Training with gaps increased mean later state accuracy in eight SEEG recordings with 20 percent simulated sample loss

We tested whether simulated gaps in training improve a later SEEG reader. SEEG means stereoelectroencephalography. We used eight recordings from OpenNeuro ds004473 version 1.0.2. Each trial supplied a rest window and a window before a response. We fitted readers on 1,332 earlier trials and tested them on 674 later trials. Each mask removed 100 of 500 samples. We used linear interpolation. Training mixed intact input with five repaired copies. Total training weight stayed fixed. The primary test used five new gap positions. Mean balanced accuracy increased from 67.56 to 69.97 percent. The paired gain was 2.41 percentage points. Its person-bootstrap interval was 0.16 to 5.62 points. Six recordings improved. The two-sided person sign-flip p value was 0.102. This does not establish a population gain. An early-data score-offset correction gave 70.63 percent mean accuracy in a secondary test. It left AUC unchanged. Gap training slightly reduced mean AUC. The package contains all selected voltage inputs, code, predictions, figures, source records, and a retrieval recipe. No actual stimulation, online control, or transfer between sessions was tested.

stereoelectroencephalography · sample loss · gap training · signal processing · balanced accuracy · score calibration · brain-computer interfaces