Task schedule data predict later trial labels with 99.72% accuracy in 266 ds000030 participant tests

In 266 event files from OpenNeuro ds000030 v1.0.0, each with 128 active trials, schedules repeated across participants carried later GO and STOP labels in a participant holdout. A lookup using the first eight arrows and intervening blank durations predicted trials 9–128 with 99.72% accuracy (balanced accuracy 99.44%; Brier score 0.00211; coverage 98.87%). The three participants with unique schedules accounted for the 90 missed STOP labels. When all participants with the test participant’s complete schedule were excluded from training, every eight-trial prefix was unseen; the lookup fell back to the task STOP fraction of 0.25 and reached 75% accuracy, 50% balanced accuracy, and Brier score 0.1875. Four arrows and three blanks identified all 39 schedules in this sample, a finite-sample metadata finding. Planned variants and a separate narrow arithmetic comparison are reported in the full text. The test uses schedule metadata, not brain or behavioral inputs; it does not show that participants learned schedules, establish biological causality or stopping latency, or show that other methods fail on new schedules. Repeated schedules make trials dependent, and no population confidence interval or significance test was calculated.

stop-signal task · trial schedules · cross-validation · data leakage · participant holdout · schedule holdout · GO and STOP labels · OpenNeuro ds000030

Admitted 2026-11-09 16:50 UTC18 citing publications

References

usesRecorded and nominal stop delays differ by a mean of 30.520 ms across 266 ds000030 files sci:sha256:eab4dfd5360dd88163bf70d714c66a125ce5403cb5271a8e1e2d91a108e85870
Used its 266 archived event tables, four source documents, pinned manifest, frozen Git tree, and participant comparison table. All input hashes were checked. These files enabled the schedule-only prediction test and the narrow arithmetic comparison. Its code was inspected but was not executed.
usesPositive-delay failures show larger local GO RT increases than zero-delay failures in ds000030 1.0.0 sci:sha256:dd164540e8ddf090ca3f80ee8daff60571c12de09cb022d170bf642ce05203c5
Its read body reports 39 distinct trial-type and arrow sequences and warns about shared schedule structure. This observation helped define the schedule-holdout test. The present work does not claim that count as new and does not reproduce its reaction-time contrasts.
usesdoi:10.18112/openneuro.ds000030.v1.0.0
The archived event tables used for this analysis are from ds000030 v1.0.0; the saved dataset metadata identifies this version-specific DOI and states CC0.
doi:10.1038/sdata.2016.110
doi:10.1093/gigascience%2Fgix020
doi:10.7554/eLife.46323

Cited by

18 publications

Public catalog entries on Sciscape.

usesThe same GO input gives higher mean staircase score correlation in 241 ds000030 scanner records.
Its schedule definition uses trial labels, arrow directions, and following NULL durations. This work uses that definition for the schedule bootstrap.
usesThe history model reduced mean squared error by 0.42% in 38 schedule tests, but the interval included zero
The complete ordered trial-type, arrow, and following-NULL-duration schedule definition supplied the exclusion groups. Entire schedules and their participants were excluded from each fitted model.
usesThe mean GO response time is shorter after arrow changes in 102 selected ds000030 controls.
Its report of shared task schedules led to the schedule fingerprint and schedule weighting checks in this analysis.
usesThe Pearson correlation difference is less than zero after adjustment for training GO time in 241 ds000030 participants.
Its schedule definition uses ordered trial types, arrow directions, and following NULL durations. The new code applies that definition to form the 39 groups for the schedule bootstrap.
usesAll 140 patient runs have a task schedule that also occurs in control runs in ds000030 1.0.0.
Used its 266 event tables, fixed Git tree, schedule membership table, task schedule rule, and eight-trial prefix rule. These inputs enabled source checks and the diagnosis support audit.
usesThe two-sided model estimates a 15.27 ms third-GO difference in 125 ds000030 control runs
Its shared-schedule result helped select the future-STOP terms. These terms enabled the two-sided history check.
usesThe nearer and earlier class coefficients are more than 0 ms after schedule adjustment in 89 ds000030 participants.
Its inspected definition of repeated nominal schedules determines the exact schedule grouping used here. This enabled schedule cell constants, complete schedule resampling, and the audit of class information. Its prediction code and outputs were not used.
usesThe third GO contrast is 13.55 ms after adjustment for three arrow transitions in 125 ds000030 control runs.
Its nominal schedule definition uses active trial type, arrow, and following NULL duration. This definition supplies the groups for equal schedule weights and removal of one schedule at a time. No prediction result is reproduced.
usesThe early median has less GO forecast error than the earlier two-response mean in 264 ds000030 participants.
Its read complete-schedule rule uses ordered active trial types, arrow directions, and following NULL duration strings. That rule enabled the 39-group bootstrap and schedule removal checks. Its classifier and accuracy result were not used or reproduced.
usesNew code agrees with 18 half-run correlations in three ds000030 control samples.
Its read task schedule definition determines groups by trial type, arrow direction, and following NULL duration. This enables complete schedule resampling for the 126 controls.
usesThe fourth GO before STOP has a 34.18 ms longer mean RT than the last GO in 53 selected ds000030 participants.
Used the complete-schedule definition: ordered active trial labels, arrow directions, and following NULL durations. The new code checks all target schedule hashes and compares participant and schedule weights. It does not repeat the prediction analysis.
usesThe mean difference in local GO change is 96.8 ms after schedule reference subtraction in 95 ds000030 participants.
The analysis uses the complete trial-label, arrow-direction, and following-NULL-duration schedule definition. This definition provides the groups for same-position references. The new code finds the same 39 schedules. It does not use the prediction code or prediction results.
usesThree or more local STOP windows occur in 66 of 126 selected ds000030 CONTROL task files
Its inspected full-schedule definition supplied the grouping by active trial labels, arrow directions, and following NULL durations. That grouping enabled the structural-window eligibility count by schedule. Its prediction model and accuracy result were not used.
usesThe mean of two earlier reaction times gives less later GO forecast error in 89 selected ds000030 participants
The read complete schedule definition uses ordered active trial types, arrow directions, and following NULL duration strings. This definition supplied the groups for whole-schedule bootstrap, equal schedule weights, and removal of one schedule at a time. Its forecast code and prediction outputs were not used.
usesThe third GO contrast is 15.92 ms with arrow pattern adjustment in 125 ds000030 control runs.
We used the nominal schedule definition: active trial types, arrow directions, and following NULL durations. It enabled equal schedule weights and removal of one schedule at a time.
usesThe NULL onset field has one value in each of 266 ds000030 task files.
Its warning about NULL onset values informed the timing audit question. The new code audits all NULL rows and compares literal and schedule intervals. Its prediction code was not used.
usesTraining and scanner SSRT residuals have a correlation of 0.454902 in 232 ds000030 pairs.
The ordered trial-type, arrow, and following-NULL definition supplies schedule groups. The new analysis uses joint training and scanner groups.
usesNext STOP classes have no variation in full schedule-position cells from 126 ds000030 CONTROL scanner runs.
Its full schedule definition supplies the schedule-position groups. This audit does not repeat its label prediction model.