (DGPM), a parameter-free image representation of multichannel EEG. For each channel, fifteen geometrical features are extracted from the second-order difference plot (SODP) to quantify nonlinear dynamics without embedding, delay, or decomposition parameters. Features from all 19 channels are arranged into a single image preserving the 10–20 spatial arrangement. DGPM images were classified with five ImageNet-pretrained backbones on three binary tasks (AD vs. NC, FTD vs. NC, and AD vs. FTD) under 10-fold and leave-one-subject-out (LOSO) CV using 5-second segments. Under LOSO CV, accuracies were 69.23%, 75.00%, and 64.41%, respectively, with bootstrap 95% confidence intervals spanning approximately 12% on either side, whereas all binary accuracies exceeded 96% under 10-fold CV, revealing 24%–35% optimism from segment-level partitioning. Three-class LOSO accuracy reached 60.23%, significantly above uniform chance and majority-class rates. DGPM significantly outperformed five time–frequency image representations (Wilcoxon signed-rank test, Holm-adjusted p < 0.001 ), while performing comparably to conventional classifiers using the same handcrafted features. Permutation importance and Grad-CAM identified temporal, central, and frontal regions as informative, confirming that network decisions were driven by DGPM cell values rather than resizing artefacts. DGPM provides a compact, parameter-free, and interpretable EEG representation, although external multicenter validation remains necessary before clinical use.