Linear interpolation gave higher mean balanced accuracy than pair-weighted power in eight fixed SEEG decoders at 20% expected loss of 10 ms packets
This study tested correction for lost packets in eight fixed SEEG state decoders. It used public OpenNeuro ds004473 version 1.0.2 and 674 later trials. Each buffer had 50 packets of ten samples, about 10 ms per packet. Packet loss was independent, shared by contacts, and had known probability. At 20% expected loss, linear interpolation gave 67.62% mean balanced accuracy. Pair-weighted power gave 62.30%. Complete data gave 70.69%. The paired difference was -5.32 percentage points; seven of eight people had a negative difference. The descriptive 95% person bootstrap interval was -8.50 to -2.10 points. Pair weights give exact expected complete-buffer raw power under this loss model. They can give negative estimates. About 17% of corrected 70–110 Hz powers were negative at 20% loss. Thus, low average raw power bias did not preserve the fixed log-power decoder. The result used a training-only lower bound. Its direction held across five bound settings in a check added after the result. The package contains exact digital inputs, fixed models, masks, raw powers, probabilities, proof, code, figures, source records, and reproduction steps. A full numerical rerun matched exactly. Fresh source ranges matched one trial per person. The test is offline and uses a reused clinical cohort. It does not establish live BCI or stimulation performance.
SEEG · neural signal processing · packet loss · balanced accuracy · power spectral density · linear interpolation · brain-computer interfaces · inverse probability weighting
References
5- usesEight SEEG pairs classify pre-response and rest windows at 70.7% mean balanced accuracy in later trials from eight recordings sci:sha256:c74068384c7c056974c8568aab987cf5a8eee9979a622c22efa1b67a911e77fd
This work used the eight early-trained SEEG state decoders first constructed in this publication. It preserved their pair selection, training split, scaler, coefficients, and threshold. The clean later-trial mean accuracy was reproduced before the new packet-loss test.
- usesFourfold filtered sampling reduced mean balanced accuracy by 0.58 points in eight offline SEEG decoders sci:sha256:f6f8bca9acbea141638cc5f4a86edba485bc8db75a281ef10abe65f1fc04225a
This work used the exact digital test-contact buffers, fixed model states, clean feature references, source headers, and pinned range records from this publication. These inputs enabled a packet-loss test with no new model fitting or contact selection.
- usesdoi:10.18112/openneuro.ds004473.v1.0.2
The analysis used the OpenNeuro sEEG Forced Two-Choice Task, version 1.0.2; source byte-range records and selected digital buffers are preserved.
- doi:10.1088/1741-2552%2Facae0a
The dataset README identifies this paper as the accompanying source study.
- doi:10.3389/fnhum.2022.934063
Background work on missing neural samples and uncertain gap timing.
Cited by
3 publicationsPublic catalog entries on Sciscape.
- usesA stationary Gaussian variance correction with 5 percent shrinkage reduced SEEG decoding accuracy for 20 percent simulated sample loss in eight recordings
This prior result motivated a test of fixed decoder accuracy when a power estimator has a valid mathematical identity. It tested pair weighting with a different mask family. The present result adds the Gaussian conditional variance test.
- usesInterpolation gave lower mean theta coherency error than pair weighting in seven people with simulated 50 ms SEEG gaps every 250 ms and uniform origins
Its finite quadratic power correction supplied prior context for sample-pair probability weighting. We tested periodic common gaps and complex band coherency in a different cohort.
- usesA stationary Gaussian variance correction with 5 percent shrinkage reduced SEEG decoding accuracy for 20 percent simulated sample loss in eight recordings
This prior result motivated a test of fixed decoder accuracy when a power estimator has a valid mathematical identity. It tested pair weighting with a different mask family. The present result adds the Gaussian conditional variance test.