Detecting drug-induced EEG alterations with machine learning: A methodological proof-of-concept study in rats treated with MK-801.

INTRODUCTION: Although deep learning methods for EEG analysis are rapidly advancing, architectures developed for human multichannel recordings may not be suited to animal EEG, where recordings typically involve substantially fewer channels. This study investigated whether deep learning could detect EEG alterations induced by the NMDAr antagonist MK-801 (dizocilpine), recorded in resting rats with auditory stimulation, to determine the potential of this method for brain circuit-related efficacy and safety assessment. METHODS: EEG recordings were obtained from male Sprague-Dawley rats implanted with telemetry implants studied in a randomized crossover treatment design. Signals from frontal and temporal electrodes were segmented into 0.5-s epochs and processed using a minimal preprocessing pipeline. Several candidate deep-learning architectures were evaluated on classification tasks. RESULTS: EEG Conformer provided the best balance between performance and robustness. On the recording-level test set, the final model reached an F-score close to 0.90 on the recording-level test set. In a complementary leave-one-subject-out cross-validation, it reached a mean weighted F1-score of 0.646 ± 0.034 (n = 5 folds), above the reference chance level and consistent with transfer of a learned MK-801-related EEG signature to unseen animals. In MK-801-treated rats, abnormal-class outputs increased with dose, from 0.17 at 0.025 mg/kg to 0.62 at 0.2 mg/kg. Auditory stimulation increased abnormal classification relative to non-stimulated epochs, indicating improved sensitivity under stimulation conditions. CONCLUSION: These findings support the feasibility of using deep learning to detect compound-related EEG alterations in low-channel rat recordings with limited preprocessing. Two complementary forms of generalization were assessed: transfer to unseen treatment conditions within known subjects in the crossover setting, and generalization to entirely unseen animals in the leave-one-subject-out (LOSO) setting. The lower LOSO performance is consistent with the more stringent validation conditions. Together, these results suggest that the approach may complement conventional analytical pipelines and support more automated, standardized preclinical EEG workflows for both safety and efficacy pharmacology studies.