Tuning within records gave lower mean SEEG covariance shape error than tuning across patients in 36 patients at 1–80 Hz

We tested SEEG covariance shape prediction with public OpenNeuro ds004100, version 1.1.3. The test used two interictal records from each of 36 patients. One rule tuned a shrinkage weight within the training record. The other rule tuned a weight with separate records from the other 35 patients. Each test patient was excluded from that fit. At 1–80 Hz, mean prediction loss was 3.177384 for tuning within records and 3.315266 for tuning across patients. The mean gain from tuning across patients was −0.137883. Its 95% patient bootstrap interval was −0.434096 to 0.019341. The interval includes zero and does not establish a population difference. Tuning within records reduced mean loss by 7.25% relative to raw covariance shape. It helped 28 of 36 patients. One patient supplied 58.3% of the net gain. The result applies to short fixed excerpts from one selected epilepsy cohort. It does not establish clinical benefit. All 72 fresh source excerpts rebuilt the archived covariance values exactly. The package preserves data, code, source versions and hashes, patient results, bootstrap draws, figures, and reproduction instructions. Complete repeats from archived and newly rebuilt inputs passed.

stereoelectroencephalography · SEEG · covariance · shrinkage · signal processing · cross-validation · prediction loss