(8/9) Testing the spectral fit of the stimulus-unrelated ongoing EEG activity suggests the following choices:
• PS: Power-law fit (equivalent to mean fit but less biased for low frequencies).
• EPS: Mean fit.
• ITC: No fit.
(7/9) The same analysis on real data (12 adult subjects presented with an on-off checkerboard sinusoidally flickering at 0.8 Hz) fully confirms the results of the simulated data. Importantly, using the optimal sliding window length, all spectral measures accurately discriminate SSEP between stimulation and rest already with as short as 8 stimulation cycles (10 s).
(6/9) For longer data length (>20 cycles of stimulation frequency) and low SNR, the performance of EPS and ITC overcomes the one of PS.
(5/9) Simulations from a generative model matching the SNR of the real SSEP and the spectral profile of the ongoing EEG show that for short data length the three measures are equivalent, and the crucial parameter is the length of the sliding window over which each measure is computed: the longer the better for PS and EPS, whereas the opposite occurs for ITC.
(4/9) Here's an example for one subject with an on-off checkerboard oscillating at 0.8 Hz: While for 16 oscillation cycles the peak in frequency is clear, for shorter data the peak is progressively shallower and spectral profiles become very noisy (first row). At the same time, data at rest becomes increasingly noisy too, and spurious spectral peaks emerge (second row).
