Distributional Latent Pooling for Learning from Variable-Length Clinical Time Series: A Study on EEG-based Dementia Detection

Low-data EEG cohorts challenge sequence models, since variable-length recordings and limited subjects make explicit temporal modelling difficult to optimize. We study distributional latent pooling for Alzheimer’s disease (AD) versus cognitively normal (CN) classification using frozen single-channel EEG representations summarized into fixed-length subject descriptors. Under leakage-controlled subject-level LOOCV, maximum pooling outperformed an attention-based sequence model (0.692 vs. 0.446 accuracy; permutation p = 0.005), though robustness checks show this advantage is partly driven by recording-length variability and raw-extremum behavior rather than a broadly robust signal. A classical power-spectral-density baseline exceeded both (0.723 accuracy), so our contribution is the pooling-versus-attention comparison itself, not state-of-the-art performance. Frontotemporal electrodes were consistently identified as discriminative across independent analyses.