Student t forecasts reduced later log size loss after five fit pulses at 83 SEEG sites in eight people
We tested forecasts after five fitting pulses at 83 SEEG sites in eight people. The public data are from OpenNeuro ds004696 version 1.0.1. The primary target was log RMS bipolar voltage from 20 to 500 ms after stimulation. Student t reduced mean later-pulse log loss from 0.4324 to 0.2290 nats per shaft. The person bootstrap 95% interval for the gain was 0.0951 to 0.3349. All eight people had a positive gain. Student t also beat a Gaussian with the same variance. Marginal coverage increased, but the marginal interval score was worse than the primary Gaussian score. The normal reference shows that five-sample estimation uncertainty can explain much of the coverage shortfall. The test reused clinical sessions and retrospective quality labels. The package preserves code, compact inputs, all forecasts, scores, figures, and source hashes. Two complete public source files were rebuilt. A separate replay matched all 12 analysis outputs.
Student t forecast · log loss · prediction uncertainty · SEEG · electrical stimulation · brain-computer interface
References
5- usesStudent t limits scored worse than normal limits for centered SEEG voltage at 86 selected sites after five fit pulses sci:sha256:012478d084fb27a8d92d12089b83be731e42bdd06cd15828d578bf1698fc0e58
This prior work found higher Student t interval scores on raw voltage. It supplies context for the distinction between interval score and the log-size density score tested here.
- usesNominal 90 percent log-size intervals from five early pulses covered 87.1 percent of later pulses at selected SEEG pairs in eight sessions sci:sha256:a25b1738aeba69d31fd48a52ce1f7144c1e0e9886146691b3c65be5eac4b2c9d
This prior work tested log-size interval coverage after five pulses on response-shape-selected pairs. It defines a related endpoint and a different contact selection scope.
- usesResponse correlations gave no clear loss reduction with five fit pulses and 50% shrink at 83 SEEG sites in eight sessions sci:sha256:0bef106b1ca9c41fac09e6a5802d56ecb71132bf5c518aba74fbeecbb337ddd0
This work supplies the exact trial products, contact rules, source metadata, and extraction code used here. Its Gaussian scores provide an independent numerical reference.
- usesdoi:10.18112/openneuro.ds004696.v1.0.1
The analysis used OpenNeuro ds004696 version 1.0.1, including its public CRP MATLAB derivatives.
- supportsdoi:10.1523/JNEUROSCI.2201-22.2023
This source study provides the collection methods for the SEEG dataset used here.
Cited by
3 publicationsPublic catalog entries on Sciscape.
- usesA shared scale in a Student t model with four degrees of freedom raised mean 4–30 Hz SEEG joint forecast loss at 100–500 ms after five fit pulses at 82 sites in eight sessions
Its five-pulse Student t forecast motivates the fixed univariate mean, predictive shape, and degrees of freedom. The new test holds those margins fixed and compares joint densities. It uses band power and all later pulses rather than the earlier voltage size endpoint and final-five test set.
- Exact normal limits reduced log width and gave 85.7 percent mean coverage for five-pulse batches at 83 SEEG sites in eight people
[Legacy relation: reproduces] All 15,600 pulse features matched. The matched Student t single-pulse all-shaft coverage and width agreed for 32 person and window scores. This check does not reproduce the earlier density loss comparison.
- usesVariance cap 10 decreased SEEG log-size loss but increased mean RMS error against Gaussian forecasts in eight people
We retained its exact compact SEEG inputs, geometric choices, 83-site cohort, five fit and five test pulses, source event-order correction, and reference-model benchmark. We verified its imported files and reproduced its 96 person loss values.