InFACT: Bridging Visual Variance With Semantic Invariance Via Informed Adapter Tuning for Cross-Subject Video Seizure Detection

Timely seizure detection is vital for patient intervention. Video-based approaches offer a non-contact, patient-friendly alternative to traditional EEG monitoring. However, their effectiveness is fundamentally hindered by the significant visual variance of seizure behaviors both within and across patients. While existing methods attempt to mitigate this via complex feature extraction, they struggle to eliminate the impact of variance, particularly under drastic appearance shifts caused by complex poses or environmental clutter. Crucially, we observe that although pixel-level appearances fluctuate, the clinical semantics of seizure actions can remain invariant across patients sharing similar semiological patterns. To this end, we propose InFACT (Informed Fusion Adapter with Consistent Text), a dual-branch framework that treats text as a semantic anchor to calibrate visual representations. Specifically, an Informed Fusion Adapter fine-tunes a VideoMAE, a pretrained video masked autoencoder, specializing generic motion priors into seizure-specific pathological patterns under textual guidance. Simultaneously, a Consistency-aware Text Encoder extracts stable behavioral descriptions from a Multimodal Large Language Model, filtering out task-irrelevant visual noise. Furthermore, a cross-modal contrastive learning mechanism aligns divergent visual features with these invariant semantic anchors, ensuring tight clustering of seizure classes regardless of patient identity. Experimental evaluations show that InFACT achieves 86.1% sensitivity, 89.6% specificity on the public SAH-ZU dataset, and 94.3% sensitivity, 96.5% specificity on our in-house FAH-USTC dataset, yielding consistent performance improvements over state-of-the-art methods and proving its robustness against cross-subject variability.