Multichannel EEG is a firehose, and squeezing it down to a compact representation before a classifier sees it is standard practice for seizure detection. Which squeezing method, though? This paper runs a head-to-head between four: PCA, Dynamical Component Analysis, Dynamic Mode Decomposition, and Average Volatility Dimensioning, all on the Temple University seizure corpus.

To keep the comparison fair, an autonomous AI framework tunes each method's classifier architecture and hyperparameters separately rather than assuming one setup fits all. The verdict favors the variance-based methods: AVD and PCA land at about 88 and 86 percent ROC-AUC, more than ten points above the dynamics-based DMD and DyCA, with AVD also generalizing most cleanly from validation to test. Tellingly, the best classifier shape changed depending on the representation, which is the paper's argument for tuning the two together.

It's one corpus and one task, so the paper is worth reading before generalizing the ranking.