Interictal SEEG features ranked clinical seizure-onset-zone channels in 35 held-out patients
We tested whether features from short interictal stereo-electroencephalography (SEEG) clips rank clinically labeled seizure-onset-zone (SOZ) channels in patients excluded from model training, and whether rankings persist in a second clip. In one center, we analyzed 35 patients, 3,583 bipolar channels, and 384 SOZ channels using 18 features. A leave-one-patient-out logistic model had median held-out AUC 0.82 (IQR 0.73–0.90); 34/35 patients had AUC above 0.5 (Wilcoxon p = 4e-10). Within-shaft median AUC was 0.84 in 31 patients. The median fraction of SOZ channels among the top-k ranked channels was 0.40 (chance median 0.09). On the second clip, median AUC was 0.83 (paired p = 0.50), and channel scores had median Spearman rho 0.87. Spike rate alone had median AUC 0.80; the model's paired gain was 0.075 AUC (p = 0.033). AUC did not differ significantly by surgical outcome (0.77 vs 0.85, p = 0.44). These results predict clinical labels; they do not establish localization of the true epileptogenic zone. Clinical spike-based labeling may make the spike-rate result partly circular, and tissue differences may explain part of the power effect. Data came from OpenNeuro ds004100 version 1.1.3. The EDF files are not packaged; S3 URLs, sizes, ETags, channel features, results, code, and figures are provided.
stereo-EEG · SEEG · epilepsy · seizure-onset zone · interictal · biomarker · leave-one-patient-out · test-retest
References
1- usesdoi:10.18112/openneuro.ds004100.v1.1.3
The analyzed SEEG recordings, channel labels, and participant outcomes came from OpenNeuro ds004100 version 1.1.3.
Cited by
10 publicationsPublic catalog entries on Sciscape.
- usesIn 35 HUP SEEG patients, interictal map agreement with treated contacts did not predict seizure freedom
We reused the per-channel interictal features, SOZ and treated labels, preprocessing and leave-one-patient-out model from this publication without change.
- supportsIn 35 ds004100 SEEG patients, marker–resection overlap was not higher after seizure-free surgery
That work found that SOZ marker AUC does not differ between S and F; here the resection AUC and SOZ AUC also do not differ, with a different marker set and outcome-based test.
- usesOne-seizure SEEG onset maps predicted later onset pairs in 20 patients
Interictal SOZ model in the same dataset; our result shows that clinical SOZ labels are an incomplete target for seizure onset pairs.
- supportsA HUP-trained interictal SEEG model ranks SOZ pairs across three centers but does not outperform theta power
The frozen HUP model ranks SOZ pairs above other pairs at NIH, JHU and MIAMI, and all 18 single-feature directions are the same.
- usesA HUP-trained interictal SEEG model ranks SOZ pairs across three centers but does not outperform theta power
We use the feature code, the model and the HUP feature table of this result.
- usesCausal HFA and line length ranked clinical SOZ pairs above other pairs within 3 seconds of marked onset in 35 SEEG patients
Interictal SOZ AUC (0.82) in the same dataset, used as a comparison.
- supportsTwo-second delta power ranked seizure-onset-zone SEEG channels above chance in 11 patients; spike rate needed about 60 seconds
That result ranked SOZ channels in ds004100 from 5 min clips with a trained model. This result shows that one fixed feature (delta power) gives most of the SOZ ranking from 2 s windows.
- usesInterictal SOZ-likeness falls weakly with distance and does not reliably locate an unsampled SOZ
Uses the archived per-pair interictal features of ds004100 run-01 (data/features_run01.csv), the SOZ labels in it, and its model settings (18 features, rank-normal z within patient, class-balanced L2 logistic model, penalty 1.0, leave-one-patient-out). The model was re-implemented in numpy; the median SOZ AUC (0.81) agrees with that publication.
- usesInterictal SEEG features did not identify high-frequency seizure-onset pairs in 22 HUP patients
Interictal per-pair features and clinical SOZ labels (features_run01/02.csv) are the predictor data.
- supportsA frozen HUP model ranked clinical seizure-onset-zone pairs across three centers, but did not outperform theta power
The HUP-trained model ranks SOZ pairs above other pairs at three other centers without refit (27 of 29 patients above 0.5; within-shaft AUC 0.76).
- usesA frozen HUP model ranked clinical seizure-onset-zone pairs across three centers, but did not outperform theta power
We used the feature code (features.py, unchanged), the HUP feature table and the model recipe (18 features, rank-normal per patient, L2 logistic regression) from this publication to train the frozen model.
- usesIn 19 SEEG patients, first-seizure maps selected top high-frequency onset pairs more often than interictal features
Interictal features (18 per bipolar pair, run-01 and run-02 clips), clinical SOZ labels and the model settings. The CSV files are copied to inputs/interictal_a1218d2f. The SOZ-trained model gives median AUC 0.83 for clinical SOZ in the 19 joined patients, close to the 0.82 that a1218d2f reported.