Under review · WACV 2027 · Applications track
A Large-Scale Empirical Study of Video and EEG Deep Learning for Automated Rodent Seizure Detection and Severity Grading
1 Institution withheld for double-blind review
What the camera sees
Twelve seconds around one seizure, cropped to the animal's cage with the box the trainer used.
What the electrode records
The same twelve seconds on the paired biopotential channel, onset marked. The cursor tracks the video frame by frame — drag anywhere on the trace to scrub both together. The discharge reaches 2.97× this channel's baseline amplitude from thirty minutes earlier.
Video and EEG are widely used to study seizures in animals. However, manual scoring remains a huge bottleneck. Making this automatic is a big challenge that few have solved for animal models. This study tackles the problem using data from twenty rats, analyzing about 24000 paired clips from 601 sessions. Experts graded the seizures from a mild Stage 2 to a severe Stage 5. Ten video models and seven EEG models were tested under strict rules to see how well they could spot seizures and grade their exact severity. For basic detection, both tools did well, while EEG was slightly better, scoring 0.974 compared to video's 0.968. Both tools also worked well on unseen rats. The gap grew when grading the exact seizure levels. In the five-level test, video scored 0.657, easily beating EEG's 0.560. The two methods also made different types of mistakes. Video usually just confused stages that were next to each other. EEG was more likely to miss seizures entirely or mix up very different stages. The size of the video model did not matter much. A tiny model did almost as well as a huge one. Instead, tracking motion over time was the real key. An additional test on the outside RodEpil dataset confirmed these results, with the model score reaching 0.979. Overall, we demonstrate that cameras offer a great, pain-free way to track animal seizures and work very well alongside typical EEG tools.
At a glance
Task suite
Detection asks seizure or not. Grading asks how severe — 3-class folds the scale into none / mild (2–3) / severe (4–5); 5-class keeps every stage. These are the stages, as the camera sees them:
Head nodding and facial clonus. The mildest graded class, and the one models most often confuse with baseline.
Forelimb clonus. Motion becomes unambiguous — this is where the camera starts to have a clear advantage.
Rearing with sustained forelimb clonus. Whole-body posture changes.
Rearing and falling, loss of postural control. The rarest class, and it carries most of the remaining error.
Why it works
Stage 3 seizure
Rearing and forelimb clonus — the behavioral signature a camera can score.
Matched non-seizure
Same animal, same cage, same length, drawn from a window overlapping no event — so appearance alone cannot separate the classes.
Where the evidence is · Stage 5 seizure
Class-evidence map (HiResCAM) from X3D-M, the grading champion above, for the class it predicts: severe, at p=1.000. The red core ignites on the convulsing animal through the clonic burst, tracks it, and fades as the animal recovers — evidence localized in space and in time.
And where it is not · matched control
The same map, same class, same model, on the duration-matched control from the same animal and session — rendered on one shared color scale. Its severe-evidence peaks at 6% of the seizure's reference, so the clip plays essentially unheated: nothing in the calm cage reads as seizure evidence. Maps are temporally smoothed for display.
Leaderboard
Matched 12-epoch budget, identical cohort and split. Macro-F1 as mean ± sd over five seeds. Click a column to sort.
| Backbone | Detection | 3-class | 5-class |
|---|---|---|---|
| SlowFast-R50 | 0.973 ± .003 | 0.789 ± .009 | 0.675 ± .008 |
| MViTv2-S | 0.972 ± .001 | 0.786 ± .005 | 0.660 ± .016 |
| R(2+1)D-18 | 0.971 ± .001 | 0.783 ± .002 | 0.667 ± .010 |
| X3D-M | 0.969 ± .002 | 0.794 ± .005 | 0.586 ± .003 |
| MViTv1-B | 0.969 ± .002 | 0.779 ± .011 | 0.653 ± .008 |
| S3D | 0.967 ± .002 | 0.764 ± .004 | 0.636 ± .009 |
| Video Swin-S | 0.966 ± .003 | 0.771 ± .009 | 0.645 ± .011 |
| Video Swin-T | 0.963 ± .001 | 0.766 ± .008 | 0.632 ± .002 |
| VideoMAE-B | 0.958 ± .004 | 0.729 ± .030 | 0.636 ± .017 |
| TimeSformer | 0.945 ± .001 | 0.735 ± .013 | 0.601 ± .009 |
| Model | Detection | 3-class | 5-class |
|---|---|---|---|
| TCN | 0.978 ± .001 | 0.767 ± .007 | 0.563 ± .012 |
| GRU (reference) | 0.974 ± .003 | 0.750 ± .010 | 0.560 ± .005 |
| EEG Conformer | 0.972 ± .001 | 0.729 ± .010 | 0.523 ± .005 |
| LSTM | 0.971 ± .008 | 0.698 ± .053 | 0.468 ± .052 |
| EEGNet | 0.923 ± .012 | 0.620 ± .015 | 0.409 ± .039 |
| XGBoost | 0.914 ± .000 | 0.642 ± .000 | 0.466 ± .000 |
| Random Forest | 0.852 ± .001 | 0.529 ± .001 | 0.332 ± .002 |
External validation
On the public RodEpil benchmark the same backbone reaches 0.979 subject-wise F1. Getting there required finding something first.
The raw release leaks its labels
Trained on the raw clips, the model scores 0.9996 — essentially perfect, across every fold. That is not skill: the class label is encoded into low-level clip encoding, so the network can read it without watching the animal. Re-encoding removes the shortcut; the honest score is 0.979. Anyone using that benchmark should re-encode first.
The cohort
12,228 seizure clips and 12,344 duration-matched non-seizure clips, mean 77.0 s — each event plus a 10 s buffer either side. Training uses 19,208 clips against 5,289 held out, split so no session appears on both sides; the generalization tables additionally hold out whole animals.
A caveat that bounds every EEG number: 19 of the 20 subjects carry a single ECG lead rather than EEG. The electrode arm is a strong biopotential reference, not multichannel electroencephalography.
@article{anonymous2026seeingseizures,
title = {Seeing Seizures: A Large-Scale Empirical Study of Video and EEG
Deep Learning for Automated Rodent Seizure Detection and
Severity Grading},
author = {Anonymous},
note = {Under review, WACV 2027. Citation updated on de-anonymization.},
year = {2026}
}