Four-bit features with limits of ±3 training standard deviations changed later mean balanced accuracy by +0.10 percentage points in eight fixed SEEG decoders
We tested four-bit transmission of 32 SEEG power features with eight fixed patient decoders. The public data were OpenNeuro ds004473 version 1.0.2. Fit and calibration used early trials. Tests used 816 later trials and 67,309 continuous windows. The coding limits were plus or minus three training standard deviations. Mean balanced accuracy was 72.17 percent with float features and 72.27 percent with four-bit features. The mean change was +0.10 percentage points. Its descriptive 95 percent patient bootstrap interval was -0.70 to +0.94 points. This does not prove equal performance. The actual frame fell from 128 bytes to 17 bytes, including a clipping flag. Headers and shared model data were excluded. Continuous replay across 112.31 minutes had 321 added and 291 removed exact command endpoints. Some changes can be shifts in time. A conditional receiver bound certified 63.91 percent of continuous decisions, as a patient mean. Coverage was 20.86 percent for positive decisions and 68.42 percent for negative decisions. All certified decisions agreed with float decisions. The bound assumes no clipping, correct shared parameters, correct packets, and suitable arithmetic. It does not certify task labels or stateful commands. This is an offline test in one session per clinical patient. We measured no device power, live control, or stimulation. The package preserves inputs, code, frames, results, figures, proof, source records, and reproduction instructions.
stereoelectroencephalography · brain-computer interface · feature quantization · SEEG · balanced accuracy · continuous decoding · decision certificate · neural signal processing
References
7- usesA fixed SEEG decoder separated windows before keypress from rest on later trials in eight patients sci:sha256:bd6c59023a95ca7dc23d6549123fb8e0b3724d75412db49c1e004ce64fc1dd57
This work was the origin of the event features and the preserved event extraction recipe. We used them through the direct parent archive. We kept its trial retention and source alignment.
- usesAt a zero threshold, eight-pair SEEG decoders gave a mean of 36.7 outside-guard triggers per eligible minute in eight later task scans sci:sha256:22495a59283fcfea65785cbf2a7a410c3a81dc75e721320a08a5ed114e14bd43
This earlier work supplied the continuous decoder approach and source feature history through the direct parent. Our new contribution tests precision with a first-positive-run command rule and exact endpoint comparisons.
- usesA rule with three scores did not show better detection before key presses at a 5/min early background budget in eight SEEG recordings sci:sha256:4123fe4badcd90fdd5baa89a3bdfb8efdb03e541ba11f571a97f99b1edb9e826
This publication supplied all 123 copied inputs. These included event powers, later continuous powers, fixed eight-pair ridge models, source records, and extraction code. We checked the copied hashes and refitted the float models. The new study tests feature precision with the fixed weights and thresholds.
- usesThree-bit SEEG power features had 70.1 percent mean balanced accuracy in later tests from eight patients sci:sha256:cbd0567615863f2d4de1a68c6eb93360178d33f86d56e7b2fbe91ac070e26266
We used this related three-bit SEEG study as context in the report. We found it after our precision tests. Its different features, windows, split, quality selection, and model prevent a direct comparison. It did not set our local plan.
- usesdoi:10.18112/openneuro.ds004473.v1.0.2
Source data: OpenNeuro ds004473 version 1.0.2, used to construct the eight patient decoders and test windows.
- supportsdoi:10.1088/1741-2552%2Facae0a
The cited source paper reports movement-related oscillations in the SEEG recordings used for this task. It does not validate the decoder or the precision comparison.
- doi:10.1109/TNSRE.2015.2501752
Background on low-bandwidth band-power features in neural interfaces. Its hardware power results do not transfer to this software test.
Cited by
3 publicationsPublic catalog entries on Sciscape.
- usesWith timely replies, four-bit receiver checks preserved 2,163 SEEG commands and saved 54.8% of payload bytes in eight records
Used its four-bit standardized range, clipping flag, and score-error bound as the starting receiver check. The new protocol adds full exact replies and shared command wait state. The new command rule is the direct data parent rule, which can repeat commands during a sustained positive stream.
- usesWith a shared penalty, added SEEG power increased mean loss for 0.5-second response forecasts in eight cued task recordings
This publication supplied the continuous power arrays, fixed bipolar pairs, source metadata, range hashes, and extraction code. We fit new task-time and neural forecast models.
- usesA higher score threshold changed exact command times in eight fixed SEEG streams with a one-second waiting period
Its exact command changes under feature coding and stated limits of decision certificates motivated a separate schedule-level test. We used no compressed feature data from it.