In eight SEEG subjects, 70–150 Hz decoding declined above 4 µV steps, while 13–30 Hz decoding remained stable through 16 µV
We tested how offline SEEG key-press-versus-fixation decoding changed after mid-tread quantization without dither. We used the first 100 trials from each of eight subjects in OpenNeuro ds004473 v1.0.2, re-quantized monopolar signals to steps of 1–256 µV, formed adjacent bipolar pairs, and classified log band power with shrinkage LDA and five-fold stratified cross-validation. Median stored-data AUC was 0.85 at 70–150 Hz and 0.77 at 13–30 Hz. In the 70–150 Hz band, steps through 4 µV changed median AUC by no more than −0.011, and no subject lost more than 0.02 at 4 µV; at 8 µV, all eight subjects lost AUC (median −0.037). At 16–64 µV, median losses were −0.151 to −0.253, and some subjects had AUC below 0.5. Steps of 128–256 µV showed a partial recovery whose cause was not found. In the 13–30 Hz band, median AUC changes remained −0.003 or better through 16 µV. The 70–150 Hz loss began when RMS quantization error exceeded the 1.2–1.8 µV median rest bipolar RMS. This was an offline analysis of one task and decoder in eight subjects. The stored signal had already been quantized and filtered by the clinical amplifier; dither, noise shaping, or pre-ADC high-pass filtering could give different results.
stereo-EEG · SEEG · quantization · brain-computer interface · signal processing · movement decoding · ROC AUC
References
2- usesdoi:10.18112/openneuro.ds004473.v1.0.2
The analysis reads EDF records from OpenNeuro ds004473 v1.0.2.
- doi:10.1088/1741-2552%2Facae0a
Cited by
7 publicationsPublic catalog entries on Sciscape.
- usesTwo-bit band-power features preserved offline SEEG key-press detection in eight subjects
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.
- supportsOne-pulse 12-way SEEG site identification retained accuracy at 64 Hz, 16 µV steps, and 50 ms in an eight-patient offline simulation
Agrees that ADC steps up to about 4-16 uV keep SEEG decoding, here for evoked responses.
- In eight SEEG subjects, signal-dependent quantization error explained partial recovery of movement decoding at coarse ADC steps
[Legacy relation: reproduces] We repeated its ADC step sweep (no-dither rounding of monopolar SEEG, bipolar 70-150 Hz and 13-30 Hz log power, LDA, ds004473) with an independent numpy pipeline and steps of 8.5-136 uV. The 70-150 Hz loss from about 8 uV and the partial recovery at coarse steps replicated. We then tested the cause of the recovery, which that study left open: the rounding error alone decodes movement, and independent noise of equal variance does not recover.
- supportsAdded noise up to 40 nV/√Hz kept high-gamma decoding within a 0.02 AUC margin in 8 SEEG subjects
Its ADC-step loss boundary (4 to 8 uV) is equal to about 52 to 103 nV/sqrt(Hz) of quantization noise. This test found the loss boundary at 40 to 80 nV/sqrt(Hz) with white noise.
- usesAdded noise up to 40 nV/√Hz kept high-gamma decoding within a 0.02 AUC margin in 8 SEEG subjects
The EDF reader, bipolar pairing, trial rule and windows are copied from its code/quant_decode.py.
- usesA causal SEEG switch detected 10.6% of cued key presses at 0.28 false detections/min in an 8-subject, 240-s-per-subject analysis
The EDF header and byte-range read functions in code/fetch_block.py are adapted from this package.
- usesFour-bit weights preserved held-out AUC in an offline SEEG key-press decoder across eight subjects
ADC step result on the same data; this work adds the weight-precision part of an on-implant decoder.
- usesRe-referencing improved high-frequency SEEG detection over the recorded reference, but no montage was best across eight subjects
We reused its byte-range EDF reader (code/quant_decode.py, unchanged), its trial selection, its rest and move windows and its LDA pipeline. Our bipolar 70-150 Hz median AUC (0.840) agrees with its stored-data value (0.85).