ExSEnt for explainable dementia detection: disentangling temporal and amplitude-driven complexity boosts EEG-based classification

Early detection of dementia enables timely intervention and better care planning. Electroencephalography, being accessible and noninvasive, offers a practical avenue for monitoring pathological alterations in neural activity. Classical biomarkers like the theta-to-alpha power ratio TAR along with complexity measures, are common methods that are usually evaluated and used for dementia detection. In this study, we aimed to assess the discriminative ability of a novel entropy-based family of measures, Extrema-Segmented Entropy (ExSEnt), summarized by multiple robust statistics per subject for dementia, along with classical measures, and evaluate the incremental value of these metrics. We analyzed an EEG dataset comprising healthy controls, individuals with Alzheimer's disease, or frontotemporal dementia. Following preprocessing and group-level analyses of independent components, we focused on source-space activity from the prefrontal cortex and visual association cortices, regions implicated in early disease. From these sources, we computed complexity metrics: Sample Entropy, Katz Fractal Dimension, Higuchi Fractal Dimension, and Hurst exponent and ExSEnt metrics along with TAR and band-limited power at delta, beta, low, and high gamma bands. Using stability-based selection with elastic net logistic models, we identified a reliable set of discriminative features and quantified their cross-subject robustness. This framework isolates interpretable and trustworthy source-local biomarkers from single-region time series. We observed that the alpha/theta temporal entropy measures (ExSEnt) are selected as the most reliably informative metrics in the left prefrontal cortex, yielding a classification performance comparable to what was recently reported with high-dimensional deep learning methods for this dataset, with a simple logistic regression model on a single brain source.