Temperature scaling and prior correction lower Brier error below a constant forecast under a 5% response weight in later task windows from eight SEEG recordings
We tested response probabilities in later task windows from eight public SEEG recordings. The source was OpenNeuro ds004473, version 1.0.2. Each test trial supplied one rest window and one pre-response window. We assigned 5% weight to response windows. This fraction was assumed, not measured. Models used eight bipolar pairs per patient. Early trials fitted the model and a separate temperature map. A standard prior correction then changed the probability odds. On 674 later trials, mean weighted Brier error was 0.043244. A constant 5% forecast gave 0.047500. The paired patient mean difference was -0.004256 (95% resampling interval -0.008933 to -0.001297). All eight patient scores improved. Raw prior correction alone gave 0.050402 and did not beat the constant forecast. At a 0.5 threshold, response detection was 10.98%, and four patients had no calls. At a 1% response weight, corrected error also did not beat the constant forecast. These offline results test probabilities for selected task states. They do not test online intention or brain stimulation. The package preserves code, all features, predictions, three figures, provenance, and a retrieval recipe.
SEEG · brain-computer interface · probability calibration · prior correction · Brier score · rare response · temperature scaling
References
6- 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 study uses all eight preserved feature files, trial records, source records, and the feature extractor. They support the rare-response probability test.
- usesEarly confidence scaling decreased mean Brier error from 0.193 to 0.185 in eight offline SEEG patient tests sci:sha256:f3d79f04937a343945e7a53f5b54b00a0284174fca051959a818f1c79d41a7d1
This study uses the separate model, temperature, and later test split. It fits new models and tests weighted response fractions.
- usesdoi:10.18112/openneuro.ds004473.v1.0.2
The analysis uses eight recordings from this exact OpenNeuro dataset version.
- usesdoi:10.1162/089976602753284446
The study applies the established class-prior correction described by Saerens and colleagues.
- supportsdoi:10.1371/journal.pone.0118432
This work is cited for how class imbalance affects positive predictive value and classifier call quality.
- doi:10.1088/1741-2552%2Facae0a
Related source study for the public SEEG dataset and task context.
Cited by
1 publicationPublic catalog entries on Sciscape.
- supersedesTemperature scaling and prior correction lower Brier error below a constant forecast under a 5% response weight in later task windows from eight SEEG recordings
This version fixes the package file check. It excludes the metadata folder added by sci fetch. All scientific results, figures, features, models, and predictions are unchanged.