RodEpil: A Video Dataset of Laboratory Rodents for Seizure Detection and Benchmark Evaluation

TL;DR

RodEpil presents a curated video dataset of 19 laboratory rodents with 10,101 normal and 2,952 seizure clips, designed for non-invasive, video-based seizure detection. The authors establish a robust baseline using TimeSformer, employing strict subject-wise five-fold cross-validation to prevent data leakage and achieve reliable generalization. They compare input modalities (RGB, AbsDiff, Optical Flow) and demonstrate that RGB with Kinetics pretraining yields the strongest performance (average accuracy 98.55%, F1 97.00%), highlighting the value of transfer learning and rich visual context. The dataset and baseline code are released to enable reproducible research in preclinical epilepsy, with future directions including longer temporal windows and multi-modal integration. The work advances non-invasive automated phenotyping for seizure monitoring in preclinical studies and provides a standardized benchmark for future methods.

Abstract

We introduce a curated video dataset of laboratory rodents for automatic detection of convulsive events. The dataset contains short (10~s) top-down and side-view video clips of individual rodents, labeled at clip level as normal activity or seizure. It includes 10,101 negative samples and 2,952 positive samples collected from 19 subjects. We describe the data curation, annotation protocol and preprocessing pipeline, and report baseline experiments using a transformer-based video classifier (TimeSformer). Experiments employ five-fold cross-validation with strict subject-wise partitioning to prevent data leakage (no subject appears in more than one fold). Results show that the TimeSformer architecture enables discrimination between seizure and normal activity with an average F1-score of 97%. The dataset and baseline code are publicly released to support reproducible research on non-invasive, video-based monitoring in preclinical epilepsy research. RodEpil Dataset access - DOI: 10.5281/zenodo.17601357

Figures (3)

• Figure 1: Video pre-processing pipeline
• Figure 2: Comparison of average F1-scores for models trained on RGB, Optical Flow, and Absolute Difference inputs (224x224, 8 frames, SGD, Kinetics pretrain).
• Figure 3: Average performance metrics for the best performing model configuration across 5 folds.