Adaptive EEG Processing · Krigolson Technologies
Adaptive EEG Processing
Clinical EEG is cleaned by a fixed sequence of steps, applied identically to every recording whatever that recording needs. About half the data ends up discarded as artifact.
One cleaning step
Forty per cent of this recording was unusable. One step of the sequence runs, and bad epochs fall from 40.6% to 1.1%. This is what the fixed pipeline is built to do.
Measured · April 2026
The same step
The same step, the same settings, a recording too corrupted for it to work on. It did not clean it. 80.3% → 99.9%. On another, 52.8% → 97.5%.
Five of thirty recordings were made worse this way
Why it happens
The same steps in the same order for every patient. The two recordings above take the identical path, and nothing anywhere in it can tell the rescue from the wreck. No step is ever asked whether it helped.
What we built
Every step is scored against the signal that went into it, and committed only if the recording actually improved. If it did not, the step is discarded and the recording goes on without it. The decision is made per recording, not per protocol.
Thirty recordings
Each recording measured, each one treated differently. In 8 of 30 the pipeline ran the tests and declined to do anything at all — the most valuable thing it does.
Across all thirty
3,347 epochs recovered · none left worse
What this does not yet prove
A pipeline that only ever keeps a lower count cannot report a higher one. Settling it takes a held-out test.
What it is for
Every percentage on this stage is measured, from thirty clinical recordings. The arrangement of epochs within a recording is illustrative.
Thirty clinical EEG recordings, nineteen channels, April 2026. Every before-and-after percentage on screen is from that run. Recordings are lettered rather than identified.
That no recording is made worse follows from the selection rule rather than from the data, and the +33.3% is the best of a large search scored on the same epochs it selected from. Both need a held-out set before they are claims.
Implemented in MATLAB, Python and C with verified numerical parity, fast enough to process a very large dataset on one workstation. Held-out validation is the next piece of work.