Cross-validation retained disjoint precision links in independent contact controls from 36 SEEG patients

Five-fold cross-validation retained disjoint precision links in independent contact controls from 36 public SEEG patients. The main test used 512 independent training samples. On average, 98.3% of disjoint pairs had an absolute fitted partial correlation of at least 0.05. The planned grid included zero penalty. A check added after inspection excluded zero and retained 94.9%. The model predicted separate test samples better than a model with independent bipolar outputs. The main mean gain was 0.222 nats per channel and sample. An exact four-contact calculation gave a removal penalty of 0.134 for equal variances. Source independence was imposed in these Gaussian controls. The results do not estimate false links in recorded neural signals. The package contains source records, covariance inputs, code, fold scores, selected models, checks, and reproduction instructions.

graphical lasso · precision matrix · partial correlation · bipolar subtraction · stereoelectroencephalography · SEEG · Gaussian controls · cross-validation · OpenNeuro