A single-channel EEG classification system for multiscale characterization of mouse vigilance state

Abstract

Long-term analysis of mouse sleep is constrained by the dependence of conventional scoring on expert interpretation of electroencephalographic (EEG) and electromyographic (EMG) recordings. We developed a channel-agnostic, EEG-only framework that combines cross-animal sleep-stage classification, causal temporal organization, and probabilistic hypnodensity analysis from a single cortical EEG signal. Motor, somatosensory, and visual cortical recordings were treated independently by a convolutional-recurrent classifier, and generalization was evaluated using nested leave-one-mouse-out cross-validation in eight mice, with each test animal excluded from training, normalization, and model selection. The primary model achieved 0.897 {+/-} 0.058 accuracy and 0.856 {+/-} 0.076 macro-F1 across previously unseen animals while preserving the principal features of expert EEG/EMG-supported sleep architecture. Causal temporal smoothing reduced fragmented predictions and restored physiologically coherent episode durations, counts, and transition structure. Beyond categorical staging, the 4-s causal EEG window was advanced in 1-s steps to generate continuous Wake, NREM, and REM hypnodensity profiles. This representation preserved overall classification performance while revealing increased probability ambiguity and state mixing around expert-defined sleep transitions. The framework was subsequently deployed without supervised adaptation in six additional mice with 32-33 recorded days per animal, where it retained organized daily sleep architecture and probabilistic sleep structure over extended recordings while remaining sensitive to changes in recording conditions. Together, these results establish a single-channel EEG framework for robust cross-animal sleep staging, physiologically structured long-term analysis, and second-by-second characterization of sleep-state probabilities in mice.