Three to four bits preserve SEEG feature decoding, while low ADC resolution can replace high-gamma features with quantization noise

In an offline analysis of three subjects performing cued key presses, we tested how quantization affects decoding from stereo-electroencephalography (SEEG). A shrinkage linear discriminant decoder used log 70–150 Hz power (excluding 115–125 Hz) from 103–104 adjacent-contact bipolar pairs. Quantizing each feature to 3–4 bits kept held-out AUC within 0.008 of the float-feature result in all three subjects (AUC 0.880, 0.822, and 0.819 at 4 bits; 0.884, 0.818, and 0.823 at 3 bits). The estimated feature link rate fell from about 6,600 bit/s to 620–830 bit/s at two updates per second. In simulated re-quantization of native 16-bit ADC values over the same ±15.65 mV range, 14 bits retained feature correlations of at least 0.96 with the 16-bit reference. At 12 bits, correlations were 0.58, 0.64, and 0.91; at 10 bits or fewer, they were at most 0.26. Yet at 8 bits or fewer, held-out AUC rose again to 0.76–0.89, because quantization-noise power followed large low-frequency press-related changes. Thus decoder AUC alone did not establish that a low-bit front end preserved high-gamma information. This offline study used three subjects, one time-ordered split, and simulated re-quantization; it reports no confidence intervals, and the simulated ADC omitted analog noise.

stereo-electroencephalography · brain-computer interface · signal quantization · analog-to-digital conversion · high-gamma power · linear discriminant analysis · classification AUC