In eight SEEG subjects, signal-dependent quantization error explained partial recovery of movement decoding at coarse ADC steps

Question: sci 94418672 showed that 70-150 Hz (hg) movement decoding in SEEG gets worse when the ADC step is 8 uV or more, and that AUC partly recovers at coarser steps. That study did not find the cause. We tested the cause and three implant design options. Method: 8 subjects of OpenNeuro ds004473 v1.0.2, first 100 key-press trials, rest and move windows of 0.5 s, bipolar pairs, log power at 70-150 Hz and 13-30 Hz, shrinkage LDA with 5-fold CV, ROC AUC. We rounded the monopolar int16 data to steps of 8.5-136 uV (12-8 bits at +-17.4 mV) without dither (QM). Controls: independent uniform noise of the same variance (NM, an ideal subtractive dither), and the rounding error alone (QE). Results: the hg loss replicated (median dAUC -0.058 at 8.5 uV, -0.229 at 34 uV; 7/8 subjects; sign-flip p = 0.0078, Holm 0.016), and hg lost more than 13-30 Hz (Holm p = 0.016). At 136 uV, QM gave higher hg AUC than NM in 7/8 subjects (mean +0.12, Holm p = 0.012). The error alone gave hg AUC above 0.5 in 8/8 subjects (median 0.63, up to 0.88; Holm p = 0.0078). Thus at least part of the recovery does not come from neural 70-150 Hz activity. Without dither, the rounding error depends on the signal, so it can carry movement information. Design results: rounding the bipolar difference (a differential ADC) gave higher hg AUC than monopolar rounding in 8/8 subjects at 34 uV (median +0.12), but still lost 0.09 against native data. Sampling at 500 Hz did not change AUC. 4 bits per log band-power feature lost no more than 0.02 AUC in any subject. Limits: offline simulation on stored clinical data, one task and decoder, 8 subjects. T3 and T4 were added after we saw the sub-1 QM results.

SEEG · ADC quantization · movement decoding · signal-dependent quantization error · high-gamma activity · feature telemetry