Michigan State Beats Deep Networks Detecting Seizures With Simple Math
A 55-patient iEEG study shows that topological features plus dimensionality reduction hit 80% accuracy, and classical models keep pace with deep nets.
ProTakeaways
- TDA features hit 80% balanced accuracy on 3-class iEEG seizure classification across 55 patients.
- Logistic regression on Carlsson coordinates reached 79.17%, matching deep networks.
- Dimension-reduced pipelines beat channel-preserving ones (80% vs 59.5% max accuracy).
- Factor analysis and t-SNE were the strongest dimensionality reduction choices.
- Topological features consistently outperformed entropy-based baselines like Renyi and Tsallis.
- Data from the public HUP iEEG dataset; no code release linked.
A team from Michigan State and Tampere University has taken one of the more exotic branches of applied mathematics, topological data analysis (TDA), and stress-tested it on a genuinely hard clinical problem: telling apart interictal, preictal, and ictal brain states from intracranial EEG recordings. Their new paper runs a large ablation over feature representations, dimensionality reduction methods, and classifiers, then lands on a finding that will feel counterintuitive to anyone reflexively reaching for a deep network.
Seizure detection from EEG signals remains hard because neural activity is high-dimensional and nonlinear, with likely stochastic components. The authors ask whether features derived from topological data analysis can improve classification of preictal, ictal, and interictal iEEG in multichannel recordings from epilepsy patients. Scale matters here: the study covers 55 patients, well above the typical patient-specific setups in prior work.
Why shape beats spectrum
Traditional EEG pipelines lean on frequency bands, wavelet coefficients, and entropy measures. TDA instead asks a geometric question: as you sweep a threshold across a signal, when do peaks and valleys appear and merge? Each such event becomes a (birth, death) pair, and the collection of pairs forms a persistence diagram, a compact fingerprint of the signal's shape across scales. Because those diagrams live in an awkward space for standard ML, they get vectorized. The paper compares three popular options: Carlsson coordinates (five hand-crafted polynomial summaries), persistence images (Gaussian smoothing over the diagram), and template functions (tent-shaped or polynomial basis functions over the birth-death plane).