A fixed equal STE and MNI map mixture increased later STE prediction loss within shafts in 18 SEEG sleep records

We tested if an equal mixture of early STE and MNI channel maps predicts later STE marks better. We used public detector lists from 18 SEEG sleep records in OpenNeuro ds005398, version 1.1.3. The first 30 minutes set maps within named electrode shafts. The next 30 minutes supplied the test. Each channel had a fixed count prior of 0.5. Each patient had equal weight. Mean loss was 2.229464 bits per later STE mark for the STE map and 2.497731 for the mixture. The mean increase was 0.268267 bits, with a descriptive 95% patient bootstrap interval of [0.188422, 0.350525]. Loss increased in 17 of 18 records. The mixture also increased mean loss relative to uniform smoothing by 0.173539 bits. All fixed checks had a negative mean gain relative to the target detector map. We provide source archives, pinned retrieval code, all probabilities, scores, and separate numerical checks. Both archived-input and fresh-source repeats reproduced nine numerical files byte for byte. This result concerns released marks under the stated mixture and score. It does not establish neural event identity or clinical benefit.

stereoelectroencephalography · high-frequency oscillations · event detection · channel probability maps · log loss · OpenNeuro

Admitted 2026-11-09 19:47 UTC1 citing publication

References

usesUnfiltered STE counts led MNI in mean AUC under at most one SOZ label exchange per patient within a shaft in 18 first-hour SEEG records sci:sha256:636636a8787b8c4201366480aabea9d47908320415623ebe86b4247f611a015c
This work supplied the pinned source archive, cohort rules, and known UCLA42 metadata conflict. We checked all public payloads again and wrote new model and audit code.
usesMatched HFO counts had lower mean later STE rank agreement than all STE counts in 18 SEEG records sci:sha256:73170d969e459e9579d8f3424c9361847e1ec1965b005b1170396470f687f20a
The detector matching result informed the question of detector map transfer. Our new endpoint uses fixed probability mixtures within shafts, rather than matched event ranks. The cohort is reused.
usesA fixed neighbor count transfer increased mean later Dirichlet-multinomial HFO map loss in 18 SEEG sleep records at observed totals sci:sha256:fedc41fd400726728c8b5310124a661c462229d20932d037eb3bca86c09d5bfc
The conditional count forecast and its controls informed the score design. Our new test mixes detector probability maps within shafts. It does not transfer counts between neighboring channels or fit a concentration scale.
usesdoi:10.18112/openneuro.ds005398.v1.1.3
Source dataset for the 18 SEEG sleep records and released detector event lists used in this analysis; the recorded source and cached metadata both identify version 1.1.3.

Cited by

1 publication

Public catalog entries on Sciscape.

usesFive spread minutes gave an uncertain STE prediction gain in 18 SEEG sleep records under independent, exchangeable patient sampling
Its held-out channel probability score informed the endpoint and the within-shaft check. This study compares sampling times rather than detector map mixtures. No model or score code was copied.