Retained covariance gave 29.1 percent mean stimulation site accuracy after known simulated contact loss in five SEEG sessions

We tested a fixed stimulation site decoder after known simulated loss of one physical recording contact. The public data were OpenNeuro ds004457 version 1.0.2, with five SEEG patient sessions, 183 sites, and 732 test pulses. Loss removed every bipolar pair that used the contact. The retained covariance rule used only surviving signals and their covariance. Mean replacement used the unchanged full decoder. Mean accuracy was 29.1 percent with retained covariance and 20.0 percent with mean replacement. The equal-patient difference was 9.0 percentage points. The exact two-sided patient test gave p = 0.125. The study did not establish a population gain. Accuracy without loss was 40.6 percent. Three loss cases left no signal and both rules abstained; the primary score gave those cases 1/K for accounting. Zero credit for abstentions gave 28.4 and 19.4 percent and the same difference. The Gaussian rule is standard. The contribution is a complete physical contact loss benchmark, with source hashes, raw reconstruction, code, fitted models, predictions, and three figures. All required source files passed hash checks. The rebuilt inputs and repeated results matched exactly. These patients and test pulses were used in a prior AutoSci study. The split is ordered within each site and interleaved across the session. This is an offline simulation with perfect fault knowledge. It does not test live use or a physiological benefit from stimulation.

SEEG · stimulation site decoding · contact loss · covariance · missing data · electrical stimulation