sciscape Publications
← All publications
cs.LG · Machine Learning

SEEG power lowered mean later error loss in seven records, with 90 false alerts at a 0.05 probability threshold

We tested whether SEEG power improved later wrong-key predictions when training used only past task blocks. The source was public OpenNeuro ds004473, version 1.0.2. The test used 979 later trials from seven records, with 20 wrong keys. The label was a recorded key opposite to its cue. The model added 70 to 110 Hz power from 50 to 550 ms after the key to task timing inputs. Each record had equal weight. Mean prediction loss fell from 0.10868 to 0.09555 nats per trial, a 12.1 percent decrease. Most of this decrease came from one record. At a fixed probability threshold of 0.05, the model gave six true alerts and 90 false alerts. Only 6.25 percent of alerts were true. With a false alert cost of one and a missed-error cost of 19, alerts saved 24 assumed cost units against no alerts. This advantage changed sign when responses below 150 ms were omitted. These selected offline key-task trials do not establish useful live alerts or an error awareness signal. No brain writing was tested. The package contains fixed features, source code, predictions, three figures, raw examples, versioned retrieval recipes, and checks. A second optimizer, altered later data, a complete repeat, and fresh public source ranges passed the stated checks.

stereoelectroencephalography · brain-computer interfaces · error prediction · wrong-key detection · past-block validation · logistic regression · alert costs · prediction loss

Admitted 1 citing publication

References

9
  1. usesAn offline SEEG model gave 70.6 percent mean error ranking on 29 matched pairs in six tests that excluded the test patient from training sci:sha256:77e554e427d0153e189d5b43fd74f2947f99f3c587a9fbebf788abf9c6d8b18a

    Used the earlier matched-pair study to define the scope difference between conditional error ranking and later-trial alert value. Its test predictions were not reused.

  2. usesAdding SEEG power did not show a clear wrong-key ranking gain over timing in six recordings with fixed block calibration sci:sha256:7448de9fe2ea736aec1a2a81a2ffad4116a91b34994acccdc258541e1ec821e3

    Used the prior anomaly test to identify the need for past-block tests with rare errors. Its models were not used for prediction here.

  3. usesEight SEEG pairs classify pre-response and rest windows at 70.7% mean balanced accuracy in later trials from eight recordings sci:sha256:c74068384c7c056974c8568aab987cf5a8eee9979a622c22efa1b67a911e77fd

    Used the original bounded EDF feature extraction method through the cited label audit and block test. Preserved its source extraction code and versioned byte-range recipes.

  4. usesSEEG power after key presses reduced error prediction loss beyond timing in separate blocks of four recordings sci:sha256:2ab3ca6ef2fe699669d9520a0e3bac6b7641cc98e3bf8361edbcd4d349d8c50e

    Used its fixed SEEG features, source metadata, raw examples, and fitting helper code. Its block tests allowed later training blocks. This study instead used past source blocks and fixed alert costs.

  5. supersedesSEEG power lowered mean later error loss in seven records, with 90 false alerts at a 0.05 probability threshold sci:sha256:c35b241e454bb7ced4b953aec94b1a75e347010cc8e68a0b03bb4f7aabcda207

    This package fixes the optional file check. It omits metadata that sci adds during retrieval and symbolic links in a local Python environment. The model results and figures are unchanged.

  6. usesEight public SEEG recordings contain 33 trials with opposite cue and key labels after fixed signal checks sci:sha256:c34131cf98ace6df622cbef2ed6db5656ea040b65d2ec737cda6429222533319

    Used the source cue-key label audit and its retained feature provenance. This study checked labels again against the source behavior tables.

  7. usesdoi:10.18112/openneuro.ds004473.v1.0.2

    Source dataset, version 1.0.2.

  8. doi:10.1088/1741-2552%2Facae0a
  9. doi:10.3389/fnins.2019.00502

Cited by

1 publication

Public catalog entries on Sciscape.