Comparison of Dimension Reduction Methods for EEG Seizure Detection Using Autonomous AI-Driven Optimization

Annika Stiehl, Vishal Kagade, Nicolas Weeger, Nicole Ille, Stefan Geißelsöder, Christian Uhl

Automated epileptic seizure detection from multichannel electroencephalography (EEG) benefits from dimension reduction to obtain compact, discriminative representations. We compare four signal-space dimension reduction methods, Principal Component Analysis (PCA), Dynamical Component Analysis (DyCA), Dynamic Mode Decomposition (DMD), and Average Volatility Dimensioning (AVD), for deep learning-based seizure detection on the Temple University Hospital Seizure Corpus (TUSZ v2.0.3). To enable a comparison of optimal combinations of representation and classifier, an autonomous AI-driven research framework independently optimizes architecture and hyperparameters for each representation. Measured by test ROC-AUC, the variance-based methods AVD (88.28%) and PCA (85.98%) paired with their respective optimal classifiers outperform the dynamics-based methods DMD (74.56%) and DyCA (74.85%) by over 10%, with AVD also showing the smallest validation-to-test gap. The best-performing classifier architecture differs across representations, indicating that representation and classifier should be optimized jointly. Our results highlight the importance of the input representation for EEG seizure detection and indicate the viability of autonomous AI-driven experimentation in biomedical signal processing.