Adaptive data compression and reconstruction for memory-bounded EEG continual learning

Electroencephalography (EEG) signals provide millisecond-level temporal resolution, but their analysis is limited by substantial noise and inter-subject variability, making robust personalization difficult when annotations are scarce. Unsupervised Individual Continual Learning (UICL) addresses this practical challenge: a model pretrained on a labeled cohort must adapt online to unlabeled subject streams under strict memory constraints. However, existing UICL methods typically store full past samples, which weakens the continual-learning goal of avoiding expensive retraining and unbounded storage. Motivated by the well-structured morphologies present in EEG, we propose A daptive Da ta Co mpression and Re construction (ADaCoRe) for UICL. ADaCoRe is a memory-efficient pipeline composed of saliency-driven keyframe protection, rational polyphase compression, adjoint reconstruction with verbatim overwrite on protected indices, and prototype-confidence selection for adaptive exemplar maintenance. Across three representative benchmarks and five repeated subject orders for the primary BrainUICL comparison, ADaCoRe improves ACC by , , and points on ISRUC, FACED, and PhysioNet–MI, respectively, while using only a compact replay buffer. Targeted ablations and diagnostics quantify component effects, memory usage, replay fidelity, and deployment cost, and visualizations confirm that key EEG morphology is preserved during compression and reconstruction.