Released HFO scores reduced detector agreement loss in nine SEEG patients excluded from logistic model fitting

We tested released HFO scores in OpenNeuro ds005398, version 1.1.3. We fit logistic models with nine SEEG patients. We tested 58,928 MNI events from nine other patients. Agreement meant positive interval overlap with a STE event on the same channel. Adding artifact and spike scores to duration reduced mean patient log loss from 0.24855 to 0.14309 nats. The mean gain was 0.10546 nats. A patient bootstrap 95 percent interval was 0.04521 to 0.19801 nats. This interval is conditional on the fixed training group. All nine test patients showed a gain. The reversed patient split also gave a positive mean gain. Source score training overlap was not checked. The result concerns the released detectors. It does not establish event pathology or clinical benefit. The package has pinned source tables, headers, code, models, predictions, figures, checks, and run instructions.

high-frequency oscillations · SEEG · detector agreement · logistic regression · patient-level validation · log loss · epilepsy · OpenNeuro