Countering Neural Activity Drift: Sustained Long-term Seizure Prediction Using an Evolutionary Machine-Learning Framework on Continuous Intracranial EEG

Objective

Seizure prediction in drug-resistant epilepsy remains a major biomedical challenge. Traditional machine learning approaches rely heavily on segmented offline testing, which suffers from artificial class rebalancing, hides neural activity drift over time, and severely inflates performance estimates. This study introduces an online evolutionary framework designed for realistic brain-computer interface validation and addresses performance degradation caused by neural drift.

Methods

We present the first publicly available, continuous long-term stereoelectroencephalography dataset for seizure prediction, tracking 16 patients across 664.9 hours of data and 121 seizures. An offline combinatorial analysis evaluated pipeline decisions (referencing, frequency bands, functional connectivity metrics, and classifiers). The highest-performing offline elements, namely monopolar referencing, cross-correlation-based connectivity matrices, and Random Forests classifiers, were evaluated under realistic class imbalances in a pseudo-prospective online framework. Candidate models (N = 1,000 per patient) were evolved across continuous streams, while a data-driven approach targeted patient-specific epileptogenic networks for input channel reduction.

Results

Transitioning from offline to online evaluation demonstrated a prominent performance drop, with mean AUROC degrading from 90% to 57.9%, confirming offline metrics conceal temporal neural drift. However, our online evolutionary framework identified models achieving complete event-level seizure prediction in 75% of patients (and all but one seizure in 87.5%). Restricting inputs to the estimated epileptogenic network identified using our previously proposed approach reduced implanted contact requirements by 79.12% while providing superior predictive performance over clinically resected areas.

Significance

This work demonstrates that actionable, deterministic seizure prediction is achievable when paired with modern computational capability. Because 1,000 candidate models represent a conservative proof-of-concept search budget, scaling search spaces in production environments can further expand optimal interictal sampling and prediction yields. By providing both a continuous benchmark dataset and a standardized online evaluation protocol, this framework offers a foundation for self-adapting closed-loop devices capable of post-event retraining and long-term deployment in clinical neuromodulatory

Key Points

• First publicly available continuous long-term iEEG dataset for seizure prediction, with the largest patient and channel count to date.

• The proposed framework predicted all seizures in 75% of patients, and all but one seizure in 87.5% of patients.

• Online evaluation reveals substantial performance overestimation in offline testing, highlighting the need for realistic BCI validation.

• Epileptogenic network-based channel selection achieves comparable predictive performance with around 79% fewer implanted contacts.

• Classifier choice and functional connectivity metric are the dominant drivers of seizure prediction performance across the pipeline.