Pulse variance correction reduced three component 50 to 250 ms SEEG prediction error in two of five patients with separate stimulation shafts

This study tested pulse variance subtraction before time basis learning in public SEEG data from OpenNeuro ds004457 version 1.0.2. Three time patterns learned from other stimulation shafts predicted five unused pulses from five training pulses. The primary window was 50 to 250 ms. All 252 complementary pulse groups were tested at 1080 links in five patients. Correction reduced error in two patients and increased error in three. The equal patient mean relative improvement was 0.11%. Descriptive 95% patient resample limits were -2.27% and +3.27%. This result does not establish a general benefit. The separate before-pulse check improved by 18.11% and remains secondary. The matrix proof requires independent pulses with fixed eligibility, weights, and scales. It does not guarantee better prediction. A fresh raw rebuild matched all 3477760 waveform values. Separate validation checked every main split loss. The package preserves code, pulse arrays, source hashes, fitted matrices, results, and reproduction instructions. This is new work on a reused cohort.

SEEG · pulse variance correction · signal prediction · intracranial EEG · low-rank time basis