A separate covariance model gave lower mean site accuracy in five SEEG records with four fit pulses and 50% shrinkage at 100–500 ms

We compared a full covariance reader with a separate recording-pair and time-bin covariance reader for identifying known electrical stimulation sites from SEEG responses. The fixed primary test used 138 eligible sites in five existing records, with four fit pulses and four later test pulses per site, a 100–500 ms response window, and 50% shrinkage. The separate model gave 15.90% equal-person mean accuracy; the full model gave 18.03%, a difference of −2.14 percentage points. A descriptive 95% whole-person bootstrap interval was −3.80 to −0.47 points, while the two-sided sign reference was 0.25. These five records do not establish a population loss. The separate covariance representation used 74% to 96% fewer entries in the four records with multiple recording pairs; the count does not measure deployed reader memory. The planned 20–500 ms check reversed the direction, favoring the separate model by 1.35 points. This is an offline method comparison on the same five clinical records used in prior work, not a new cohort or independent replication.

SEEG · electrical stimulation · covariance model · brain-computer interface · site identification · shrinkage