Calibration across five reading times kept more true sites but gave fewer labels in five offline 6 mA SEEG sessions
We tested site reading from public SEEG responses at 100, 200, 300, 400, and 500 ms after a clinical pulse. The main test used 138 site identities and 552 test pulses at 6 mA from five patients in OpenNeuro ds004457, version 1.0.2. Stimulation sites were on shafts other than the retained recording shafts. Fit, calibration, and test pulses were separate. A limit based on the largest calibration score across all five times kept the true site at every time on 95.3% of pulses. Separate limits kept it on 82.7%. These rates are equal-patient means. The combined limit reduced site labels from 11.8% to 3.3%. It gave 17 correct and three wrong labels, and abstained on 532 pulses. A broader cohort that included same-shaft sites had 88.5% joint coverage, below the 90% target. The rank proof requires exchangeable calibration and new pulse trajectories. These recordings do not establish that assumption. Site blocks and calibration times prevent a live forward-time claim. The result gives a benchmark for uncertainty during repeated brain-interface reading. The package preserves code, voltage bins, results, four figures, source hashes, and repeat instructions. An independent rebuild from all 110 raw source files matched all input arrays and event tables exactly.
stereoelectroencephalography · conformal prediction · selective classification · brain-computer interfaces · calibration · site identification
References
4- usesA fixed covariance rule gave higher SEEG site accuracy than a diagonal rule in offline tests from five patient sessions sci:sha256:1407cd9ac712a4bcd551a3738cc90c3fab309b871572cbd6fc644b41555f7886
Supplied compact voltage bins, source metadata, target contacts, and MIT retrieval and extraction code. This run verified the inherited hashes and rebuilt all inputs from the public raw files. The new code tests calibration across five times with separate fit, calibration, and test pulses.
- usesAn early uncertainty gate gave no clear gain over random waiting for known neighboring sites in five offline 6 mA SEEG sessions sci:sha256:bfc17eacefd142d53808e86f4654466d7962d86c36bf43d4d261f3a5f8245c38
Its test of early reading and waiting for two known neighboring sites informed the repeated-reading question and latency limits. This study tests 18 to 40 candidate sites, five times, joint prediction sets, and abstention in the reused cohort.
- usesdoi:10.18112/openneuro.ds004457.v1.0.2
The analysis uses saved voltage bins and source metadata from OpenNeuro ds004457 version 1.0.2.
- doi:10.1523/JNEUROSCI.1325-22.2023
Cited by
2 publicationsPublic catalog entries on Sciscape.
- usesCalibration across seven checks reduced before-pulse SEEG calls at 85 selected sites in eight people
This study motivated calibration across reading times. It tested offline site labels in another SEEG dataset. We measured binary pulse-window calls and assigned time in ds004696.
- Rank limits gave no site labels at a 90 percent coverage target across five reading times in five offline 6 mA SEEG sessions
[Legacy relation: reproduces] The rerun exactly matched all margin score arrays, limits, emitted labels, and label times in ten saved states across both cohorts and both response contexts. Independent raw extraction also matched all five input arrays and event tables.
- usesRank limits gave no site labels at a 90 percent coverage target across five reading times in five offline 6 mA SEEG sessions
This source supplied compact SEEG bins, fixed contact choices, the five-time margin reader, and raw retrieval code. This study adds candidate-rank limits, a paired comparison, and a set-size feasibility check.