Working with EEG is the fastest way I know to stop believing that biology and physics are different subjects.
You present a stimulus. Some milliseconds later a voltage deflection arrives at your electrode. Between those two events is a chain that crosses every level of description a university keeps in separate buildings: ion channels opening under an electrochemical gradient, current spreading through a volume conductor, thousands of postsynaptic potentials summing into something large enough to detect at the scalp at all. Whether you file that under chemistry, physics, or biology is an administrative question. The signal doesn't know.
The electrode is part of the experiment
Conventional EEG has a spatial resolution problem that everyone acknowledges and then mostly works around. The scalp, skull, and cerebrospinal fluid act as a volume conductor that smears the signal, so a conventional disc electrode integrates over a large and poorly-specified region. You are not measuring "activity at that spot." You are measuring a spatially low-pass-filtered version of activity across a wide area.
Tri-polar concentric ring electrodes attack this directly by computing a surface Laplacian in hardware — the ring geometry approximates a second spatial derivative, which suppresses the distant contributions and sharpens what remains. The interesting research question isn't whether the signal looks cleaner. It's whether it carries different information, and that's a question conventional averaging can't answer.
Why MVPA rather than averaging
Classical ERP analysis averages across trials and compares waveform amplitudes at chosen electrodes and latencies. It's robust and it throws away most of the structure. Information distributed across channels — where no single channel is individually discriminative but the pattern is — is invisible to it by construction.
Multivariate pattern analysis asks a different question: can a classifier decode the condition from the spatial pattern of activity? If a decoder succeeds on TCRE data where it fails on conventional data, that's evidence the electrode geometry is recovering information the standard setup loses, not merely cosmetic improvement.
Framing it as decoding rather than as description also imports the whole machine learning hygiene apparatus — cross-validation, held-out subjects, permutation testing for the null. Which you badly need, because EEG datasets are small, high-dimensional, and extraordinarily easy to overfit into a publishable-looking result.
What it pointed at
The honest conclusion from a couple of years at the rig was about the limits of my own working memory. Reasoning about the coupling from ion channels to population activity is not something a human context window does. You hold two or three levels and lose the rest.
Computation is the only tool I know for holding a structure larger than the head that built it — provided the knowledge is codified well enough for a machine to reason over. That proviso has occupied most of my time since.