SPECTRA-ICA
A team-built EEG artifact remover that strips only artifact-characteristic frequencies during detected events, and beat plain ICA on all 7 P300 metrics.
Standard EEG cleaning throws away a whole independent component when it thinks that component carries a blink. The component also carries alpha and beta, so the neural signal inside it goes too. In the benchmark, binary ICA came out at 0.499 OCI, about half the neural power in the 1 to 30 Hz band removed, on average.
The method we built does two things instead. It gates in time, detecting blinks, muscle bursts, heartbeats and channel noise per component, and it gates in frequency, estimating a per-component weight from the STFT of that component’s own artifact windows against its own clean windows. Removal is then scaled by the ICLabel probability, so a component nobody is sure about is barely touched.
Across 10 subjects on a P300 oddball BCI benchmark it beat standard ICA on all 7 metrics, with balanced accuracy 0.6270 to 0.6505 (+2.35 pp), AUC +1.89 pp, Cohen’s d of 0.703, and 6 of 10 subjects improved. The spectral path switches itself off for muscle components because their profiles did not survive testing. Working with Team DMLS, it took first place in the machine learning category at SURGE NeuroHack 2026, and the repository has not moved since.
- 1 Temporal gate Blink, muscle, cardiac and channel-noise events detected per component.
- 2 Spectral profile Each component's own STFT decides which frequencies are artifact.
- 3 Gated subtraction Removal scaled by the ICLabel probability instead of a binary cut.
- 4 10-subject LOPOCV 7,200 P300 epochs, Euclidean Alignment then logistic regression.
Overview
What I built
- +Beat standard ICA on all 7 P300 BCI metrics across 10 subjects, Cohen's d = 0.703, a medium effect.
- +Found 28 more P300 targets and produced 11 fewer false alarms than binary ICA exclusion.
- +Type-specific event detectors and a per-component spectral weight, with no hardcoded frequency bands.
What I rebuilt
- ~Dropped the spectral path for muscle components after it measurably degraded neural preservation.
- ~Grew the sample benchmark from 6 subjects to the full 10-subject leave-one-person-out result.
Known limitations
- !The spectral path needs at least 5 STFT frames of artifact; fewer than 5 blinks or heartbeats falls back to broadband.
- !The Gini quality threshold of 0.05 was never tuned systematically against the full dataset.
- !One dataset only, so nothing is known about other EEG hardware, electrode layouts or clinical populations.
Decisions
- 1. Gate in time and frequency I chose remove only artifact-characteristic frequencies, only inside detected artifact windows, Instead of subtracting whole artifact components, as binary ICA does, Because binary exclusion removes artifact energy in every sample and scored OCI 0.499, roughly half the neural power gone..
- 2. Subtract the uniform gain first I chose subtract the mean per-frequency power ratio before normalising it into a weight, Instead of using the raw artifact-minus-clean power difference, Because a blink lifts the whole component's amplitude, so the raw difference marks alpha and beta as artifact too..
- 3. Derive the bands, do not write them I chose estimate a continuous weight per component from its own STFT, in seconds off the sampling rate, Instead of Bailey 2025's hardcoded delta and high-gamma bands, Because no other method in the benchmark restricted removal per component using frequencies derived from that subject's data..
Timeline
| Version | Date | Description |
|---|---|---|
| v0.1.0 | 2026-03 | Implementation, proof script and the report with the full 10-subject LOPOCV results. |
| v1.0.0 | 2026-03 | Competition report dated 22 March, same day as every code commit. |
| v1.0.0 | 2026-05 | NeuroHack slides added to the repository, and the NeuroHack PDF deleted. |