In eight SEEG subjects, an SPRT reduced errors and windows versus fixed-N stopping but exceeded its nominal 0.05 error in five

We assessed whether a Wald sequential probability ratio test (SPRT) calibrated on trials 1–100 retained its nominal error rate on later trials in an offline, cued SEEG key-press task. We used archived high-gamma (70–150 Hz) and beta (13–30 Hz) bipolar log-power features from eight subjects in OpenNeuro ds004473 v1.0.2. A Ledoit–Wolf shrinkage LDA was trained on trials 1–100, and a Gaussian log-likelihood ratio model was fit to blocked cross-fitted scores. The SPRT (a = b = 0.05) was evaluated on trials 101 onward, with a 20-window cap and forced decisions at the end of data. Mean realized error was 0.046 above nominal; error exceeded 0.05 in five of eight subjects (median 0.082), but the primary one-sided sign-flip test did not reject (p = 0.086). Errors were asymmetric: mean miss rate was 0.148 and mean false-alarm rate was 0.043. Compared with a fixed-N rule designed for the same nominal error, the SPRT had lower realized error (mean difference −0.027, p = 0.031) and used fewer windows (mean difference −1.99, p = 0.016). Shuffling trial order did not change error (p = 0.34). These offline results from one cued task, one session, and eight epilepsy patients suggest that SPRT can reduce errors and attempts relative to fixed-N stopping, while its nominal error is not a reliable guarantee after calibration. The primary test was not significant; secondary tests were uncorrected. Trials were about seven seconds apart, sequences overlapped, and the cap and end-of-data decisions affected some subjects. Recalibration or threshold margins may be needed where an error guarantee is required.

brain-computer interface · sequential probability ratio test · stereoelectroencephalography · error rate · sequential decision · shrinkage linear discriminant analysis