In 8 SEEG subjects, an SPRT used fewer windows and errors than fixed-N, but exceeded nominal error in 5 (p = 0.086)

A sequential probability ratio test (SPRT) can stop a brain-computer interface (BCI) after repeated attempts when accumulated log-likelihood evidence crosses a threshold. We tested whether an SPRT calibrated on trials 1–100 kept a nominal 0.05 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 8 subjects in OpenNeuro ds004473 v1.0.2. A shrinkage LDA decoder and a Gaussian log-likelihood model were fitted using calibration trials; the SPRT was evaluated on trials 101 onward, with at most 20 windows. Mean realized error was 0.046 above nominal, with error above 0.05 in 5 of 8 subjects; the exact one-sided sign-flip test was not significant (p = 0.086). Mean miss rate was 0.148 and mean false-alarm rate was 0.043. Against a fixed-N rule designed for the same nominal error, the SPRT had lower mean error (difference −0.027, p = 0.031) and used fewer windows (difference −1.99, p = 0.016). Shuffling test order did not change error (p = 0.34). These results suggest that the SPRT was more efficient than fixed-N in this dataset, but its nominal error rate was not a reliable guarantee after calibration. The study was offline and used one cued task, one session, 8 epilepsy patients, one decoder, windows from separate trials about 7 s apart, overlapping sequences, and forced decisions at a 20-window cap or end of data. The primary test was not significant; secondary tests were not corrected.

sequential probability ratio test · brain-computer interface · stereo-EEG · SEEG · shrinkage LDA · error rate · calibration · OpenNeuro