Chybowski, B.; Gonzalez-Sulser, A.; Escudero, J.

Study Objectives Automated sleep staging underpins clinical sleep assessment and translational neuroscience, yet most data analyses work addresses human and animal data separately. We tested whether a seizure-oriented machine learning framework can be repurposed for three-state sleep staging in humans and rats, and whether models trained solely on rodent data can be applied directly to human recordings using an explicit cross-species montage. Methods We used the PySeizure, a standardised EEG preprocessing and seizure-detection framework, together with TinySleepNet as the core classifier. Models were trained and evaluated on the Sleep-EDF expanded Sleep Cassette subset (three classes: wake, non-rapid eye movement sleep, rapid eye movement sleep), then applied without fine-tuning to the Sleep Telemetry subset. The same pipeline was used on a SYNGAP1 rat dataset with analogous three-state labels. A novel human-rat electroencephalography montage mapped rat electrodes to putative human scalp homologues, enabling direct application of rat-trained models to Sleep Cassette. Results Within Sleep Cassette, the accuracy in three-stage sleep classification was 0.95. Applying this model directly to Sleep Telemetry yielded an accuracy of 0.89. On the rodent dataset, accuracy was 0.78. When the rat-trained model was applied directly to Sleep Cassette, accuracy was 0.68. Conclusions A single deep learning pipeline can support robust three-state sleep staging in humans and rodents and retains meaningful performance under both human cross-subset and rat-to-human transfer without any retraining or fine-tuning. The rat-trained model's above-chance performance on human data, achieved without human training samples, shows that rodent-derived representations can contribute directly to human sleep staging when constrained by an anatomically informed montage, linking preclinical rodent recordings and clinical human sleep studies.