Twu31/ssvep-jfpm-12target

SSVEP calibration recordings — 12-target JFPM, 10 parietal/occipital channels

Calibration sessions from a self-built SSVEP BCI aimed at short decoding windows. Single participant (the author), consenting to public release. No personal identifiers are present in the files — the metadata holds only stimulus and recording parameters.

Sessions

File	Trials	Classes	Samples/trial	Frequencies
ssvep_20260725_150031.npz	120	12	750 (3.0 s)	8.0–15.7 Hz
ssvep_20260725_164720.npz	120	12	500 (2.0 s)	8.0–15.7 Hz
ssvep_20260725_171821.npz	60	4	500 (2.0 s)	8.571–15.0 Hz
ssvep_20260725_160927.npz	3	1	750 (3.0 s)	8.0 Hz (smoke test)

Format

Each .npz:

Key	Shape	Meaning
X	(trials, channels, samples)	float, microvolts
y	(trials,)	target index
filled	(trials,)	fraction of samples reconstructed by interpolation — see below
fs	scalar	250.0 Hz
freqs, phases	(targets,)	stimulus frequency (Hz) and phase (rad)
names	(targets,)	command label
ch_names	(10,)	P7 P3 PZ P4 P8 O1 OZ O2 PO3 PO4
meta_json	scalar	full session parameters as JSON

A matching .json carries the same metadata in plain text.

import numpy as np
d = np.load("ssvep_20260725_150031.npz", allow_pickle=True)
X, y, fs = d["X"], d["y"], float(d["fs"])          # (120, 10, 750), (120,), 250.0
clean = d["filled"] < 0.02                          # trials worth trusting

Stimulus

Joint frequency-phase modulation: 12 targets at 8.0–15.7 Hz in 0.7 Hz steps, phases stepping by 0.5*pi. The spacing is chosen so no target's second harmonic lands on another. Luminance was computed from the wall clock rather than a frame counter, so a dropped frame costs amplitude but never shifts frequency or phase — which matters here because JFPM carries information in the phase. Trials are phase-aligned at recording onset, as the public benchmarks do, so they can be averaged into templates.

filled — read this before using a trial

Acquisition is over WiFi/UDP at 250 Hz. A lost packet means a missing sample, and simply skipping it would compress the time axis: every SSVEP decoder assumes sample i sits at i/fs. At 250 Hz one lost frame is 4 ms, which is 14 degrees of phase at 10 Hz, and the error accumulates within a window.

Missing samples are therefore reconstructed by linear interpolation so the sample index stays true to real time, and filled records what fraction of each trial was reconstructed. Filtering to filled < 0.02 is a reasonable default.

Caveats

• One participant, one day. Nothing here supports claims about across-subject performance.
• Dry electrodes. Occipital contact is the known weak point of this rig; some trials are noisier than others, which is what filled and bad_trials in the metadata are for.
• Recorded with a 32-channel ADS1299 system; only the 10 parietal/occipital channels relevant to SSVEP are retained.

Related

Stimulus and decoders (CCA / FBCCA / eTRCA / TDCA): https://github.com/twu3202/EEG_Drone Rokid Glasses display measurements and AR stimulus: https://github.com/twu3202/rokid-glasses-ssvep