One scalar weight predicted separate-file SEEG correlation maps better than repaired link-specific weights in 36 patients

In 36 patients with epilepsy, this study compared one reliability weight for a correlation map with a separate weight for each link followed by nearest correlation matrix repair. The primary test used 1–80 Hz bipolar SEEG links between named shafts and predicted correlations in a separate recording file. Repaired link-specific weights increased group mean squared error by 4.54% relative to the scalar method (patient bootstrap 95% interval 2.77 to 6.26%; higher error in 31 of 36 patients). Three of 36 primary link-weight matrices failed the prespecified positive semidefinite limit before repair; all repaired matrices passed. Repair changed mean squared error by only 0.0000004825 and did not close the gap. Planned checks had the same mean direction except that the 30–80 Hz interval included both directions. All 72 public-source covariance rebuilds matched the archived inputs exactly. The finding applies to the specified two-block rule and these recordings; it does not show that all link-specific methods have higher error, establish neural coupling, or demonstrate clinical value.

stereoelectroencephalography · correlation matrix · shrinkage · prediction error · matrix repair · bootstrap · epilepsy