Two-bit band-power features preserved offline SEEG key-press detection in eight subjects
A fully implanted stereoelectroencephalography (SEEG) brain-computer interface could calculate band power on the implant and send the resulting features over a low-power radio link. This study asked how many bits per feature retained offline key-press detection and whether, at a fixed bit budget per decision, many low-bit features performed differently from fewer high-bit features. We used archived high-gamma (70–150 Hz) and beta (13–30 Hz) log band-power features from eight subjects in OpenNeuro ds004473 v1.0.2. A shrinkage linear discriminant analysis decoder was calibrated on trials 1–100 and tested on later trials in time order. Each feature was quantized using equal-occupancy thresholds from calibration data, and ROC AUC measured cued key-press versus fixation-window classification. For the primary comparison, two-bit high-gamma plus beta features were non-inferior to float32 features under a 0.02 AUC margin (mean AUC change −0.007; exact one-sided sign-flip p = 0.027); one subject lost 0.033 AUC. One-bit features did not show non-inferiority (mean change −0.023; five subjects lost more than 0.02). At 16–32 bits per decision, two bits per feature gave the highest mean AUC. At a 64-bit budget, 16 features at four bits gave mean AUC 0.831, near 0.828 for float32 with all features; the planned contrast between 32 two-bit and eight eight-bit features was not significant (p = 0.34). Assuming one decision every 0.1 s, a 64-bit budget corresponds to 640 bit/s, about 3,100 times less than raw streaming. This was an offline analysis of cued windows in eight subjects, one session each; it did not test continuous self-paced control or across-day drift. Exact on-implant band-power calculation, no radio bit errors, and the 0.1 s decision rate were assumptions.
stereoelectroencephalography · brain-computer interface · feature quantization · band power · key-press detection · linear discriminant analysis · ROC AUC · signal processing
References
3- usesBeta power increased later-trial SEEG key-press AUC in 6 of 8 subjects sci:sha256:b76113878b22b13b8fce6b4b57cf321bb22ce16b618b55ae02c2ef42b972f816
We use its archived band-power feature files (SHA-256 checked), its LDA code and its time-ordered split. Our float32 baseline reproduces its AUC values (mean 0.828, median 0.869).
- usesIn eight SEEG subjects, 70–150 Hz decoding declined above 4 µV steps, while 13–30 Hz decoding remained stable through 16 µV sci:sha256:4798ae98a792c8b2ec8fe52fa59ed047650bafc5786cbf69b0fe1c8562b4c59a
It showed that ADC steps up to 4 uV keep 70-150 Hz decoding on the same data. We use this as the basis for our assumption that the implant calculates band power exactly; this work studies the next stage, the quantization of the sent features.
- usesdoi:10.18112/openneuro.ds004473.v1.0.2
The analysis uses archived band-power feature files derived from OpenNeuro ds004473 v1.0.2. The publication records feature-file SHA-256 values matching the prior SCI feature package.
Cited by
2 publicationsPublic catalog entries on Sciscape.
- supportsIn eight SEEG subjects, 12 of 15 decoder changes reduced trial-level AUC; four were non-inferior at a 0.02 margin
2-bit features at training quartiles were non-inferior at 0.02 (mean -0.008, CI lower limit -0.018).
- usesFour-bit weights preserved held-out AUC in an offline SEEG key-press decoder across eight subjects
That work quantized the features to 2 bits; this work quantizes the decoder weights on the same features and split.