In two SEEG patients, a 15–95 ms window retained 88% and 105% of 15–295 ms site accuracy, but the 90% claim failed in one patient

We measured site-identification accuracy and mutual information after single SEEG pulses in two patients from OpenNeuro ds004457 v1.0.2, using archived single-trial features from SCI contribution 313c7b97. A leave-one-trial-out nearest-centroid decoder classified 39 sites in sub-1 and 43 in sub-3. The preregistered criterion required the lower 95% confidence bound for the 15–95 ms versus 15–295 ms accuracy ratio to be at least 0.90 in both patients. The ratios were 0.877 (95% CI 0.789–0.973; McNemar p=0.027) in sub-1 and 1.046 (0.951–1.144; p=0.43) in sub-3, so the criterion was not met. The 15–55 ms window achieved 0.323 and 0.419 accuracy, respectively, and 1.11 and 1.69 bias-corrected bits per pulse. Under the assumptions that a new pulse can start at the window end and responses do not change with repetition, its information-rate upper bounds were 20.2 and 30.6 bit/s; these are not measured rates. An exploratory analysis found that the 15–25 ms bin contributed to short-window accuracy in sub-1 and may contain residual artifact or an early neural response. Results are limited to two patients, a simple decoder, offline analysis, and no test of adaptation or overlap at fast pulse rates.

SEEG · electrical stimulation · site identification · brain-computer interface · response window · mutual information · nearest-centroid decoding · intracranial EEG