In 8 SEEG patients from OpenNeuro ds005448, two-step CCEP paths added 0.082 AUC to distance but 0.001 after site and contact N1 rates (6/8, p = 0.09)

Question: for a BCI that writes to the brain with SEEG, do two-step CCEP paths (site A causes an N1 at site B, and B causes an N1 at contact c) predict a direct N1 from A to c? Do they predict it better than distance and the N1 rates of the site and of the contact? Data: OpenNeuro ds005448 1.0.0, 8 SEEG patients (STREEF06-13), expert-checked N1 files, 335 stimulation sites, 29,788 site x contact entries 10 mm or more from the pair, 2,533 N1. We did not read raw signals. Method: per-patient logistic models; 5-fold x 4 random-entry cross-validation. We masked the test labels before we calculated any feature. T = observed fraction of two-step paths through other stimulated sites; G = the fraction that distance and node rates predict. Primary contrast (fixed before results): held-out AUC of M3 (distance, shaft, site and contact rates, G, T) minus M2 (the same without T). We used exact sign-flip tests over the 8 patients. Results: T added to distance increased AUC by +0.082 (8/8, p = 0.008). Site and contact rates increased AUC by +0.114 (8/8, p = 0.008). After the rates and G, T added only +0.0013 AUC (range -0.0006 to +0.0053, 6/8, p = 0.086) and did not decrease log loss (p = 0.36). In each T band, the rate model was almost calibrated (observed/expected 0.83-1.10). Conclusion: node propensity explains almost all of the apparent transitivity. For mapping the effect of stimulation, distance, shaft and two node rates are sufficient. Limits: one center, 8 patients, clinical 1-2 mA pulses, expert labels, only stimulated sites can be intermediate sites, and the result applies to entry completion and not to new sites. The package has the code, 30 input files with MD5 and URLs, the results, a figure and the commands to run again.

SEEG · CCEP · N1 response · two-step paths · brain connectivity · stimulation · node propensity · AUC · epilepsy