Early-train response does not predict late-train response better than a fixed population ratio (SEEG stimulation trains)

We used archived per-channel pulse-response amplitudes from two OpenNeuro SEEG stimulation datasets: ds005169 (15 subjects, 150 responsive channels) and ds006519 (7 subjects, 113 channels). Trains were about 5 s at 43-50 Hz. We fixed the plan before the analysis. We predicted the amplitude in the last 1.0 s of each train from the first 1.0 s, with leave-one-subject-out fits. Model M0 multiplies the 0-0.5 s amplitude by a population ratio. Model M1 multiplies the 0.5-1.0 s amplitude by a population ratio. Model M2 is a linear fit on both bins. The primary hypothesis was that M1 is better than M0. The data did not support it. In ds005169, M1 was better in 4 of 15 subjects (exact sign-flip p = 0.37). In ds006519, M1 was better in 4 of 7 subjects (p = 0.67). The subject median error was 0.36-0.53 log2 units (about 30-45 %) for all models. Only 39-46 % of channels were in +/-25 %. This error is 2-4 times the split-half noise floor of the target. Thus, each channel has late-train change that the first 1.0 s does not show. A fixed population ratio (about 0.63) is as good as the early measurements, and it transfers between the two datasets. An exploratory analysis showed that in ds006519 the error decreases near the noise floor only when a bin at 2.5 s or later is the predictor. For a closed-loop stimulation BCI, this result means that dose control cannot use only a short early probe. It must monitor the response during the train. Limits: one amplitude measure (half-sum RMS, 5-18 ms), source-selected channels, fixed-parameter trains only.

SEEG · intracranial EEG · cortical stimulation · pulse response · stimulation train · prediction · closed-loop brain-computer interface · leave-one-subject-out · OpenNeuro