Research log

Why I care about state-space models for EEG

A quick note on using HLDS and related models to capture structure in EEG, instead of treating everything as i.i.d. features.

Writing / Why I care about state-space models for EEG
Nov 1, 2025 · eeg · state-space models · signal processing
Diagram representing a state-space modeling pipeline for EEG.

Entry

EEG is genuinely dynamical. Adjacent windows are not independent, latent state matters, and a pipeline that ignores temporal structure usually pushes the hardest scientific question out of the model and into post-hoc interpretation.

State-space models force a cleaner separation between what is observed, what is hidden, and what is drifting over time. That matters when the goal is not only prediction but also understanding whether a result is stable across subjects, sessions, and recording conditions.

I keep returning to HLDS-style models because they give me a vocabulary for debugging. If the learned state is unstable, the issue may be preprocessing, segmentation, artifact handling, or model mismatch. That is more informative than a black-box score moving up or down without any clue about what changed.

For EEG research at UF, this translates into practical choices: cleaner windowing, clearer diagnostics, and a better bridge between mathematical structure and the real biological variability that shows up in lab data.

Key takeaways

  • Temporal dependence is part of the signal, not only a nuisance to remove.
  • Latent-state models make debugging and interpretation easier than static feature pipelines.
  • Good research tooling matters because the value of a state-space model depends on reproducible preprocessing and evaluation.