In five SEEG subjects, 12-bit quantization largely preserved high-gamma features, while 10-bit quantization reduced them

We re-quantized partial session-1 SEEG recordings from five subjects in OpenNeuro ds004752 to estimate how ADC step size affects band-power feature fidelity and fixation-versus-maintenance decoding. The recordings were the first 24,000,000 bytes of five EDF files from an unversioned S3 mirror; they contained 9–18 usable trials per subject. With a ±5 mV full scale, 12-bit quantization (2.4 µV step) retained 70–150 Hz log-power correlation r ≥ 0.9 in 151/171 bipolar channels (88%); all-band decoding AUC changes were −0.025 to +0.018. At 10 bits (9.8 µV step), only 33/171 channels (19%) met this fidelity threshold, and all-band AUC decreased by 0.093 and 0.123 in two subjects. At 8 bits, the median 70–150 Hz AUC change was −0.098. A white quantization-noise model agreed with the measured error spectrum and estimated median requirements of 11.6 bits for 70–150 Hz and 9.8 bits for 30–70 Hz at ±5 mV. These results suggest a step of about 1–2.5 µV (12–13 bits at ±5 mV) for high-gamma SEEG band-power features in this sample; bands below 30 Hz tolerated 8–9 bits. Two of three pre-specified predictions were not supported. The analysis covered five subjects, partial recordings, and one task contrast; AUC estimates are variable, and the original recording’s amplifier noise served as dither. The exact immutable dataset release corresponding to the unversioned mirror files is not established.

SEEG · ADC · quantization · band power · high gamma · brain-computer interface · signal fidelity · working memory