Gaussian EM fitting did not establish lower prediction loss than pairwise fitting in 13 SEEG patients at 20 percent shrinkage

We tested whether expectation maximization (EM) on incomplete SEEG trials improves channel feature prediction. We used OpenNeuro ds006136 version 1.0.1, with 143 channels in 14 records from 13 epilepsy patients. We sorted source trial IDs and fitted on the first 65 percent. IDs were assumed to give trial order. The test hid each available channel value in turn. The primary feature was early delay gamma power change from fixation. Both rules used shared fit scales and 20 percent shrinkage. Mean Gaussian prediction loss was 1.376628 nats for EM and 1.375759 for repaired pairwise correlations. The mean gain was -0.000868 nat. Its descriptive 95 percent patient bootstrap interval was [-0.003762, 0.001626] nat. Six patients had a positive gain. The exact sign-flip reference gave p = 0.5544. This test did not establish an EM benefit or equal performance. Both joint models had lower mean loss than independent channels. Longer fitting changed the primary group gain by less than 0.0000001 nat. The test concerns values that passed source rejection. It does not establish recovery of rejected signals, live decoding, neural connections, or stimulation effects. The package has three figures, all derived inputs, fitted models, conditional predictions, source hashes, a signal retrieval recipe, and code. Features rebuilt from fresh source signals matched exactly. A full replay matched all scores and 1,316 model arrays. Independent checks tested conditional predictions and separation of fit and test values.

stereoelectroencephalography · missing data · conditional prediction · Gaussian expectation maximization · pairwise correlation · working memory