Machine Learning Framework for Integration of Video and EEG Reveals Distinct Seizure Phenotypes in Mouse Models of...

Video-electrocorticogram (VEEG) recordings are the gold-standard diagnostic tool for detecting and characterizing seizures in rodent models of seizure and epilepsy. Manual/visual scoring remains the prevailing strategy despite significant limitations. Alternatively, automated and semi-automated analyses have been applied to either EEG or video features alone but integrating these disparate data forms presents a challenge that has not been sufficiently addressed. Here, we introduce an integrated machine learning pipeline that jointly leverages EEG and behavioral video data to enable scalable and objective event classification. Integrated video-EEG analysis was initially validated using a well-established pharmacologically induced seizure paradigm in male mice, where overt generalized motor seizures provide a clear benchmark. We then tested the pipeline on a genetic mouse model of developmental and epileptic encephalopathy (DEE) that displays a complex spike-wave discharge (SWD) seizure phenotype that poses significant challenges for traditional scoring methods. Machine learning-powered integrated analysis of video-EEG recordings outperformed unimodal EEG- or video-based approaches and performed favorably to experienced rodent EEG electrophysiologists while also uncovering distinct phenotypic relationships, including classes of atypical behavioral and typical non-behavioral SWD events. This framework provides a flexible tool for high-throughput, precise analysis of VEEG data, with potential to accelerate discovery in epilepsy and seizure research. Furthermore, this approach represents a generalizable strategy for the integration of diverse neurobehavioral time series data types.Significance Statement Seizure detection in animal models of epilepsy has relied heavily on manual analysis of video-EEG recordings, a process that is slow and subjective. We present a machine learning pipeline that integrates EEG and video data from mice to classify seizures with high accuracy. This work was validated across two models, first a pharmacologically induced generalized seizure model, and second, a spontaneous spike-wave discharge (SWD) model in genetic epilepsy. Our approach detected events missed by human reviewers and revealed previously uncharacterized SWD subtypes distinguished by their behavior. By combining open-source tools in a flexible framework, this work lowers the barrier to high-throughput, unbiased seizure phenotyping, and also offers a broadly applicable strategy for linking neural activity to behavior in neuroscience research.