Application of a multivariate seizure detection and prediction method to non-invasive and intracranial long-term EEG recordings

Ariane Schad, Kaspar Schindler, Björn Schelter, Thomas Maiwald, Armin Brandt, Jens Timmer, Andreas Schulze‐BonhageClinical 2007 · DOI

TL;DR

This study evaluates a multivariate EEG analysis method for detecting and predicting epileptic seizures using both scalp (non-invasive) and intracranial (invasive) recordings. The findings suggest that the method performs promisingly for seizure detection in practical EEG monitoring and shows significant seizure prediction capabilities, with a slight advantage for scalp EEG in some aspects.

What Problem Does This Paper Address?

Epilepsy is a neurological disorder characterized by recurrent, unpredictable seizures, significantly impacting patients' quality of life. While antiepileptic medications are the primary treatment, some patients are refractory, necessitating surgical intervention. Diagnosing these patients involves long-term EEG monitoring, often with both scalp and intracranial electrodes, to pinpoint seizure onset zones. Manually analyzing these extensive EEG datasets is time-consuming and labor-intensive. Therefore, automated seizure detection and prediction methods are crucial for improving epilepsy management, enabling timely interventions, and potentially developing seizure warning systems. A key question in this field is whether non-invasive scalp EEG or invasive intracranial EEG provides superior data for these automated methods, and this paper directly addresses this by comparing the performance of a specific multivariate algorithm on both types of recordings.

Methods

The researchers employed a multivariate method based on simulated leaky integrate-and-fire neurons, originally developed for non-invasive EEG. This method extracts spatio-temporal information from multi-channel EEG data, capturing synchronized, fast-transient, and rhythmic activities characteristic of epileptic events. Two key measures were derived:

  • Spiking Rate (SR): This measure reflects the level of spatio-temporal synchronized bioelectrical activity in the brain. It's calculated by integrating time-differentiated EEG signals across channels and detecting when this integrated signal exceeds a threshold.
  • Feature Fsz(t): This feature is derived from the average spiking rate over time and is interpreted as representing the gradual increase and decrease of spatio-temporal synchronization. It was modified from its original formulation to avoid using a posteriori information (i.e., knowledge of the seizure occurrence) during its computation.

The algorithm's parameters (thresholds T₁ and T₂) were adjusted individually for each patient to optimize performance. For seizure detection, the algorithm identifies a seizure when the spiking rate exceeds a fixed threshold. For seizure prediction, it identifies pre-seizure changes when the Fsz(t) feature crosses a threshold. The performance was assessed retrospectively using metrics like sensitivity and false positive rates.

Datasets

The study utilized long-term EEG data from six patients with pharmacorefractory focal epilepsy who underwent pre-surgical video-EEG monitoring. This included:

  • Simultaneous Scalp and Intracranial EEG Recordings: Data were collected continuously over several days from each patient.
  • Total Data: 423 hours of EEG data encompassing 26 seizures.
  • Electrode Configuration: For each patient, 30-90 focal and extrafocal intracranial electrodes were used, along with 21 scalp electrodes placed according to the international 10-20 system.
  • Patient Characteristics: Detailed information on sex, age, seizure types, electrode types, number of seizures analyzed, and outcome was provided in Table 1.

Key Insights and Ideas

  • Multivariate Approach Advantage: The chosen multivariate method avoids the need for preselecting specific electrodes, circumventing the problem of multiple testing that can arise when comparing different electrode combinations.
  • Similar Dynamical Behavior: Features extracted from both scalp and intracranial EEG showed remarkably similar dynamical patterns, suggesting that the underlying brain dynamics related to seizures are captured by both recording modalities.
  • Seizure Detection Promise: The algorithm demonstrated a promising performance for seizure detection, particularly for practical applications in continuous EEG monitoring.
  • Seizure Prediction Potential: The study indicated a significant seizure prediction performance, although further modifications might be needed for clinical utility.
  • Scalp vs. Intracranial EEG: While both modalities yielded comparable results, there were subtle differences. Scalp EEG showed a tendency towards better detection performance in some cases, while intracranial EEG sometimes allowed for earlier detection of seizures.
  • Influence of Sleep: Circadian rhythms and sleep stages were observed to influence the features used for prediction, potentially impacting algorithm performance and suggesting the need for adaptive thresholds.

Key Results

  • Seizure Detection: Sensitivities of up to 73% for scalp EEG and 62% for invasive EEG were achieved with an upper limit of 0.15 false detections per hour.
  • Seizure Prediction: Up to 59% of seizures could be predicted from scalp EEG and 50% from invasive EEG, given a maximum of 0.15 false predictions per hour.
  • Performance Comparison: For seizure detection, a tendency towards better performance with scalp EEG was observed. For seizure prediction, scalp EEG generally showed slightly higher sensitivities (average 59% vs. 50% for invasive EEG).
  • Latency: In some cases, intracranial EEG allowed for earlier seizure detection compared to scalp EEG, likely due to direct access to the seizure onset zone.

Limitations and Future Work

  • Clinical Validity of Prediction: While statistically significant, the seizure prediction performance obtained was not yet sufficient for direct clinical application, requiring higher sensitivities and specificities for practical intervention systems.
  • Feature Sensitivity to Non-Seizure Activity: The Fsz feature, used for prediction, was found to be sensitive to non-seizure related brain activities like changes in vigilance and sleep, which can lead to false alarms.
  • Fixed Thresholds: The study used fixed thresholds for detection and prediction. The authors suggest that using dynamical thresholds, which adapt to changing brain states (like sleep), could significantly improve prediction performance by reducing false alarms.
  • Electrode Location Dependence: For seizure detection, the performance could be influenced by the specific location of invasive electrodes, as they measure localized synchronization, whereas scalp EEG captures broader brain activity.
  • Future Directions: The authors propose further investigations into the effects of sleep, vigilance changes, and subclinical seizures on detection and prediction performance. They also emphasize the importance of using dynamical thresholds and collecting more time-continuous, simultaneous scalp and intracranial EEG data to gain deeper insights into seizure generation and propagation mechanisms.

Summary updated

BibTeX
@inproceedings{schad2007application,
  author = {Ariane Schad and Kaspar Schindler and Björn Schelter and Thomas Maiwald and Armin Brandt and Jens Timmer and Andreas Schulze‐Bonhage},
  title = {{Application of a multivariate seizure detection and prediction method to non-invasive and intracranial long-term EEG recordings}},
  booktitle = {Clinical},
  year = {2007},
  doi = {10.1016/j.clinph.2007.09.130},
}
Figure