Three-bit SEEG band-power codes kept mean decoding AUC within 0.003 of float in two offline tasks; 2-bit non-inferiority was shown only in exploratory recall

An implanted BCI must send its features through a low-power wireless link. I measured the decoding loss when each SEEG log band-power feature is coded with 1-8 bits instead of a 32-bit float. Primary set (pre-registered): OpenNeuro ds004752, 15 patients, set size 8 vs 4, 70-280 features per patient, archived features of dd73fe97. Exploratory set: ds004789 FR1, 19 patients, recalled vs not, 704-1384 features, archived features of c4b8c884. Decoder: shrinkage LDA, stratified 5-fold CV, quantizers fit on the training fold only. Primary quantizer: equal-probability bins with bin-mean code values. Results, ds004752 (float AUC 0.705): 3 bits lost 0.003 AUC (95% CI -0.018 to 0.014). 2 bits lost 0.013; non-inferiority at margin 0.02 was not shown (p = 0.31). 1 bit lost 0.027 (p = 0.28). Results, FR1 (float AUC 0.611): 3 bits lost 0.002 (CI -0.009 to 0.006). 2-bit non-inferiority was shown (p = 0.004). 1 bit gained 0.014 (p = 0.34). A uniform +/-3 SD quantizer at 2-3 bits did not lose AUC in either set. At the median 265 features, 3-bit codes need 795 bits per decision instead of 8480 (11x less). Limits: offline, one session, 17-35 trials per patient in ds004752, the quantizer is fit within the session, one decoder, no entropy coding. Thus 3-4 bits is a safe default for scalar-coded band power; 1-2 bits must be checked per task.

SEEG · brain-computer interface · band power · quantization · decoding · AUC