A forecast with previous labels and a known target count reached 61% balanced accuracy in 129 PURSUE records
We tested target forecasts without EEG in PURSUE visual oddball event records. We used OpenNeuro ds007056 version 1.1.1. A fixed split assigned 129 people to training and 129 to testing. The primary model used trial order and three previous labels. It increased test log loss by 0.005805 nats per trial. The 95% person bootstrap interval for this increase was 0.004236 to 0.007385. A secondary forecast used correct previous labels and the known count of eight targets in 40 trials. It reached 61.04% balanced accuracy. Uniform order with the same count gives 60.95% expected balanced accuracy. These values describe a protocol baseline, not neural prediction. The package preserves all source event files, 51,600 derived trials, code, predictions, bootstrap samples, a count proof, and reproduction instructions.
PURSUE · visual oddball · target prediction · event history · trial order · balanced accuracy · protocol baseline · machine learning
References
2- usesdoi:10.18112/openneuro.ds007056.v1.1.1
The analysis uses the PURSUE P300 Visual Oddball event tables from OpenNeuro ds007056 version 1.1.1, fixed to the recorded Git commit.
- doi:10.1016/j.neuroimage.2020.117465
Cited by
2 publicationsPublic catalog entries on Sciscape.
- Mean response slopes were positive for previous target gaps and smaller for next gaps in 49 EEG and 16 fMRI adults during visual tasks
Fixed target counts motivate caution about future gap information.
- usesDuplicate target columns reduced test Pz error at ridge penalty 100 in 68 ds006018 adults
This publication supplied the known-count probability rule and its protocol interpretation. We apply that rule to checked ds006018 blocks and test later EEG voltage rather than target-label forecast accuracy.