Two selected pairs gave lower sound amplitude error skill than eight pairs in seven of ten offline SEEG readers at a 400 ms delay

We tested whether fewer recording pairs can reduce an offline SEEG feature stream. OpenNeuro ds004703 version 1.1.0 supplied ten people who heard speech. Each reader used 70 to 110 Hz log power, five input times, 10 Hz outputs, and a 400 ms delay. Earlier passages set pair ranks, penalties, and model weights. Later passages supplied the test scores. Two pairs gave 5.41% mean error skill; eight gave 5.35%. Skill compares squared error with a training-mean baseline. Two-pair skill fell in seven people. The mean paired change was +0.06 percentage points. Its 95% person bootstrap interval was -3.15 to +4.25 points. These results do not establish equal performance. Two pairs reduced the nominal scalar feature payload by 75% under a 32-bit value assumption. No device energy was measured. One early validation passage selected four pairs on average. Its mean later change was -1.78 points, with an interval of -5.17 to +1.27. The eight-pair control exactly matched the prior publication. Independent fits and fresh first-passage source checks passed. The package preserves all 95 feature excerpts, source records, code, models, predictions, checks, and four figures. This exploratory study tests heard sound amplitude. It does not test intended speech or brain stimulation.

stereo-electroencephalography · ridge regression · feature selection · sound amplitude · brain-computer interface