CSOPE-Net

A stereo-EEG (SEEG) evaluation for drug-resistant focal epilepsy implants around 150 to 200 recording contacts, and an epileptologist decides by eye which handful sits inside the seizure onset zone. That judgment determines what gets resected and whether the patient becomes seizure-free, and it is also the source of the labels any supervised model would learn from. Those labels are weakest exactly where the stakes are highest: when a resection fails, seizure-generating tissue lay outside what was labeled, so a classifier fitted to the annotations learns the annotation rather than the physiology.

CSOPE-Net (Contrastive Seizure-Onset Pattern Encoder), developed at the Cleveland Clinic Epilepsy Center with the Indian Institute of Technology Indore, avoids that circularity by not using labels during pretraining at all. Each contact's peri-ictal recording becomes an adaptive-superlet spectrogram, and a CNN–Transformer encoder is trained with an InfoNCE objective in which two randomly time-shifted views of the same contact's seizure form a positive pair and every other sample in the batch is a negative. The encoder is then frozen at a single checkpoint and reused unchanged for contact classification, unsupervised phenotype clustering, and retrieval of contacts resembling known onset tissue.

The catalog's other intracranial models — Brant and BrainWave — are broad backbones covering detection, forecasting and disease screening. CSOPE-Net is deliberately narrower: it operates on a fixed 60-second window centered on a clinically annotated ictal onset, not on continuous monitoring data, and buys with that constraint a representation built around the presurgical localization decision itself.

#Key Features

• Label-free pretraining over the full contact pool: Contrastive pretraining runs across every contact in the 119-patient development cohort irrespective of downstream label, so ambiguous and mislabeled contacts contribute signal rather than corrupting the objective.
• Adaptive superlet time–frequency input: Sixty-second peri-ictal windows are represented as 150 × 600 spectrograms spanning 1–150 Hz, capturing slow pre-ictal shifts and fast ictal activity in a single image.
• Frozen encoder, three downstream uses: One checkpoint supplies the 128-dimensional embedding used for classification, clustering and retrieval, with no per-task fine-tuning and no per-dataset refitting.
• Cross-institution transfer without adjustment: The frozen encoder and the development-cohort classifier were applied unmodified — no retraining, no threshold change — to the public HUP iEEG cohort (OpenNeuro ds004100), recorded at a different center on different hardware.
• Phenotype families rather than a bare score: Clustering the embedding recovers six reproducible seizure-onset morphologies, from sparse low-frequency buildup to abrupt sustained broadband onset.

#Technical Details

The encoder stacks three 3×3 convolutional blocks with channel widths 32, 64 and 128, pools across the frequency axis into a 600-step time sequence, passes it through four pre-norm transformer layers with 8 attention heads and feed-forward dimension 512, then applies attention pooling and a two-layer MLP to yield a 128-dimensional L2-normalized embedding. Pretraining ran 200 epochs of AdamW (learning rate 1e-4, cosine schedule) at temperature 0.1 and batch size 256 in mixed precision on a single NVIDIA A100; the epoch-100 checkpoint was selected without reference to downstream labels. Data came from 149 Cleveland Clinic patients recorded at 1,000 Hz between 2017 and 2023, split a priori into 119 development and 30 held-out patients. A Random Forest fitted on the frozen embeddings reached aggregate ROC-AUC 0.854 (95% CI 0.807–0.897) under leave-one-subject-out cross-validation, 0.935 on the 18 held-out good-outcome patients, and 0.822 (0.766–0.879) on the external HUP cohort. Precision is low at the Youden-optimal operating point (F1 0.25–0.27) because clean onset contacts are only 3.9% of those sampled, so the authors favor higher-specificity thresholds surfacing 20 to 25 contacts per patient.

#Applications

The immediate use is triage during presurgical SEEG review: ranking contacts by onset-zone likelihood so an epileptologist reads the most suspicious ones first, and describing each seizure's onset as a reproducible phenotype family rather than an opaque score. The more distinctive application is retrospective: in patients whose surgery failed, the model flags contacts clinicians had labeled non-SOZ but whose morphology resembles confirmed onset tissue. In a blinded review of 609 stimuli, three experts endorsed those contacts as showing ictal-onset morphology at roughly 15-fold higher odds than matched non-SOZ controls (OR 15.1, 95% CI 11.5–19.7).

#Impact

CSOPE-Net shows that contrastive pretraining on a single center's intracranial archive can yield an embedding that transfers to another institution's recordings with a fixed checkpoint and a fixed threshold — the property separating a reusable representation from a per-cohort model. The limits are acknowledged by the authors: pretraining covers 119 patients on one acquisition system, one external site is not multi-center validation, blinded inter-rater agreement was only moderate (Fleiss' κ = 0.46), and the Random Forest head is fitted separately rather than shipped with the encoder. The work remains a preprint, and neither code nor weights are released; the authors state both will follow publication, with weights under an academic-use license.