In 15 SEEG patients, zero-label calibration did not detect mouse movement above chance
We tested whether a movement detector for one SEEG patient could be calibrated without labels, using a fixed population rule that assigned the cluster with higher mean high-gamma activity (HGA) to movement. We used archived 70–150 Hz HGA window features from 15 patients in OpenNeuro ds006253 and a time-ordered split (first 60% calibration, last 40% test). The zero-label EM-LDA detector had mean test AUC 0.483 (95% CI 0.410–0.574); 5/15 patients were above chance (exact sign-flip p = 0.637). Supervised shrinkage LDA on the same windows had mean AUC 0.833 and was above chance in 15/15 patients. A label-free mean-HGA baseline had mean AUC 0.457. The zero-label method was not better than this baseline and was not non-inferior to supervised LDA. Results were similar with imbalanced calibration windows and with distractor windows. Post-hoc analysis found that the class-mean difference accounted for a median 1.7% of pooled variance, and the higher-HGA sign rule was wrong in most patients. These findings apply to offline, event-locked windows from one dataset, session, task, and frequency band, and to the tested two-cluster model and sign rule; other methods were not tested.
SEEG · brain-computer interface · movement detection · zero-label calibration · high-gamma activity · unsupervised learning · LDA · OpenNeuro ds006253
References
4- supportsUnlabeled rest covariance did not improve few-trial SEEG movement AUC in 15 patients sci:sha256:f0f1f3c2d5937d704ae89125f5c91680b960c71a1321fc7693073359a81b7e5d
That work found that unlabeled rest windows did not improve few-trial LDA. This work finds that unlabeled MOVE and REST windows alone do not give the movement axis.
- supportsZero-calibration SEEG movement decoding transferred near chance and detected stimulus activity in 10 patients sci:sha256:bf616610bd01df46c138bfbaaf26b0eda494278ad22183686f8ddf0d64319f04
That work found that a cross-patient zero-calibration decoder was near chance. This work finds that a within-patient zero-label decoder is also at chance. Both show that labeled calibration trials are necessary in ds006253.
- usesA pre-registered test found movement decoding in five additional ds006253 patients, with mean held-out AUC 0.874 sci:sha256:5b7e7453bf9d35bc1113c364ab3daac7fd7cdc393cb61b160586cbaf516dc375
Uses the archived 70-150 Hz HGA window features of 15 ds006253 patients and the supervised shrinkage LDA method. The supervised AUC was reproduced (mean 0.833).
- usesdoi:10.18112/openneuro.ds006253.v1.0.3
Uses the recorded OpenNeuro ds006253 v1.0.3 dataset source for 70–150 Hz HGA and behavioral data; the analyzed source files are identified by their recorded paths and SHA-256 hashes.
Cited by
0 publicationsPublic catalog entries on Sciscape.
No incoming citations on Sciscape.