cap32-ssvep — SSVEP calibration on a 32-channel dry-electrode cap

Four SSVEP calibration sessions from one subject on a low-cost 32-channel dry-electrode cap (TI ADS1299, 250 Hz, WiFi/UDP). Recorded for a drone-control project; published because the failure mode is more informative than the accuracy.

Code and analysis: github.com/twu3202/EEG_SSVEP_Drone Motor-imagery data from the same cap: Twu31/cap32-mi-eeg

⚠ These files hold 10 posterior channels, not 32. Channel selection was applied at record time, so these sessions cannot be re-analysed with a different montage. meta_json.channels_full lists the 32 the cap was actually wearing.

Quick start

import numpy as np
d = np.load("data/ssvep_20260725_171821.npz", allow_pickle=True)
X, y = d["X"], d["y"]        # (60, 10, 500) float32 µV;  (60,) int32 target index
d["freqs"]                   # (4,) Hz          d["phases"]  (4,) rad
d["names"]                   # ('forward','back','left','right')
d["ch_names"]                # ('P7','P3','PZ','P4','P8','PO3','PO4','O1','OZ','O2')
d["filled"]                  # (60,) fraction of each trial reconstructed across UDP drops
import json; json.loads(str(d["meta_json"]))   # paradigm, timing, link and stimulus stats

Signals are raw µV, unfiltered, not re-referenced. Trials are already epoched.

Sessions

file	trials	targets	window	frequencies (Hz)	stimulus fps / jitter	frame loss
ssvep_20260725_150031	120	12	3.0 s	8.0 … 15.7, 0.7 steps	404.2 / 0.11 ms	0.087 %
ssvep_20260725_160927	3	1	3.0 s	8.0	433.1 / 0.21 ms	0 %
ssvep_20260725_164720	120	12	2.0 s	8.0 … 15.7	142.8 / 0.72 ms	0 %
ssvep_20260725_171821	60	4	2.0 s	8.57, 10, 12, 15	125.0 / 0.58 ms	0 %

Trial structure: cue 1.0 s → flicker (window) → rest 0.5 s, with a measured 0.14 s display latency already accounted for. Targets are drone commands (forward / back / left / right / up / down / rotate cw / rotate ccw / takeoff / land / flip / hover).

ssvep_20260725_164720's link_received is recorded as 0, which is a metadata bug — the signal is present. Its stimulus ran at 142.8 fps with 0.72 ms jitter, the worst of the four.

What the data shows

Decoding is only just above chance:

session	targets	best decoder	accuracy	chance	95 % CI
150031	12	FBCCA, 2.4 s	0.150	0.083	[0.10, 0.22]
171821	4	TRCA-CCA, 1.8 s	0.450	0.250	[0.33, 0.58]

Statistically above chance, practically unusable. The useful finding is how it fails.

Errors fall to low frequencies — and it is not alpha

Misclassifications piled up at 8 Hz whatever the true target. This subject's IAF is exactly 10.00 Hz, so alpha is the obvious suspect, but it is the second effect, not the cause.

CCA/FBCCA's ρ is a variance ratio, so argmax over candidates compares absolute power. Under an S(f) ∝ f^-α aperiodic background the noise floor goes as ρ_noise(f) ∝ f^(-α/2) — monotonically decreasing, so the lowest candidate wins by default. Measured slope −0.95 to −1.11, matching an independently fitted α = 1.79–2.34 to within 0.06. The 8 Hz floor is 1.7–1.9× higher than 15.7 Hz.

Alpha does sit on top: in the one confound-free session the excess over the power law peaks at exactly 10.0 Hz, matching the measured IAF — but that excess is +19.7 % against +71 % from the 1/f floor.

Derivation and measurements: docs/lowfreq_bias_report.pdf (中文).

Practical consequence: with log-spaced or high-only stimulus frequencies, or with the 1/f floor divided out before argmax, this bias largely disappears. It is a scoring artifact, not a property of the subject.

Caveats

1. 10 channels only — see the warning above.
2. Dry electrodes over the occiput seat poorly. SSVEP lives on Oz/O1/O2/POz; this is the hardest region for a dry cap, and it bounds everything here.
3. Timing is critical. At 12 Hz one cycle is 83 ms, so a 4 ms epoch-boundary error is 14° of phase; eTRCA/TDCA collapse to chance when that drifts. Dropped UDP frames are reconstructed rather than skipped, and filled marks how much of each trial that was.
4. The cap's sample clock runs +3235 ppm fast (measured later against an external generator). Negligible for band power, but a 0.4 % error rotates a 12 Hz tone by a full cycle in ~2 s.
5. Single subject, single day. All four sessions are from 2026-07-25.

License

CC BY 4.0. Please link back to github.com/twu3202/EEG_SSVEP_Drone.

Recordings are from a single consenting adult subject (the author). No clinical or identifying information is included.